JP2564405B2 - Knowledge acquisition method and process operation support method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of application] The present invention relates to driving support control of a general process,
In particular, it relates to automatic acquisition of fuzzy rules for fuzzy application systems, and driving support and control using them.

[Conventional technology]

As a conventional process control method, a method of modeling human thought is used when mathematical modeling of a process is difficult to construct. A method applying fuzzy logic to one of the methods is known. When applying fuzzy logic, the point is to deal with human vague sense quantitatively, and this quantification function is called membership function. Also, using Aimana rules (knowledge),
You can make inferences and draw ambiguous conclusions like humans.

[Problems to be Solved by the Invention]

In the fuzzy application system, the setting of the membership function, which is the point, exists by trial and error, and the logical decision method has not been established. Also, regarding fuzzy rules, the ambiguous "if ... if ..." relation is expressed using certainty (this is called "fuzzy rule with certainty"), and this method is performed through interviews with operators. There is. It is difficult to apply the membership function and the fuzzy rule with certainty obtained in this way to control unless the two are well matched.

Therefore, before the operation of the system, the above-mentioned work requires experience and much labor. Furthermore, after system operation, much effort is required for software maintenance such as addition, deletion and improvement of membership functions and fuzzy rules.

As an example, at a sewage treatment plant, when operating the plant, the operator pays attention to past experience and operating records, and also takes into consideration the current situation, which makes complicated and continuous monitoring and manual control indispensable. It was

[Means for solving the problem]

The first object of the present invention is to automatically generate a fuzzy rule with a certainty factor when a human being sets a membership function among the fuzzy rules with a certainty factor of an unknown membership function. The second purpose is to perform support control. To achieve this,
Uses a neural network (neural circuit model).

The present invention is characterized by performing a neural network for plant evaluation and operation (control) based on past results.

In the present invention, specifically, a neural network model having a hierarchical structure having an input layer and at least one intermediate layer and an outgoing layer is used, and the past of the process is measured in the input layer at an arbitrary time t and The measured values and the manipulated variables at a plurality of time points are converted back into the membership value and inputted, and the teacher signal value is obtained by converting the manipulated variable at the time t of the process into the membership value. The relationship between the input value and the teacher signal value is learned by modifying the magnitude of the weighting coefficient so that the difference between the output value from the output layer of the neural network model and the teacher signal value becomes small.

Then, based on the weighting coefficient from the input layer to the intermediate layer and the weighting coefficient from the intermediate layer to the output layer of the learned neural circuit model, knowledge between the input signal and the teacher signal is acquired as a fuzzy rule. .

Based on the obtained fuzzy rules, inference is executed to support or control driving operation.

[Operation] In the present invention, the learning / prediction ability by the neural network is applied to an automatic acquisition method of a fuzzy rule with confidence in fuzzy control and its use.

In the present invention, the membership value converted by using the preset membership function is input to the neural net to learn the past history, so that the vague reasoning possessed by the operator can be automatically performed. Is.

〔Example〕

The present invention first converts disturbance, process state amount, evaluation amount, process operation amount at past multiple time points into a membership value by a preset membership function,
This converted value is input to the neural network to learn the relationship between. A fuzzy rule with certainty factor is derived from the weight coefficient distribution of the learned neural network, and inference is performed based on this to support or control the driving operation. This execution process is as follows.

(1) Converting the value of ~ into a membership value with a membership function (2) Learning process by a learning neural network (3) Process of acquiring a fuzzy rule with certainty from a learned neural network (4) Fuzzy rule diagnosis process ( 5) Fuzzy inference process (6) Prediction process of learned neural network (7) Process operation control process.

An embodiment in which the present invention is applied to a sewage treatment process will be described with reference to FIG.

