JP3186898B2 - Support system using neural network - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic additional learning method for reconstructing a neural network in a system using a neural network. The present invention relates to a plant operation support system capable of performing highly predictive guidance.

[0002]

2. Description of the Related Art The operation management of various plants including a water treatment process and a sewage treatment process requires experience and know-how of operation management by a skilled person. However, in recent years, because of the shortage of skilled personnel at the plant operation management site and the lack of experience of young operation managers, empirical knowledge of skilled operators is extracted and used together with the inference mechanism, a so-called knowledge engineering method. Is increasingly being applied.

A driving support system using such a knowledge engineering technique is logically clear in the inference process and can present the basis of the guidance by the inference, so that it is persuasive to the operator. However, on the other hand, there is a problem that acquisition of such empirical knowledge and maintenance are complicated. So, as a solution to these problems,
Driving support technology based on a neural network model has been introduced. A plant operation support system using this neural network builds a network by learning based on past history (experience), and a neural network (neural network) built from the past experience , The prediction result is calculated. In other words, each time a new net configuration is learned like a human, the influence of the past net configuration gradually diminishes (oblivious), and a pattern that is repeatedly learned forms a neural network with a strong causal relationship. is there.

That is, in the driving support system based on such a neural network model, the knowledge corresponding to the experience and intuition of the operator buried in the driving history data is learned, and the network obtained by the learning is recalled. Can provide only relevant guidance. This is described, for example, in Japanese Patent Laid-Open No. 4-133164.
As described in Japanese Patent Application Laid-Open Publication No. H10-209, if there is driving history data, it means that driving support can be realized by automatic additional learning without complicated steps of acquiring knowledge. As a method of learning in an automatic control device using such a neural network, Japanese Patent Application Laid-Open No.
According to Japanese Patent Application Publication No. 101-101 and the like, there is known an apparatus for re-learning a neural network when a state monitoring unit detects a control state.

[0005]

On the other hand, in a driving support system using a neural network model according to the prior art described above, especially in a neural network constructed by a conventional automatic additional learning method, the average recall error is reduced. In some cases, learning and recall takes time and is inefficient, and depends on the performance of the hardware (the burden on the hardware is large). This requires a general-purpose computer instead of a computer dedicated to neural network control. There is a problem that it cannot be realized with a system that performs processing while performing operation control and other processing, and it is impossible to actually adopt the system.

Further, the above prior art disclosed in Japanese Patent Laid-Open No.
In Japanese Patent No. 0101, although re-learning is performed according to a change in the state of the plant, there is no idea to reduce the average recall error.

Further, as will be described later in detail, the above-mentioned conventional technique can be used even in an inexperienced situation or a sudden situation change.
Precise predictions were difficult.

In view of the above, the present invention has solved the problem of the conventional automatic additional learning by taking advantage of the fact that it is possible to make a prediction based on the experience and intuition of the operator of the neural network model, that is, to reduce the average recall error. It is an object of the present invention to provide a learning method for automatically adding a neural network with less burden on hardware, and a plant operation support system using the learning method.

Further, the present invention provides a plant operation support system capable of providing highly accurate prediction guidance even in an inexperienced situation or a sudden situation change, which is difficult with the conventional automatic additional learning method. That is its purpose.

[0010]

SUMMARY OF THE INVENTION The present invention utilizes a neural network constructed while learning the past history automatically with the operation of the system, and assists the operator who operates the system in a changing situation. A support system for calculating and displaying a predicted value for the purpose of automatically performing additional learning during a learning period in a combination of a learning period and a learning determination history period that minimizes an average recall error, and performing the automatic additional learning. A support system using a neural network, characterized in that a combination of a learning cycle and a learning determination history period that minimizes the average recall error is determined according to a rapid change in situation. I do.

[0011]

In other words, according to the support system using the neural network, the neural network automatic additional learning method, and the plant operation support system according to the present invention, a plant or the like to be monitored described in detail below is used. By setting an optimal value of the learning period and the number of days for learning determination according to the characteristics, and thereby performing learning and recall, it is possible to efficiently and reduce the average recall error.

Also, by enabling efficient learning and recall, the dependence on hardware is reduced, and even in hardware where learning with a small average recall error is difficult in the conventional learning method. Learning and recall with a small average recall error can be performed.

