JP5038882B2 - Fuzzy inference control apparatus, plant system, and fuzzy control method - Google Patents

Description

The present invention relates to a fuzzy inference control device, a plant system, and a fuzzy control method.

  A rolling mill (rolling machine) processes a rolled sheet by passing a rolled sheet such as a metal material between rotating rolls. The processing result of the rolled sheet material is affected by the shape or state of the surface of the rolling roll. Therefore, the nozzle of the rolling mill changes the shape or state of the surface of the rolling roll during rolling by injecting a coolant (cooling medium) onto the rolling roll. The rolling mill actuator connected to the nozzle controls the coolant flow rate.

  As a technique for controlling the coolant flow rate for determining the shape control of the cold rolling mill, as shown in Patent Document 1 and Patent Document 2, fuzzy control is applied. Shape control using fuzzy control converts the absolute value of shape deviation, shape time deviation, and shape position deviation from the shape detector into a qualitative concept using a membership function, and fuzzy inference (Fuzzy inference) Running Inference). Control rules and membership functions are determined by engineers based on the mechanical characteristics of the rolling mill, the control operation method of skilled workers, and the like. Therefore, it is a common-sense control rule and membership function.

  However, the actual rolling phenomenon changes depending on the rolling speed, the type of rolling oil, the temperature, the state of the rolling roll, etc. in addition to the thickness, width, and steel type of the material to be rolled. Use of the membership function causes inconvenience in terms of response speed and control accuracy.

In order to solve this problem, gene expression of control rules and membership functions is applied as in Patent Document 3, and a genetic algorithm (GA) is applied, and the gene of the child generation obtained is expanded from that gene. An automatic tuning means that performs fuzzy inference using the rules obtained in this way is considered.
Japanese Patent No. 2515028 Japanese Patent No. 2548818 JP 2000-339168 A

  In the prior art, improvement of control accuracy and shortening of the adjustment period are problems. In other words, if the control rules and membership functions are automatically tuned using the conventional technology, GA computations are required for all optimization targets, resulting in unnecessary computations (duplication) and inconsistent control rule modifications. Or Hereinafter, the shape control of the rolling mill will be described as an example.

  FIG. 16 is an explanatory diagram showing a general concept for performing fuzzy inference for each of a plurality of coolant nozzles corresponding to each zone of the shape detector.

  The shape detector detects the plate width of the rolled plate material for each zone. The shape detector is delimited by sensors. Each sensor is distinguished by Channel (hereinafter abbreviated as ch). Based on each parameter detected by the shape detector, a control amount of fuzzy control is calculated. And the state of a rolled sheet material changes with the result of fuzzy control, and a shape detector detects each changed parameter again.

  Taking the shape control of a rolling mill as an example, the shape control device includes a plurality of cooling medium injection nozzles arranged side by side in the length direction of the rolling roll and injecting a roll cooling medium to control the shape of the surface of the rolling roll. is there. Each of the plurality of nozzles is controlled by an actuator. This actuator converts the numerical value of the input control amount into a nozzle control amount. Therefore, it is necessary to calculate the control amount input to the actuator.

  A control amount of fuzzy control is calculated as a result of fuzzy inference for each actuator for operating a control device (a nozzle that ejects coolant). There are a plurality of actuators, and each actuator is distinguished from other actuators by Channel (hereinafter abbreviated as ch). “(I)” in the fuzzy control (i) is an index representing the position in the length direction of the work roll corresponding to the number of cooling medium spray nozzles.

  Here, since fuzzy control is performed from each channel, a control rule having n different spaces is required. When the control rule from each channel is optimized, it is necessary to optimize n times, and the overlapping portion (space) is optimized many times, and there is a problem that the correction of the control rule is inconsistent. For example, when optimizing a 1ch control rule and then optimizing a 2ch control rule, if the 1ch and 2ch control rules overlap, a contradiction occurs in optimizing the optimized control rule again. Note that Patent Document 1 looks at the entire space of the control rule base 6. In Patent Document 3, since there is only one control object, there is only one space for the control rule base 6.

  FIG. 17 is an explanatory diagram showing a method for generating and optimizing genes for each control rule. As a result of being controlled by a plurality of nozzles, a plurality of control rules (optimization target rules) extracted from the control rule base 6 with respect to the plate width (zone) for each ch detected by the shape detector (FIG. 16) ( In order to control using black squares in the figure, a gene is generated for each control rule and optimized for each channel. As a result of optimization, qualitative expressions (fuzzy expressions) such as “NB, NS, ZO, PS, PB” indicating the degree of control amount of fuzzy control are appropriately changed. The meaning of abbreviations for qualitative expressions such as NB will be described in Table 1 below.

