JPH10143343A - Associative plant abnormality diagnosis device - Google Patents

Abstract

(57) [Summary] [Problem] To provide an apparatus for diagnosing the cause of an abnormality in a system such as a plant, and accurately estimate the cause of the abnormality even when inputting symptom data in which noise is superimposed or missing. . SOLUTION: Means 1 for storing symptom data of an abnormal state corresponding to time-series data of observation items and known cause data thereof as pair information, and a cross-correlation matrix as an associative memory model created from the pair information Means 2 for storing
Means 3 for estimating the cause of the abnormality as the current state of the system using the associative memory model and the symptom data corresponding to the time-series data indicating the current state of the system.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an abnormality diagnosis apparatus for diagnosing the cause of an abnormality in a system such as a plant, and more particularly, to the correspondence between symptom data as an abnormal state of a plant and cause data as a cause of the symptom. The present invention relates to an associative plant abnormality diagnosis device that estimates the cause of a symptom from the symptom data before and after a certain abnormality occurs in a plant using an associative memory model.

[0002]

2. Description of the Related Art For example, when an abnormality is detected during the operation of a plant, it is necessary to search for the cause of the abnormality and take a subsequent measure for the abnormality.
If possible, it is desirable to search for the cause of the abnormality before the abnormality occurs, and to take necessary precautionary measures.

For this reason, for example, in a conventional plant abnormality diagnosis apparatus, a method of performing abnormality diagnosis by paying attention to a deviation between an output as a response from a plant and an output of a dynamic characteristic model thereof is used. However, the operation of a relatively simple plant can be modeled with practical accuracy, so that it is possible to diagnose anomalies using such a method. It was difficult to diagnose the cause of the abnormality.

As described above, in a large-scale plant, a knowledge base expressing expert's empirical knowledge in the form of a conditional proposition and an inference mechanism for executing a diagnosis are used for knowledge engineering to estimate the cause of an abnormality. Techniques have also been used. In such a method, a causal relationship between a symptom / sign and an abnormal cause is regarded as empirical knowledge, and "if the symptom / sign is B, the cause is A
It is. In the form of a knowledge base, and search for the cause, etc. from the current symptoms and signs.

[0005]

However, the knowledge of diagnosis obtained from specialists is often obtained as a causal relationship such as "if cause A occurs, symptoms and signs B appear as a result." For this reason, building a knowledge base for abnormality diagnosis is equivalent to generating knowledge that solves the inverse problem of causality rather than acquiring expert knowledge. To be
There was a problem with the existence and uniqueness of the solution.

That is, the cause of an unknown symptom / sign is unknown, and a solution may not be obtained. In the case where one symptom / sign has a plurality of causes, the aforementioned diagnostic knowledge However, there is a problem that the cause cannot be uniquely determined by using.

According to the present invention, a cross-correlation matrix as an associative memory model is created in advance on the basis of a known relationship between observed values of symptoms / signs such as time-series data input from a monitoring target and causes of abnormalities. Provided is a plant abnormality diagnosis device that accurately associates and infers the cause of a corresponding abnormality even when an unknown symptom / sign, for example, symptom / sign data in which noise is superimposed or missing is input, is provided. The purpose is to:

[0008]

FIG. 1 is a block diagram showing the principle configuration of the present invention. FIG. 1 is a block diagram showing the principle configuration of an associative plant abnormality diagnosis device for diagnosing the cause of an abnormal state of a plant based on observation data input from a monitoring target.

In FIG. 1, symptom / cause data storage means 1
Stores, as pair information, data of a symptom as an abnormal state and data of a known cause for the abnormal state, for example, corresponding to the respective time-series data for a plurality of observation items. The abnormal state as the symptom data stored in the symptom / cause data storage means 1 has a known cause, and the cause data is given by, for example, an operator.