The structure and operation of the sewage treatment process will be described below. First, in the sedimentation basin 9, a part of the suspended matter in the inflowing sewage is removed by gravity sedimentation. First, the sewage overflowed from the settling tank 9 and the returned sludge from the returned sludge pipe 16P1 flow into the aeration tank 15. Air is supplied to the aeration tank 15 from the blower 17 through the control valves 17A and 17B, and the sewage and the returned sludge are mixed and stirred. The returned sludge (activated sludge) absorbs oxygen in the supplied air, decomposes the organic matter in the sewage, and guides it to the final settling tank 16. In the final settling tank 16, the activated sludge is settled down by gravity and the supernatant water is discharged. The activated sludge settled in the final settling tank 16 is pulled out, and most of the activated sludge is returned as sludge from the returned sludge pump 16C1.
Returned to the aeration tank 15 through 16P1. The rest is discharged as excess sludge by the excess sludge pipe 16P2 and the excess sludge pump 16C2.

Next, the measuring instrument will be described. First settling tank 9, exposure tank
15 and final sedimentation tank 16 are equipped with measuring instruments 9M, 15M and 16M, respectively. Measurement items include inflow sewage, suspended solids concentration, chemical oxygen demand, pH, nitrogen concentration, ammonia concentration, nitrate nitrogen concentration, nitrite nitrogen concentration, phosphorus concentration, dissolved oxygen concentration, sludge interface height (SVI : Sludge Volume Index, etc.), and image information of microorganisms and airborne substances.

Next, the configuration of the predictive operation control device 80 will be described. The computer system of the predictive operation control device 80 is shown in FIG. 1 to facilitate the explanation of the present invention. First, the history pattern data file 71F sequentially stores the data of the measuring instruments 9M, 15M, 16M, and outputs the selected data string to the membership conversion process 69. The membership conversion step 69 converts the value of to a membership value and outputs this signal to the learning neural net 71.
The prediction neural network 72 is a learning neural network 71.
From which signals 71S1 and 71S2 are received and a signal 72S is output. On the other hand, the fuzzy rule acquisition step 73 receives signals 71S1 and 71S2 from the learning neural network 71. The fuzzy rule diagnosis step 74 receives a signal from the communication means 46 and a signal from the fuzzy rule acquisition step 73, and stores it in the fuzzy rule candidate base 60B or the fuzzy rule base 60A. The fuzzy logic mechanism 61 receives signals from the fuzzy rule bases 60A and 60C and outputs a signal 61S to the operation control process 75. The operation control process 75 receives the signal 61S and the signal 72S and outputs the signal 75S,
Excess sludge pump 16C2, Return sludge pump 16C1, Blower 1
7. Control valves 17A and 17B. The communication means 46 is a history pattern data file 71F, with the intervention of the operator 101.
Learning neural network 71, prediction neural network 7
2. Communicate with the operation control process 75 and the fuzzy rule diagnosis process 74. In addition, the part described as "network" in the drawing indicates a large number of wirings (or information communication paths) corresponding to a solid line 702 in FIG. 5 described later.

 Next, the operation of the predictive operation control device 80 will be described.

The history pattern data file 71F is a disturbance characteristic measured by the measuring instrument 9M, a process state amount measured by the measuring instrument 15M, a process evaluation amount measured by the measuring instrument 16M, a surplus sludge pump 16C2, a returning sludge pump 16C1, a blower 17, a control. The process operation amount of the valves 17A, 17B, etc., and time series data of the above are stored.

The membership changing step 69 is selected from the history pattern data file 71F automatically or by the instruction 71S from the communication means 46 from the outer pipe characteristic, the process state amount, the process evaluation amount, the process operation amount, and each of the above items. The converted data strings D1, D2, D3, D4 are received and converted into membership values. Although each data string actually includes a plurality of items, in the present embodiment, the representative symbol D1 is used for ease of explanation.
It shall be represented by ~ D4.

The learning neural network 71 executes learning using the converted membership value, and the execution results 71S1 and 71S2 are
Fuzzy rule acquisition process 73 and prediction neural network 72
And output to.