  The control rule base 6 is illustrated in a lattice shape, but this lattice is a simplified version of the two-dimensional matrix described later in FIG. Each white cell of this lattice cell indicates a control rule (only the consequent part) that is not used for control. The set of grid cells of black cells indicates a control rule space which is a set of control rules (only the consequent part) used for control.

  Accordingly, an object of the present invention is to improve control accuracy and shorten a control rule adjustment period in a plant system that performs fuzzy control based on a control rule.

To solve the above problems, the present invention includes a plurality of control target plant, and the state of each plant that is a control object, the control content of each plant according to the state of the each plant, a combination of qualitative A fuzzy inference control device having a plurality of fuzzy control devices for fuzzy control of each of the plurality of plants based on a set of control rules expressed in simple expressions,
A control rule base for storing the control rules;
A control rule space determining mechanism that determines a control rule space that is a combination of control rules used in a fuzzy control device from the state of the plurality of plants, and the control rule base ;
A deterioration control rule space determining mechanism for determining a deterioration control rule space in which the state quantity has deteriorated due to control from the state quantity change before and after control;
From deteriorating control rules space where the deterioration control rule space determining mechanism has determined, the control rule extracted mechanism to create a single gene showing the control rules for the plant, Filter control rule,
An evaluation mechanism for evaluating a control result obtained by controlling each plant based on the control rules narrowed down by the control rule extraction mechanism, using a predetermined evaluation function;
And a genetic operation mechanism for optimizing a gene created by the control rule extraction mechanism by a genetic algorithm during control of each plant based on the evaluation result of the evaluation mechanism.
Other means will be described later.

  According to the present invention, in a plant system that performs fuzzy control based on a control rule, it is possible to improve control accuracy and shorten a control rule adjustment period.

  Hereinafter, an embodiment of a coolant control system to which the present invention is applied will be described in detail with reference to the drawings. As in the following embodiment, in coolant control in which fuzzy control is introduced, it is possible to optimize control rules by genetic operations.

  FIG. 1 is a block diagram showing a coolant control system provided with a fuzzy control control rule optimization mechanism according to the present embodiment. The coolant control system is a system in which when the fuzzy inference control device 1 controls the control target (rolling phenomenon) 12, the result is reflected in the state quantity 15 of the control target. The operation of the fuzzy inference control device 1 is determined by the mode switching mechanism 16.

  The fuzzy inference control device 1 includes a fuzzy control device 13 and a valve ON / OFF pattern determination mechanism 14. The fuzzy inference control device 1 is configured as a computer including at least a memory serving as storage means used when performing arithmetic processing and an arithmetic processing device that performs the arithmetic processing. The memory is constituted by a RAM (Random Access Memory) or the like. Arithmetic processing is realized by an arithmetic processing unit configured by a CPU (Central Processing Unit) executing a program on a memory.

  The fuzzy inference control device 1 includes a control rule space determination mechanism 7, a control rule extraction mechanism 8, a control rule evaluation mechanism 9, a genetic operation mechanism 10, a deterioration control rule space determination mechanism 11, a fuzzy control device 13, And a valve ON / OFF pattern determination mechanism 14. The fuzzy control device 13 includes a classification mechanism 2, an inference mechanism 3, an evaluation mechanism 4, and a control rule base 6. The classification mechanism 2 uses a membership function 5.

  The classification mechanism 2 classifies the state quantity 15 to be controlled using the membership function 5, converts it into a qualitative concept, and calculates a certainty factor. The state quantity 15 to be controlled that the classification mechanism 2 receives input is defined for each ch, such as (i-1) ch, (i) ch, (i + 1) ch.

  The inference mechanism 3 refers to the control rule base 6 and performs fuzzy inference using categorized qualitative concepts. Specifically, the inference mechanism 3 discriminates a control rule to be applied using the certainty factor calculated by the classification mechanism 2, extracts it from the control rule base 6, creates an “IF-THEN” rule, and creates a fuzzy Make inferences.

  The evaluation mechanism 4 quantifies the fuzzy inference result of the inference mechanism 3. Specifically, the evaluation mechanism 4 calculates Δα (i) and calculates α (i). Note that (i) is an index representing the position in the length direction of the work roll corresponding to the number of cooling medium spray nozzles.