Further, according to the automatic additional learning method proposed in the present invention, by performing the switching process of the automatic learning period, the responsiveness to sudden changes in the situation is improved, and the average recall error is always improved. It is possible to provide the operator with less stable, optimal prediction guidance,
It is possible to execute highly reliable driving support.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Before describing embodiments of the present invention, a conventional neural network (neural network) model which forms the basis of the present invention and constitutes the most important part of the present invention will be described. The analysis and consideration of problems by the present inventors in a system (operation support system for a plant or the like), and further, matters that have been confirmed and recognized by the present inventors regarding solutions to the problems will be described in detail below.

A system based on a conventional neural network model (particularly, an operation support system for a plant or the like)
Had the following issues.

First, as a first problem, in order to reduce the average recall error, the learning time increases as the number of learnings per learning increases, and a network with good responsiveness to a situation change is constructed. For example, a method such as minimizing the learning period is used, and the dependence on hardware performance becomes extremely high.

[0017] Here, the "average recall error" is, for example, in neuro system shown in FIG. 10 (a), gives 10 times the teacher data as shown in FIG. 11, the actual teacher output data y _i at that time This is the average value of the difference (y _i −Y _i ) from the output Y _i (see FIG. 10B), and is defined as in Equation 1 below.

(Equation 1)

Further, this is generally defined as the following equation (2).

(Equation 2) Here, Y _i is an output value obtained within the permissible learning time at the i-th time and a value that is a set value smaller than the permissible error value. In general, as shown in FIG. 10 (c), the error (| Y _i −y _i |) gradually decreases as time elapses (learning progress) from the start of learning (t ₀ ), and the error is allowed. in learning time t _p, and terminates the learning by selecting alpha _i of when the error value becomes smaller than the tolerance value alpha _th (construction of a neural network). It should be noted that the allowable learning time t _p
Do not learn beyond.

According to the present inventors, factors for reducing the above-mentioned average recall error include a learning period and a learning determination history period, and between these learning periods and the learning determination history period. It is recognized that there exists an optimal combination according to the characteristics of each plant to be controlled, for example. However, conventionally,
The relationship between the learning determination history period and the learning cycle is not taken into account so much, so that there is a problem that the difference between the average recalls does not easily converge.

Here, in the present invention, when performing the automatic learning of the neural network, the time interval at which the automatic learning is performed is referred to as an "automatic cycle". The condition for updating the neural network is N days of past history data. The number of days or hours for determining whether to determine the average recall error is referred to as a “learning determination history period”. That is, the “learning determination history period” refers to, for example, the number of past days (however, there may be hours other than the number of days) of how much data in the past is used as teacher data.

As a second problem, conventionally, in a system in which data is normally stable in a normal state, a network corresponding to a stable input condition is constructed by repeatedly learning the stable data. In response to the rapid change in the situation, the network that responds to the situation change cannot be quickly rebuilt, which makes it difficult to provide appropriate prediction guidance (performs first-order lag prediction guidance). is there.

On the other hand, as described above, in a system using a general neural network (for example, a plant operation system), the average recall of the predicted value calculation is made in accordance with the controlled object (for example, plant characteristics). In order to minimize the error, the present inventors have confirmed and recognized that an optimum value exists in the relationship between the automatic additional learning period and the learning determination history period. For example, FIG. 3 shows a learning determination history period when learning / recall is performed in a neural network constructed in a certain plant with automatic addition learning at a fixed period (however, in FIG. It is assumed that the period is the number of days, and is referred to as “learning determination days”) and the average recall error. As can be seen from FIG. 3, when the automatic additional learning is performed at a constant cycle,
There is an optimal learning determination history period (days) in which the average recall error is minimized. That is, in order to minimize the average recall error, the optimal learning determination history period (number of days) is determined by determining the automatic additional learning period. Further, the relationship between the learning cycle and the learning determination history period (days) for minimizing the average recall error is such that the learning cycle and the learning determination history period (days) are proportional as shown in FIG. .

Therefore, the present inventors use the optimum value in the relationship between the automatic additional learning period and the learning determination history period for minimizing the above-mentioned average recall error in the calculation of the predicted value, thereby solving the above problem. It is the first finding that it is possible to eliminate.