In the fuzzy rule acquisition step 73, the fuzzy rule candidates related to the above are converted into symbols or words based on 71S1 and 71S2. In the fuzzy rule diagnosis process 74, the fuzzy rule obtained in the fuzzy rule acquisition process 73 is used as a communication means 46.
It is stored in the fuzzy rule candidate base 60B or the fuzzy rule base 60A according to the instruction from. The fuzzy inference mechanism 61 receives the fuzzy rule base 60A and the fuzzy rule base 60C input in advance, executes inference, and outputs a signal 61S to the operation control process 75.

On the other hand, the prediction neural network 72 selects data required for prediction from the history pattern data file 71F and uses it for prediction. The prediction result signal 72S is displayed on the communication means 46 and is output to the operation control process 75.

The operation control process 75 receives the prediction signal 72S and the signal 61S and outputs the signal 75S to output the excess sludge pump 16C and the return sludge pump.
It controls the process operation amount of 16C1, blower 17, control valves 17A, 17B, etc. At the same time, the control target value signal 75S is displayed on the communication means 46, and the actual control amount is corrected by the selection of the operator 101 if necessary. The correction value is output again.

Next, the operation of the present invention will be described in detail with reference to FIG.
2 and subsequent figures will be used for the description.

First, a method of storing data in the history pattern data file 71F will be described with reference to FIG. Time t =
The history pattern data D1 (0) to D4 (0) of 0 are stored in the history pattern data file 71F. By repeating this, the data of t = 0, -1, -2, ... Is sequentially stored. The time interval is, for example, one hour, but the implementation of the present invention is not restricted by this time setting.

The membership conversion process 69 will be described with reference to FIG. FIG. 3 shows a conversion example of a preset membership function. FIG. 3 illustrates the case where the representative of D2 is the dissolved oxygen concentration (hereinafter abbreviated as DO), and the states of “high”, “normal”, and “low” of the DO are respectively converted into member shift functions 691H, 691M,
Converted by 691L to obtain converted values D2H, D2M, D2L. The converted values D2H, D2M, D2L are collectively referred to as a symbol DF2. The horizontal axis of the conversion function shows DO which is the process state quantity, and the vertical axis shows the degree with a value of 0 to 1. To explain the membership conversion function 691L as an example, the membership value = 0.9 when the DO value = 0.5. This means that if the DO value = 0.5, "DO is low" with a degree of 0.9 (membership value). Represent In this way, membership functions are defined in advance for all variables, and data examples D1, D2, D3, D
Convert to a membership value of 4 to get DF1, DF2, DF3, DF4. It should be noted that in this example, there are three categories of “high”, “normal”, and “low”, but the number of categories is arbitrary.

Then, the learning neural network 71 in the learning process.
The operation of will be described below. The learning neural network 71 receives the membership value and executes learning. This data selection and learning method will be described below with reference to FIG. Fourth
As shown, DF1 (ie D1H, D1M, D1L) ~ DF4
(That is, D4H, D4M, D4L), _first t ₁ -1, t ₁ -2, ... Similarly, t ₂ −1, t ₂ − based on t = t ₂ (t ₂ ≠ t ₁ ).
The pattern data of 2, ... Are learned, and a total of q pieces of pattern data are learned. It is desirable to select a typical pattern in the past for selecting q patterns. Since the time ti is an arbitrary time,
If you constantly learn, you can automatically acquire new fuzzy rules corresponding to new situations.

Learning is performed by dividing these groups of data into input data and teacher data. As shown in FIG. 4, the input layer 71
DF1 at any time t (where t = t ₁ , t ₂ , ...) Is 0
(T), DF2 (t) and DF going back to the past from time t
1 (t-1) to DF4 (t-1) are sequentially input. here,
Note that DF3 (t-1) and DF4 (t-1) are added. The same applies to t = t−2, t−3, ... On the other hand, DF3 (t) and DF4 (t) are input to the teacher layer 750. Learning is performed by the learning neural net 71 having a structure including an input layer 710, an intermediate layer 720, an output layer 730, a comparison layer 740, and a teacher layer 750.

