Multi-level intelligent control. Methods and systems of intellectual control of technological equipment Intellectual control within the framework of applied semiotics

INTRODUCTION

The operating conditions of modern technological complexes lead to the need to take into account in the process of control and management the following types of uncertainty:

1. Low accuracy of operational information received from control objects, arising due to the large error of sensors for measuring technological parameters (flow, pressure, etc.), their low reliability, communication channel failures, large delay in the transmission of information across control levels, the lack of the possibility of measuring parameters at all points of the technological process required for models.

2. Inaccuracy of models of control and management objects, caused by: non-equivalence of solutions of systemic multilevel hierarchical models and individual local problems used in practice; incorrectly carried out decomposition of the general control task, excessive idealization of the technological process model, breaking essential links in the technological complex, linearization, discretization, replacement of the actual characteristics of the equipment with passport ones, violation of the assumptions made when deriving equations (stationarity, isothermality, homogeneity, etc.).

3. Fuzziness in the decision-making process in multilevel hierarchical systems, due to the fact that the presence of clear (accurate) goals and coordinating decisions at each level of control and management, and for each local control device, complicates the coordination process and predetermines the long iterative nature of coordinating decisions.

4. The presence of a human operator, including a dispatcher in the control loop and conducting the coordination process in a real production system in a natural language, leads to the need to take into account the difficulties of representing the dispatcher's knowledge in the form of algorithms and the consistency of the solution obtained by the computer with its assessment.

“The excessive desire for precision has begun to have an effect that nullifies control theory and systems theory, since it leads to the fact that research in this area focuses on those and only those problems that lend themselves to exact solution. Many classes of important problems in which data, goals, and constraints are too complex or ill-defined to permit accurate mathematical analysis have been, and remain, left out for the sole reason that they are not amenable to mathematical treatment.

L.Zadeh

Among modern production processes, there are many that have a complex of qualities that are unexpected for the classical theory of automatic control (TAU). This "uncomfortable" or, as they are commonly called, "weakly structured" or "ill-defined" objects have properties such as uniqueness, lack of a formalized goal of existence and optimality, nonstationarity of the structure and parameters, incompleteness or almost complete absence of a formal description of the object.

Conceptual Foundations

management under uncertainty

Uncertainty factors, which are understood as sources of uncertainty, are rather conventionally divided into the following three large groups:

1. uncertainty and incomplete information about the situation, which is used to make a decision on assessing the quality of functioning or forming the management of the functioning of the system - system and environment uncertainty factor;

2. factors generated by uncertainty, fuzzy thinking and human knowledge- uncertainty that manifests itself in the interaction of a person with the system and its environment;

3. uncertainties, fuzziness(inaccuracy) accumulated knowledge, concentrated in the knowledge bases of artificial intelligent systems, the uncertainty of operating this knowledge in the process of implementation those or other logical and logical-algebraic procedures for collecting and processing information, developing, choosing and making managerial decisions.

Classification of factors (sources) of uncertainty, requiring their consideration in the study of complex systems, is shown in Fig. B.1.

Fig.B.1. Classification of uncertainty factors

Methodology for analysis and accounting for uncertainty factors in

management in complex organizational and technical systems...

(ACS with DSS and DSS decision support systems and decision making systems)

1. Problems and generalized formalization of the tasks of developing and applying

making managerial decisions under conditions of uncertainty….

2. Deterministic game approach to decision making under conditions

in the face of uncertainty …………..……………..……………………..

3. Stochastic approach to solving problems of decision making in

conditions of uncertainty ……………………………………………

4. Probabilistic - statistical approach to decision-making in the us-

conditions of uncertainty ……………………………………………..

5. Probabilistic approach to decision-making under conditions of uncertainty

laziness ……………………………………………………………………

6. Fuzzy - stochastic approach to decision making under conditions

uncertainty ……………………………..………………………..

7. Possibility theory and the problem of decision making under conditions

uncertainties …………………….……………………………………

8. Fuzzy - possibilistic approach to decision making in conditions

uncertainties ……………………………………………………….

9. Linguistic approach to decision-making under conditions of uncertainty

divisions..………………………..………………………………….

The control of semi-structured objects is, from the point of view of the classical TAU, a rather complex, practically unsolvable task. This is due to the fact that when building a traditional automatic control system (ACS), it is necessary to first formally describe the control object and form control criteria based on the mathematical apparatus that operates with quantitative categories. If it is impossible to give an exact mathematical description of the object and the criteria for managing it in quantitative terms, the traditional TAU turns out to be inapplicable.

For example, the classical TAU with deterministic and stochastic systems is successfully used to build ACS for aircraft, power plants, etc., but attempts to extend traditional methods to areas such as biosynthesis, multi-phase chemical and technological processes associated with roasting, smelting, catalysis etc., did not give tangible practical results, despite the increasingly complicated mathematical methods of their description.

However, in practice, such semi-structured objects are quite successfully managed by a human operator, who is helped out by the ability to observe, analyze and remember information, draw certain conclusions, etc., and, as a result, make the right decisions in an environment of incomplete and fuzzy information. Thanks to your intellect, a person can operate not only with quantitative(which to a certain extent the machine can), but also with qualitative non-formalized concepts, as a result of which it copes quite successfully with the uncertainty and complexity of the management process. Therefore, the construction of models of approximate human reasoning and their use in ACS is today one of the most important directions in the development of TAU.

There is no doubt that a significant increase in the efficiency of managing complex objects lies in the creation of intelligent ACS capable of reproducing to some extent certain intellectual actions of a person related to the acquisition, analysis, classification of knowledge in the subject area of ​​process control, as well as operating knowledge accumulated by the human operator or the system itself in the course of practical activities to manage the facility.

The need to work in these conditions makes it difficult to use standard automation systems and process control systems. Particularly difficult is the description of the areas of permissible modes of operation of equipment in such conditions when the setting of strict (clear) restrictions for process control systems and automation systems leads to automatic or manual shutdown of these systems. Therefore, it is extremely important to be able to use for the description and formalization of the areas of permissible operating modes of equipment theories of artificial intelligence (AI) and intelligent systems (IS).

Due to the rapid development of computer technology in recent the use of new methods of intelligent control in industry began. And although the first applications of intelligent ACS took place in Europe, such systems are most intensively implemented in Japan. Their range of applications is wide: from controlling industrial robots, distillation plants and blast furnaces to washing machines, vacuum cleaners and microwave ovens. At the same time, intelligent automatic control systems allow improving product quality while reducing resource and energy costs and provide higher resistance to disturbing factors compared to traditional automatic control systems.

An intelligent system is(K.A. Pupkov) a set of technical means and software united by an information process, working in conjunction with a person (a team of people) or autonomously, capable of synthesizing a goal, making a decision for action and finding rational ways to achieve goals.

The main architectural feature that distinguishes intelligent control systems (IMS) from "traditional"is a mechanism for obtaining, storing and processing knowledge to implement its functions.

The creation of intelligent control systems is based on two principles: situational control (control based on the analysis of external situations or events) and the use of modern information technologies for knowledge processing (expert systems, artificial neural networks, fuzzy logic, genetic algorithms, and a number of others).

Economic decisions, depending on the certainty of possible outcomes or consequences, are considered within the framework of three models:

choosing a decision under conditions of certainty, if for each action it is known that it invariably leads to some outcome;

choosing a decision at risk if each action leads to one of the many possible particular outcomes, and each outcome has a calculated or expertly estimated probability of occurrence;

the choice of decisions under uncertainty, when one or another action has as its consequence many particular outcomes, but their probabilities are unknown.

Probabilistic methods provide suitable conditions for decision making and meaningful guarantees of the quality of the choice. It is based on the assumption that judgments about meanings, preferences, and intentions are valuable abstractions of human experience and can be processed to make decisions. While judgments about the likelihood of events are qualified by probabilities, judgments about the desirability of actions are represented by concepts. Bayesian methodology considers the expected utility U(d) as an estimate of the quality of the solution d. Accordingly, if we can choose either action d 1 or d 2 , we calculate U(d 1),U(d 2) and choose the action that corresponds to the largest value. The semantics of utility is to describe risk.

Risk is commonly understood as the probability (threat) of a person or organization losing part of its resources, shortfall in income or the appearance of additional expenses as a result of the implementation of a certain financial policy.

The level of risk is understood as the objective or subjective probability of losses. Objective is a quantitative measure of the possibility of a random event, obtained using calculations or experience, which makes it possible to assess the probability of detecting this event. The subjective one is a measure of the confidence and truth of the stated judgment and is established by an expert.