FIG. 2 shows the current control rule base 6. The control rule base 6 shows control rules in a matrix (table) format. Each matrix corresponds to three evaluation indexes shown in the following (1) to (3). These evaluation indices are parameters detected by the sensor for each channel by the shape detector shown in FIG.
(1) The control deviation A (i) is a deviation from the target shape. Corresponds to the matrix row elements (ANB to APB).
(2) The shape change rate B (i) is obtained by differentiating the deviation from the target shape with respect to time. This corresponds to the matrix column elements (BNB to BPB).
(3) The local elongation evaluation index C (i) is obtained by differentiating the deviation from the target shape with respect to the plate width (space). It corresponds to each of the three matrices (C = CF, CS, CB).

  Table 1 shows the meaning of the abbreviations written in each matrix.

  (Formula 1) shows calculation formulas of three evaluation indexes.

Hereinafter, terms related to fuzzy control are defined.
The control rule is equivalent to the “inference rule” of Patent Document 1. That is, the control rule is expressed by a set of an antecedent part (IF˜) and a consequent part (Then˜). That is, when the antecedent part is satisfied, the control content is implemented by the consequent part. The antecedent part is expressed using the antecedent part variables A (i), B (i), and C (i). The consequent part (Then ~) is a part that is optimized by GA, and is expressed as an output amount (NB, NS, ZO, PS, PB).
The control rule base 6 arranges the consequent part (output amount: NB, NS, ZO, PS, PB) using the antecedent part variables A (i), B (i), and C (i) of the control rule. It is a thing.
The control rule space is a three-dimensional representation of the three matrices for each antecedent variable C (i) shown in FIG.

  FIG. 3 is an explanatory diagram showing how to read the coordinates from the control rule base 6. When the state quantity 15 to be controlled is classified using the membership function 5 and the certainty factor is calculated, the control rule belonging to each nozzle can be determined. The coordinates of the control rule are read, and one gene is created while arranging them in the coordinate order. Overwrite the coordinates of overlapping control rules.

The row components (A _{1 to} A ₅ ) in FIG. 3 correspond to the row components (ANB to APB) in FIG. For example, the row component _{A 1} is ANB.
The column components (B _{1 to} B ₅ ) in FIG. 3 correspond to the column components (BNB to BPB) in FIG. For example, the row component _{B 1} represents a BNB.
The three matrices (C _{1 to} C ₃ ) in FIG. ₃ correspond to the three matrices (C = CF, CS, CB) in FIG. For example, the matrix “C ₁ ” is the matrix “C = CF”. A hatched rectangular portion of the matrix “C ₁ ” indicates a control rule to be used. By extracting this control rule, a control rule of the “RULE (C _i , B _j , A _k )” format can be expressed.

Hereinafter, terms related to GA are defined. GA is used for the purpose of optimizing control rules.
An individual is a gene in which the consequent parts of corresponding parts are collected as shown in FIG. That is, the parts (spaces) determined from the control rule space determination mechanism 7 in FIG. 1 are collected.
-The chromosome which comprises a gene is one consequent part (output amount: NB, NS, ZO, PS, PB) of the control rule base 6. For example, if the part (space) determined from the control rule space determination mechanism 7 of FIG. 1 is CF, BNS to BZO, ANS to AZO, the gene is NS (CF, BNS, ANS), NS ( CF, BZO, ANS), NS (CF, BNS, AZO), ZO (CF, BZO, AZO).
The gene fitness is divided into fitness (ΔS) for the control rule space and fitness (ΔJ) for the individual by switching between the learning mode and the optimization mode. The fitness (ΔS) for the control rule space is obtained by converting the difference between the target shape and the plate shape into an area, and the unit is “I-unit”. The fitness (ΔJ) for an individual is calculated using the control deviation Ai, the shape change rate Bi, and the local elongation evaluation index Ci, and the unit is “I-unit”.

  The control rule space determination mechanism 7 in FIG. 1 uses the qualitative concept classified by the classification mechanism 2 to control rules that may be used in control from the control rule base 6 of the fuzzy control device 13. Determine the space. That is, the control rule space determination mechanism 7 discriminates a control rule that may be used in the control using the certainty factor calculated by the classification mechanism 2 while the inference mechanism 3 performs the inference, and the control rule base 6 The range is determined from the above, and the control rule space is determined.

