JPH05265509A - Learning control device - Google Patents

Abstract

PURPOSE:To improve control speed by reducing unnecessary input term and calculation volume at the time of an operation mode. CONSTITUTION:A computer processor 10A performs the learning of a target generation device 100 by switching to a learning mode side at first. In this learning mode, an operation changeover switch is connected with the side of a learning control device 400 and a learning is executed by a measuring data corresponding to educator data. The operation changeover switch is switched to the side of an influence degree evaluation means 500 when an error becomes less than a fixed value epsilon by this learning, the influence degree evaluation is performed and the evaluation result is delivered to an omission means 600. The omission means 600 deletes input terms having little influences, based on the evaluation result. Afterwards, the learning is performed by switching to the learning mode again. When this learning which was performed again is terminated, the control for a control object 300 is performed by switching to an operation mode.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pump control system for a rainwater drainage pump station and a learning control system for various plants which are difficult to model and are controlled by a skilled operator.

[0002]

2. Description of the Related Art In a device controlled by experience of an operator such as control of a rainwater drainage pump station, it is difficult to accurately model the watershed to be controlled. For such a control target, conventionally, a neural network is made to learn the operation result of a skilled operator and controlled.

In this case, since it is unclear what kind of information the operator controls the pump based on, it is necessary to use the observed value that seems to affect the output as an input term.
The operation amount of the operator is adopted as the output. When the actual operating data of the operator is learned as teacher data, a judgment model of the operator is generated in the neural network, and the same control as the operator becomes possible.

FIG. 4 shows an example of the configuration of a conventional learning type control device in this case, and FIG. 5 shows a schematic flow chart at that time. In FIG. 4, reference numeral 10 is a computer processing device for realizing a neural network, and the internal function is composed of a target generation device 100 and a learning control device 400. Further, mode changeover switches 21 and 22 for switching between the learning mode and the operation mode are provided on the input side and the output side of the target generation device 100, respectively. When the learning mode is switched by the mode changeover switches 21 and 22, the learning control device 400 is connected to the target generation device 100, and when the operation mode is switched, the target generation device 100 is the control device 200, the control target. Connected to 300.

The above target generation device 100 is called a perceptron type multilayer feedforward combination type neural network and is often used at present. Each layer of the target generation device 100 has a multi-input 1 called a neuron.
It is composed of an output nonlinear element. One neuron I
i has n inputs, Iij (j = 1 to n), a weighting coefficient applied to the input value Iij is Wij, and an output is Oj, the total sum Ii of the inputs is represented by Ii = ΣIij * Wij, For this input, a nonlinear function f (x) = 1 / (1 + exp (-x + di)) called a logistic function is used for the output. With x = Ii, the output value Oj is calculated by Oi = f (x). Note that di is a threshold value.

At the start, the operation mode of the computer processor 10 is switched to the learning mode by the mode changeover switches 21 and 22, and the target generator 100 is set to the learning controller 40.
Connect to 0 for learning. In this learning, learning method BP
A method called (Back Propagation, Back Propagation) is used to compare the output value with the teacher data for the input,
The weighting factor Wij is corrected until the error becomes equal to or less than the determined value ε (for example, 0.001). Then, when the error becomes equal to or less than ε, the learning is terminated and the mode changeover switches 21 and 2 are
The operation mode is switched by 2 and the target generation device 100 causes the control device 200 to output the target value with respect to the measured value in the same manner as the operator performed, and the controlled object 300 is controlled.

[0007]

By learning the actual operating data of the operator as the teacher data as described above, the same control as the operator becomes possible. However, if all the input terms that are likely to be relevant to the operator's judgment are used for learning, there is a problem that the neural network becomes large and the time required for execution becomes long.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a learning control device capable of reducing the amount of calculation in the operation mode and reducing the control speed by reducing unnecessary input terms. ..

[0009]

According to the present invention, there is provided a learning type control device using a multi-layered neural network having multiple inputs and multiple outputs, a means for evaluating the degree of influence of an input value on the output, and this means. And a means for omitting an item having a small influence degree from the input item based on the evaluation result.

[0010]

In the learning control device, all the factors that seem to affect the control during learning are input to the learning control device to perform learning, and after the learning is completed, the influence evaluation means determines the influence that affects the output of the input term. Then, based on this degree of influence, those having a small influence on the output are omitted from the input term. Then, learning is performed again to reconfigure the learning control device. As a result, unnecessary input terms can be reduced, the amount of calculation in the operation mode can be reduced, and the control speed can be improved.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows the overall configuration of a learning control device when the present invention is applied to pump control of a rainwater drainage station, and FIG. 2 is a flow chart showing its operation.