Further, when the relation between the automatic additional learning period and the learning determination history period mentioned in the first problem is considered as a precondition, the third problem is to follow a change in an input value or a recall value. In order to construct a good-quality neural network, it is conceivable to speed up the reconstruction of the neural network by increasing the automatic additional learning period. However, in order to reduce the average recall error from the above prerequisites, it is necessary to shorten the learning determination history period by increasing the learning period. In this case, increasing the learning period imposes a heavy burden on hardware. (Depends on hardware performance).

On the other hand, when an attempt is made to construct a neural network with good follow-up performance by accelerating the automatic additional learning period, the follow-up performance of predicted data with respect to actual measurement data is improved (a network that responds quickly). By reducing the number of judgment days, a network that is not easily affected by past history data is constructed, and it is easy to fall into prediction data that traces actual data without control. Specifically, by recalling a network built under the influence of disturbances (noise), even when the data returns to stable data in normal times, it is possible to provide guidance for predicting a large accumulated error due to abnormal data. There is a problem that there is.

Further, as a fourth problem, in order to improve the third problem, it is necessary to increase the number of days for learning and to construct a network taking into account past history data. To reduce the error, it is conceivable to increase the learning period. However, when an abnormality is caused by a disturbance, a network corresponding to the sudden change cannot be constructed, and the prediction data performs a temporally delayed prediction with respect to the actual data. That is, the past history data greatly affects the causal relationship of the neural network constructed under the condition that the learning determination days are large (forgetting is slow).
Due to the large learning period, the network cannot be quickly reconstructed in response to a sudden change, and a problem arises in that only networks having poor followability (slow network reconstruction speed in response to a situation change) can be constructed. More specifically, if a sudden change in the situation does not allow predictive guidance to follow the change, and the disturbance returns to a stable state after a continuous disturbance, the predictive guidance may not be stable for a long time, A problem such as an increase in the accumulated error due to the prediction can be considered.

These problems are also solved by using the optimal value in the relationship between the automatic additional learning period and the learning determination history period for minimizing the average recall error in the above-described prediction value calculation. Has been confirmed to be possible.

As is clear from the above, in the present invention,
Neural network model systems, especially
In view of the various problems and circumstances described above in operation support systems such as plants, reduce the dependence on hardware, reduce the average recall error efficiently, and perform not only quantitative guidance, but also Also, in order to minimize the average recall error recognized by the present inventors for the first time, a practical system (particularly, a plant operation support system) having good responsiveness and always performing optimal prediction guidance is provided. Is newly created based on the relationship between the learning cycle of the learning and the learning determination history period.

According to the present invention, as means for finding the optimum combination of the learning period and the learning determination history period, which is the basis of the present invention, specifically, when constructing a new neural network, By recalling the combination of the learning period and the learning determination history period, an optimal combination of the learning period and the learning determination history period with a small average recall error is found.

Further, according to the present invention, there is provided means for performing processing switching in response to a sudden change in the state of an input value or a recall value while comprehensively considering actual problems in an actual support system. We propose a plant operation system equipped with. Also, the history data used at that time is
Recalling is divided into normal ones and abnormal ones (when roughly divided into two bands), and the learning cycle and the number of learning determinations are simultaneously switched in the switching process, so that the recalling with a small average recall error Can be performed.

As means for switching the processing in response to a sudden change in the state, there is a method for performing the automatic additional learning.
For example, when a stable / abnormal value can be set in an input value or a recall value, a determination logic for determining stability / abnormality is provided to provide a process changeover switch for performing a process corresponding to stability / abnormality. And

In a system in which the band of the stable data is already known, a trigger level is provided for the stable band, so that the processing is switched only at the time of a sudden change outside the band, so that a sudden change in the situation may occur. By quickly rebuilding a network that supports
The above-mentioned problem has been solved. Therefore, in various systems involving an operator using the automatic additional learning method for reconstructing a neural network proposed by the present invention, a sudden change in data occurs when the system is stable and abnormal, or when the system is abnormal and stable. Also in this case, by performing the automatic addition learning period of the teacher data more frequently than in normal times, it is possible to perform optimal prediction guidance with good followability corresponding to a situation change.

