JPH0395603A - Operation support method for process applying multiple decision - Google Patents

Abstract

PURPOSE:To attain the synergetic effect of intelligent ability between the logical decision and the intuitive decision by providing a logical decision rule base and an intuitive decision rule base and securing the mutual drive of rules between both rule bases. CONSTITUTION:An abnormal state countermeasure module 200 identifies the abnormality and decides the factor and the countermeasure of the abnormality based on the data received from a data base module 350 and a keyboard. Then a fuzzy neuro-inference is carried out via an inference mechanism 210 when the module 200 is started. In this inference process, the backward inference is applied to the operation states whose possibilities are decided by a monitor module 120 in the order of higher preference levels. The fuzzy neuro-inference handles the rules stored in a logical decision rule base 260 and an intuitive decision rule base 280 of a rule base 250. In this case, no discrimination is substantially required among those rules of both bases 260 and 280 in the inference. In such a constitution, the logical decision and the intuitive decision can be fused together with higher interactivity.

Description

[Detailed description of the invention]

[Industrial Application Field] The present invention is a method for more effectively supporting operator judgment in overall process operation management that requires operator judgment. [Prior Art] In water purification processes, sewage treatment processes, etc., the behavior of the process cannot be described deterministically by a mathematical model, so automatic control by a computer using a mathematical model is difficult. Therefore, for the operation management of these processes, it is essential to (1) monitor and judge the situation based on the know-how of experienced operators, and (2) monitor and judge the situation that reflects the past operation history of the process. ing. As computer-assisted methods for these, knowledge engineering methods are mainly being introduced for the former, and learning methods using neural network models are being introduced for the latter. The experiential knowledge of the operator is replaced with declarative knowledge (IF-THEN type fuzzy rules and F
By extracting it as a RAME type rule, etc.) and using it together with an inference mechanism, part of the operator's monitoring and judgment can be reproduced on the Keishin machine, replacing or supporting the operator. (For example, JP-A-63-240601
In addition, neural network models do not rely on declarative knowledge;
By learning B-history data of process operation, it is used as knowledge of the distributed representation in the form of a matrix of parameters in the model. This allows the operator to make intuitive judgments that cannot be expressed in words (declarative expressions). Furthermore, attempts are being made to integrate these two methods to obtain the benefits of both, while also compensating for the disadvantages of each. Conventional methods for integrating the two methods are: (1) implementing and using declarative rules in a neural network model; and (2) calling the neural network model as a subroutine from a knowledge engineering system. ing. (For details, see Nikkei Computer 198
P53-P61 of the January 2, 1999 issue, details about ``Neurology that is beginning to be put to practical use in combination with ES''. ) [Problems to be solved by the invention] The following problems exist in the conventional method of integrating knowledge engineering methods and neural network models. First of all, by implementing declarative rules in a neural network model (by having the parameters in the model be represented in a distributed manner), we lose the logical clarity inherent in declarative rules. This has the disadvantage that the content is difficult for the operator using the support system to understand. Second, with the method of unilaterally calling the neural network model as a subroutine from the knowledge engineering method, it is not possible to actively obtain data by actively activating declarative rules from the neural network model side. It has its drawbacks. In addition, this method allows the operator to make decisions using the right hemisphere (logical judgment).
There is also the issue of not being able to sufficiently simulate the interactive functioning of the left hemisphere (intuitive judgment). An object of the present invention is to realize a more effective process operation support method by introducing a method of combining a knowledge engineering method and a neural network model that simultaneously solves the above problems. [Means for Solving the Problems] In order to achieve the above object, the present invention uses normal forward inference and backward inference for both fuzzy rules related to processes and neural network rules expressed by neural network models. Use a reasoning mechanism that allows you to do this. This inference mechanism handles both fuzzy rules and neural net rules without any distinction, except for the difference in the method of calculating the degree to which the conclusion part of the rule holds true. In addition, neural net rules are prepared for each purpose, and these rules are selected depending on the situation to quantitatively calculate the amount of operation in the process and provide guidance. [Operation] According to the present invention, expressed by fuzzy rules,
Judgment results (or data) can be freely exchanged between logical judgments based on the operator's experience and intuitive judgments based on historical data expressed by neural network rules based on a neural network model. As a result, you can obtain the synergistic effect of your intellectual abilities, both logical and intuitive judgment. This makes it possible to provide more intelligent support to the target process. [Example] The present invention intelligently supports the operational management of processes that require judgment by operators. Accordingly, the present invention is useful for various operator-mediated processes, e.g.