The cross-correlation matrix storage means 2 associates a cross-correlation matrix created in accordance with pair information of the symptom data stored in the symptom / cause data storage means 1 and the cause data as the cause of the symptom. This is stored as a storage model.

Abnormality diagnosis means 3 using an associative memory model
Is for estimating the cause of the abnormality as the current state of the monitoring target using the symptom data corresponding to the time-series data indicating the current state of the monitoring target and the cross-correlation matrix stored in the cross-correlation matrix storage means 2. It is. Here, the time-series data indicating the current state of the monitoring target may be superimposed, for example, when noise is superimposed or the data is partially lost. Thus, the cause for the abnormality is estimated.

In the present invention, the cross-correlation matrix is created in accordance with, for example, symptom data and cause data expressed as bipolar data, and is added to the cross-correlation matrix already stored in the cross-correlation matrix storage means 2. , The result of the addition is stored again in the cross-correlation matrix storage means 2,
Alternatively, it is stored as a cross-correlation matrix different from the cross-correlation matrix already stored. Therefore, a plurality of cross-correlation matrices are generally stored in the cross-correlation matrix storage means 2 as an associative memory model.

In the abnormality diagnosis by the abnormality diagnosing means 3 using the associative memory model, the present state of the monitored object is shown from among a plurality of cross-correlation matrices as associative memory models stored in the cross-correlation matrix storing means 2. The bidirectional associative action using the cross-correlation matrix is repeated for the input of the symptom data corresponding to the time series data, and when the result of the bidirectional associative action is stabilized, the symptom data and the cause data in the stable state are compared. The cross-correlation matrix matrix with the minimum corresponding energy is selected, and the cause corresponding to the cause data is estimated as the cause for the abnormality.

As described above, according to the present invention, the cross-correlation matrix created in accordance with the cause data and the symptom data whose relation is known is used as the associative memory model, and the symptom data of the unknown cause is obtained. An estimate of the cause is made for the input.

[0015]

FIG. 2 is a system configuration block diagram of an embodiment of an associative plant abnormality diagnosis apparatus according to the present invention.
In FIG. 1, an associative plant abnormality diagnosis device 10 diagnoses the cause of an abnormality in accordance with generally observed values of a plurality of items input from a monitoring target plant 11, and outputs the diagnosis result to a display device 12.

The associative plant abnormality diagnosis apparatus 10 is realized, for example, as a single computer system. The observation value vector input unit 21 therein receives a set of time-series data, that is, a time-series pattern for each of a plurality of monitoring items for the monitoring target plant 11, and sets one set of each time-series pattern, that is, one set of each time-series pattern. It is output to the vector quantization learning unit 22 or the vector quantization recall unit 24 as an observation value vector. Here, the time-series data for one monitoring item includes n sampling times, that is, t ₁ , t ₂ ,.
.., T _n , and the observation value vector for each monitoring item is output as an n-dimensional vector.

The vector quantization learning unit 22 uses the observation value vector provided from the observation value vector input unit 21 as the time series pattern of each observation item, and performs the observation item quantization according to the learning vector quantization procedure. Is performed, the output of the network as a result of the learning is provided to the quantized data storage unit 25, and the weight of the neural network after learning is output to the weight vector storage unit 23.

Inside the vector quantization learning unit 22, a hierarchical neural network having two layers, an input layer and an output layer, is provided corresponding to each observation item for the monitoring target plant 11. FIG. 3 shows the structure of this neural network. In the figure, the number of neurons in the input layer is n, and the time series pattern of one observation item in each neuron, ie, the observation corresponding to _n sampling times t ₁ , t ₂ ,. Value x
_1, x _2, ···, it is x _n is input.

The number of neurons in the output layer is m, and the j-th neuron is connected to the neurons in the input layer by weights w _j1 , w _j2 ,..., W _jn respectively. Although the neurons in the output layer are depicted in FIG. 3 as being distributed in a plane, the concept is that the ones that can be arranged one-dimensionally horizontally like the input layer are represented by the representative points of the connection weight vector. It is shown on a plane for illustrative purposes. That is, the crosses on the plane represent inputs to the input layer neurons, and conceptually indicate that the output layer neurons represent the representative points of the connection weight vector corresponding to these inputs.