The risk level is most easily established using attributive assessments such as "high", "medium", "small". A variation of attributive risk assessment is a letter coding. In this case, in order of increasing risk and decreasing reliability, Latin letters from A to D are used.

AAA - the highest reliability;

AA - very high reliability;

A - high reliability;

D is the maximum risk.

The level of risk can be assessed using indicators of accounting and statistical reporting.

Of all possible indicators, the current liquidity ratio (CTL) is the best for this purpose - the ratio of a partner's liquid funds to his debts, which answers the question of whether the partner will be able to cover debts with his active liquid assets.

As a result of the analysis of the situation, cause-and-effect diagrams (“tree of causes”) and dependency diagrams are built. A cause-and-effect diagram is a formal display of the structure of a problem situation in the form of a hierarchically open graph, the vertices of which correspond to the elements of the problem, reflecting the causes of its occurrence, and the arcs correspond to the links between them. The relationship of elements-subproblems is displayed as a "cause - effect" relationship (Fig. 11.1).

OLTR - means of data warehousing and online transaction processing; OLAR - means of operational information processing.

A corporate database organized as a data warehouse is filled with information using OLTR and OLAR technologies. To develop and implement a DSS for semi-structured problems, the following methods and tools should be developed and adapted to its conditions:

a system of signs for registering problem situations;

methods for assessing the degree of criticality of problem situations;

cause-and-effect diagrams for diagnosing the causes of problem situations;

decision table for the formation and selection of decision options;

methods for predicting the results of decisions;

models of the functioning of the enterprise and the external environment.

Fig.11.1. Decision support system model

The most common form of identifying problems using technical and economic indicators is to compare their actual values ​​with normative and average values.

The logical analysis of the problem-causes, located at the lower levels of the hierarchy, shows that in many cases they allow you to form options for solving problems of a higher level. For example, as options for solving the problem of reducing the volume of production and sales of products, alternatives are possible:

price variation;

variation in forms of payment;

reduction in the number of employees;

reduction in the share of semi-fixed costs in the cost of production;

reduction of terms of execution of orders;

strengthening the marketing service.

When there are no statistical data necessary to calculate the objective probability of risk, resort to subjective assessments based on the intuition and experience of experts. J. Keynes introduced the concept of subjective probability. In accordance with the principle of indifference, equally plausible events or judgments must have the same probability, which is mathematically written as:

A ~ B ≡ P(A) = P(B),

where ~ is a sign expressing the attitude of indifference or tolerance.

A more plausible event or judgment must have a greater probability, i.e. if A>B, then P(A)>P(B). Subjective estimates of probability link verbal and quantitative values ​​(Table 4).

Table 4

When carrying out transactions in the securities market, there are forms of risks:

Systematic risk is the risk of a fall in the securities market as a whole. Not associated with a specific security.

Unsystematic risk is an aggregate concept that combines all types of risks associated with a particular security.

Country risk - the risk of investing in securities of enterprises located under the jurisdiction of a country with an unstable social and economic situation, with unfriendly relations with the country of which the investor is a resident. In particular, political risk.

The risk of legislative changes is the risk of losses from investments in securities due to changes in their market value caused by the emergence of new or changes in existing laws.

Inflationary risk is the risk that when inflation is high, the income received by investors from securities will depreciate.

Currency risk is the risk associated with investments in foreign currency securities due to changes in the foreign exchange rate.

Industry risk is a risk associated with the specifics of individual industries.

Regional risk - the risk inherent in single-product areas (agriculture, military, heavy, light industries).

Enterprise risk is the risk of financial loss from investing in the securities of a particular enterprise.

Credit risk is the risk that the issuer of securities will be unable to pay interest on them.

Liquidity risk is the risk associated with the possibility of losses in the sale of a security due to a change in its valuation.

Interest rate risk is the risk of losses that investors may incur due to changes in interest rates.

Capital risk is the risk of a significant deterioration in the quality of a securities portfolio.

There are several popular approaches to making decisions regarding the choice of investment portfolio and allocation of funds. The simplest, conforming, requires that the portfolio be designed to meet the specific requirements of the investing company. In accordance with this approach, the investor makes contributions of a fixed amount in various categories of securities. An assessment of the quality of a firm's securities can be based on the size of the company's capital, its performance indicators, and the contributions of other organizations.

Strategies in which assets are mixed according to the phases of the national and global economy are called tactical asset allocation. Tactical placement of assets is conformal, while funds are invested in those assets that have fallen in price. When the proportion of funds invested in different asset classes is based on some forward looking estimates of macroeconomic parameters, this approach is called scenario allocation.

The most widely used approach to portfolio selection is the mean-variance approach proposed by Harry Markowitz. The main idea is to consider the future income brought by a financial instrument as a random variable, that is, the income on individual investment objects randomly change within certain limits. Then, if in some way to determine for each investment object well-defined implementation probabilities, it is possible to obtain a distribution of income generation probabilities for each investment alternative. To simplify, the Markowitz model assumes that returns across investment alternatives are normally distributed.

According to the Markowitz model, indicators are determined that characterize the amount of investment and risk, which allows you to compare different alternatives for investing capital in terms of goals and thereby create a scale for evaluating various combinations. In practice, the most probable value is used as the scale of the expected income from a number of possible incomes, which, in the case of a normal distribution, coincides with the mathematical expectation.

At the heart of the Markowitz model, portfolio selection appears to be an optimization problem:

under restrictions

,

,

where n– number of available securities; part of the portfolio contained in securities of the type i;R i= E(r i) is the expected value of income on securities i;R p =E(r p) – target level of expected portfolio income; σ ij– covariance of income on securities i and j;V p is the variance of portfolio income.

This problem is a quadratic programming problem.

TOPIC 13. INTELLIGENT CONTROL SYSTEMS

A new generation of systems - intelligent systems (IS) - brought to life other principles for organizing system components, other concepts, terms, blocks appeared that were not previously encountered in developments and, therefore, in the scientific literature.

Intelligent systems are able to synthesize a goal, make a decision for action, provide action to achieve the goal, predict the values ​​of the parameters of the result of an action and compare them with real ones, forming feedback, adjust the goal or control

Figure 13.1 shows a block diagram of the IS, where two large blocks of the system are highlighted: the synthesis of the goal and its implementation.

In the first block, based on the active evaluation of information received from the sensor system, in the presence of motivation and knowledge, a goal is synthesized and a decision is made for action. Active evaluation of information is carried out under the influence of trigger signals. The variability of the environment and the system's own state can lead to a need for something (motivation), and if knowledge is available, a goal can be synthesized.

The goal is understood as an ideal, mental anticipation of the result of an activity. By continuing to actively evaluate information about the environment and the system's own state, including the control object, when comparing options for achieving the goal, you can make a decision for action.

Further, in the second block, the dynamic expert system (DES), based on current information about the environment and its own state of the IS, in the presence of a goal and knowledge, performs an expert assessment, makes a decision on management, predicts the results of an action, and develops a management.

The control presented in coded form is converted into a physical signal and fed to the actuators.

The control object, receiving a signal from the actuators, performs one or another action, the results of which, presented in the form of parameters, are fed through the feedback loop 2 to the DES, where they are compared with the predicted ones. At the same time, the parameters of the result of an action, interpreted in accordance with the properties of the goal and entering block I, can be used for an emotional assessment of the result achieved: for example, the goal was achieved, but the result is not liked.

If the goal is achieved in all respects, then management is reinforced. Otherwise, a control correction occurs. When the goal is unattainable, the goal is corrected.

It should be noted that with sudden changes in the state of the environment, or the control object, or the system as a whole, it is possible to synthesize a new goal and organize its achievement.

The IS structure, along with new elements, contains traditional elements and connections; a dynamic expert system occupies a central place in it.

Block 1 - goal synthesis Block II - goal realization

Figure 13.1 - Structural diagram of the IP

Formally, IS is described by the following six expressions:

T X S M T ;

T M S ST ;

C T S R T;

T X= (A T)X T+(B T)U T;

T Y = (D T) X T;

T R Y With T ,

where T is a set of time points;

X, S, M, C, R and Y - sets of states of the system, environment, motivation, goal, predicted and real result;

A, B and D - matrixes of parameters;

Intelligent conversion operators using knowledge.

This description combines representations of system objects in the form of a set of values, or a set of statements, or some other forms.

The dynamic properties of the IS can be described in the state space. Intelligent operators that implement perception, representation, concept formation, judgments and inferences in the process of cognition are a formal means of processing information and knowledge, as well as making a decision. All these aspects should form the basis for the construction of DES, functioning in real time and in the real world.

