JPH04138502A - Manipulated variable sequential forecasting operation system and control variable sequential forecasting - Google Patents

Abstract

PURPOSE:To enable an unexperienced operator also to execute proper operation by presenting operation contents to be successively executed with liklihood, also reliability to the operator based upon a result obtained by measuring the state of a target and the history of the operation contents executed by the operator. CONSTITUTION:This system is provided with a neural circuit network having learned past successful examples obtained by properly controlling an operation target or a control target, and while sequentially forecasting a manipulated variable or control variable, the recommended value of the manipulated variable is presented together with its recommended liklihood. Namely, the 1st neural circuit network 40 outputs operation contents similar to the successful examples executed by an experienced operation by means of the operation system. The 2nd neural circuit network 50 for outputting the forecasting value of measuring information uses the measuring information in the operation success example used for the learning of the network 40 as teacher's data and inputs the history of the past measuring information collected several times prior to the measuring information concerned and the history of the operation information as example learning. Consequently, an unexperienced operator also can execute proper operation.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to a driving system that operates a driving object based on measurement information obtained from the driving object, and a driving system that operates a driving object based on measurement information obtained from the controlled object. This invention relates to an automatic control system that sequentially controls a controlled object.

(Prior Art) In a driving system in which an operator controls a driving object toward a certain goal, determining the operation content is an important point. This determination of the operation content is often made based on the result of direct monitoring of the condition of the object by the operator, or based on the measurement result indicating the condition of the driving object obtained by some measuring means. The operator determines the content of the operation and operates the system so that the observed state of the operating target becomes the target state. On the other hand, the driving object is influenced by the environment in which it is placed, and the state of the driving object changes depending on the interaction between the contents of the operations performed by the operator and the influence from the environment. If this influence from the environment is very weak and can be ignored, the details of the operation will be directly reflected in the state control, and the system operation by the operator will be highly accurate.

However, it is rare that the influence from the environment can be ignored, and the operator is required to have driving techniques that take this influence into account. Generally, the influence from the environment is often unknown to the operator, and the operator operates the system while estimating the influence from the environment. In particular, if the influence from the environment is fixed and unchanging, and if the fluctuation trend is constant or periodic and stationary, it is relatively easy to estimate the influence, so You can expect accurate driving. In an actual driving system, the influence from the environment fluctuates from moment to moment, and the tendency of the change is often not fixed, making it extremely difficult to estimate the influence from the environment. This is where the difference between a skilled operator and an inexperienced operator lies, and it can be said that a skilled operator can estimate the influence of the environment relatively well. Therefore, in the driving systems described so far, the skill level of the operator is an important factor, and the training of skilled professional operators has been essential for the widespread use of the system.

In addition, traditionally, driving operation support and automation have been referred to as zero.
Expert systems that incorporate the specialized knowledge of experienced operators have been developed, but because most of the operating knowledge is based on experience, it is not systematized, making knowledge acquisition difficult and not always successful. It's hard to say.

Furthermore, even in automatic control systems such as robot control, it is important to estimate the influence of the environment in which the controlled object is placed, and the quality of this estimate is thought to affect the performance of the system. . Generally, the controlled object is often modeled, including the influence from the environment, and incorporated into the system. However, it is almost impossible to incorporate all possible situations into the model, and situations that deviate from the model may result in incorrect control, causing failure or loss of control. For this reason, there is active development of sub-groove mechanisms that guarantee robustness, but the problem has not yet been fundamentally solved.

[Invention or problem to be solved]

As explained above, in conventional driving systems in which the operator operates based on the results of observing the state of the target, the operator or his/her expert knowledge and experience predict and determine the amount of operation that is considered to be optimal. Because I was performing an operation,
There has been a drawback in that there is a large difference in quality, such as driving accuracy, between operation by an inexperienced operator and operation by a skilled operator.

Furthermore, as is clear from the above, the operator of such a driving system requires a certain level of knowledge and experience as an expert, and training for experts is an important issue in disseminating the driving system itself. There is.