Note that the number m of neurons in the output layer corresponds to the number of categories into which the time series pattern of one observation item to be monitored is to be classified, and the output of the m output neurons is the quantization information of symptoms / signs. Is output to the quantized data storage unit 25 corresponding to the symptom / cause data storage unit 1 in FIG. For example, if the observation item is temperature, the number of categories such as high temperature, slightly high,... Is the number m of neurons in the output layer.

FIG. 4 is a flowchart of the learning vector quantization process in the vector quantization learning unit 22. As described above, in each of the two-layer hierarchical neural network corresponding to each observation item, learning of the connection weight vector is performed using the flowchart of FIG.

When the process is started in FIG. 4, first, in step S1, an initial value of the weight between the input layer neuron and the output layer neuron is set using, for example, a random number, and the network is initialized. Subsequently, in step S2, an input vector given by the following equation is input to the input layer neuron. Note that vector characters in the text are expressed by adding "vector" characters.

Vector X = (x ₁ , x ₂ ,..., X _n ) ^T ... (1) Subsequently, in step S3, the distance between the connection weight vector of each neuron of the output layer and the input vector is calculated. Is calculated. The distance between the j-th neuron in the output layer and the input vector is calculated by the following equation.

[0024]

(Equation 1)

The calculation of the distance and the control of the learning vector quantization process are performed by a learning control unit (not shown) inside the vector quantization learning unit 22. Step S4
Then, a neuron in which the distance between the connection weight vector and the input vector is minimized from the neurons in the output layer is selected. This neuron is called the winner neuron.

In accordance with the selection of the winner neuron, learning of the connection weight vector is performed in step S5. In the learning of the connection weight vector, that is, the updating of the weight, the updating method is different depending on whether the network has correctly recognized or incorrectly recognized. When the network correctly recognizes, that is, when the winner neuron and the given neuron have the largest teacher value, the weight of the winner neuron is updated based on the following equation.

Δw _ji = + η (x _i −w _ji )
(3), where η is a learning constant When the network misrecognizes, that is, when the winner neuron and the given neuron with the largest teacher value are different, the neuron with the second smallest distance is the neuron with the largest teacher value. If the condition that the difference between the winner neuron and the second neuron is small is satisfied, the weights for the winner neuron and the second neuron are updated by the following equation.

Δw _ji = −η (x _i −w _ji ) (4) After the weight is updated, it is determined in step S6 whether or not there is still input data to be learned. In step S6, the processing after step S2 is repeated for different input data, and the processing ends when it is determined in step S6 that there is no input data. Note that the different input data is a time series pattern of one observation value corresponding to different n sampling times, and the processes of steps S2 to S5 are repeated by the number of times corresponding to the number of the patterns to learn the weight of the network. Is performed.

In FIG. 3, the output of the output layer neuron is a general one given by the following equation using a weight matrix, an input vector, and a sigmoid function F. F (vector W vector X) (5) The quantized data storage unit 25 in FIG. 2 corresponds to the output of the vector quantization learning unit 22 after the learning is completed, that is, the time series data of a plurality of observation items. The output pattern of each neural network, ie, the quantized data of the symptoms / signs given as the outputs of the m output neurons, and the data of the cause corresponding to the symptoms / signs as a pair, the value of which is −1, Or, it is stored as bipolar information represented by any of +1. Here, the cause corresponding to the symptom / sign data output from the vector quantization learning unit 22 is known, and the cause data is given, for example, by the operator of the associative plant abnormality diagnosis device, and the quantization data storage unit 25 Is stored. Note that the output 0 of the neuron is −
1, the output 1 is converted to +1 as it is. The reason why the output 0 is converted to −1 is that a value of 0 in a matrix operation (multiplication) described below makes the meaning of the operation result unclear (any value multiplied by 0 becomes 0).