A dynamic expert system is a kind of complex entity capable of assessing the state of the system and the environment, comparing the parameters of the desired and real results of an action, making a decision and developing a control that contributes to the achievement of the goal. To do this, DES must have a stock of knowledge and have methods for solving problems. The knowledge transferred to an expert system can be divided into three categories:

1) conceptual (at the level of concepts) knowledge is knowledge embodied in the words of human speech or, more specifically, in scientific and technical terms and, of course, in the classes and properties of environmental objects behind these terms. This also includes connections, relationships and dependencies between concepts and their properties, and abstract connections, also expressed in words and terms. Conceptual knowledge is the sphere, mainly, of the fundamental sciences, given that the concept is the highest product of the highest product of matter - the brain;

2) factual, subject knowledge is a set of information about the qualitative and quantitative characteristics of specific objects. It is with this category of knowledge that the terms "information" and "data" are associated, although such use of these terms somewhat diminishes their significance. Any knowledge carries information and can be represented as data; factual knowledge is what computers have always dealt with and what they have dealt with the most so far. The modern form of data accumulation is usually called databases. Of course, to organize databases, to search for the necessary information in them, one must rely on conceptual knowledge;

3) algorithmic, procedural knowledge - this is what is commonly called the words "skill", "technology", etc. In computing, algorithmic knowledge is implemented in the form of algorithms, programs and subroutines, but not all, but those that can be transferred from the hands in the hands and used without the participation of the authors. Such implementation of algorithmic knowledge is called a software product. The most common forms of a software product are application software packages, software systems, and others focused on a specific area of ​​DES application. The organization and use of application packages is based on conceptual knowledge.

It is clear that conceptual knowledge is a higher, defining category of knowledge, although, from the point of view of practice, other categories may seem more important.

This is probably why conceptual knowledge is rarely embodied in a form that can be processed on computers. And if it is embodied, it is most often incomplete and one-sided. In most cases, the person remains the bearer of conceptual knowledge. This slows down the automation of many processes.

Representations of conceptual knowledge, or rather, systems that implement all three categories of knowledge, but highlight conceptual knowledge to the fore and work on the basis of its intensive use, are called knowledge bases.

The creation and widespread use of knowledge bases in IS is one of the most urgent tasks. The conceptual part of the knowledge base will be called the domain model, the algorithmic part - the software system, and the factual part - the database.

The next function of DES is problem solving. A problem can be solved by a machine only if it is formally stated - if a formal specification is written for it. The latter should be based on some knowledge base. The domain model describes the general environment in which the task arose, and the specification describes the content of the task. Taken together, they make it possible to establish what abstract connections and dependencies, in what combinations and in what sequence should be used to solve the problem.

Application programs are the concrete tools behind these dependencies and also contain algorithms for solving the resulting equations. Finally, the database supplies all or part of the initial data for the execution of these algorithms; the missing data must be contained in the specification.

These three parts of knowledge bases correspond to three stages of solving the problem:

1) construction of an abstract solution program (including the emergence of a problem, its formulation and specification);

2) translation of the task into a suitable machine language;

3) broadcasting and execution of the program.

The construction of an abstract program is associated with the representation and processing of conceptual knowledge in IS and, by definition, is the property of artificial intelligence.

Artificial intelligence is associated with the processing of texts, oral messages in natural language, with the analysis and processing of information (recognition of all types of images, theorem proving, logical inference, etc.).

The functions of the DES are also the evaluation of the results of solving the problem, the formation of parameters of the future result of the action, the decision to control, the development of control and the comparison of the parameters of the desired and actual results. It provides for the modeling of processes to assess the possible consequences and the correctness of the solution of the problem.

Note that in real cases there is a problem of describing the objects under study. Such a description should not be considered part of the specification of the task, since, as a rule, many tasks are assigned to one object, which, of course, must be taken into account when forming the knowledge base. In addition, it may turn out that the problem that has arisen cannot be completely solved automatically, for example, due to the incompleteness of the specification or description of the object.

Therefore, in the IS, at certain stages, an interactive mode of operation with DES is expedient. It must be remembered that the domain model describes the general environment (knowledge), and the specification describes the content of the task. Very important problems are the creation of a unified software environment and the synthesis of algorithms directly according to the problem statement.

Depending on the goal facing the IS, the knowledge base, algorithms for solving a problem, making a decision, developing control can, of course, have a different representation, depending, in turn, on the nature of solving problems. Accordingly, three types of DES can be seen. The structure of the DES of the first type is shown in Figure 13.2.

Figure 13.2 - The structure of the DPP of the first type

It is assumed here that conceptual and factual knowledge accurately reflects the processes and information related to a certain subject area.

Then the solution of the problem that arises in this area will be obtained on the basis of rigorous mathematical methods, in accordance with the formulation and specification. The results of the decision study and forecast are used to obtain an expert opinion and decide on the need for management. Then, on the basis of a suitable control algorithm available in the knowledge base, a control action is formed.

The effectiveness and consistency of this impact, before it enters the control object, is evaluated using a simulation mathematical model. Evaluation should be faster than real processes in IS.

However, DES implementing decision making are complex software systems designed for automatic decision making or to assist decision makers, and in the operational management of complex systems and processes, as a rule, they operate under severe time constraints.

Unlike DES of the first type, designed to find the optimal solution and based on rigorous mathematical methods and optimization models, DES of the second type are mainly focused on solving problems that are difficult to formalize in the absence of complete and reliable information (Fig. 13.3). Here, expert models are used, built on the basis of the knowledge of experts - specialists in this problem area, and heuristic methods for finding a solution.

One of the main problems in designing a DES of the second type is the choice of a formal apparatus for describing decision-making processes and building on its basis a decision-making model that is adequate to the problem area (semantically correct). Production systems are usually used as such apparatus. However, the main research is carried out in the context of an algorithmic (deterministic) interpretation of a production system with its inherent sequential scheme for finding a solution.

The resulting models are often inadequate for real problem areas characterized by non-determinism in the process of finding a solution. The way out of this situation is parallelism in search.

In reality, one should focus on combining DES of the first and second types into a computational and logical DES of the third type, where the knowledge base combines description in the form of strict mathematical formulas with expert information, and also, accordingly, mathematical methods for finding a solution with non-strict heuristic methods, and the weight of one or the other the component is determined by the possibility of an adequate description of the subject area and the method of finding a solution (Fig. 13.4).

Figure 13.3 - The structure of the DES of the second level

During the development of DES, the following problems arise:

1. determination of the composition of the knowledge base and its formation;

2. development of new and use of known theories and methods for describing information processes in IS;

3. development of ways to represent and organize the use of knowledge;

4. development of algorithms and software with parallelization and the use of "flexible logic";

1. finding suitable computing environments for the implementation of parallel algorithms in the formation of DES.

Figure 13.4 - Structure of the DES of the third level

Along with the above, it is important to note that DES should have the property of adapting to a dynamic problem area, the ability to introduce new elements and relationships into the description of situations, change the rules and strategies for the functioning of objects in the process of decision-making and development of control, work with incomplete, fuzzy and contradictory information and etc.

Dynamic expert systems function as part of ISs with feedback, and therefore it is important to ensure the stable operation of such ISs.

From traditional positions, we can assume that the duration of the DES response to input actions, i.e. the time spent on processing the input information and generating the control action is a pure delay. On the basis of frequency analysis, it is possible to estimate the change in the phase properties of the system and thereby determine the stability margin. If necessary, the system can be corrected by means of filters.

However, from the point of view of the classical control theory, ISs are multi-object multiply connected systems, the stability analysis of which by conventional methods is very difficult.

At present, the theory of robust control (-control theory, -control) is one of the intensively developing branches of control theory. Relatively young (the first works appeared in the early 1980s), it arose from urgent practical problems in the synthesis of multidimensional linear control systems operating under conditions of various kinds of disturbances and parameter changes.

It is possible to approach the problem of designing the control of a real complex object operating under uncertainty in a different way: do not try to use one type of control - adaptive or robust. Obviously, one should choose the type that corresponds to the state of the environment and the system, determined from the information available to the system. If it is possible to organize the receipt of information during the functioning of the system, it is advisable to use it in the management process.

But the implementation of such a combined control, until recently, encountered insurmountable difficulties in determining the algorithm for choosing the type of control. The progress achieved in the development of artificial intelligence problems makes it possible to synthesize such an algorithm.

Indeed, let us set the task: to design a system that uses adaptive and robust control and selects the type of control based on artificial intelligence methods. To do this, we will consider the features of both types and, taking into account their specific qualities, we will determine how a combined control system can be built.

One of the basic concepts in the theory of robust control is the concept of uncertainty. The uncertainty of an object reflects the inaccuracy of the object model, both parametric and structural.