A signal processing method for learning will be described below with reference to FIG. The configuration of FIG. 5 and the signal processing method are publicly known except for the input data and the setting method of the teacher layer 750 described above. That is, the configuration and signal processing method shown in FIG. 5 are well-known techniques devised by Rumelhart (for details, refer to the reference: Parallelell Distribut
ed Processing, MIT Press, Vol.1, (1986)).

The configuration and operation of FIG. 5 will be described. In FIG. 5, a circle indicates a neuron element model 701 having an integration operation and a sigmoid transformation function, and a solid line 702 connecting the circle and the circle indicates that information is exchanged between the neuron element models 701.
Here, each layer is composed of a finite number of neuron element models, and the neuron element models of adjacent layers are all connected. The intermediate layer 720 may have a plurality of layers, but in this embodiment, an example in which the number of intermediate layers is one is shown for simplicity of explanation. Also, in FIG. 5, the output layer 730, the comparison layer 740, and the teacher layer 750.
Are DF3 (t) and DF4 (t), but a plurality of cases are shown as general expressions. The configuration of FIG. 5 is called a neural network (neural circuit model).

Next, the basic operation of the neuron element model 701 will be described with reference to FIG. Batch time series of each of the data DF1~DF4 applied to the input layer 710 (total is n is) and referred to by the 6 n-number of variable values Y ₁ -Yn as shown in FIG. Each of the input signal values Y _{1 to} Yn is multiplied by the weighting coefficient Wji, and the addition operation (sum of products operation) is calculated by the equation (1).

Where Yi (1): Yi value of the input layer (first layer), Wji
(2 ← 1): Weighting coefficient from the i-th variable of the input layer (first layer) to the j-th neuron element model of the intermediate layer (second layer), Zj (2): Intermediate layer (second layer) Is a sum total value of inputs to the j-th neuron element model of.

In the neuron element model 701, the output value here is calculated by the equation (2) (Sigmond transform) according to the magnitude of Zj (2).

Yj (2) = 1 / (1-e- ^{Zj (2)} ) (2) The calculation contents of equation (2) are non-linear transformations as shown in FIG. The effect can be obtained. The calculated value Yj (2) is further sent to the output layer, and the output layer also performs the same calculation.

Next, the outline of the calculation method in the neural network will be described. The above-mentioned variable value Yi (1) is the input layer 71 of FIG.
This signal value is input to 0, and this signal value is output to the neuron element model of the intermediate layer 720. In the neuron element model of the intermediate layer 720, these output values Yi (1) and weighting factors Wij (2 ← 1)
The sum of products Zj (2) is calculated by the equation (1), and the output value Yj (2) to the output layer 730 is determined by the equation (2) according to the magnitude. Similarly, the output value Yj (2) of the intermediate layer 720 is further the product sum Zj (3) of the weight coefficient Wij (3 ← 2) of the intermediate layer (second layer) 720 and the output layer (third layer) 730. ) Is calculated by the equation (3).

Here, Yi (2): value of the intermediate layer (second layer), Wji (3
← 2): Weighting coefficient from the i-th variable of the intermediate layer (second layer) to the j-th neuron element model of the output layer (third layer), Zj (3): of the output layer (third layer) It is a sum total value input to the j-th neuron element model.

Further, the output value Yj (3) to the output layer 730 is calculated by the equation (4) according to the magnitude of Zj (3).

Yj (3) = 1 / (1-e- ^{Zj (3)} ) (4) In this way, the calculated value Yj (3) of the output layer is obtained. Yj (3) is DF3 (t) *, DF4 (t) * in this embodiment, and * represents the output signal of the neural network.