A is the cause data stored in the quantized data storage unit 25, and B is the symptom / sign data corresponding to the cause A.
Then, B has m elements as the output of the vector quantization learning unit 22, but the cause is that A has 1 (ell) elements and is represented by a matrix of one row each. Shall be performed. All elements of these matrices have a value of +1 or -1.

Vector A = (a ₁ a ₂ ... _Al ) (6) Vector B = (b ₁ b ₂ ... B _m ) (7) Distributed associative memory of FIG. The model learning unit 26 creates a cross-correlation matrix using the pair information of the quantized data of the symptom / sign stored in the quantized data storage unit 25 and the corresponding cause data, and generates the cross-correlation matrix. Storage unit 27
Is to be stored.

This cross-correlation matrix is a bidirectional associative memory model as a model of the relationship between the cause of the abnormality and the symptom / sign, and is used to estimate the cause from symptom / sign data generated from an unknown cause. That is, the cross-correlation matrix created based on the known cause and the symptom / sign pair information of the cause stored in the quantized data storage unit 25 is stored in the cross-correlation matrix storage unit 27, An abnormality diagnosis is performed using the associative memory model.

When the cross-correlation matrix is created from the relationship between the cause of the abnormality and the symptom / sign for the cause, the following method is used because the storage capacity of the cross-correlation matrix storage unit 27 is limited. That is, when a cross-correlation matrix as a bidirectional associative memory model is created using paired information of new symptom / sign quantized data and cause data stored in the quantized data storage unit 25, cross-correlation matrix storage is already performed. The relationship between the pair information already used for creating the correlation matrix already stored is associated with the bidirectional associative action using the matrix obtained by adding the correlation matrix to be newly created to the correlation matrix stored in the unit 27. Then, a cross-correlation matrix as a new bidirectional associative memory model is created only when no association is made.

FIG. 5 is a detailed flowchart of a process of creating a cross-correlation matrix from paired information of quantized data of a new symptom / sign and its corresponding cause data by the distributed associative memory model learning unit 26. When the process is started in the figure, first, in step S11, the initial value of the maximum number L of the cross-correlation matrix is set to 1. Then, in step S12, the read count _k of the pair information (A _k , B _k ) of the cause and symptom is read.
Are initialized and k = 0 is set. Then, for all the new pair information to be expressed as a cross-correlation matrix (hereinafter, referred to as memorized), steps S13 to S3
The processing up to 1 is repeated.

Here, the pair information to be recorded in the cross-correlation matrix includes (A ₁ , B ₁ ), (A ₂ , B ₂ ),.
It is assumed that there is a kmax set of (A _kmax , B _kmax ). First, in step S13, the value of k is incremented, and is set to 1 here.
In step S14, a cross-correlation matrix vector U for the k-th, here, first pair information (A _k , B _k ) is generated using the following equation.

Vector U = Vector A ^T Vector B = (a ₁ a ₂ ... A _l ) ^T (b ₁ b ₂ ... B _m ) (8) In step S 15, the cross-correlation matrix already created and stored As the initialization of the read count of the correlation matrix vector W _q stored in the unit 27, the value of q is set to 0,
In step S16, the value of the flag indicating whether or not the pair information (A _k , B _k ) can be stored in the correlation matrix vector W _q is set to zero. Step S17 is performed for each of the correlation matrices already stored in the cross-correlation matrix storage unit 27.
Steps S28 to S28 are repeated.