On the other hand, when the data fluctuates suddenly from the time of stability to the time of abnormality or from the time of abnormality to the time of stability, the past history of the teacher data (reference data) is shortened to reduce the past history. It is possible to construct a net that has little effect on the causal relationship that the data creates in the neural network, that is, a fast forgetting network. A stable prediction with good tracking performance can be performed by an integrated system in which the learning period is shortened and the number of days for determination is reduced. Further, a plant system in which a situation change is classified into several patterns has a multiplex neural network, and by switching the network, the cumulative error of the prediction guidance is reduced. With the above, it is possible to construct a neural network logic that can respond quickly and quickly to stable data during sudden fluctuations while having experience at the time of abnormalities, thereby maximizing the utility value of the neural network. You can do it.

In order to more specifically explain the present invention described above, an embodiment in which a plant operation support system using the present invention is applied to, for example, operation management of a water purification process will be described. FIG. 1 is a block diagram showing the overall configuration of such a plant operation support system. The configuration and the procedure of the water purification process are described below.
This will be briefly described below with reference to the drawings.

In FIG. 1, source water from a river or lake is led to a landing bowl 100. In the rapid mixing pond 110 into which water from the landing well 100 is introduced, the coagulant in the coagulant tank 111 is injected by the coagulant injection pump 112, and the stirring blade 113 is driven by the stirrer 114 to stir. Here, an alkali agent that promotes floc formation may be injected.

Subsequently, the water from the rapid mixing pond 110 is led to the floc forming pond 120. This floc forming pond 12
A stirring paddle 121 is arranged at 01 and is gently rotated. Thereafter, the formed flocs are transferred to the sedimentation basin 130.
, And the supernatant is filtered through a filtration pond 140.

Next, the measurement system will be described. First,
In order to measure the quality of raw water taken from rivers and lakes,
The landing well 100 is provided with a measuring device 100M. Measurement items are water temperature, turbidity, alkalinity, electric conductivity, p
H, residual chlorine, chlorine demand, water volume, water level, etc.
A measuring instrument 120M is installed in the floc formation pond 120, and the measuring instrument 120M
In addition to the items measured at 00M, it includes an imaging means 160 such as an underwater camera.
Similarly, a measuring instrument 130M is provided. Further, as shown in the figure, a measuring device 110M similar to the measuring device 120M is installed in the rapid mixing pond 110, and
It is also possible to install a measuring device 140M on the 40.
The measurement items by these measuring instruments are the same as the measurement items by 100M and 120M. Then, the measurement data from the above-mentioned measuring instrument is transmitted to the support system 2
The data is taken into the operation history database 180 in 00.
The data obtained by the imaging means 160 is processed by the image processing means 170, and then is taken into the operation history database 180.

Next, an intelligent driving management support system (supporting system) 200 for performing driving management support will be described. The support system 200 includes the following subsystems. That is, (1) neural network application support system (neuro system) 230, (2) image processing means 170, and (3) operation history database 180.
And, the neurosystem 2 which is a feature of the present invention
30 includes learning means 250, quantitative guidance and display means (guidance display means) 380, and neural network base 400.

The neuro system 230 first selects data considered effective from the past history data from the driving history database 180 as teacher data. The teacher data is learned by the neural network in the learning means 250, and is recalled. In addition, the role of the neural network after learning is to find quantitative guidance necessary for driving management by recall and send the result to the quantitative guidance and display means 380. Further, the knowledge extracted here is stored on the CRT 190. Is displayed as quantitative guidance and extracted knowledge.

Next, input and output of various data to and from the support system 200 will be described. First, the input will be described. The measuring instruments 100M, 110M, 120M, 130
In the M, 140M, and the imaging means 160, data is sampled at required time intervals, and the sampled data is sent to and saved in the operation history database 180 of the support system 200. In addition, for example, hand analysis data and visual observation data, which cannot be measured online by a measuring instrument, are separately communicated via input means such as a keyboard 195 shown in the figure with reference to a message displayed on the screen of the CRT 190. Input.