It can be applied to sewage treatment processes, water purification processes, chemical reaction processes, bioprocesses, nuclear processes, financial processes such as stock prices and exchange rates, etc. Examples of the present invention will be described below. FIG. 1 is an overall configuration diagram when the present invention is applied to operation management of a sewage treatment process. First, we will briefly explain the flow of the sewage treatment process. Note that the solid lines in the figure indicate the flow of materials (or the flow of system execution), and the dotted lines indicate the flow of signals (or the flow of data). Sewage flows into the initial settling tank 5 from a sewage inflow pipe 20. In the initial settling tank 5, some of the impurities and suspended solids in the sewage are removed by gravity settling. The sewage overflowing from the initial settling tank 5 and the return sludge from the return sludge pipe 40 flow into the aeration 110 . In the aeration tank 10, air is blown from the blower 50 through the aeration device 65, and the sewage and returned sludge are mixed and stirred. The returned sludge, or activated sludge, absorbs oxygen from the supplied air, decomposes dissolved organic matter in the sewage through aerobic metabolism, and decomposes it into carbon dioxide and water. A portion of the removed organic matter is used for the growth of activated sludge. The sewage from which activated sludge and organic matter have been removed is led to the final settling tank l5. In the final settling tank 15, activated sludge is separated into solid and liquid by gravity settling into activated sludge and treated sewage, and the treated sewage is discharged through a treated water discharge pipe 30. On the other hand, the activated sludge that has settled in the final settling tank 15 is pulled out from the sludge extraction pipe 35, and the multiplied portion is discharged as surplus sludge by the surplus sludge pump 60. The remaining activated sludge is returned to the aeration tank 1 through the return sludge pipe 40 from the return sludge pump 55 as return sludge.
Returned to 0. Next, measuring instruments for the sewage treatment process will be explained. A measuring device 70 is installed in the sewage inflow pipe 20 to measure the water quality of the inflowing sewage. Measurement items include inflow sewage volume, suspended solids concentration, chemical oxygen demand, pH. These include nitrogen concentration, phosphorus concentration, normal hexane extract concentration, and cyanide compound concentration. A measuring device 75 is also installed in the initial settling tank 5, and measures water quality items similar to those of the measuring device 70, as well as the interface height of settled sludge. Furthermore, a measuring device 80 and an image information measuring device 85 such as an underwater camera are installed in the aeration tank 10. In addition to the same water quality items as the measuring device 70, the measuring device 80 also measures
Dissolved oxygen concentration etc. are measured. Image information measuring device 85
Then, the distribution and color of the activated sludge in the aeration tank 10, the size of flocculating microorganisms (flocs) constituting the activated sludge, and the shape and amount of filamentous microorganisms and protozoa are measured. The final settling tank 15 includes a measuring device 90 and an image information measuring device 9.
5 will be installed. In addition to the water quality items similar to those in the measuring device 70, the measuring device 90 measures the interface height of activated sludge that has settled due to gravity. In the image information measuring device 95, in addition to the measurement items similar to those of the image information measuring device 85, the final settling tank 15
The presence or absence of a hydrophobic microbial film (scum) on the water surface is measured. In addition, the treated water discharge pipe 30 includes a measuring device 1.
00 is installed, and the same water quality items as the measuring instruments are measured for treated water. Next, the structure of the computer system 110 that provides operational management support (multi-judgment type operational management support system, hereinafter abbreviated as support system) will be explained. The support system 110 includes (1) a monitoring module 120;
(2) Abnormality countermeasure module 200, (3) Maintenance module 300, (4) Database module 350
, (5) configured from the image processing module 370. The abnormality countermeasure module 200, which is a feature of the present invention, is composed of an inference mechanism 210 and a rule base 250, and the rule base 250 is further composed of a logical judgment rule base 26o and an intuitive judgment rule base 280. ing. Details of these will be described later. Next, input/output of various measurement data to the support system 110 and interactive data input/output with an operator will be explained. First, input will be explained. Said measurement 70, 75
, 80, 90, 100, and image information measuring devices 85, 9
5, data sampling is performed at regular time intervals. The online data of these analog signals is taken into the input/output device 101, converted into digital signals, and then taken into the support system 110 through the input port 102. On the other hand, input of (1) manual analysis data that cannot be measured online, (2) data that cannot be measured with an image information measuring device, and (3) data that the support system deems particularly necessary is connected to the support system 110. This is done interactively using the keyboard 350 that has been created, while referring to the CRT 360. Next, the output will be explained. The support system 110