First, in step S17, the value of the number of read counts of the correlation matrix is incremented. Here, q = 1 is set first. Then, it is determined in step S18 whether or not the value of the flag is 0. If the value is 0, the value of the flag is set to 1 in step S19, and the correlation matrix vector W _q , here, the vector W ₁ and the step S14 The sum with the cross-correlation matrix vector U created in step (1) is calculated, and the result is defined as a vector W _q '. In step S20, the value of the read count number z of the pair information already stored in the correlation matrix vector W _q , here, the vector W ₁ is initialized to 0,
Processing of step S21~ step S25 is repeated for each of the memorization by pair information on the correlation matrix vector W _1.

In step S21, the value of the read count number z of the pair information is incremented, and here the value is set to 1. Then, in step S22, the pair information (C _qz , D _qz )
, A bidirectional associative action by the cross-correlation matrix vector W _q ′ is performed, and as a result of the associative action (recall operation), stable state pair information (X, Y) is obtained (output). This bidirectional associative action (recall operation) will be described in detail with reference to FIG.

It is determined in step S23 whether the stable state outputs X and Y obtained in step S22 satisfy the following conditions. X = C _qz and Y = D _qz (9) If this condition is satisfied, the pair information already stored in the cross-correlation matrix vector W _q (here, (C ₁₁ ,
D ₁₁ ) is also associated with the new correlation matrix vector W _q ′ created in step S 19, so that the process proceeds to step S 25 without performing step S 24.

On the other hand, if this condition is not satisfied, the pair information (C ₁₁ , D ₁₁ ) will not be associated with the newly created correlation matrix vector W _q ′. Step S1 from W _q '
The correlation matrix vector U created in step 4 is subtracted, the value is returned to the original correlation matrix vector _Wq , the value of the flag is set to 0, and the process proceeds to step S25.

[0041] At step S25, it is determined whether or not reached to the number k _q of pair information read count z already memorisation has been paired information to cross-correlation matrix vector W _q is already memorisation is reached If not, step S2
The processing after 1 is repeated. That is, step S21
Here, the value of z is set to 2, an associative action is performed on the pair information (C ₁₂ , D ₁₂ ) in step S22, an output in a stable state is obtained, and then the processing after step S23 is performed. .

[0042] When step S25 the number read count of pair information that is already memorisation the correlation matrix vector W _q in is determined to have reached the number k _q of pair information that is memorisation, flags in step S26 It is determined whether the value is 1 or not. Unless the value of this flag is set to 1 in step S19 after being set to 1 in step S19,
It is kept at 1. The the value of the flag is set to 0 in step S24, plus the pair information already memorisation the correlation matrix vector W _q, among the k _q sets in total, the correlation matrix vector U to W _q matrix It is limited to the case where even one pair of pair information not associated with the vector W _q 'exists.

Therefore, when the flag is 1,
The pair information on which the correlation matrix vector U generated in step S14 is based, in this case, (A ₁ , B ₁ ) is converted into a vector W _q ′ including the pair information recorded in the cross-correlation matrix vector W _q. Step S2 to be able to memorize
7 The number of pair information that is memorisation on the correlation matrix is incremented, in that the incremented number, i.e. pair information k _q +1 th pair information is already memorisation the correlation matrix vector W _q In addition, the processing shifts to the processing of step S28.

[0044] If the flag is not 1 at step S26, it is not possible to memorisation pair information is the basis of the correlation matrix created in step S14 to the correlation matrix vector W _q, performs the processing of step S27 The process moves to step S28 without any processing.

In step S 28, it is determined whether or not the read count number q of the cross-correlation matrix vector W _q has reached the number L of correlation matrices already stored in the cross-correlation matrix storage unit 27. Here, the value of q is 1, and generally does not reach L yet, so the processing after step S17 is repeated. That is, the value of q is 2 in step S17, already step S18 and subsequent processing is repeated for the second matrix vector W ₂ of the correlation matrix being created, pair information (A _1, B ₁₎ is and it determines whether it memorisation the correlation matrix vector W _2, processing corresponding to the determination result is repeated.