Next, the output relationship will be described. The support system 200 determines the content of guidance based on the input data. That is, the guidance obtained from the neuro system 230 in the support system 200 is displayed to the operator through the screen of the CRT 190. It should be noted that the CRT 190 can also serve as a monitor for displaying an image from the imaging means 160 as necessary. The operator refers to the guidance displayed on the CRT 190 and inputs a change or the like that is deemed necessary for the operation amount to the control unit 150 via the keyboard 195 as the input unit. Therefore, the control means 150
Each device of the process is controlled according to the input data. Note that the guidance from the support system 200 may be directly input to the control unit 150. The above is the outline of the overall configuration and operation of the present embodiment.

Next, a neural network (neural network) performed in the neural system 230 which forms the core of the present invention.
The model learning and knowledge extraction (recall) method will be described step by step below. Learning of the neural network model is performed by modifying the weight coefficient matrix so as to reduce the average recall error. As the learning algorithm, the conventional BP method that is most frequently used is adopted, that is, the sum of squares of the average recall error is defined as an evaluation function, and the weighting coefficient matrix is modified so that the evaluation function decreases. .

Next, the neuro system 230 of this embodiment is described.
A specific method of the learning means 250 according to the present invention will be described by taking, as an example, the case of operating in a water purification process. First,
In the water purification plant, since the quality of the inflow water and the characteristics of the measuring instrument change with time, the coagulant injection method changes accordingly. That is, if the neural network performs additional learning flexibly in response to this situation change, driving support having self-growth can be performed. FIG. 2 shows an outline of such an automatic additional learning method.

The concept of this operation procedure is as follows: (1) Only data that has good processing and does not largely conflict with the past operation history are reflected on the net. (2) Reflect new data most strongly while respecting past driving histories. (3) Automatic learning without maintenance. By doing so, it is necessary to always perform optimal prediction guidance in adaptation to the structural change of the plant data group. That is, the prediction guidance is performed while performing the automatic additional learning based on the procedure of the automatic additional learning described below.

Specifically, first, a new neural network is constructed before operation (260), and automatic learning is performed after operation of the actual machine. That is, learning is performed by the learning means (280), and thereafter, it is determined whether or not it is a guidance cycle (370). As a result, "N
o ", the learning means (2
Returning to 80), if it is determined to be “Yes”, the guidance is displayed via the CRT 190 (39).
0).

FIG. 5 shows details of the new neural network construction 260 shown in FIG. This new neural network construction 260 is a procedure for constructing a neural network suitable for the actual machine before operation of the actual machine. The neural network constructed by learning is used as an initial network in the neural network base 400 of the neural system 200 described above. Store.

Details of the new neural network construction 260 are as follows. First, in steps 265 and 268, learning / recall is performed using existing data at an arbitrary initial value, and learning / recall is performed until the average recall error becomes smaller than the target error. Repeat the recall. Subsequently, in step 270, the network in which the average recall error has become sufficiently small is stored in the neural network base 400. Further, in the following step 275, the network constructed up to the previous step 270
“Learning cycle” and “Learning judgment history”, in which learning / recall is performed by combining “Learning cycle” and “Learning judgment history period (days)”, and the average recall error is minimized for each case such as normal / abnormal. Set the optimal combination with "Period". Note that the setting of the “learning cycle” and the “learning determination history period” is switched by processing switching described later.

Here, the most important fact for the present invention, the grounds for the existence of the optimum combination of the learning period and the learning determination history period (days) will be described again with reference to FIG. Will be described. The bold line A drawn in the middle of FIG. 3 indicates that the network constructed in steps 265 and 268 in FIG. 5 is learned and recalled at a constant learning cycle, and at the same time, the number of learning determination days is changed as a parameter. , The average recall error. This graph shows that when a “learning cycle” is determined to a predetermined value in a certain network, an optimal learning determination day (learning determination day at which the error is minimized) is determined in order to minimize the average recall error. It shows that. At the same time, when this “learning cycle” is set to be large (longer than the bold line), as shown by the curve b in the figure, the optimal learning determination days at which the average recall error is minimized increase, and conversely,
When the learning period is set to be small, the optimal learning determination days also become small, as indicated by the curve a in the figure.