A learning means using a neural network will be described. First, the comparison layer 740 compares the signal 740S of the output layer 730 with the teacher signal 750S of the teacher signal layer 750. For example, DF which is the output signal 730S
The magnitudes of 4 (t) * and DF4 (t), which is the teacher signal 750S, are compared. The magnitudes of the weighting factors Wji (3 ← 2) and Wji (2 ← 1) are corrected so that the error between them becomes small. Using this modification, the calculation of equations (1)-(4) is executed again to obtain a new DF4 (t) * value. This process is repeated until the error becomes equal to or less than a predetermined value. Initially, since the weighting coefficient is given randomly by the generation of random numbers, the error is large, but the output signal value gradually approaches the teacher signal value. At this time, since only the weighting coefficient value Wji is changed in the equations (1) to (4), the learning result is Wji.
Note that it will be reflected in the distribution of values. This Wji value is used in the prediction process.

A method of correcting an error in this way is called an error back-propagation method, and the known technique devised by Rumelhart et al. Is used. Although this learning method is known, in the present invention, in particular, the history pattern data groups at different times (t = t ₁ , t ₂ , t ₃ , ..., T _q ) are converted into membership values and then repeatedly learned. At this time, especially note that the teacher signal is an analog value. As a result, the pattern grasping ability comparable to the operator's past experience is accumulated and stored in each weight coefficient Wji of the neural network, and the same effect as the operator's past experience learning is provided.

A method of learning the history pattern data group at a plurality of times t = t ₁ , t ₂ , t ₃ , ... T _q will be described below. A plurality of history pattern data groups are learned by arbitrarily changing the time t. This day
There are cases in which the operator selects t ₁ , t ₂ , t ₃ , ..., T _q and cases in which they are automatically selected, and both will be described below.

The case where the operator selects is a typical pattern of the variable values Y _{1 to} Yn that is desired to be reflected in the operation later or an abnormal pattern that is to be referred to at a later date. The time is set by the operator through man-machine conversation via the communication means 46. For example, 12 patterns are learned from a typical pattern in January to a typical pattern in December. Especially when the typical inflow water quality is
Learn the pattern when the treated water quality measured in M (concentration of suspended solids, concentration of organic matter, etc.) is good (minimum criteria are met). That is, what kind of operation is performed when the processing is good is learned. Needless to say, it is possible to selectively learn only when there is an abnormality. Abnormality learning is suitable for abnormality diagnosis.

On the other hand, when it is automatically performed, the statistical analysis of the data string is performed in advance. That is, the pattern with the highest occurrence frequency is obtained by statistical analysis, and it is considered as a typical example in a normal time to learn it. On the other hand, the case with a low occurrence frequency is considered as a typical example in a pattern in an abnormal case and these are learned. . Although the number of neural networks is one in the embodiment shown in FIG.
It is more effective if they are separately applied to the normal time neural network and the abnormal time neural network.

Next, in the fuzzy rule acquisition step 73, the fuzzy rules between the above are converted into symbols or words based on 71S1 and 71S2. Variable i set in input layer 710 and output layer 730
The confidence factor of the rule with respect to the variable j to be set to The equation (5) was independently derived by mathematical analysis.

Where Iji is the certainty factor and m is the number of neuron element models in the intermediate layer.

The settlement of the equation (5) is a calculation of integrating the weighting coefficients of all the paths from the input layer to the output layer. The rule representing the relationship between the i-th input and the j-th output is established according to the certainty factor of Iji. Iji is calculated for all combinations of input layer and output layer, and the results are converted into Japanese as fuzzy rule candidates. For example, [fuzzy rule candidate]: “If the i-th input value is large, the j-th output value is large with the certainty factor Iji.” This conversion is output in order from the rule with the largest certainty factor Iji. It can also be converted into a rule candidate combined with.

[Fuzzy rule candidate]: “If the i-th input value is large and the k-th input value is also large, the j-th output value is large with certainty factor = (Iji + Ijk) / 2.” Fuzzy rule diagnosis step 74 Then, the validity of the fuzzy rule candidates obtained in the fuzzy rule acquisition step 73 is judged by an instruction from the operator 101 and the communication means 46, and the fuzzy rule candidates judged to be valid are fuzzy rule bases.
It is stored in 60A, and the other rules are stored in the fuzzy rule candidate base 60B. If the fuzzy rule candidates stored in the fuzzy rule candidate base 60B occur once more than once, the occurrence frequency is counted, and when the number of occurrences exceeds a predetermined number (for example, 2 times), the communication means 46 again inquires of the operator 101, and fuzzy The rule diagnosis process 74 is repeated.