In FIG. 1, reference numeral 10A denotes a computer processing device for realizing a neural network, which includes a target generation device 100, a learning control device 400, and an impact degree evaluation means 500 for assessing the impact degree on the output of an input value. The omission means 600 is configured to omit the items having a small influence degree from the input items based on the evaluation result by the means 500. Further, mode changeover switches 21 and 22 for switching between the learning mode and the operation mode are provided on the input side and the output side of the target generation device 100, respectively. Further, the computer processing device 10A is provided with the operation changeover switch 23 on the input side of the learning control device 400 and the influence degree evaluation means 500. This operation changeover switch 23 connects the learning control device 400 to the target generation device 100 during learning in the learning mode, and when the learning is completed, the influence degree evaluation means 500 is connected to the target generation device 1.
Switch to 00 and connect.

When the mode changeover switches 21 and 22 are switched to the operation mode, the target value output from the target generation device 100 is sent to the control device 200 and the controlled object 3 is controlled.
00 is controlled. Then, detection signals for the various state quantities of the controlled object 300 are input to the target generation device 100.

Next, the operation of the above embodiment will be described with reference to the flowchart of FIG. FIG. 2 shows the processing operation of the computer processing apparatus 10A. When the operation is started, the computer processing device 10A first determines whether or not it is the first operation (step A1), and if it is the first operation, switches the mode changeover switches 21 and 22 to the learning mode side to generate a target. Learn the device 100 (step A)
2). In this learning mode, first, the operation changeover switch 23 is connected to the learning control device 400 side, and learning is performed by the measurement data corresponding to the teacher data. Then, the value of the output with respect to the input of the target generation device 100 is compared with the teacher data, and the weighting coefficient is corrected until the error becomes equal to or less than the determined value ε.

By the above learning, it is judged whether or not the error is equal to or less than the determined value ε (step A3), and the error is ε.
When it becomes the following, the operation changeover switch 23 is changed over to the influence degree evaluation means 500 side and the influence degree is evaluated (step A4). The impact degree evaluation means 500 uses the target generation device 10
For each input item of 0, the degree of influence on the output is evaluated by, for example, the following formula. Wi = Σj | Woj | * | Wji | where Wi: contribution to the output value of the i-th input neuron Woj: link weight from the j-th neuron of the intermediate layer to the output layer Wji: i-th neuron of the i-th input layer Link weight to

The value of Wi allows comparison of important and less affected inputs. FIG. 3 shows an example of the evaluation result by the influence degree evaluation means 500. From the evaluation results shown in FIG. 3, the pump well water level,
It can be seen that the contributions of the rainfall amount and the predicted inflow amount are large, and that it is possible to perform equivalent control with these three measured values.
The evaluation result by the influence degree evaluation means 500 is sent to the omission means 600.

The omission means 600 deletes the input term having a small influence based on the evaluation result (step A5).
After that, the operation changeover switch 23 is set again to the learning control device 4
The learning mode is set by switching to the 00 side (step A6).
As a result of this re-learning, it is judged whether or not the error is below a certain value (step A7), and if it is below a certain value, the learning is terminated.

After the above learning is completed, the mode selector switches 21 and 22 are switched to the operation mode (step A).
8). In this operation mode, the target generation device 100 causes the control device 200 to output the target value for the measured value, and the control target 300 is controlled. Even in this operation mode, it is judged whether the deviation is large (step A9), and if the deviation is within a certain value, the control target 3
The control operation for 00 is continued. However, if the deviation becomes large, the process returns to step A2 via step A10, and the above learning is performed again.

[0019]

As described above in detail, according to the present invention, the degree of influence of the input value on the output is evaluated, and the input item having the small degree of influence is omitted based on the evaluation result.
It is possible to reduce unnecessary input terms, reduce the amount of calculation in the operation mode, and improve the control speed.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of a learning control device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of the embodiment.

FIG. 3 is a diagram showing an output example of an impact evaluation unit.

FIG. 4 is an overall configuration diagram of a conventional learning control device.

FIG. 5 is a flowchart showing the operation of the conventional device.

[Explanation of symbols]

10, 10A ... Computer processing device, 100 ... Target generation device, 200 ... Control device, 300 ... Control object, 400 ... Learning control device, 500 ... Impact evaluation means, 600 ... Means for omitting input.

Claims (1)

[Claims] 1. A learning-type control device using a multi-layered neural network having multiple inputs and multiple outputs, a means for evaluating the degree of influence of an input value on the output, and a means for evaluating the degree of influence based on the evaluation result by this means. A learning type control device comprising means for omitting a small one from an input term.