That is, the relationship between the “learning cycle” and the “learning determination history period (number of days)” for minimizing the average recall error, ie, the optimum, is substantially proportional to each other as shown in FIG. You can see that it has become. Therefore, in the present invention, these "learning cycles"
And a plurality of combinations of “learning determination history period” and an optimal value of “learning cycle” and “learning determination history period” are set. Here, when the process switching (SW function) described later is used, it is necessary to set these optimal combinations as combinations for stability / abnormality (when the input value is roughly divided into stable and abnormal). is there.

Next, the details of the learning means 280 and the guidance display 390, which are the procedures of the automatic learning after the operation of the actual machine, shown in FIG. 2 will be described below. In addition,
In the present embodiment, as an additional learning method of the neural network, automatic additional learning is performed.

FIG. 6 shows a schematic block diagram of the learning means 280 of FIG. FIG. 6 is a block diagram showing a series of processes from learning and recalling of data (database without history data) measured at a fixed cycle by a plurality of instruments. First, in the input data 285, a history database or data (for example, pH, turbidity, water temperature, etc.) is fetched for each learning cycle. next,
In the input value determination logic 290, a flag (trigger switch) corresponding to the change in the input value is provided, and processing is switched by setting / resetting the flag in order to reconstruct a neural network corresponding to a sudden change in the input value. Next, in the neural network additional learning process 300, the quality of the previous process is determined. If the previous process is good, the additional learning of the previous data is performed as an initial value of the current net to form a new net. Both of the nets are recalled for N days of learning determination days, and when the average recall error of the new net is smaller, the new net is stored in the neural net base 400 as the current net.

Next, in the recall 330 of the current operation amount, the operation amount and the like are recalled based on data obtained from a measuring instrument or the like on the current net, and guidance is displayed. Next, in the recall determination logic 340, the input value determination logic 29
Similarly to 0, the SW 350 is set / reset to switch the process due to a sudden change in the recall value. Finally, in SW350, the input value determination logic 29
With reference to 0 and the flag set / reset by the recall value determination logic 340, the stability / abnormality is determined while considering the relationship between the optimal “learning cycle” and “learning determination history period” which is the basis of the present invention. The processing is switched according to the time (stable band / abnormal band) (set of the learning period and the number of days for learning determination), and the automatic additional learning is performed.

Here, the process switching, which is the second feature of the present invention, will be described in detail by taking the above-mentioned water purification plant as an example. In a system such as a water purification plant (including other general plant systems), there are stable input values in a steady state, and a network is constructed accordingly,
Predictive guidance is provided by recalling the network. However, on rare occasions, the input value may suddenly change due to the influence of disturbance (change in weather conditions) or the like.

However, in a conventional driving support system using a neural system, depending on the automatic additional learning method, as shown in FIG. Network is formed, and a network with little experience of disturbances is formed.) It is difficult to quickly build a network that responds to sudden changes in input, and it is difficult to predict optimal guidance by recall. Cases arise.

Therefore, in the input value determination logic 290 and the recall value determination logic 340 shown in FIG. 6, the processing is switched according to a change in the input value or the recall value, whereby the input value is rapidly changed due to disturbance or the like. Like rebuilding a neural network that responds to changes,
It is proposed to provide a flag switch (trigger switch) corresponding to a change in the input value, and to switch the processing by setting / resetting the flag switch. Here, the flag switch means that the set / reset sets a trigger level (a level set by edge sensing and level sensing) with respect to a change amount of input or teacher data.
This is a software switch that sets a flag when the level is exceeded (passed).

With regard to the trigger level setting reference, for example, the stable band described in FIG. 7 is set as a trigger level for level sensing. In the edge sensing for switching the processing for a rapid change, a differential value level (change amount) such as an input value is set, and a flag for switching the processing is set when the level exceeds each set level.

A method for performing the process switching, which is the second feature of the present invention, will be described with reference to FIG.
First, in a stable state, the number of days for learning determination is increased in order to construct a network in consideration of past history data (past experience). Accordingly, the learning period is increased to perform automatic learning (see the relationship between the learning period and the number of learning determination days in FIG. 4). On the other hand, when the state transitions from the stable state to the abnormal state or from the abnormal state to the stable state due to the influence of disturbance or the like, it is necessary to reconstruct the network faster than normal. Therefore, the learning period is reduced,
In order to reduce the influence of the past history data, automatic additional learning is performed for a certain period of time until the network corresponding to the state is reconfigured under the condition that the learning determination days are reduced.