Let us consider in more detail the forms of setting uncertainty in robust control theory using a simple system - with one input and one output (Figure 13.5).

The signals have the following interpretation: r - setting input signal; u - input signal (input) of the object; d - external disturbance; y - output signal (output) of the object, measured.

Figure 13.5 - System with one input and one output

In -control theory, it is convenient to specify uncertainty in the frequency domain. Let us assume that the transfer function of a normal plant is P, and consider a perturbed plant whose transfer function,

,

where W is a fixed transfer function (weight function);

is an arbitrary stable transfer function that satisfies the inequality .

Such a perturbation will be called admissible. Below are some options for uncertainty models:

(1+W)P; P+W; P/(1+WP); P/(1+W).

Appropriate assumptions must be made for the quantities and W in each case.

The uncertainty of the input signals d reflects the different nature of external disturbances acting on the object and the controller. An indefinite object can thus be viewed as a set of objects.

Let us choose some characteristic of systems with feedback, for example, stability. Regulator C is robust with respect to this characteristic if any of the set of objects defined by uncertainty has it.

Thus, the concept of robustness implies the presence of a controller, a set of objects, and the fixation of a certain characteristic of the system.

In this work, we will not touch upon the entire set of problems solved within the framework of control theory. We will only touch on the problem of minimum sensitivity: the construction of such a controller C that stabilizes the closed system and minimizes the influence of external disturbances on the output y, in other words, minimizes the norm of the matrix of transfer functions from external disturbances to the output y.

One of the features of solving this, and indeed the whole set of robust control problems, is the fact that in the process of designing the controller, we set restrictions on the input actions and the uncertainty of the object in the form of inequalities in advance.

During the operation of a robust system, information about the uncertainties in the system is not used for control.

Naturally, this leads to the fact that robust systems are conservative and the quality of transient processes sometimes does not satisfy the developers of these systems.

Like a robust adaptive control system, it is built for objects, information about which or about the impacts on which is not available at the beginning of the system operation. Most often, the property of adaptation is achieved through the formation in an explicit or implicit form of a mathematical model of an object or an input action.

This is the difference between both exploratory adaptive control, which is based on the search and retention of the extremum of the control quality index, and searchless, which is based on compensation for the deviation of the actual changes in the controlled coordinates from the desired changes corresponding to the required level of the quality index. Further, according to the refined model, the adaptive controller is adjusted.

Thus, the main feature of adaptive control systems is the ability to obtain information in the process of operation and use this information for control.

Moreover, in adaptive systems, a priori information about the uncertainty in the system is always used. This is the fundamental difference between the adaptive approach and the robust one.

Consider the simplest adaptive control system that provides tracking of the input signal in the presence of noise at the input of the object (Figure 13.6).

Picture. 13.6 - Adaptive control system

The formal difference from the circuit in Figure 13.5 is the adaptation block A, which, based on the output signal of the object and the signal characterizing the given quality, generates a signal for adjusting the coefficients of the adaptive controller.

Keeping in mind the shortcomings of each of the regulators, it is advisable to try to use their advantages by proposing a combined scheme for managing the object. An adaptive system, with the help of an adaptation block, generates some information about the state of the external environment. In particular, in the case under consideration, one can obtain information about the external perturbation d. The control algorithm C a corresponds to the current state of the external environment, according to the criterion laid down in the adaptation block. But the adaptive system requires the input signal r to have a sufficiently wide frequency range and imposes severe restrictions on the value and frequency spectrum of the external disturbance signal d. Therefore, adaptive systems can operate only in narrow ranges of the input signal r and external disturbance d. Outside these ranges, the adaptive system has a low quality of control and may even lose stability.

The above properties of robust and adaptive control lead to the conclusion that in the process of system operation, in some cases it is advantageous to use robust control, in others - adaptive, i.e. be able to combine control depending on the state of the external environment.

Combined management. The main question in the design of combined control systems is how, on the basis of what knowledge (information) to select one or another type of control.

The greatest opportunities for this are the methods of artificial intelligence. Their advantage over simple switching algorithms is the use of a wide range of data and knowledge to form an algorithm for choosing the type of control.

If we formally combine the circuits shown in Figures 13.5, 13.6, we get a combined control circuit (Figure 13.7).

As can be seen from the figure, the control signal and must switch from a robust controller to an adaptive one and vice versa - as the environment changes during the operation of the system. Using the methods of the theory of intelligent systems, it is possible to ensure the transition from one type of control to another, depending on the operating conditions of the system.

Figure 13.6 - Combined control scheme

Let us first consider what information can be used for the operation of the intelligent unit of the system. As is known, systems with one input and one output are well described in the frequency domain. Therefore, it is natural to use the frequency characteristics to organize the decision-making process when choosing the type of control.

As mentioned above, the frequency response of a system with robust control corresponds to the worst combination of parameters in the region of uncertainty. Therefore, robust control can be taken as one of the boundaries of the selected control.

Another limit is determined by the capabilities of the system under study (speed of the drive, power-to-weight ratio, etc.). Between these two boundaries is the area where it is reasonable to use adaptive control.

Figure 13.7 - Combined control scheme

Since the adaptive algorithm is sensitive to the initial stage of the system operation, it is advisable to use robust control at this stage, which is sufficiently insensitive to the rate of change of external noise. But its disadvantage is the long duration of transient processes and large allowable values ​​of the output coordinate under the influence of interference.

After some time, it makes sense to switch robust control to adaptive control.

Adaptive control allows you to more accurately track the input signal in the presence of information about the interference. Adaptive control is demanding on the richness of the spectrum of the input signal, and, for example, with slowly changing signals, adaptation processes may be disrupted or they may be severely slowed down. In such a situation, it is necessary to switch back to robust control, which guarantees the stability of the system.

It follows from the foregoing that for the functioning of the system it is necessary to have information about the frequency spectrum of the useful interference signal and about the signal-to-noise ratio.

In addition, preliminary information is required about the frequency spectrum on which the adaptive system operates, and about the particular characteristics of the control object at the boundaries of the uncertainty region. From this information, it is possible to form a database in which information, individual for each class of objects, is entered in advance. Information about the frequency spectrum of the useful signal, interference and signal-to-noise ratio enters the database as the system operates and is constantly updated.

The contents of the database can be used in the knowledge base, which is formed in the form of rules. Depending on the specific properties of the system, two types of control switching can be set. The required rules are formed in one of the logical systems suitable for the case under consideration.

Having databases and knowledge, it is possible to develop a decision-making mechanism that will ensure the correct choice of the type of control depending on the operating conditions of the system.

Figure 13.8 - Structural diagram of a system with an intelligent unit (IB)

The intellectual part of the system operates discretely, at specified time intervals. Figure 13.8 shows a block diagram of a system with an intelligent IS block that provides a choice of control type.

The input of the block receives the signal r and the measured output signal of the object y. In the pre-processing block of the information of the BPOI, the frequency characteristics of the input signal r(w) and external disturbance d(w), the relative position of the spectra r(w) and d(w), and the characteristic signal-to-noise ratio r(w)/d(w). All this information enters the database database. The BPR decision block, using the generated KB knowledge base and DB data, develops a decision, according to which one of the control types is switched on. At the next interval, the process is repeated using new data.

Program No. 14 of fundamental research of OEMMPU RAS

"ANALYSIS AND OPTIMIZATION OF FUNCTIONING OF SYSTEMS OF MULTILEVEL, INTELLIGENT AND NETWORK CONTROL UNDER UNCERTAINTY"

1. Rationale for the Program

1.1. Scientific and practical significance

The intensive development of technology (networking, miniaturization of computers, increasing their speed, etc.) imposes new requirements on modern control systems and opens up new opportunities both at the level of embedded control systems (at the level of large dispatch centers) and at the network level (communication network, group) interaction of decentralized multi-agent systems. Control systems are increasingly acquiring the character of information and control systems and are being studied at the intersection of theories of control, computing and communication. Thus, taking into account the properties of communication channels (communications) is necessary, for example, in decentralized (multi-agent) systems, and the characteristics of the built-in computer are important when implementing such intellectual functions in multilevel control systems as technical vision, action planning, training, multi-criteria decision making, reflection and etc. In particular, the intellectualization of control is designed to increase the degree of autonomy of the functioning of systems, when the absence of quantitative models of dynamics or disturbances in the functioning of the control object, causing the loss of the adequacy of quantitative models (for example, equations describing the evolution of a complex system), enhance the role of qualitative (the so-called " knowledge”, for example, logical-linguistic) models of the object and environment used at the upper levels of the control system.