  The control rule extraction mechanism 8 extracts a control rule from the deterioration control rule space determined by the deterioration control rule space determination mechanism 11 to create one gene, and sets the generated gene as an optimization target. The control rule extraction mechanism 8 selects a control rule part that overlaps with the gene to be optimized every sampling, and changes only the control rule at random.

  The control rule evaluation mechanism 9 evaluates the control rule space to be applied, and calculates the fitness using the evaluation function for the control rule extracted by the control rule extraction mechanism 8 and the generated control rule. The control rule evaluation mechanism 9 evaluates the control rule space using the evaluation function ΔS in the learning mode. In the optimization mode, the control rule evaluation mechanism 9 calculates the fitness of the gene created by extracting the control rule using the evaluation function ΔJ and the actual data (state quantity 15 to be controlled). Note that an individual that is a member of a population and a gene having a control rule is updated to the control rule base 6 before evaluating the gene.

FIG. 4 is an explanatory diagram showing a method by which the control rule evaluation mechanism 9 calculates the fitness for each control rule space based on the evaluation function ΔS. In shape control, the state quantity 15 to be controlled is represented by I-Unit. (K-1) to represent the area of the sampling results shape the area of the actual shape of S _{k-1, k} sampled at S _k, the evaluation function ΔS using the two areas can be calculated by (Equation 2) . In this way, the fitness can be calculated for all the control rule spaces, and the control rule space can be evaluated based on the change rate of the control result.

  FIG. 5 is an explanatory diagram showing a method by which the control rule evaluation mechanism 9 calculates the fitness for each individual based on the evaluation function ΔJ. In the shape control, the plate shape is represented with respect to the target shape, and therefore is represented by a control deviation A (i), a shape change rate B (i), and a local elongation evaluation index C (i). Using the three evaluation indexes, the evaluation function ΔJ can be calculated by (Equation 3). In this way, the fitness with respect to the target value (target shape) is calculated for all individuals, and optimization is performed so that the deviation from the target shape becomes zero.

  The genetic operation mechanism 10 in FIG. 1 performs a genetic operation using the goodness of fit calculated by the control rule evaluation mechanism 9, and selects a good individual for all individuals for which the genetic operation has been performed. The control rule base 6 is updated.

FIG. 6 is an explanatory diagram showing control rule optimization means by the genetic operation mechanism 10. The control rule extraction mechanism 8 uses the result of control by the valve ON / OFF pattern determination mechanism 14 in the fitness function (evaluation function) ΔJ and the actual data (state quantity 15 to be controlled) based on the gene The evaluation is performed, and the result is set as an evaluation value (Fit _{1 to} Fit _p ) for every p rules “ _{1 to p”} . The genetic operation mechanism 10 performs control by updating the gene changed based on the evaluation value to the control rule base 6, and evaluates the individual generated using the control rule evaluation mechanism 9. The above processing is performed up to the set value (p-1) times. That is, (p-1) genes are created and controlled (p-1) times.

  The deterioration control rule space determination mechanism 11 in FIG. 1 searches for and determines a control rule space in which control is likely to deteriorate from the control rule space determined by the control rule space determination mechanism 7. That is, the deterioration control rule space determination mechanism 11 searches for a control rule space in which control may be deteriorated by using the evaluation by the control rule evaluation mechanism 9 for the control rule space applied up to m sampling, and the result Is output to the control rule extraction mechanism 8. The deterioration of control will be described in detail with reference to FIG.

  FIG. 7 shows a method in which the deterioration control rule space determination mechanism 11 searches for a control rule space to be optimized, and a control rule extraction mechanism 8 extracts a control rule from the control rule space and expresses it as a gene. It is explanatory drawing which shows the method of producing | generating.

  The plate shape changes every time. That is, the portion (space) to be used moves each time in the control rule base 6 that represents the regulation of the control amount by changing the plate shape. When controlling up to m (set value) times, there should be m parts (spaces). Each part (space) is evaluated (ΔS). The part (space) with the worst evaluation is selected.

  According to FIG. 7, control is performed three times (m = 3), and three control rule spaces “A, B, C” are obtained. The control rule space with the worst evaluation is the control rule space “C”, and an individual is generated from the space. Then, (p-1) copies are made from the generated individuals (see FIG. 9 described later). The individuals generated from the control rule space “C” and the copied (p−1) individuals are the population. The following process is repeated (p-1) times for each of the generated p individuals for each control timing.