Determine the guidance content based on the input information. This guidance is displayed to the operator via CRT360. Note that this CRT 360 can also serve as a monitor that displays images from the image information measuring devices 85 and 95, if necessary. Further, changes in the target values and operation amounts of the blower 50, return sludge pump 55, and surplus sludge pump 60, which are devices for controlling the process, are performed from the keyboard 350, and after being recorded in the support system 110, Output port 1
04 to the control device 103. Control signals are sent from the control device 103 to each device. The above is the overall configuration of this embodiment. Next, the specific operation of the support system 110 of this embodiment will be explained with reference to FIG. In FIG. 2, solid lines indicate the flow of activation between modules, and dotted lines indicate the flow of data. In the operational management of sewage treatment processes, the type and order of tasks performed by operators are modeled by simulating the thought process (
Management business model) Then, as shown in Figure 2, (1) the monitoring module 120, (2) the abnormality countermeasure module 200, (3) the maintenance module 300, (4) the database module 350, and (5) the image processing module 370. be. The relationships and execution order of each module in the above management business model created by the inventors are as follows. First, the monitoring module 120 is activated, reads online data from the database module 350 at a predetermined cycle, and, if necessary, requests the operator to input data through the CRT 390.
Receive input from 80. The monitoring module 120 constantly monitors the operating status of the process based on this data. The monitoring results are displayed on the CRT 390 and sent to the determination section 199. When the determination unit 199 recognizes that there is a possibility of an abnormality in the driving situation, the abnormality countermeasure module 200
is started. Based on the data from the database module 350 and the keyboard 380, the abnormality countermeasure module 200
Identify abnormalities, determine their causes, and countermeasures (quantitative manipulated variables, etc.). The inference results are displayed as guidance through the CRT 390. The maintenance module 300 can be activated from the monitoring module 120 and the abnormality countermeasure module 200 at any time. Execution is performed by transferring the right to interact with the operator from the program that started it. This module displays various tools for maintenance of the support system 110, manual knowledge such as maintenance guidelines necessary for daily maintenance work, search and display of past driving history, and seasonal fluctuation commands. A keyboard 380 and CRT 39 are used for calculation tools such as extraction of data and definition of membership functions in fuzzy logic for various data, as well as various simulators.
Provides functionality to run interactively through 0. Database module 350 structures and stores all data read into support system 110 through input port 102 . Furthermore, input data from the keyboard 380 and execution results in each module are also stored in the database module 350 as necessary. Of the data read from the input port 102, the image processing module 370 takes in only the data from the image information measurement devices 85 and 95 shown in FIG. do. By introducing the above-mentioned management work model, it becomes possible to provide highly compatible support that is closely related to the operator's work. Furthermore, detailed operations within each module will be explained in order. First, the specific procedure of the monitoring module 120 will be explained using FIG. 3. FIG. 3 is a flowchart showing the operation of the monitoring module 120. During execution of the support system 110, the monitoring module 120 reads online data from the database module 350 at predetermined intervals, and receives interactive input from the operator using the keyboard 380 as needed. These data are converted into a predetermined format in a data input step 130 and evaluated in a data evaluation step 140. Knowledge regarding evaluation is stored in the monitoring knowledge base 121 and is referred to as necessary. The evaluation here determines several operating conditions of the process (
Select the possible operating conditions from among (normal, bulking, nitrification, sludge putrefaction, etc.). The results are sent to the determination section 199. If there is at least one selected operating condition (excluding normal), the abnormality countermeasure module 200 is activated. As a method for implementing the data evaluation technique 8140, here, knowledge engineering and fuzzy reasoning using fuzzy logic, which is a known technology, is used. In other words, this is a method in which the operator's experiential knowledge regarding situational judgment is described as fuzzy rules, and these rules are activated based on input data. For more information on fuzzy inference, see other references (Nishida, Takeda (1978)
) Mathematics Library 48 Fuzzy Sets and Their Applications, Morikita Publishing, etc.), but the basics are to find the membership function of the conclusion part using equations (1) and (2), and then quantify the conclusion part by calculating the centroid of the functional form. It is about finding the value.・If rule condition parts A and B are logical and combinations, MF (
A, B) =: min {MF(A), MF(B))
-Formula (1)・When rule condition parts A and B are logical or combination MF(A, B) = n+ax(MF(A), MF
(B)} - Formula (2) In this case, data input step t3