The program is aimed at solving fundamental problems arising in the priority areas of science, technology and engineering of the Russian Federation. The task is to obtain new fundamental and applied results in the field of control theory for complex technical, human-machine and other systems, taking into account the uncertainty and lack of initial information, including: the theory of analysis and synthesis of stochastic systems, the theory of creating motion control systems and technological processes, with current diagnostics and control over the technical condition, as well as the theory of creating automated design systems and intelligent control based on modern information technologies.

Due to the variety of use of control theory, analysis and optimization in various applications (transport, logistics, manufacturing, aviation and space systems, submarines and surface ships, etc.), it is necessary to take into account a large number of complexity factors, such as:

multilevel management,

decentralization,

non-linearity

The multiplicity

Distribution of parameters

variability of processes in space and time,

high dimension,

Heterogeneity of the description of subsystems,

multimodality,

The presence of impulsive influences,

the presence of coordinate-parametric, structural, regular and singular perturbations,

the use of deterministic and probabilistic models for describing the uncertainty of information about the state vector and system parameters, about the properties of measurement errors and the external environment,

the presence of delay effects in the control or object,

· general structural complexity of modern control systems.

To achieve the set goal and solve the main tasks, the Program includes research and development in the following main areas:

1. Analysis and optimization of functioning in different time scales of multilevel control systems with incomplete information.

2. Management and optimization in multilevel and decentralized systems of organizational and technical nature.

2.1. Control and optimization in network-centric systems.

2.2. Intelligent control of moving objects.

2.3. Modeling and optimization of multilevel real-time information and control systems.

Direction 1. Analysis and optimization of functioning in different time scales of multilevel control systems with incomplete information

The complexity of many modern control systems often does not allow obtaining a complete description of the processes occurring inside the system and its interaction with the environment in advance. As a rule, real systems are described by nonlinear equations of dynamics, and quite often mathematical models of control systems take into account only the permissible ranges of changes in the parameters and characteristics of individual elements without specifying these parameters and characteristics themselves.

In addition, in some systems, in particular, micromechanical and quantum ones, the use of classical methods of description in continuous or discrete time is difficult because the emerging internal and/or external interaction forces, as well as control actions, are transient, impulsive in nature and cannot be accurately calculated. . The system, as it were, functions on different time scales: real (slow) and fast (impulse). Such temporal multiscaleness is an internal property of many modern control systems, including systems with multilevel control, in which the upper levels use qualitative and discrete models, and the lower levels use more often quantitative models with continuous time.

For this reason, the development of methods for mathematical formalization of the description of the functioning of such systems in hybrid (continuous-discrete) time, the study of their properties for controllability and stability under conditions of incomplete information, counteraction and non-standard restrictions on controls and phase variables is an urgent task. The same urgent task is the development of methods for the synthesis of optimal control of such continuous-discrete systems, both deterministic and stochastic.

In addition, under conditions of uncertainty and lack of a priori information, the problems of optimizing the process of collecting and processing information (observation management and optimal filtering) are very relevant.

Direction 2. Management and optimization in multilevel and decentralized systems of organizational and technical nature

2.1. Control and optimization in network-centric systems

Modern complex organizational and technical systems are characterized by high dimensionality, decentralization, multi-level management, the need for effective planning of activities, taking into account training, the multicriteria of decisions made and the reflection of controlled entities.

Problems of planning and control of discrete and continuous distributed multi-connected systems of large dimension are also characterized by processes of different scales not only in time, but also in distribution and scale in space and represent one of the most complex and time-consuming classes of optimization problems. For this reason, it is advisable to develop research methods and approaches to finding exact and approximate solutions, as well as simulation tools for use in decision support systems for planning, designing and managing complex technical, organizational (including transport and logistics) and information systems.

To manage the group interaction of components of decentralized organizational and technical systems (network-centric systems, production systems, computing, telecommunication and other networks, etc.) under conditions of restrictions on communication channels and the complexity of calculations, the characteristics of information processing processes, as well as restrictions on decision-making time, computational capabilities and throughput of communication channels. Therefore, it is relevant to develop methods for optimizing (taking into account the above limitations) the structure of complex organizational and technical systems, including taking into account many criteria at the same time: the detail of the initial data, the efficiency of collecting information, planning and reflective decision-making, the limited performance of individual computers, reducing duplication of work , as well as the proportion of auxiliary calculations associated with the maintenance of data transmission.

Multi-level and decentralized systems are characterized by distributed decision-making in real time in the face of information counteraction, as well as incompleteness and heterogeneity of information, often of a multi-criteria qualitative and subjective nature. For this reason, it is necessary to develop methods for creating adequate information support systems and supporting the adoption of strategic and operational decisions in conditions of incomplete information and counteraction. For this, it is advisable, in particular, to develop: multi-agent models of dynamic organizational and technical systems, including network models with conflicting agents, models of group behavior and its prediction, assessment of the balance of interests and the formation of coalitions in these systems, as well as the development of information technologies and information presentation tools about the external environment and knowledge of intelligent agents.

2.2. Intelligent control of moving objects

To solve the tasks set, it is far from always possible to create quantitative models, therefore, along with traditional methods, the Program uses artificial intelligence methods. Artificial intelligence, as a field of knowledge, has undergone a huge leap over the past fifty years both in the development and refinement of the very concept of intelligence, and in the field of practical application of artificial intelligence in various fields of human activity: in technology, economics, business, medicine, education, etc. Many theoretical provisions and methods of artificial intelligence have been transformed into applied intelligent technologies based on knowledge.

The peculiarity of the modern generation of intelligent systems is that they are based on a complex model of the external environment, which takes into account both quantitative information and qualitative models - knowledge about the possible behavior of various objects of the external environment and their interconnections. The use of such models became possible due to the development of methods for representing knowledge, methods for integrating data from different sources, a significant increase in the speed and memory of computers.

The presence of a model of the external environment allows modern intelligent control systems for moving objects to make decisions in conditions of multi-criteria, uncertainty and risk, and the quality of these decisions can exceed the quality of decisions made by a person who is in conditions of information overload, limited time and stress.

In this regard, an urgent task is to develop new tools and methods for the development of intelligent control of moving objects in the presence of the factors listed above.

2.3. Modeling and optimization of multilevel real-time information and control systems

The relevance of research in this direction is due to the need to develop methods for the analysis and synthesis of multilevel open modular real-time information and control systems (IMS RT) of multi-mode and multi-purpose objects operating under conditions of uncertainty, structural disturbances and emergency situations (ESS). Among these objects of control are critical objects and systems of responsible use that determine the security of the state.

It is obvious that the problems and tasks of creating systems of this class can be successfully solved on the basis of the development of a unified theory and applied software-oriented methods of dynamic and scenario analysis and synthesis of the structure of such systems, their algorithmic, software and information support, and mechanisms for developing effective management actions. These, first of all, include the development of a formalized methodology for designing open information and control systems, including models and methods for synthesizing an optimal, according to various efficiency criteria, modular structure of object-oriented IMS RT with an open architecture. Based on the results obtained at the stage of dynamic analysis, the optimal functional modular structure of data processing and control is synthesized, i.e., the optimal composition and number of RT IMS modules are determined, the system interface is synthesized, and the structure of its software and information support for processing input flows of applications is determined.

To plan actions and support decision-making under conditions of uncertainty, structural disturbances and emergency situations, it is advisable to use the methods of scenario analysis and synthesis of effective control actions in the IMS RT. In this case, the mathematical model of the propagation of structural disturbances and emergency situations will be formed in the language of weighted or functional sign graphs. Based on this model, rational scenarios for managing objects will be synthesized using the concepts of operability potential, durability and survivability of their constituent elements. Synthesis of scenarios for eliminating the causes and consequences of NSS in multi-mode target facilities will be carried out taking into account dynamically determined time and resource constraints. It is also necessary to develop formulations and methods for solving inverse problems of managing the survivability of multi-mode and multi-purpose objects operating under conditions of uncertainty, structural disturbances and emergency situations.

The specificity of control systems and objects noted above, the scientific and practical significance of solving control, analysis and optimization problems for them allow us to formulate the following main goals and objectives of the Program.

1.2. Main goals and objectives

The main goal of the Program is to solve the fundamental problems of control theory that hinder the implementation of promising projects of national importance in the field of control of complex dynamic and intelligent systems with applications to control the movement of technical objects and processes in technological and organizational systems.

Research will be conducted on the following generalized topics.

Direction 1

· Development of methods for stabilizing nonlinear systems in situations of incomplete measurement of coordinates and restrictions on the admissible structure of control forces.

· Development of methods for robust and adaptive monitoring and control under conditions of deterministic, probabilistic and other models of uncertainty of the parameters of the control object and the operating environment.