On the other hand, in an automatic control system, it is necessary to take measures such as quickly switching to manual control when the controlled object falls outside the controllable range. Therefore, it is important to be able to determine whether the controlled object is within the controllable range of the system, or whether it is outside the controllable range, by some method such as observing the situation of the controlled object. It has become a challenge.

A first object of the present invention is to provide an operating system that does not require expert knowledge or experience.

A second object of the present invention is to provide an automatic control system that can detect at an early stage whether or not the object to be controlled is within the controllable range of the system.

[Means to solve the problem]

The driving system of the present invention includes a first system that sequentially captures the history of measurement information obtained by the measurement means.
a second shift register that sequentially captures a history of operation information, which is the operation content performed by the operation means; a first neural network that receives the contents of each stage of the first and second shift registers as input and outputs an amount corresponding to a predicted value of the current operation content based on the learning result; and a history of the measurement information. The relationship between the operation information and the history is learned in advance.1. The contents of each stage except the first stage that holds the latest measurement information of the first shift register and the contents of each stage of the second shift register are input, and based on the learning results, the latest measurement information is a second neural network that outputs a quantity corresponding to a predicted value of the measurement information; and a comparison means that compares the output of the second neural network with the latest measurement information measured by the measurement means. , evaluate the comparison results by the comparison means, and apply the evaluation results to the first
and evaluation means for outputting additional information indicating the recommendation accuracy of the predicted value of the operation content, which is the output of the neural network.

The automatic control system of the present invention includes a first system that sequentially captures the history of measurement information obtained by the measurement means.
a second shift register that sequentially takes in the history of control information that is the control content performed by the control means; a second shift register that is made to learn in advance the relationship between the history of the measurement information and the history of the control information; a first neural network that receives the contents of each stage of the first and second shift registers as input and outputs an amount corresponding to a predicted value of the current control content based on the learning result; and a history of the measurement information. The contents of each stage except the first stage which holds the latest measurement information of the '11 shift register, whose relationship with the history of the control information has been learned in advance;
! a second neural network that receives the contents of each stage of the shift register of J2 as input and outputs an amount corresponding to a predicted value of the latest measurement information based on the learning result; and an output of the second neural network. and a comparison means for comparing the latest measurement information measured by the measurement means; and a comparison means for evaluating the comparison result by the comparison means, and
and evaluation means for outputting as additional information indicating the recommendation accuracy of the predicted value of the control content which is the output of the neural network.

[Effect]

In order to achieve the above object, the present invention provides a driving system or an automatic control system with a neural network that has learned past successful examples of appropriately controlling an operation target or control target to control the operation amount or control amount. It makes predictions sequentially and presents recommended values of manipulated variables or controlled variables with recommended accuracy.

The first neural network that outputs the recommended value of the operation amount uses the operation contents in successful cases of operation of the driving system by a skilled operator as training data, and also the measurement information of the target involved in the operation and the information prior to the operation. The history of measurement information for the past several times and the history of operation information are used as case study inputs, and the operation content equivalent to a successful example performed by a skilled operator using the driving system is output. Further, the recommended accuracy of the recommended operation value is determined by comparing the predicted value of the measurement information obtained by the second neural network with the actual measurement result. The second neural network that outputs the predicted value of the measurement information uses the measurement information as training data from the operation success case used in the learning of the first neural network that outputs the recommended value of the operation amount, and uses the measurement information as training data. The history of measurement information and the history of operation information from the past several times are used as case study inputs, and the influence of the environment on the target of the driving system is simulated in the form of fluctuations in measurement information.

In an automatic control system, instead of the operation information that shows the details of the operation carried out by the operator in the above-mentioned driving system, it is sufficient to consider the corresponding control information that shows the details of the control performed by the system. Outputs control content equivalent to past control success cases with accuracy.

