JPH05151373A - Fuzzy reverse diagnostic method using neural network - Google Patents

Abstract

PURPOSE:To provide the fuzzy reverse diagnostic method using a neural network which can execute a fuzzy reverse diagnosis at a high speed. CONSTITUTION:This method is provided with a first process for expanding a fuzzy relational expression for expressing knowledge of a fuzzy reverse diagnosis for estimating fault phenomena UP1-UPj from fault causes UC1-UCi to a neural network in which plural fault cause neural units become an input layer, and plural fault phenomena neural units become an output layer, and a second process for deriving a difference of a degree that a fault phenomenon outputted from the expanded neural network occurs and a degree of an actual fault phenomenon as an error, deriving a degree of a fault cause for minimizing the error with regard to plural fault causes, respectively by a reverse inference by a back propagation method, and outputting a degree of conviction of each fault cause as a result of fuzzy reverse diagnosis from a final error corresponding to the degree of each fault cause derived in such a manner.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention uses a fuzzy inverse diagnosis for inferring a cause of a failure from a failure phenomenon by using a neural network by using a fuzzy relational expression expressing the correlation between the cause of failure and the phenomenon of failure. Regarding how to do.

[0002]

2. Description of the Related Art In 1976, the fuzzy inverse problem proposed by E. Sanchez and its solution method have recently come to be widely used for fault diagnosis problems and the like. This is based on the knowledge that the fuzzy reverse diagnosis problem system does not perform the diagnosis based on the knowledge to judge the failure cause from the failure phenomenon like the conventional failure diagnosis system, but on the contrary to the knowledge to judge the failure phenomenon from the failure cause. This is because diagnosis can be performed.

Generally, the former knowledge of normal diagnosis is the knowledge possessed by the user who uses the machine, and is obtained from the experience of failure diagnosis over a long period of time. On the other hand, the knowledge of the latter reverse diagnosis is the knowledge possessed by the designer who designed the machine, which is obtained from the structure of the machine.

Therefore, by using this reverse diagnosis system, it is possible to perform a suitable diagnosis even when the machine has just been manufactured and has little experience in failure diagnosis.

[0005]

However, in the conventional fuzzy inverse diagnosis system, the operation for solving the inverse problem for predicting the possible cause of the failure from the input failure phenomenon is extremely complicated, and the diagnosis There is a drawback that the required calculation time becomes very long. For this reason, in this conventional reverse diagnosis system, especially when real-time diagnosis is required, the slow calculation speed becomes a problem.

The present invention has been made in view of the above circumstances, and an object thereof is to provide a fuzzy reverse diagnosis method using a neural network capable of performing high speed fuzzy reverse diagnosis.

[0007]

According to the present invention,
A first step of developing a fuzzy relational expression expressing knowledge of fuzzy reverse diagnosis for estimating a failure phenomenon from a failure cause into a neural network having a plurality of failure cause neural units as input layers and a plurality of failure phenomenon neural units as output layers And, the difference between the degree of occurrence of the failure phenomenon output from the expanded neural network and the degree of the actual failure phenomenon is obtained as an error error, and the cause of the failure that minimizes the error error by inverse inference by the backpropagation method. A second step of obtaining a degree for each of the plurality of failure causes, and outputting a certainty factor of each failure cause as a result of fuzzy reverse diagnosis from a final error error corresponding to the obtained degrees of each failure cause. To do so.

[0008]

According to the present invention, a fuzzy relational expression expressing knowledge of fuzzy inverse diagnosis for estimating a failure phenomenon from a cause of a failure is developed in a neural network, and then fuzzy inverse diagnosis by backpropagation is performed. Infer the certainty factor of the cause of failure that has the closest pattern to the failure phenomenon.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described in detail below with reference to the embodiments shown in the accompanying drawings.

This fuzzy reverse diagnosis system is roughly divided into the following two functions.