· Development of methods and algorithms for qualitative and quantitative analysis of continuous, discrete and multilevel continuous-discrete dynamic models and control synthesis based on the reduction method with vector and matrix comparison functions and model transformations.

· Investigation of the problem of optimal control of a new class of mechanical systems moving in resisting media by changing the configuration or movement of internal bodies.

· Development of methods for mathematical formalization and solution of problems of impact interaction of mechanical systems in the presence of dry friction.

· Development of optimal control methods for discrete-continuum and impulsive dynamical systems.

· Development of methods for guaranteed control of non-linear objects exposed to uncontrolled disturbances in the form of dynamic games.

· Development of the theory of control of quantum systems.

· Development of methods and algorithms for analyzing dynamic properties such as stability, invariance, dissipativity for state estimation and synthesis of multilevel control systems with a heterogeneous description of the dynamics of processes at different levels.

Direction 2.1

· Methods for solving control problems for large-scale network-centric systems with distributed parameters and multi-scale (in space and time) processes.

· Models and methods of communication-network decentralized intelligent control of distributed projects and programs.

· Methods for optimizing the structure of multilevel and decentralized systems.

· Methods and structures of computer implementation of network-centric control in a mathematically homogeneous space of distributed and parallel computing.

· Models and methods of group decision-making based on incomplete, heterogeneous, qualitative and subjective information.

· Models and methods of planning and managing complexes of interrelated operations in complex technical and transport and logistics systems.

· Development of principles, architecture, methods and algorithms for creating distributed intelligent software systems based on multi-agent technologies.

· Development of models and methods of information management in multi-agent network structures.

Direction2.2

· Development of generalized models of situational control, reflecting the features of inclusion in the structure of models of fuzzy, neural network and logical-dynamic elements.

· Development of a route planning method that provides the property of communication stability of a group of controlled dynamic objects, heterogeneous (quantitative-qualitative) in their model representation.

· Development of methods for analysis and synthesis of adaptive real-time modeling platforms that take into account non-linearity, multi-connectivity, high dimensionality of control objects with application to marine moving objects.

· Optimization of intelligent systems for multilevel control of moving objects in a conflict environment, taking into account their group interaction, multicriteria, uncertainty and risk.

· Development of methods for providing technical vision for intelligent control systems.

· Development of methods for intelligent control of dynamic objects performing complex maneuvering based on the organization of forced movement in the space of system states.

Direction2.3

· Models and methods for analyzing and optimizing the modular structure of object-oriented multilevel real-time information and control systems with an open architecture under conditions of uncertainty and structural disturbances.

· Methods of analysis and optimization of modes of electric power systems and their management.

· Models and methods of the scenario-indicator approach to the search for points of vulnerability for management tasks.

· Methods of modeling, analysis and optimization of multi-mode control processes for moving objects.

· Development of methods and algorithms for intelligent identification of non-linear non-stationary objects to improve management efficiency by forming a technological knowledge base based on a priori information about the control object.

· Geoinformation technologies for modeling natural and technogenic complexes in the problems of managing the ecosystems of megacities.

· Analysis and optimization of information support for navigation and control systems.

· Models and methods of managing production processes.

The results of the developed theory and methods of analysis and synthesis of control systems will be used in the following areas:

· traffic control in aviation and astronautics, land and sea facilities, vehicles;

· multi-agent network-centric systems, production systems, computing, telecommunication and other networks ;

· transport and logistics systems ;

· global energy, gas transportation and other large-scale infrastructure systems;

· systems of information support for the tasks of management and support for the adoption of strategic and operational decisions in conditions of incomplete information and counteraction.

The fundamental problems of the theory of construction of control systems require their intensive development. The development of research in this direction will allow:

The development of the theoretical foundations for solving the complex triune problem of control-computing-communication (problem - " control- Computation- communication") for complex information and control systems, including in conditions of restrictions on communication channels and failures of subsystems;

Solve the problems of managing fundamentally new objects and processes related to moving objects, special-purpose objects, technological and organizational systems;

Create effective methods for functional diagnostics and ensuring the fault tolerance of control systems for aircraft and other moving objects, as well as the dynamic stability of electric power systems;

To improve the quality, speed up and reduce the cost of developing design solutions through algorithmization and automation of the process of developing control systems.

Hereinafter, control is understood in a broad sense, including communication-network, group, distributed control (in English literature - control in networks, control over networks, distributed control, etc.)

One of the ways out in this situation is to use intelligent control methods, which involve the rejection of:

• - the need to obtain an accurate mathematical model of the object;
• - orientation to the use of “rigid” (as a rule, linear) algorithms for the formation of control actions;
• - striving at all costs to use the methods of synthesis known to the developer, which previously positively proved themselves for other, simpler classes of objects.

Before moving on to intelligent control, one cannot fail to note the world-recognized and classic trends in machine control developed by domestic scientific schools. These are the works of Balakshin B.S., Bazrov B.M., Bzhozovsky B.M., Gornev V.F., Morozov V.P., Kolosov V.G., Ratmirov V.A., Solomentsev Yu.M., Pusha V.E., Sosonkina V.L., Timiryazev V.A., Zakovorotny V.L., Tugengold A.K. and etc. . In particular, works are devoted to the creation of adaptive control systems for machine tools in terms of product manufacturing quality. Flexible automated production (FAP), which makes it possible to reduce costs, increase reliability and flexibility (ability to reconfigure) functioning with frequent changes in control programs, are described in detail in the works. The creation of CNC programs and the features of their implementation as part of an integrated automated production are considered in.

A detailed analysis of modern methods of controlling technological processes and equipment is given in the work, which shows the achievements of modern control theory, in particular, the use of methods for the analytical design of regulators (Letov A.M. and others), modal control (Pospelov G.S. and others. ), inverse problems of dynamics (Krutko P.D. et al.), invariant control (Shchipanov G.V., Kulebakin V.S., Petrov B.N. et al.), adaptive control (Tsipkin Ya.D. and etc.), etc. It is noted that the pinnacle of representations in the synthesis of control systems is the analytical design of controllers. A fundamentally different way of building control systems, based on the synergetic control theory, is described in. The principles of evolutionary transformations and self-organization, based on a synergistic approach, in the control system, the state coordinates of which interact with the environment, are set out in the works. All the mentioned approaches and principles of synthesis of control systems have their advantages and disadvantages, but the common thing for all of them is that they are based on the mathematical model of the control object obtained in one way or another, and the mathematical model is a system of differential or difference equations that describe the physical the essence of processes or objects. A fundamentally different approach to control is to use mathematical models of knowledge about the controlled object, i.e. use of intelligent control methods. With regard to TO, this area of ​​research is presented in the works.

Intelligent control is based on the idea of ​​building highly organized ACS based on the use of models of variable complexity and uncertainty, with the performance of such intellectual functions inherent in a person as decision making, behavior planning, learning and self-learning in a changing environment. Learning should be understood as the ability of the system to improve its behavior in the future (in relation to TO - to improve the quality of processing), based on the experimental information that it received in the past, about the results of interaction with influencing factors. Self-learning is learning without external adjustment, i.e. without instructions from the "teacher" - the operator.

An intelligent control system (IMS) is one in which knowledge about the unknown characteristics of the control object and the environment is formed in the process of learning and adaptation, and the information obtained from this is used in the process of automatic decision making so that the quality of control improves.

A necessary feature of the IMS is the presence of a knowledge base containing information (facts), models and rules that allow you to clarify the task of management and choose a rational way to solve it. Intelligent systems are often referred to as knowledge-based systems. Depending on the nature of the implemented intellectual functions, i.e., on the level of intelligence, IMS are distinguished, intellectual “in big” and “in small”.

According to control systems, intelligent "in the big" are systems organized and functioning in accordance with the following five principles (in their entirety).

• 1. Interaction with the real outside world using information communication channels.
• 2. Fundamental openness of the system in order to increase intelligence and improve their own behavior.
• 3. The presence of mechanisms for predicting changes in the external world and the system's own behavior in a dynamically changing external world.
• 4. The presence of a multi-level hierarchical structure built in accordance with the rule: increasing intelligence and reducing the requirements for model accuracy as the level of hierarchy in the system increases (and vice versa).
• 5. Persistence of functioning (possibly with some loss of quality or efficiency, i.e. with some acceptable degradation) when links are broken or control actions are lost from higher levels of the hierarchy of the control structure.

Control systems that are intelligent “in the small” do not satisfy the principles listed above, but use knowledge (for example, in the form of rules) in the course of their functioning as a means of overcoming the uncertainty of input information, inaccuracy in the description of the controlled object or its behavior.