0, numerical data is converted into qualitative data (fuzzy variables, their values, and their membership values; for example, water temperature = high, 0.857'.j, etc.) by the membership function. Another implementation method is to use a simulator based on a process phenomenon model. For example, prepare a model related to treated water quality, input actual measured data into it, simulate the treated water quality (mainly predicting the future), and check whether the value is within an acceptable range. As yet another implementation method, a method may be used in which the neural network model is made to learn historical data regarding past driving conditions. This method determines whether the same driving situation as the learned driving situation has occurred by recalling actual measured values. If none of the model's recalled values for a certain actual measured input value reaches a predetermined value, it is determined that this situation has never occurred in the past (or it has not been learned),
New learning will be conducted [Realization of a learning-based self-growth support system]. The neural network model used here and its learning algorithm will be described later. The above is the operation of the monitoring module 120. The driving status monitoring results from this module are sent to the determination section 199, and if at least one possible abnormal condition is recognized, the abnormality countermeasure module 200 is activated. If not, the result is stored in the database module 35.
0 and repeats the next monitoring module execution. Next, the abnormality countermeasure module 200 will be explained with reference to FIG. FIG. 4 is a flowchart showing the operation of the abnormality countermeasure module 200. While the previous monitoring module 120 was centered around online data, this module requires more data to be entered interactively. When this module is activated, it first performs a fuzzy neuroinference step 211. In this step, regarding the operating situation that is determined to be possible by the monitoring module 120,
Perform backward inference in order of priority (for example, highest membership value). Determine which driving situations are really likely. The fuzzy neuro-inference algorithm used here is based on rules with certainty or inference used in fuzzy rules (known technology: Winston + P.).
l. (1977) Artificial Intel
ligence, Addison Iiesley, et al. ) has been extended so that neural net rules can also be applied. Before explaining fuzzy neuroinference, we will briefly explain the outline of the neural network model. What is the basis of the neural network model? rq to tf
It is a nerve cell (neuron) that performs ytIi. Mathematical models of single neurons that reflect this physiological knowledge of neurons are widely used. The model shown in FIG. 5 is expressed by a function having threshold characteristics of large force and one output as shown in equation (3). y=f (Σwi-xi-θ)...Equation (3) However, y
: Output signal strength of neuron wi: Weighting coefficient xi of i-th input: Shinchin strength θ of i-th input: Threshold In other words, the signal xi from another neuron i to the neuron of interest is Wi is multiplied according to the connection strength (weighting coefficient, synaptic strength). The idea is that when the sum Σwi-xi of all connections of these weighted input signals wi-xi exceeds a certain threshold value θ, the neuron is excited and outputs an output signal y. As the function f that determines the threshold characteristic at this time, a sigmoid function shown in equation (4) is used. f (u)=L/ (1+e-υ)...Equation (4) The neural network model is a circuit model whose basic elements are the neurons described above. Two types of models are widely known: a hierarchical model in which neurons are connected hierarchically, and a mutually connected model in which neurons are connected in a network. In the present invention, the former hierarchical model (also known as Rul
emhart type model). This model consists of three layers, an input layer, a middle layer, and an output layer, as shown in Figure 6.
It has a layered structure. During learning, the weighting coefficients are modified from the output layer toward the input layer so that the error between the signal from the output layer (recall value) and the teacher signal when a learning pattern is input to the input layer becomes smaller. (Details of the algorithm can be found in Rumelhar
t, D. E. , ate (1986) Learni
ngRepresentations by Back
Propagating Errors, Nature
, vol. 323, etc.). In fuzzy neuro-reasoning, logical judgment rule base 260. of rule base 250. The rules stored in the intuitive judgment rule base 280 are handled. Each rule in both rule bases is an IF-T as shown in Figure 7.
They are written in the same HEN format, and there is almost no difference between the rules of either rule base.
can be inferred. In other words, fuzzy rules can activate neural net rules, or the inference results of neural net rules can be used to advance inference, and vice versa. The only difference between the two rules is the method used to calculate the degree to which the conclusion part of the 5 rules holds true. In the case of fuzzy rules in the logical judgment rule base 260, calculations are made using fuzzy inference, a known technique, as shown in FIG. In other words, the calculation is performed by taking the logical product (min operation) of the membership functions in the condition part, making it the membership function in the conclusion part, and taking the center of gravity of this membership function to find its degree. The neural network rules in the intuitive judgment rule base 280 are rules for causal relationships between data, which operators and experts are not clearly aware of, by learning historical data using a neural network model. . As shown in FIG. 9, the rule condition part corresponds to the neuron of the input layer, and the conclusion part corresponds to the neuron of the output layer. To calculate the degree to which the conclusion holds true, recall is performed using a neural network model. During recall, [fuzzy variables] derived from the database module 350 and fuzzy rules are added to the input layer, as shown in the example below.