Therefore, according to the present invention, the operator of the driving system refers to the presented recommended operation value with recommended accuracy, and
Since the actual operation details can be determined, there is no need for advanced knowledge or deep experience as an expert.In addition, in automatic control systems, the recommended control value with recommended accuracy and the control amount generated by the system itself can be compared. Since it becomes possible to target the situation, trends that deviate from the controllable range can be detected at an early stage.

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of a manipulated variable sequential prediction driving system showing one embodiment of the present invention, and FIG. 2 is a diagram showing an operating procedure of the driving system of FIG. 1.

The operational variable sequential prediction driving system of this embodiment has a measuring means (
(not shown), an operating means (not shown), an input interface section 11.12, a data preprocessing section 21.22,
The first shift register 31 and the second shift register 3
2, the first neural network 40, and the second neural network 5o
, a data post-processing section 60.70, an output interface section 80.90, and a data display device (not shown).

The measuring means measures the state of the driving object (not shown). The operating means is for the operator of the driving system to operate the driving object. The input ink face section 11 receives measurement information indicating the state of the driving object obtained by the measurement means. The input interface unit 12 receives operation information that is the content of the operation performed by the operator in order to drive the driving object toward the target. The data preprocessing units 21.22 each input the measurement information and operation information received by the input interface unit 11.12, perform processing such as normalization, and convert them into a data format that can be input by the neural circuit 1111140.50. . The first shift register 31 has an (n+1) film configuration, and the data preprocessing section 2! When the first measurement information, which is the latest measurement information, is input, the stored information is sequentially shifted and the oldest (i-th
−n−1)th measurement information is discarded, and as a result, the past (
It holds measurement information for n+1) times. The second shift register 32 has an n-stage configuration, receives operation information sequentially from the data preprocessing section 22, and operates in the same manner as the shift register 31, and holds operation information for the past n operations up to the previous time. The first neural network 40 includes an input layer 401 and an intermediate layer 402.
It is composed of three layers: a middle layer 403 and an input layer 401.
Each unit 400 is connected to each stage of the shift registers 31 and 32, and is trained in advance on the relationship between the history of measurement information and the history of operation information.
and 32, that is, the latest i-th measurement information and the measurement information and operation information for the past n times from the (i-1)th time, and based on the learning results, the i-th operation content is input. Outputs the amount corresponding to (predicted value of operation content). The data post-processing unit 60 converts the output result of the first neural network 40 into a data format of the original operation amount of the driving system. The output interface unit 80 is the data post-processing unit 60
The output is displayed on a data display device to present to the operator of the driving system as recommended operation information. The second neural network 50 includes an input layer 501, a middle layer 502, and an output layer 5.
03, each unit 5 of the input layer 501
00 indicates each stage of shift registers 31 and 32 (however,
(excluding the first stage that holds the latest measurement information of the shift register 31), the relationship between the history of measurement information and the history of operation information is learned in advance, and each of the shift registers 31 and 32 The information of the stage, i.e. the (i-1)th
Measurement information and operation information from the past n times are input, and an amount corresponding to the predicted value of the i-th (latest) measurement information is output based on the learning results. The data post-processing unit 70 re-transforms the output result of the second neural network 50 into a data format equivalent to the input data format converted by the data pre-processing unit 21,
This enables comparison with the i-th measurement information actually obtained. The output interface section 90 corresponds to a comparison means and an evaluation means, and inputs the output of the data post-processing section 70 and the i-th (latest) measurement information held in the first stage of the shift register 31, and inputs both. The comparison is performed, and prediction accuracy data corresponding to the comparison difference is converted and generated, and the output interface unit 8
The recommended operation information is output as recommended accuracy information of the recommended operation information outputted from 0, and similarly to the recommended operation information, it is displayed on a data display device for presentation to the operator of the driving system.

Next, in this embodiment, a process of presenting recommended operation information to which recommended reliability information is added to an operator will be explained based on the operating procedure of the driving system shown in FIG. 2.

The operating procedure of this driving system is a step 101 of measuring the state of the driving object, and a step 102 of evaluating the measurement results.
The basic operation consists of four steps: step 103 for determining the content of the operation, and step 104 for implementing the operation, and these steps are repeated to drive the object to the target state. Here, the subject is constantly influenced by the environment in which it is placed.