By performing such a process switching, a network is constructed reflecting abundant experience in the past when the system is stable, and the automatic additional learning period is made faster than that in the steady state when the system is abnormal. It is possible to quickly reconstruct a network corresponding to the above, and to expect highly reliable guidance prediction.

FIG. 9 shows details of a specific procedure for switching the processing. In the flowchart of FIG. 9, first,
When a rapid change in the input value is used as a condition for switching the processing by the switch function, the amount of change in the input value is determined in step 291 in the processing block 290 for setting the switch flag based on the input value. "Yes")
In step 293, a stable flag is set in order to perform the process at stable time. Conversely, when an abnormality occurs (in the case of "No"), an abnormality flag is set in step 295, and the process is switched.

Next, the neural network additional learning processing 30
0, additional learning is performed in step 305,
The learning result is used as a new net, and the current net and the new net are recalled in step 310. Subsequently, in step 315, the average recall error is compared. As a result of this comparison, if the new net has a smaller average recall error, the new net is used as the current net and stored in the neural net base 400.

In step 330, the current operation amount is recalled by the current net.

Subsequently, in step 340 in which the process is switched according to the change of the recall value, it is first determined in step 341 whether or not the recall value is stable. As a result, if the recall value is stable (“Yes”). Step 34)
If the stability flag is set in step 3 and the predicted recall value is abnormal ("No"), an abnormal flag is set in step 345, and the process is switched.

Finally, a processing block 290 for setting the flag based on the change in the input value and a processing block 34 for setting the flag based on the change in the recall value.
The switch switching block 35 depending on the flag status of 0
At 0, processing switching of the learning period and the number of days for learning determination, or, if there is a multiplex network, network switching processing according to the situation.

That is, in step 360, the stability of the input value and the recall value is determined, and if it is determined that the input value and the recall value are stable (“Yes”), the stable net processing is performed in step 363. If it is determined to be “(No)”, an abnormal-time net process is performed in step 365.

Here, in the process switching based on the change of the input value or the change of the recall value, it is not always necessary to provide both the determination switches, and the change of the input value or the change of the recall value depends on the system. It is also possible to provide only the switch function.

In these procedures, the automatic addition learning period, the number of days of learning determination, and the processing switching of the multiplex system are switched as described above to improve the learning efficiency and improve the hardware efficiency based on the feature of the reconstruction of the neural network. In addition to reducing the burden on the hardware, it is possible to rebuild a network that can quickly respond to sudden changes in input values, etc., along with this, it is always possible to provide optimal and highly reliable guidance display The system can be established.

Further, in the above-described embodiment, as a specific application example of the present invention, the operation support system for the operator who manages the water purification treatment process has been described in detail. In a system, by inputting data obtained from an instrument or the like to an input layer of a hierarchical neural network model, it can be used as a model for outputting a predicted value with a small average recall error to an output layer, and furthermore, an operator In addition to the water purification system, which is a process in which the determination is important, various processes involving an operator, for example, a sewage treatment process, a chemical reaction process,
It is needless to say that the present invention can be applied to a bioprocess, a nuclear power generation process, a financial process such as stock price / exchange, and the like.

[0068]

As is apparent from the above detailed description of the present invention, that is, in the support system using the neural network method of the present invention, the operator is provided with quantitative and less predictive guidance for the operator. In particular, in a plant operation support system,
By knowing the optimal values for the learning determination days and the learning period corresponding to the plant characteristics, it becomes possible to efficiently perform prediction with a small average recall error. In addition, even in the case of a sudden input situation change, which was difficult with the prediction guidance of the conventional driving support system, the network that responds to the input situation change is reconstructed more quickly, so that the optimal recall prediction is always performed, An excellent effect of being able to provide a driving support system that provides highly predictive guidance is exhibited.