Based on all of the above, the following conclusion can be drawn. With a multitude of factors affecting the achievement of the quality of machining on metal-cutting machine tools, “fuzzy” information about these factors, the stochastic nature of the cutting process itself, as well as with a variety of methods for ensuring a given machining accuracy, intelligent control systems are a promising area of ​​research and development in machine tools.

Now the most widely used methods of intelligent control related to the following four classes:

• - expert systems (ES);
• - fuzzy controllers (NR);
• - neural networks (NN);
• - genetic algorithms (GA).

Expert systems (expert systems) deal with the tasks of artificial intelligence at the top level, working with symbolic information to obtain conclusions about the environment and form management decisions, taking into account the nature of the current (or predicted) situation. Expert systems accumulate and manipulate heuristic knowledge in an attempt to mimic the behavior of an expert.

Figure 1.2 shows an example of building an expert controller, which is a combination of an ES and a traditional controller (or a system of controllers) in relation to the management of TO. A more complex structure (with a detailed specification of individual blocks of ES) are proposed in the model of the technological system, which are based on the principle of decision-making on making a predictive correction in the manufacturing process of parts, taking into account a specific situation. The expert system, as in and in Figure 1.2, forms the upper, supervisory level of management and includes a number of subsystems.

Identification and prediction subsystem - provides finding the mathematical model of the control object directly in the process of operation, based on the results of observations of its input/output variables. That is, the tasks of the block include the acquisition of information necessary for decision-making. This block performs software adjustment of the movements of the working bodies, measures and identifies the parameters of the state of the external environment - F, control actions - U, the results of the AIDS system - Y.

The database contains continuously updated data (previous, current, forecast) on the characteristics of the AIDS system and the environment, as well as information on the boundary (critical, maximum allowable) values ​​of the relevant parameters. The knowledge base contains knowledge about the specifics of the work of a particular TO, goals, strategies and control algorithms, about the results of identifying and predicting the characteristics of the AIDS system.

The logical inference subsystem selects a rational (the most appropriate at the time of processing a certain part for maintenance) structure and parameters of the controller, as well as, possibly, identification and prediction algorithms.

The interface subsystem is designed to organize an interactive mode for filling the knowledge base with the participation of an expert (learning mode) and providing communication with the user-operator (professional worker), including an explanation of the mechanism for making certain management decisions (operation mode).

The difference between the architecture of the expert system shown in Figure 1.2 and the architecture of conventional (static) expert systems is that it provides the following important functions:

• * building a dynamic model of the object and its environment;
• * maintaining contact with the outside world (sensors, DBMS, regulators, other ES).

This circumstance allows us to refer the considered expert system to the class of dynamic (“active”) expert systems, or real-time expert systems, capable of making up for the lost contribution of a professional worker with his experience, knowledge and skills in achieving the quality of processing.

Fuzzy controllers. The ideas of fuzzy logic, first expressed in 1964 by the American L. Zadeh, a well-known specialist in the field of systems theory, found their first application in problems of controlling real technical objects in Europe. In 1974, the works of the English scientists E.Kh. Mamdani and S. Assilian, devoted to the problem of regulating a steam generator plant with the help of specially designed fuzzy rules (productions).

A typical structure of an IMS with HP is shown in Figure 1.3. For simplicity, we will assume that the control object (for example, a feed drive based on a DC motor (DC motor)) is one-dimensional, i.e. it has one input (control signal - u) and one output (motor shaft rotation speed - y). The control error e, which is the difference between the setting action and the output of the object (controlled variable) y, is fed to one of the inputs of the fuzzification block. The other input of this block receives a derivative signal calculated using a differentiating device (DU).

The purpose of the fuzzification block is to convert the values ​​of the error signals e and its derivative into linguistic variables determined by membership functions. Here A i and B j , respectively, are the values ​​(terms) taken by the linguistic variables “Control error” and “Error derivative”. An example of the construction of membership functions and is shown in Figure 1.3, where the following notation is used:

Z - “Close to zero” (zego);

MR - “Middle positive” (middle positive);

LP - “Large positive” (largepositive);

MN - “Medium negative” (middlenegative);

LN - “Large negative” (largenegative).

The knowledge base stores knowledge in the form of rules, the left parts of which contain conditions regarding the above values ​​of the linguistic variables “Control Error” and “Error Derivative”, and the right parts contain statements regarding the values ​​of the linguistic variable “Control Increment” (the index k here means k -th moment of time t k). These rules can take the following form:

• 1) IF (Control error = Close to zero) AND (Error derivative = Close to zero), THEN (Control increment = Close to zero);
• 2) IF (Control Error = Positive Average) AND (Error Derivative = Large Negative), THEN (Control Increment = Positive Average), etc.

It is assumed that the implementation of these rules guarantees the fulfillment of certain requirements for the system related to the provision of the desired type of its transient function (given speed, monotonicity, weak oscillation of the transient process, for example, in terms of control and disturbing influences for the servo drive of a metal-cutting machine). A possible option for setting the membership functions that determine the basic values ​​of the linguistic variable “Increment of the control action” in the form of single-point fuzzy sets (singletons) is shown in Figure 1.4.

The operation of the inference engine is based on the maximum-minimum method or the maximum-product method. The application of these methods makes it possible to obtain the resulting membership function of the linguistic variable “Increment of the control action” (Figure 1.4), taking into account specific (i.e. measured at the moment t k) values ​​of the signals e k and - inputs of the fuzzy controller.

And, finally, the transition from the resulting fuzzy set, described by the membership function, to a single (clear) value of the output variable is carried out in the defuzzification block using the center of gravity method.

For the case considered in Figure 1.4, this value is calculated as

where - the values ​​of the membership function at the points -c 1 , -c 2 , 0, -c 1 , -c 2 , called the activity levels of the corresponding rules and calculated using the inference mechanism.

The fuzzy controller output u k is found by the formula

where u k-1 - the previous value of the control action u; - increment calculated on the k-th cycle of the controller.

Another type of fuzzy controller is the Sugeno-type controller. In this case, only the left parts of the rules (conditions) contain linguistic variables; the right-hand side of these rules (outputs) are linear combinations of controller input variables plus a constant component (offset). For example, fuzzy rules might look like this:

1) IF (Control error = Close to zero) AND (Error derivative = Close to zero), THEN

2) IF (Control error = Average positive) AND (Error derivative = Large negative), THEN

Here - given (selected by the expert) numerical coefficients; are the values ​​of the error signal and its derivative measured in the k-th cycle. The resulting output is the weighted average of the outputs of each rule (1.3)

where is the activity level of the i-th rule; N is the number of such rules; - increment calculated using the i-th rule for specific values.

The main advantage of using fuzzy controllers in the management of maintenance is the ability to effectively control complex dynamic objects that are part of the AIDS system under the conditions of uncertainty in their characteristics by modeling the knowledge processing mechanism by analogy with the behavior of a highly skilled worker (operator).

Neural networks. The history of artificial neural networks (artificialneuralnetworks) begins with the work of American scientists W. McCulloch, W. Pitts (1943 - a model of a formal neuron) and F. Rosenblatt (1958 - a single-layer neural network, which he called a perceptron). Today, neural networks (NNs) refer to parallel computational structures that model biological processes commonly associated with those of the human brain. NNs have the ability to acquire domain knowledge by learning from examples and adjusting their weights to interpret the multidimensional data presented to them.

Figure 1.5 shows a block diagram of a direct propagation NS - a multilayer perceptron. Circles (vertices) denote elementary information converters - neurons, and arrows (arcs) - connections between them that have different “strength” (weights of synaptic connections). As can be seen from Figure 1.5, the considered perceptron consists of several layers of neurons:

• * input layer to which a set of input signals is applied;
• * one or more "hidden" (intermediate) layers;
• * output layer of neurons.

The essence of the NN learning process is to perform the following multi-step procedure.

Step 1. The training set (“task book”) is set

the elements of which are training pairs. In this case, the 1st input vector (or the 1st input image) presented to the neural network; - vector of reference (required) reactions of NS in response to the 1st input vector; L is the number of different training pairs.

Step 2. The initial state of the neural network is established by assigning some random (small) values ​​to all its weights. is the weight of the connection connecting the output of the i-th neuron of the k-th layer with the input of the j-th neuron of the (k + 1)-th layer.

Step 3. The input vector is fed to the network input; the responses of neurons in the output layer are determined.

Step 4. Calculate the difference between the desired network response and its actual output, i.e., as well as the total squared error

Step 5. The neural network weights are corrected in such a way as to reduce the error.

Step 6. Steps 3-5 are repeated for each pair of training sets until the error on the entire set reaches a small, predetermined value E*.

The result of training is such a tuning of the weights of synaptic connections, in which the network associates the required (or close to it) output with each input vector.