Assume now that the i-th basic operation of the driving system has started to be performed. After the 3Ii-th measurement information is obtained by the measurement means and received by the input interface unit 11, it is transformed by the data preprocessing unit 21 into the input data format of the neural network 40.50, and is input to the shift register 31. This shift register 31 stores the latest measurement information, i.g.
When the i-th measurement information is input, the held information is sequentially shifted, and the oldest (i-n-1)th measurement information overflowing from the (n+1)th stage is discarded, and as a result, the past Measurement information for (n+1) times is held sequentially. On the other hand, the operation information at this point is the previous (i-1)
The operation information indicating the operation content performed in the first basic operation is
The data is received via the input interface unit 12, transformed into the input data format of the neural network 40, 50 by the data preprocessing unit 22, and input to the shift register 32. The shift register 32 sequentially holds operation information for the past n operations up to the previous one in the same manner as the shift register 31 operates. In this way, when determining the operation content of the i-th basic movement, the neural network 40 has the following:
The information of each stage of the shift registers 31 and 32, that is, the latest i-th measurement information and! Measurement information and operation information for the past n times are input from the (i-1)th time. The neural network 4o inputs this information and, based on the learning performed in advance, '! The amount (predicted value) corresponding to the Js-th operation is output. This output result is converted by the data post-processing unit 60 into the format of the original operation amount of the driving system, is output as recommended operation information via the output interface unit 8o, and is displayed by the operator of the driving system on the data display device. will be presented.

On the other hand, among the information in each stage of the shift registers 31 and 32, the neural network 50 stores, except for the latest i-th measurement information,
Measurement information and operation information for the past n times are input from the (i-1)th time. The neural network 50 receives this information and outputs an amount corresponding to the predicted value of the i-th measurement information based on the learning performed in advance. This output result is re-transformed by the data post-processing section 70 into a data format equivalent to the input data format converted by the data pre-processing section 21. The output interface unit 90 compares this with the actually obtained i-th measurement information, converts and generates prediction accuracy data corresponding to the comparison difference, and outputs the predicted accuracy data as recommended accuracy information.

Similar to the recommended operation information, this recommended accuracy information is also presented to the operator of the driving system by the data display device. The operator of the driving system can refer to the presented recommended operation information and recommended accuracy information to determine the actual operation content.

In the above explanation, the shift register 3!32 is a method for importing past history into the neural network 40.5.0, and is a method for importing past history into the neural network 40.5.0, and is a measurement result including past operation history and its reaction, that is, operation information. It becomes possible to incorporate into the neural network 40.50 the influence from the environment reflected in the history of the measurement information and the history of the measurement information. The number of stages in the shift registers 31 and 32 is determined by how long the environmental influence lasts, and the more stages the data is stored, the more likely it is to reflect the long-term fluctuation trend of the environmental influence. can. Further, since the neural networks 40 and 50 perform processing related to the operation of the same driving system, operation examples must be used in learning of each neural network. Note that although an example using a three-layer neural network 40.50 is shown here, the neural network need not be limited to the three-layer structure.

A specific example of the above explanation is, for example, operating a small diameter pipe propulsion machine that constructs a tunnel without excavation and buries a pipe body, and each of the above steps 101 to 104 is
■ Step of measuring the position and attitude angle indicating the state of the tip device of the small diameter tube propulsion device, ■ Step of evaluating the measurement results based on the planned line of pipe burial work using the small diameter tube propulsion device, ■ This corresponds to the step of determining the propulsion direction, which is the specific operation content, and the step of carrying out a correction operation of the small diameter tube propulsion direction, in which the small diameter tube propulsion is carried out and the state of the tip device is actually changed. The environmental influence corresponds to the influence of earth pressure when propelling pipes etc. into the soil.