As the fuzzy rule base of this embodiment, there is obtained a relation of what kind of disturbance and process state quantity, what kind of evaluation quantity, and how to operate the process as a result.

The fuzzy inference mechanism 61 receives the fuzzy rule base 60A and the fuzzy rule base 60C, executes inference, and outputs a signal 61S to the operation control process 75. The fuzzy rule base 60C stores the production rules or fuzzy rules previously acquired by an interview, which is a conventional method, with respect to the operator 101 and uses them as an auxiliary when necessary. The fuzzy reasoning mechanism 61 performs rule-based forward or backward inference.

Next, the prediction process by the prediction neural network 72 will be described. The structure of the prediction neural network 72 is shown in FIG. As shown in FIG. 1, the prediction neural network 72
Then, the learning result in the learning neural network 71, that is, the values of the weighting coefficient values Wji (3 ← 2) and Wji (2 ← 1) are received as signals 71S1 and 71S2. Although the learning neural network 71 and the prediction neural network 72 are separately illustrated in FIG. 1 for the purpose of explaining the processing flow, it goes without saying that the same neural network may be used.

Next, the operation of the prediction neural network 72 will be described with reference to FIG. As shown in FIG. 8, the prediction neural network 72 has a configuration in which the comparison layer 740 and the teacher layer 750 are removed from the learning neural network 71. The prediction neural network 72 predicts the unknown, corresponding to the above. For this purpose, first, as the input layer pattern in the input layer 710, the variable value Yj (DF
1 (0), DF2 (0)) and a variable value Yi (DF1 (i) to DF4 (i), i = -1, -2, ...) Set based on t = -1, -2, ...
Into the input layer 710. This variable value Yi is, as shown in FIG. 8, (1) known at present, (2) past (t =
In -1, -2, ...). Please note that these are all actual values or known data. Based on these values, the calculation of the above equations (1)-(4) is executed, and the unknown process state evaluation amount (treated water quality) DF3
(0) * and the process operation amount (return / excess sludge amount, aeration air amount) DF4 (0) * are output to the output layer 730. For the above, guidance is displayed on the communication means 46, and for the above, the control signal 72S is output to the operation control step 75.

Next, the operation control process 75 will be described below. In the operation control process 75, the signal 72S of DF5 (0) * and the fuzzy inference mechanism 61
First, the signal 61S is received and the matching between the two is checked. If the signal 72S does not conflict with 61S, the signal 75S is used as the target value of the process operation amount, and the excess sludge pump 16C and the return sludge pump 16C
Output to C1, blower 17, control valves 17A, 17B, etc.
On the contrary, if there is a contradiction, the operator 101 is notified through the communication means 46 to make a correction. The control frequency is 1 in this embodiment.
Although it is every hour, this time unit can be set arbitrarily. Of course, if the time interval is small, the prediction system will improve. When the chlorine injection amount in a short time (for example, 1 minute) cannot be predicted because the set time interval (1 hour in this embodiment) is long, it is predicted by mathematical interpolation. At the same time, the injection amount target value signal 75S of the target value (prediction result) is displayed on the communication means 46, and the actual operation is corrected by the selection of the operator 101 if necessary.

Although the present invention has been described by using the sewage treatment process as an example, it goes without saying that it can be applied to a general process.