  At a certain control timing, a control rule space “D” is obtained. The control rule space “C” and the control rule space “D” are compared to determine an overlapping space. Read the coordinates of the overlapping space and change the corresponding control rule (only the consequent part) using mutation. Control and evaluate (ΔJ) using the changed individuals. The above process is repeated up to (p-1) times and evaluated (ΔJ) for individuals who are population members. On the other hand, if there is no overlapping space, the individual is deleted from the population.

  The controlled object (rolling phenomenon) 12 in FIG. 1 is controlled by a control output from the valve ON / OFF pattern determining mechanism 14, and the control result is a controlled object state quantity 15.

  The valve ON / OFF pattern determination mechanism 14 inputs the fuzzy inference result output from the evaluation mechanism 4 and converts it into a control output for the controlled object (rolling phenomenon) 12. The valve is attached to a nozzle that injects coolant into the rolling mill, and the amount of coolant injection is adjusted by opening and closing (ON / OFF).

  The state quantity 15 to be controlled is used for the purpose of calculating the fitness of the individual target value for each of the p individuals using the evaluation function ΔJ. In order to calculate the fitness for p individuals, the control is repeated up to p times. When the fitness is obtained for all individuals, genetic manipulation is performed using the fitness.

  The mode switching mechanism 16 switches from the normal control (learning) mode to the optimization mode. Each of these modes specifies an object to be evaluated by the control rule evaluation mechanism 9 and an evaluation function used for the evaluation.

  FIG. 8 shows that the control rule extraction mechanism 8 re-creates one gene while discriminating and extracting rules to be applied from input information for all control targets, and the genetic operation mechanism 10 is optimized by GA. It is explanatory drawing which shows the method to do. Compared to FIG. 17, in FIG. 17, a plurality of control rule spaces are created for each plate width (zone), whereas in FIG. 8, a single control rule space (a collection of a plurality of control rule spaces) is created. has been edited. As a result, the control rules can be automatically adjusted during the control, and optimization for all the control targets can be performed.

  Further, the control rule space is evaluated by the evaluation function, and a control rule space in which the control may be deteriorated is searched. Concentrate on the searched control rule space and optimize by GA.

  FIG. 9 illustrates a method in which the deterioration control rule space determination mechanism 11 searches for a portion where control may be deteriorated, and a method in which the control rule extraction mechanism 8 extracts and generates an individual from the control rule space that is a search result. It is explanatory drawing which shows. Deterioration of control refers to an event in which a deviation from a target value increases during control. According to the fitness (ΔS) for the control rule space, the target value is deviated due to erroneous control, and ΔS increases as compared with the previous value.

The control rule space determination mechanism 7 retrieves all the calculated certainty factors and extracts control rules that may be used in the control from the control rule base 6. The L-shaped black figure of the control rule base 6 is a control rule space used at the sampling time “t”. A control rule (only the consequent part) is extracted from an L-shaped black figure. The control rule extraction mechanism 8 arranges the extracted control rules into one gene and generates an individual.
Further, by copying (p−1) pieces from the generated individual, up to p set values of individuals having an array having the same length as the number of extracted control rules are generated. Search for the overlapping part of the control rule space and the individual used at the time of control execution, and change the control rule (only the consequent part) to the part at random. We will go sequentially.

  In this way, by expressing the control rule as a single gene for all the controlled objects, it is possible to optimize during control without unnecessary computation. Further, optimization efficiency can be improved by searching for a control rule space in which control is likely to deteriorate and concentrating on the control rule space and optimizing with GA.

  FIG. 10 is a flowchart showing the processing of the inference rule optimization mechanism. This flowchart is started by operating the system to be controlled.

  First, the fuzzy control device 13 executes parameter setting (m, p setting) and mode initialization (counter k, g ← 0) (S1). Here, “←” is an assignment operator. The parameter m is a setting value for switching from the learning mode to the optimization mode. The parameter p is a setting value for switching from the generation mode for generating an individual to the GA mode for optimization using the generated individual. Counters k and g are variables for determining mode switching.

  The classification mechanism 2 performs classification using the membership function 5 from the input information for all the control objects that are the state quantities 15 of the control objects (S2), and calculates the certainty factor.

  The control rule space determination mechanism 7 determines a control rule that may be used in control using the calculated certainty factor, determines a range from the control rule base 6, and determines a control rule space (S3).