One of the first algorithms that successfully proved itself in training a multilayer NN was the backpropagation algorithm proposed in 1986 by Rummelhart (USA) and his colleagues (Bask-ProgationAlgorithm), which subsequently underwent numerous changes and improvements.

Today, more than 200 varieties of NS are known. In addition to the multilayer perceptrons mentioned above, these are:

• * dynamic (recurrent) NS;
• * networks based on radial basis functions;
• * Hopfield networks;
• * Kohonen networks;
• * neocognitrons, etc.

Figure 1.6 shows an example of using NS to solve the problem of controlling a complex dynamic object (as in the example with a fuzzy controller, maintaining a given speed for a DPT drive is considered). The NS acts in this case as a non-linear controller, which, after the completion of the learning process, ensures a minimum mismatch between the outputs of the reference model (EM) and the ACS TO as a whole.

The advantages of using multilayer NS as TO controllers are explained by a set of their properties: 1) signals in such NS, as in automatic control systems, propagate in one, forward direction; 2) the key role in the formation of the necessary nonlinear control algorithms is played by the universal approximation properties of these networks; 3) the ability of a multilayer NN to learn gives adaptive properties; 4) the ability of neural networks to parallel processing of both analog and discrete signals makes it natural to use them to control multidimensional objects. Implementation of neurocontrollers based on trained neural networks does not cause fundamental difficulties: existing microprocessor tools can fully implement the functions of neural networks. The inclusion of a multilayer NS in the control loop expands the phase space of the object and increases the number of its degrees of freedom, thereby making it possible to synthesize optimal control laws.

Genetic algorithms (geneticalgorithms). This is a large group of adaptive search and multi-parameter optimization methods that have been intensively developed in recent years both for their independent use and in combination with other intelligent control methods.

The very name of these algorithms indicates that their origin is associated with the use of the principles of natural selection and genetics. Traditional search methods usually assume differentiability of the studied objective function in terms of parameters and, as a rule, use gradient procedures. Genetic algorithms (GA) differ from conventional optimization methods in a number of ways. At its core, GAs are a method of parallel search for a global extremum based on the use in the search process of several appropriately coded points (candidates for solutions) that form a population that develops according to certain random laws. The selection mechanisms used in this case, first clearly formulated by Charles Darwin (“The fittest survives!”), Allow us to weed out the least suitable options and, conversely, highlight and then enhance the positive qualities of those options that most fully meet the goal.

Let us outline the range of tasks solved with the help of GA in relation to TO.

Optimization problems occupy one of the central places in the design of various classes of ACS TO. The reason for this is the natural desire to choose the simplest option for constructing a system or model, while meeting the specified requirements for the quality of its functioning (the problem of structural synthesis) or to find the optimal settings for the parameters of a multicomponent system with a given structure (the problem of parametric synthesis). Several examples of setting the corresponding tasks are given in.

Task 1. It is required to find the optimal algorithm for identifying and predicting the characteristics of an object, which is used as part of an IMS TO with an expert controller (see Figure 1.2). The variable parameters are the numerical coefficients of the regression model, the number of basis functions, or the order of the regression equations. The objective function is the identification and forecast error, estimated as the difference between the outputs of the control object and its model at the current (or future) moment in time.

Task 2. It is required to choose the form and mutual arrangement of the membership functions of the fuzzy controller that provides the specified quality of control processes in the ACS TO. Variable parameters - numerical coefficients a i ,b j ,c s of membership functions (see Figure 1.5); number of membership functions. The target function is a quality indicator (functional), the minimum of which corresponds to the reference transient processes.

Task 3. It is required to choose the structure (topology, architecture) of a multilayer perceptron used as a non-linear controller in the ICS TO, shown in Figure 1.6. The variable parameters are the number of layers and the number of neurons in each layer of the NN. The objective function is the network training error, which is the mismatch between the outputs of the object and the reference model of the system.

In all the above examples, the optimization problem takes the following mathematical formulation: to find such values ​​of the variable parameters V 1 , V 2 , …, V n , which provide the minimum of the objective function f(V 1 , V 2 , …, V n) provided that the indicated parameters V 1 , V 2 , …, V n satisfy some admissible area. Setting the area of ​​restrictions in each particular case is dictated by the specifics of the problem being solved. For example, in problem 2, the type of the region is determined by the choice of the boundaries of the intervals within which the required optimal parameters of the membership functions are sought. In Problem 3, the corresponding restrictions are related to limiting the maximum permissible complexity of the studied class of NN, etc.

When using traditional multi-parameter search algorithms to solve the above problems, a number of difficulties arise, which include:

• * a sharp increase in computational costs and search time with an increase in the number of variable parameters (“curse of dimensionality”);
• * the local nature of the search algorithms, associated with the need to calculate the derivatives (gradient) of the objective function at each search step;
• * the possibility of "hanging" the search algorithm in the vicinity of one of the local extrema;
• * low noise immunity of the algorithm;
• * low search efficiency in the presence of "gully" situations.

The attractiveness of GA lies precisely in the fact that they are largely devoid of these shortcomings.

According to the GA terminology, borrowed from genetics and the theory of evolution of living nature, they deal with a population of “individuals”, each of which is a contender for solving the problem under consideration. Each individual is assigned a certain “fitness (fitness) index”, depending on how successful this particular variant of solving the problem is. For example, one of the objective functions mentioned above (see tasks 1-3) can act as such a suitability index. Further, the most fit individuals are given the opportunity to "breed" by "crossing" with other individuals in the population. As a result of this, new individuals appear - “descendants”, who inherit some of the characteristics from each of their “parents”. The least fit members of the population, therefore, "die out." The resulting new population of possible solutions forms a new “generation” that retains in a much greater proportion those qualities (features) that were inherent in the best representatives of the previous generation. By applying the scheme described above from generation to generation, and encouraging crossbreeding and the exchange of traits primarily among the most fit individuals, one can consistently improve the population by preserving and increasing in it the strongest points of individuals. In other words, the search will explore the most promising, promising areas of the space of variable parameters. With the correct functioning of the GA, the population converges to the optimal solution of the problem.

It is generally accepted that genetic algorithms do not guarantee finding a global optimum, but their strength lies in the fact that they allow “quickly enough” to find “good enough” solutions to a wide range of problems, including those that are difficult to solve by other methods.

The history of the application of genetic algorithms begins with the works of R. Holstien and De Jong, in which, using a number of examples, the capabilities of GA for solving multi-parameter optimization problems were first demonstrated. In 1975, J. Holland's monograph "Adaptation in natural and artificial systems" was published, in which the theoretical substantiation of the method was given, and the basic principles underlying it were formulated. And, finally, the book “Genetic Algorithms in Search, Optimization and Machine Learning” by D. Goldberg, published in 1989 and which became a classic, gained great popularity, containing a large number of examples and possible problem statements from various areas of applications solved using GA.

In recent years, the scope of GA has expanded significantly. It is shown that these methods are effective in solving such problems as:

• * identification of complex dynamic objects;
• * selection of the optimal configuration of multi-agent robotic systems;
• * synthesis of optimal control algorithms for multilink robotic manipulators;
• * optimal control of spacecraft docking;
• * planning routes for vehicles in the face of obstacles ;
• * structural synthesis of design solutions, synthesis of schedules

and many others.

The use of GA covers not only the class of traditional optimization problems, but also quickly spreads to the problems of managing complex dynamic objects under uncertainty. Therefore, in the tasks of machine equipment control, HA can also be used to solve a wide range of problems.

To ensure the specified quality of processing at TO, it is necessary to organize intelligent control at all levels of the IMS: organizational, coordinating and tactical. This means that both the system of regulators and the block of identification and forecasting as part of the ES must have “intellectual abilities”. As regulators of nonlinear control objects, fuzzy, NS regulators and their varieties are often used, and for identification and prediction systems - neuro-fuzzy systems (ANFIS - AdaptiveNeuro-FuzzyInferenceSystem) and various types of neural networks. The ES themselves can also be built on the basis of the use of "clear" or "fuzzy" logic. Thus, ES can be developed on the basis of NN or fuzzy rules, or both at the same time. Therefore, when organizing the intelligent control of TO, it is more expedient to create neuro-fuzzy (hybrid) ES, which have ample opportunities to use the advantages of both fuzzy logic and NN. Moreover, it is necessary to use a mixed (hybrid) management strategy at all levels of IMS TO, because this will allow the fullest use of the advantages of intelligent management methods not only at the upper level of management (organizational and coordinating), but also at the lower (tactical) level, where there is the need for non-linear algorithms for various strategies for controlling actuators in real time.