More specifically, in the conventional neural network system, a method of reducing the average recall error by increasing the number of times of learning has been mainly employed. Was very highly dependent and was inefficient. Therefore, in the present invention, by setting the optimum values of the learning period and the learning timing, as described above, the load on the hardware can be reduced, and the average recall error can be reduced efficiently. Also,
Unlike digital logic, it has the advantage that it can make appropriate predictions (average recall errors for those who have no experience in the past are quite large) even in situations where there is no experience. On the other hand, the prediction of a sudden change in the situation is disadvantageous in that the neural network is not performed well and a large average recall error is likely to occur. Had the disadvantage of becoming larger.

Therefore, according to the present invention, in a plant operation support system using a neural network, when the situation suddenly changes (from a steady state to an abnormal state, or from an abnormal state to a steady state), the actual data can be more quickly obtained. In order to follow up, by using the means of automatic additional learning provided with the above-described process switching, it is possible to reconstruct a network that can perform the process switching at the time of a sudden change in the situation and follow up more quickly. . As a result, the advantages of the prediction calculation using the neural network can be maximized, and a plant operation support system that can provide highly reliable prediction guidance as a plant operation support system can be constructed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an overall configuration of an embodiment in which a plant operation support system according to the present invention is applied to operation management support of a water purification process.

FIG. 2 is a flowchart showing an outline of a learning process in the plant operation support system.

FIG. 3 is a graph showing the basis for the existence of the optimum values of the learning cycle and the learning determination history period (days) for minimizing the average recall error, which is the basis of the present invention.

FIG. 4 is a graph showing a relationship between a learning cycle and a learning determination history period (days) in the present invention.

FIG. 5 is a diagram showing a detailed flow for constructing a new neural network in the learning process of FIG. 2;

FIG. 6 is a block diagram showing a schematic configuration for performing the process switching of the present invention.

FIG. 7 is a waveform diagram showing a state of a sudden change which cannot be handled by the conventional automatic additional learning method.

FIG. 8 is a waveform diagram showing a concept of a switch trigger level in processing switching to cope with the above-mentioned rapid change.

FIG. 9 is a flowchart showing a detailed procedure for switching processing in the configuration of FIG. 6;

FIG. 10 is a diagram for explaining an average recall error, which is an important concept in the present invention.

FIG. 11 is a diagram showing an example for calculating an average recall error.

[Explanation of symbols]

REFERENCE SIGNS LIST 100 landing well 110 rapid mixing pond 111 coagulant tank 112 coagulant injection pump 113 stirring blade 114 stirrer 120 floc forming pond 121 stirring paddle 130 sedimentation pond 140 filter pond 100M, 110M, 120M, 130M, 140M
Measuring device 150 Control means 160 Imaging means 170 Image processing means 180 Operation history database 200 Support system 230 Neurosystem 250 Learning means 380 Quantitative guidance and display means 400 Neural net base

Continued on the front page (72) Inventor Kenji Baba 7-1-1, Omikacho, Hitachi City, Ibaraki Prefecture Inside Hitachi, Ltd. Hitachi Research Laboratory, Ltd. (72) Inventor Katsuo Yahagi 7-1-1, Omikacho, Hitachi City, Ibaraki Prefecture (72) Inventor Misako Obuchi 7-1-1, Omikacho, Hitachi City, Ibaraki Pref. Hitachi Research Laboratory, Ltd. (56) References JP-A-5-265997 (JP, A) JP JP-A-6-60049 (JP, A) JP-A-5-108596 (JP, A) JP-A-5-94554 (JP, A) JP-A-5-89076 (JP, A) JP-A-4-15696 (JP) , A) JP-A-2-195935 (JP, A) "Study on Extraction of Plant Operation Rule Using Neural Network" Transactions of the Institute of Electrical Engineers of Japan, D, 111, p20-28, Jan. 20, 1991 (58 ) Surveyed field (Int.Cl. ⁷ , DB name) G06G 7/60 G05B 13/02 G06F 15/18

Claims (1)

(57) [Reference number] [Claims] 1. A neural network constructed while learning a past history automatically with the operation of a system to calculate a predicted value for an operator who operates the system from a fluctuating situation to provide operational support. A supporting system for displaying, wherein a learning period for performing the automatic additional learning is determined while a learning period in a combination of a learning cycle and a learning determination history period that minimizes the average recall error is determined. A support system using a neural network, wherein a combination of a learning cycle and a learning determination history period that minimizes an average recall error is switched according to a sudden change in a situation.