  The inference mechanism 3 has the calculated certainty factor and infers with reference to the control rule base 6 (S4). The evaluation mechanism 4 evaluates and quantifies the inferred result (S5). The valve ON / OFF pattern determination mechanism 14 determines and controls the valve ON / OFF pattern (S6). As a result, the controlled object (rolling phenomenon) 12 is in the state indicated by the controlled object state quantity 15.

  The control rule evaluation mechanism 9 calculates the fitness of the gene created by extracting the control rule using the evaluation function ΔJ and the actual data (state quantity 15 to be controlled) (S7).

  The fuzzy control device 13 determines the mode switching using the parameter m and the counter k (S8). That is, a control rule that may be used in control is determined, a range is determined from the control rule base 6, and a control rule space is determined. When controlling up to m samplings, there are m control rule spaces, which are evaluated using the evaluation function ΔS and switched to the optimization mode. First, in the learning mode (k <m) in S8, the control rule space is evaluated using the evaluation function ΔS, and 1 is added to k (S8b). On the other hand, in the optimization mode (k ≧ m) in S8, the process proceeds to S22.

  If k = m (S22, Yes), the fuzzy control device 13 executes S9 and S10. The deterioration control rule space determination mechanism 11 searches m control rule spaces evaluated with the evaluation function ΔS, and determines a control rule space when the control deteriorates (S9). The control rule extraction mechanism 8 creates one gene while extracting a control rule from the determined deterioration control rule space (S10). The gene created above becomes an optimization target and becomes the basis of an individual to be generated.

  Then, the fuzzy control device 13 determines whether to perform individual generation (S10b). When satisfying “g ≦ p−1” (S10b, Yes), the control rule extraction mechanism 8 executes the generation mode. Before generating an individual, the control rule space of (k + 1) sampling is compared with the deterioration control rule space, and an overlapping space is searched (S10c).

  If there is an overlapping space (S10c, Yes), the coordinates of the overlapping space are read, and the control rules for the coordinates of the genes that are individuals are changed at random. Then, a gene having the same control rule is created (copied) from the gene created from the deterioration control rule space (S10d), and a mutated individual is generated (S11). Then, 1 is added to the counter g (S11b). On the other hand, when there is no overlapping space (S10c, No), individual generation is stopped, and control is performed using the (k + 1) sampling control rule space. Further, 1 is added to the counter k (S11c).

  The above processing is performed up to the set value (p-1) times. That is, (p-1) individuals are created and controlled (p-1) times. The gene generated by the above processing is updated to the control rule base 6 to perform control, and the individual generated using the evaluation function ΔJ is evaluated.

  On the other hand, when satisfying “g> p−1”, the control rule extraction mechanism 8 executes the GA mode. When (p-1) individuals are generated and evaluated, the control rule extraction mechanism 8 is a population that is a target of p genetic operations including the genes extracted from the deterioration control rule space and the generated individuals. Configure. Using the population, genetic operations such as mutation (S12), crossover (S12b), and elite preservation selection (S12c) are performed, and the individual selected by the elite preservation selection method is updated to the control rule base 6. Then, the fuzzy control device 13 initializes variables (counters k and g) relating to mode switching to 0 (S13). The above processing is repeated. In this way, optimization can be performed during control.

  Here, the GA executed by the genetic operation mechanism 10 will be described in detail.

  GA is used for the purpose of learning control rules in order to improve the accuracy of fuzzy inference. New control rules are not predictable, and control rules that could not be predicted at all may be optimal. Therefore, it is possible to operate the control operation end randomly and find the control result while looking at the control result. Necessary. In order to realize this, it is only necessary to change a control rule at random while performing control and learn a rule that improves control.

  FIG. 11 is an explanatory diagram showing a general concept of GA. GA is a model for gene creation methods when adapting organisms to new environments. In a living organism, mutation (S103), crossover (S104), and elite preservation selection (S105), which are genetic operations, are performed on the generated gene (S101), and the end condition is satisfied (S106, Yes). ). If the gene is not suitable for the environment, it is selected as natural, and if it matches the environment, the gene is retained and adapted to the environment.

  FIG. 12 is a flowchart showing a shape control rule optimization process using GA. A control rule used for the control process (S201) is extracted from the rule base, randomly generated (S202), and reconfigured into an individual. Control processing (S201) is performed using all the reconfigured individuals, and evaluation is performed using the fitness (ΔJ) for the individuals (S211).

  The individual is deformed using the result of the evaluation (S211). Specifically, mutation (S221, mutation means) mutation (S222) by crossover means (S223, crossover means) and crossover (S224), which are new operation methods, are executed.