〔The invention's effect〕

The effects of the present invention will be described below. Although the general process is the same, the conventional control method for sewage treatment plants is often used for analyzing a large amount of data, acquiring fuzzy rules by interviewing operators, and creating, adding, modifying, deleting, and improving inference rules. Labor was required. Further, in the conventional method, the fuzzy rule to be acquired depends on the subjective judgment of a specific person. However, in the present invention, fuzzy rules can be automatically acquired from past performance data, and driving guidance and control can be performed by prediction of a neural network. Therefore, if the present invention is applied, it is possible to easily perform “a vague operation in accordance with the actual results and the previous example” performed by the operator from an engineering point with less labor. In addition, since learning can be performed at any time, it is possible to quickly follow changes in the situation and learn and control.

Particularly, in the present invention, the physical / chemical / biological characteristics of the inflow water, the process state quantity, the process state evaluation quantity,
Since the process evaluation amount and the process operation amount can be predicted based on the past by paying attention to the process operation amount, the same operation as the operator can be performed.

[Brief description of drawings]

1 is a configuration diagram of an embodiment of the present invention, FIG. 2 is an explanatory diagram showing details of a history pattern data file, FIG. 3 is an explanatory diagram showing a membership conversion process, FIG. 4 is an explanatory diagram of a learning pattern, FIG. 5 is a configuration diagram of a neural network, FIG. 6 is an explanatory diagram showing a neuron element model, FIG. 7 is a characteristic diagram showing signal conversion in the neuron element model, and FIG. 8 is an explanatory diagram explaining a prediction process. is there. 9 …… First sedimentation tank, 15 …… Aeration layer, 16 …… Final sedimentation tank,
17 …… Blower, 16C1 …… Returning sludge pump, 16C2 …… Surplus pump, 9M, 15M, 16M …… Measuring instrument, 46 …… Communication means, 10
1 …… Operator, 71F …… History pattern data file, 69 …… Membership conversion process, 71 …… Learning neural network, 72 …… Prediction neural network, 73 ……
Fuzzy rule acquisition process, 74 …… Fuzzy rule diagnosis process, 75 …… Operation control process, 61 …… Fuzzy reasoning mechanism, 71
0 …… Input layer, 720 …… Middle layer, 730 …… Output layer, 740 ……
Comparison layer, 750 ... Teacher layer, 701 ... Neuron element model.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yasuo Yahagi 4026 Kuji Town, Hitachi City, Hitachi, Ibaraki Prefecture Hitachi Research Laboratory, Hitachi Ltd. (72) Harumi Matsuzaki 4026 Kuji Town, Hitachi City, Ibaraki Hitachi Research Institute, Hitachi Ltd. In-house (72) Inventor Naoki Hara 5-2-1 Omika-cho, Hitachi City, Ibaraki Hitachi, Ltd. Omika Plant, Hitachi Ltd. (72) In-house Mikio Yoda 5-2-1 Omika-cho, Hitachi City, Ibaraki Hitachi, Ltd. Omika Factory (72) Inventor Tomonori Kaneko 5-2-1 Omika-cho, Hitachi City, Ibaraki Hitachi, Ltd. Omika Factory (56) Reference JP-A-1-224804 (JP, A) JP-A-62- 274301 (JP, A) JP-A-2-178704 (JP, A)

Claims (2)

(57) [Claims] 1. A knowledge acquisition method for acquiring a fuzzy rule by learning a hierarchical neural circuit model having an input layer, at least one intermediate layer, and an output layer, the measurement being performed at an arbitrary time t in the past of the process. Value and the measured values and the manipulated variables at a plurality of time points back in the past are converted into membership values and input to the input layer, and the manipulated variables at the time t of the process are converted into membership values and the teacher signal And learning the relationship between the input value and the teacher signal value by correcting the magnitude of the weighting coefficient so that the difference between the output value from the output layer and the teacher signal value becomes small. Based on the weighting factors from the input layer to the intermediate layer and the weighting factors from the intermediate layer to the output layer of the circuit model, it is possible to acquire knowledge between the input signal and the teacher signal as a fuzzy rule. Characteristic knowledge acquisition method. 2. A process operation support method for supporting or controlling a process operation by using the fuzzy rule according to claim 1.