(1) Expansion of fuzzy relational expression to neural network Function for expanding fuzzy inference expressed by fuzzy relational expression to neural network (2) Fuzzy reverse diagnosis by backpropagation Neural network created by (1) Function of using the function to infer the certainty factor of the failure cause whose pattern matches the current failure phenomenon first. First, the function of expanding the fuzzy relational expression of (1) to the neural network will be described.

In this development, the following steps (1)
(2) Build a fuzzy dedicated neural network according to the order of (3). The procedure will be described below with reference to the flowchart of FIG.

FIG. 1 is a table showing a fuzzy relational expression between a failure cause i and a failure phenomenon j combined by a correlation degree Wij.

(1) In order to develop the fuzzy relational expression in the neural network, the failure cause i of the fuzzy relational expression is set to the failure cause unit UCi and the failure phenomenon j
Is the failure phenomenon unit UPj, and the failure cause unit UC
A network is created in which i is placed in the input layer and the failure phenomenon unit UPj is placed in the output layer, and these units are connected according to the magnitude of their correlation Wij (steps 100, 110, 120 in FIG. 2). .. FIG. 1 shows an example of the expanded neuron unit.

The unit means a nerve cell model (neuron) in a neural network, and in the neuron, a coupling coefficient (weight) w = {wi} corresponding to an external input x = {xi}. Is multiplied, and a unit-specific function f (xi, wi) is applied to them, and then a threshold function (non-linear function) Θ is applied to the output y. i = 1 to n.

That is, in the unit, the input x
The output y for (= x1, x2, ..., Xn) is expressed as y = .THETA. (F (xi, wi) ,. THETA.) (1) by the coupling coefficient w (= w1, w2, ..., Wn).

Note that θ is a parameter vector of the threshold function.

(2) 0 in the failure cause unit UCi of the input layer
When the certainty factor Xi of ˜1 is input, the certainty factor Xi and its correlation degree Wij are added to the failure phenomenon unit UPj of the output layer having the connection.
Is created in the failure cause unit UCi (step 130).

The evaluation function f is expressed by the following equation.

Evaluation function f; Qij = f (Xi, Wij) (2) As a specific example thereof, there are the following two expressions.

Qij = Xi * Wij (3) Qij = min (Xi, Wij) (4) (3) Output layer failure phenomenon unit UPj has all the {Qij} propagated from the input layer. A function for outputting the evaluation result Oj is created (step 130).

The evaluation function Θ at this time is expressed by the following equation.

Evaluation function Θ; Oj = Θ ({Qij}) (5) As a specific example, there is the following equation (2).

Oj = Σ (Qij) (6) Oj = max {Qij} (7) As described above, the fuzzy relational expression is developed into the neural network shown in FIG. Then, according to such a neural network, the failure phenomenon {Oj} can be predicted by inputting the certainty factor {Xi} to the failure cause unit UCi as shown in FIG.

Next, the fuzzy reverse diagnosis function by the back propagation of (2) will be described.

In this fuzzy reverse diagnosis, the certainty factor of the cause of the fault whose pattern matches the current fault phenomenon is inferred at high speed using the backpropagation method using the network created earlier.

However, in this fuzzy reverse inference,
An initial value (default is 0.5, for example) is input for each failure cause UCi, and this input value is optimized by back propagation. At this time, the failure cause UCi
Input values for failure causes other than i shall be fixed to 0.

This is because most failures in the world are considered to be caused by a single cause rather than a combination of a plurality of causes.

In the case of a failure due to a combination of a plurality of failure causes, it is the same as the conventional backpropagation solution, and a combination of certainty factors of failure causes close to the current failure phenomenon can be predicted. Depends on the initial value. On the other hand, when predicting a single cause of failure, the result does not depend on the initial value.

Here, the following procedures (a) (b) (c) (d) are carried out for each failure cause UCi to obtain the degree of the failure cause that most reproduces the current failure phenomenon, and step (e) Determines the certainty factor of failure cause considering error error. Hereinafter, description will be given with reference to the flowchart of FIG.