  The result controlled by the control rule that is the chromosome of the individual is evaluated using an evaluation function (ΔJ), and a next generation population is created. The result controlled by the control rule, which is the chromosome of the transformed individual, is evaluated using the evaluation function (ΔJ), and the best individual is selected by the elite preservation selection means (S225, elite preservation selection means) elite preservation selection (S226). By adopting as above and reflecting it in the update to the control rule base (S227), it becomes possible to optimize the control rule.

  FIG. 13 is a schematic view showing GA mutation means. The mutation means selects two individuals at random from all individuals, compares the fitness of the two selected individuals (x, y), selects an individual with good fitness, and suddenly has a set value. This is a method of randomly changing the control rule of an individual selected using the mutation probability. Once the mutation is over, place it in the next generation population. In this way, the mutation is repeated to reconstruct the next generation population.

  FIG. 14 is a schematic view showing a crossing means of GA. The crossover means selects two individuals (for example, m, n) at random from all individuals, compares the fitness of the two selected individuals, selects an individual with good fitness, and selects a parent (for example, m ). The above processing is performed to determine two parents. This is a method of generating a new child while crossing a control rule that is a parent chromosome selected using a crossover probability that is a set value. After crossover, place it in the next generation population. In this way, crossover is repeated to reconstruct the next generation population.

  FIG. 15 is a schematic diagram showing GA elite storage selection means. The elite storage selection means is a method for searching for individuals having good fitness for all individuals and storing them in the next generation population. If the end condition is satisfied, the corresponding control rule in the control rule base 6 (FIG. 1) is updated with the control rule that is the chromosome in the individual selected by the elite preservation selection means.

  The embodiment has been described above. The problems, configurations, and effects of this embodiment will be summarized below.

  First, in the fuzzy control of the coolant, fuzzy inference is performed for each of a plurality of coolant nozzles corresponding to each zone of the shape meter. The inference rules used in each fuzzy inference are common (using an inference database). In the tuning of inference rules, conventionally, tuning has been performed for each fuzzy inference corresponding to each nozzle. In that case, the control rule applied to each nozzle may be the same, and a wasteful calculation is performed.

  Therefore, it is more effective to collectively tune inference rules actually used for each nozzle. Therefore, the feature of the present embodiment is that the inference rules actually used before tuning are extracted and tuned together to shorten the calculation time and enable online use.

  By using this embodiment (especially, S3, S7, and S9 in FIG. 10), evaluation is performed using the actual data and the obtained fitness, and a portion that needs to be optimized is determined and optimized. Optimization is possible, and it is possible to respond automatically to secular changes.

  According to the present embodiment described above, optimization of the control rule can be executed according to the rolling state, and optimization and high accuracy of fuzzy control can be achieved. That is, by performing control while optimizing the control rule, the rule can be automatically adjusted during the control, and optimization for all control objects is possible. Therefore, it is possible to cope with a case where the product accuracy is deteriorated when changing the control rule in the actual process, and it is possible to automatically correct the control rule during the control and to optimize it.

  Such optimization can be realized in a short time, and the on-site adjustment period can be shortened. In addition, since it is possible to automatically cope with aging changes such as rolling mill machines and operating conditions, the control accuracy can be improved and the adjustment period can be shortened.