[Fuzzy variable value] 〔Effect of the invention〕

As described above, according to the present invention, the fusion of logical judgment (left brain: knowledge engineering and fuzzy logic) and intuitive judgment (right brain: neural network model), which was insufficient with conventional methods, is improved. It can be made interactive. This accurately simulates the judgments of process experts and operators, making computer assistance more intelligent. As a result, higher quality operation can be achieved than when the operator is alone, and the burden on the operator can be reduced.

[Brief explanation of drawings]

FIG. 1 is a schematic diagram showing the overall configuration of an embodiment of the present invention, and FIG.
The figure is a configuration diagram of the support system in the embodiment, FIG. 3 is a flowchart showing the operation of the module that supports monitoring, and FIG.
The figure is a flowchart of a module that supports countermeasures when a process abnormality occurs, Figure 5 is a schematic diagram of a single neuron model, Figure 6 is a configuration diagram showing a hierarchical neural network model, and Figure 7 is a schematic diagram of a single neuron model.
Figure 8 shows the format of rules used in the support system, Figure 8 is a flow diagram explaining fuzzy inference, Figure 9 is a conceptual diagram explaining neural net rules, Figure 1
FIG. 0 is a configuration diagram of a module that supports maintenance management, and FIG. 11 is a flowchart of a module that performs image processing. 5... First settling tank, 10... Aeration tank. 15...Final sedimentation tank, 50...Blower 55...Return sludge pump, 60...Excess sludge pump, 65...Diffuser, 70, 75, 80, 90, 100...Measurement Machine, 85
．． 95... Image information measuring device, 101... Input/output device, 103... Control device, 110... Support system,
120... Monitoring module, 200... Abnormality countermeasure module, 210... Reasoning machine 4t. 250.. ．． Rule base, 260...Logical judgment rule base, 2
80...Intuitive judgment rule base, 300...Maintenance module, 350...Database module, Fig. 1-21 Fig. 5 Fig. 6 Figure 3 l9'1 Figure 4 Figure 7 Figure 8 Figure 9 F+=η Fx-VEF,=v, Figure 1 350

Claims (1)

[Claims] 1. In the operation management of a process that requires operator judgment, at least a logical judgment rule base consisting of fuzzy rules regarding the process and a neural network rule based on a neural network model that can learn about the process. A process operation support method based on multidimensional judgment, characterized by having an intuitive judgment rule base and an inference mechanism capable of mutually driving rules between both of these rule bases. 2. In claim 1, the intuitive judgment rule base includes at least a neural net rule whose conclusion part relates to the amount of operation of the process, and selects the neural net rule that is most suitable for the state of the process estimated by the inference mechanism. A process operation support method based on multidimensional judgment, which is characterized by being able to provide quantitative guidance on the amount of operation by executing the following steps. 3. In claim 1 or 2, the inference mechanism uses a logical operation of a membership function for a fuzzy rule, and uses a logical operation between neurons for a neural net rule as a means for calculating the degree to which a rule conclusion part holds true. A process operation support method based on multidimensional judgment, characterized by using a product-sum operation. 4. According to claim 1, the neural network model learning algorithm is a method for modifying the connection strength distribution within the network so that the error between the output and the teacher signal for the output is reduced. How to support process operation through judgment.