(A) An initial value Xi (default is 0.5, for example) for each failure cause UCi of the fuzzy inference dedicated network shown in FIG. 4 is input and each failure phenomenon {Oj} is predicted (step 200). , 210).

Oj = Θ (f (Xi, Wij)) (8) (b) The predicted failure phenomenon {Oj} and the actual failure phenomenon {TO
j} are compared, and an error error E defined by the sum of squared deviations thereof is calculated according to the following equation (step 220).

E = Σ (Oj−TOj) (9) (c) where the error error E is a reference value (a certain positive constant) ε
(Step 230), if the error error E is not within the reference value ε, the error error E is set to Xi.
Is partially differentiated to calculate the correction amount dXi of Xi (step 240).

DXi = −η (∂E / ∂Xi) (10) Note that ∂E / ∂Xi can be calculated according to the following formula.

∂E / ∂Xi = Σ (∂E / ∂Θ2) (∂Θ2 / ∂Θ1) (∂Θ1 / ∂Xi) (11) (d) The correction amount dXi to Xi of Xi is updated ( Step 250), again the procedure of (a), that is, Step 21.
Return to 0.

Xi = Xi + dXi (12) The procedure from (a) to (d), that is, steps 210 to 210.
250 is repeated until the error error E becomes equal to or less than the reference value ε, and Xi at that time is finally determined as the degree of failure cause Xl.
i and E at that time are the final error errors Eli. The above procedure is based on the steepest descent method and the error backpropagation method.

(E) The final certainty factor XXi of the failure cause i is calculated from the above Xli and Eli according to the following formula (step 260).

XXi = Xli−Eli (13) When the above procedures (a) to (e) are carried out for each failure cause UCi, the certainty factor XXi for all causes is calculated. (Steps 270 and 280).

According to the above embodiment, an arbitrary initial value is given to the certainty factor and the certainty factor is obtained by the error back propagation method. However, the certainty factor of each failure cause is 0 to 1.
A method may be used in which the error is calculated by gradually changing between the two, and the confidence with the smallest error among them is calculated. In such a case, if the amount of change is reduced, the accuracy increases, but the number of checks increases, which impedes speeding up. Therefore, it is considered more appropriate to use the error back-propagation method and adjust the accuracy with the convergence width for speeding up.

[0040]

As described above, according to the present invention,
The fuzzy relational expression in the fault diagnosis problem is expanded to the neural network, and the fault cause is predicted from the actual fault phenomenon by the error backpropagation method using this, so the result equivalent to the result of the conventional fuzzy back diagnosis is obtained in a short time. It becomes possible to calculate.

[Brief description of drawings]

FIG. 1 is a diagram conceptually showing expansion to a neural network according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a procedure for expanding a fuzzy relational expression into a neural network.

FIG. 3 is a diagram conceptually showing a fuzzy relational expression including causes and phenomena.

FIG. 4 is a diagram conceptually showing fuzzy inference.

FIG. 5 is a flowchart showing a fuzzy reverse diagnosis procedure by backpropagation.

[Explanation of symbols]

 UCi ... Cause of failure UPj ... Failure phenomenon

Claims (1)

[Claims] 1. A fuzzy relational expression expressing knowledge of fuzzy reverse diagnosis for estimating a failure phenomenon from a failure cause is applied to a neural network having a plurality of failure cause neural units as input layers and a plurality of failure phenomenon neural units as output layers. The first step of developing, and the difference between the degree of occurrence of the failure phenomenon output from the developed neural network and the degree of the actual failure phenomenon is obtained as an error error, and the error error is most determined by inverse inference by the backpropagation method. The degree of the failure cause to be reduced is obtained for each of the plurality of failure causes, and the certainty factor of each failure cause is output as the result of fuzzy reverse diagnosis from the final error error corresponding to the obtained degree of each failure cause. Fuzzy reverse diagnosis method using neural network including