In the case of an automatic control system, a control means is provided instead of the operation means, and instead of operation information indicating the details of the operation performed by the operator in the driving system, control information indicating the details of the control performed by the system is provided. It is only necessary to consider the information in correspondence, and the control content equivalent to past control success cases implemented by the automatic control system will be output with accuracy.

- Therefore, it becomes possible to compare the recommended control value with recommended accuracy and the control amount generated by the system itself, so that a tendency to deviate from the controllable range can be detected at an early stage.

〔Effect of the invention) As explained above, the present invention has the following effects.

(1) In the driving system, based on the results of measuring the state of the object and the history of the operations performed by the operator, the next operation to be performed is presented to the operator with an accuracy that can be called its reliability. Therefore, even an inexperienced operator can operate the system appropriately by referring to this output information and operating the system. In particular, recommended accuracy information is considered to be useful additional information when referring to recommended operation information, so if you judge this information comprehensively, even an operator with no experience can Abnormal situations can be detected and driving failures can be prevented. In addition, regardless of experience or lack of experience, you will be able to perform operations with high quality such as driving accuracy.
It is possible to promote the spread of the driving system regardless of the lack of experienced personnel.

(2) In an automatic control system, based on the results of measuring the state of the controlled object and the history of the control content performed by the system, the next control content is output with an accuracy that can be called reliability within the controllable range. By doing so, it is possible to early detect a tendency for the controlled object to deviate from the controllable range.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an operational variable sequential prediction driving system according to an embodiment of the present invention, and FIG. 2 is a diagram showing an operating procedure of the driving system of FIG. 1. 11.12--input interface section, 21.22--data processing section, 31--first shift register, 32--second shift register, 40--first neural network, 50--second Neural network, 60.70-Data post-processing unit, 80.90-Output interface unit, 101-104-Steps. Patent applicant Nippon Telegraph and Telephone Corporation

Claims (1)

[Scope of Claims] 1. A driving system comprising a measuring means for measuring the state of a driving object and an operating means for operating the driving object based on measurement information obtained by the measuring means, wherein the measuring means a first shift register that sequentially captures a history of the obtained measurement information; a second shift register that sequentially captures a history of operation information that is the operation content performed by the operation means; a history of the measurement information and the operation information. The relationship between the first and second shift registers is learned in advance, and the content of each stage of the first and second shift registers is input, and based on the learning results, an amount corresponding to the predicted value of the current operation content is output. The neural network of No. 1 is trained in advance on the relationship between the history of the measurement information and the history of the operation information, and each stage except the first shift register holds the latest measurement information of the first shift register. The contents of the column and the second
a second neural network that receives the contents of each stage of the shift register as input and outputs a quantity corresponding to a predicted value of the latest measurement information based on the learning result; and an output of the second neural network. , a comparison means for comparing the latest measurement information measured by the measurement means, and a comparison result by the comparison means, and a first
1. An evaluation means for outputting as additional information indicating the recommended accuracy of a predicted value of an operation content which is an output of a neural network. 2. In an automatic control system having a measuring means for measuring the state of a controlled object, and a control means for sequentially controlling the controlled object based on measurement information obtained by the measuring means, a first shift register that sequentially captures a history of measured information; a second shift register that sequentially captures a history of control information that is control content performed by the control means; a history of the measurement information and the control information; The first shift register has been previously trained to learn the relationship between the first and second shift registers, receives the contents of each stage of the first and second shift registers as input, and outputs an amount corresponding to the predicted value of the current control content based on the learning result. A neural network is made to learn in advance the relationship between the history of the measurement information and the history of the control information, and each stage except the first stage which holds the latest measurement information of the first shift register. The content and the second
a second neural network that receives the contents of each stage of the shift register as input and outputs a quantity corresponding to a predicted value of the latest measurement information based on the learning result; and an output of the second neural network. , a comparison means for comparing the latest measurement information measured by the measurement means, and a comparison result by the comparison means, and a first
1. An automatic control system for sequentially predicting a control amount, comprising: an evaluation means for outputting as additional information indicating the recommended accuracy of a predicted value of a control content which is an output of a neural network.