It is a block diagram which shows the coolant control system provided with the inference rule optimization mechanism of the fuzzy control regarding one Embodiment of this invention. It is explanatory drawing which shows the control rule base regarding one Embodiment of this invention. It is explanatory drawing which shows how to read the coordinate from the control rule base regarding one Embodiment of this invention. It is explanatory drawing which shows the method of calculating evaluation function (DELTA) S regarding one Embodiment of this invention. It is explanatory drawing which shows the method of calculating evaluation function (DELTA) J regarding one Embodiment of this invention. It is explanatory drawing which shows the control rule optimization means by the genetic operation mechanism regarding one Embodiment of this invention. It is explanatory drawing which shows the control rule space search method which is the optimization object regarding one Embodiment of this invention, and the method of extracting a control rule from control rule space, expressing it with a gene, and producing | generating an individual | organism | solid. It is explanatory drawing which shows the method of recreating one gene and optimizing by GA, discriminating and extracting the rule applied from the input information with respect to all the control objects regarding one Embodiment of this invention. It is explanatory drawing which shows the method of extracting and producing | generating the individual from the control rule space which is a search result, and the control rule space which is a search result, regarding the part which may worsen the control regarding one Embodiment of this invention. It is a flowchart which shows the process of the inference rule optimization mechanism regarding one Embodiment of this invention. It is explanatory drawing which shows the general concept about the genetic algorithm regarding one Embodiment of this invention. It is a flowchart which shows the optimization process of the shape control rule using GA regarding one Embodiment of this invention. It is the schematic which shows the mutation means of the genetic algorithm regarding one Embodiment of this invention. It is the schematic which shows the crossing means of the genetic algorithm regarding one Embodiment of this invention. It is the schematic which shows the elite preservation | save selection means of the genetic algorithm regarding one Embodiment of this invention. It is explanatory drawing which shows the general concept which performs a fuzzy inference for every some coolant nozzle corresponding to every zone of a shape detector. It is explanatory drawing which shows the method of producing | generating and optimizing a gene with respect to each control rule.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Fuzzy inference control apparatus 2 Classification mechanism 3 Inference mechanism 4 Evaluation mechanism 5 Membership function 6 Control rule base 7 Control rule space determination mechanism 8 Control rule extraction mechanism 9 Control rule evaluation mechanism 10 Genetic operation mechanism 11 Deterioration control rule space determination Mechanism 12 Control target (rolling phenomenon)
13 Fuzzy Control Device 14 Valve ON / OFF Pattern Determination Mechanism 15 State Quantity to be Controlled 16 Mode Switching Mechanism

Claims (4)

A plurality of control target plant, and the state of each plant that is a control object, the control content of each plant according to the state of the each plant, the combination of, based on a set of control rules shown in qualitative representation A fuzzy inference control device having a plurality of fuzzy control devices for fuzzy control of each of the plurality of plants,
A control rule base for storing the control rules;
A control rule space determining mechanism that determines a control rule space that is a combination of control rules used in a fuzzy control device from the state of the plurality of plants, and the control rule base ;
A deterioration control rule space determining mechanism for determining a deterioration control rule space in which the state quantity has deteriorated due to control from the state quantity change before and after control;
From deteriorating control rules space where the deterioration control rule space determining mechanism has determined, the control rule extracted mechanism to create a single gene showing the control rules for the plant, Filter control rule,
An evaluation mechanism for evaluating a control result obtained by controlling each plant based on the control rules narrowed down by the control rule extraction mechanism, using a predetermined evaluation function;
Based on the evaluation result of the evaluation mechanism, during said control of each plant, fuzzy inference, characterized in that it has a, a genetic operation mechanism be optimized by genetic algorithm genes the control rule extraction mechanism created Control device. The plant to be controlled is an actuator for opening and closing a nozzle valve for coolant injection provided in the rolling mill,
For the target shape of the rolled sheet material rolled by the rolling mill, the state of the plant is a control deviation that is a deviation from the target shape, a shape change rate obtained by differentiating the deviation from the target shape with respect to time, and the target shape. The fuzzy control device according to claim 1, wherein the deviation is a local elongation evaluation index obtained by differentiating the deviation with respect to the plate width. 3. A plant system comprising: the fuzzy control device according to claim 1; and the plant to be controlled. A plurality of control target plant, and the state of each plant that is a control object, the control content of each plant according to the state of the each plant, the combination of, based on a set of control rules shown in qualitative representation A fuzzy control method by a fuzzy inference control device having a plurality of fuzzy control devices for fuzzy control of each of the plurality of plants,
The fuzzy control device has a control rule base for storing the control rules, a control rule space determination mechanism, a deterioration control rule space determination mechanism, a control rule extraction mechanism, an evaluation mechanism, and a genetic operation mechanism. And
The control rule space determination mechanism determines a control rule space that is a combination of the control rule base and a control rule used in a fuzzy control device from the states of the plurality of plants,
The deterioration control rule space determination mechanism determines a deterioration control rule space in which a state quantity has deteriorated due to control from a state quantity change before and after control,
The control rule extraction mechanism, the deterioration control rule space where the deterioration control rule space determining mechanism has determined to create a single gene showing the control rules for the plant, refine control rule,
The evaluation mechanism evaluates a control result obtained by controlling the plants based on the control rules narrowed down by the control rule extraction mechanism with a predetermined evaluation function,
A fuzzy control method characterized in that the genetic operation mechanism optimizes a gene created by the control rule extraction mechanism with a genetic algorithm during control of each plant based on an evaluation result of the evaluation mechanism .

Images