If it is determined in step S28 that the processing has been completed for all the correlation matrices already created,
The process moves to step S29. Alternatively, when it is determined in step S18 that the value of the flag is not 0, the process proceeds to step S29. Step S18
The flag value is determined to be not 0, that is, 1 when the flag value is determined to be 1 in step S26. For example, the flag value is generated with respect to the pair information (A ₁ , B ₁ ). and a case where the correlation matrix vector U is all pair information that is memorisation in the correlation matrix vector W _q by the vector W _q to the applied correlation matrix vector W _q 'is associative, pair information (a _1, B ₁ ) is step S27
This is the case where the information is added to the pair information already recorded. In this case, the determination of the association possibility of the already stored pair information by the correlation matrix to which the correlation matrix vector U generated for the pair information is added, and the processing corresponding thereto are discontinued, and the process of step S29 is terminated. It will shift to processing.

In step S29, it is determined whether or not the flag is 0. If the flag is 0, the process proceeds to step S30. If the flag is not 0, the process in step S31 is performed without performing the process in step S30. Move on to processing. The flag is 0, a case where one of the pair information already memorisation vector W _q by the correlation matrix vector W _q 'in step S24 is not associative,
Moreover, such a state is already stored in the cross-correlation matrix storage unit 27.
, And corresponds to the case where the read count number q of the correlation matrix reaches L in step S28. In this case, a completely new correlation matrix is created, the value of the number L of the correlation matrix already created in step S30 is incremented, and the pair information, for example, (A ₁ , B ₁ ) is already recorded. It is informed pair information,
The correlation matrix vector U generated in step S14 is used as a new correlation matrix vector WL as the cross-correlation matrix storage unit 27.
And the process shifts to the process of step S31.

In step S31, it is determined whether or not the read count number k of the pair information has reached _{kmax. When the} count has reached _kmax , the process ends. If not reached, the processing from step S13 is repeated. That is, step S
At step 13, the value of k is incremented. Here, k is set to 2, and the processing after step S14 for generating the cross-correlation matrix for the second pair information (A ₂ , B ₂ ) is repeated.

Thus, the preparation for diagnosis, that is, the learning operation in the associative plant abnormality diagnosis apparatus of the present invention is completed. When the learning operation is completed, the cause of the abnormality is diagnosed from the monitoring data of each observation item input from the monitoring target plant 11 in FIG. 2, that is, input data that may indicate an abnormal state whose cause is unknown. An abnormality diagnosis is performed.

At the time of this abnormality diagnosis, the data of the symptoms / signs inputted from the monitoring target plant 11 is given to the vector quantization recalling section 24 via the observation value vector input section 21. Like the inside of the vector quantization unit learning unit 2, the vector quantization recall unit 24 is provided with a two-layer neural network for each observation item, and the connection weight of the network is stored in the weight vector storage unit 23. Is set to the value of the weight stored in.

Each neural network provided in the vector quantization recall unit 24 outputs a pattern as the above-mentioned symptom data in the same manner as in the vector quantization unit learning unit 22. The output is converted to bipolar information and used for abnormality diagnosis by the abnormality diagnosis unit 28 based on the distributed associative memory model. That is, the data output by the vector quantization recall unit 24 is a pattern of the symptom data before being converted into the bipolar information, and the result of the conversion into the bipolar information corresponds to the aforementioned matrix vector B.

Abnormality Diagnosis Unit 2 Using Distributed Associative Memory Model
8 corresponds to A when the stable state is reached by repeating the bidirectional associative action based on the given symptom / sign pattern B using the correlation matrix stored in the cross-correlation matrix storage unit 27. The data is output from the monitoring target plant 11 to the display device 12 as a cause for a symptom or sign whose cause is unknown.

Abnormality diagnosis unit 2 based on distributed associative memory model
8 is provided with a two-layer neural network having a bidirectional connection as shown in FIG. In the figure, squares represent neurons, and symbols a ₁ ,..., A _l ,
b ₁ ,..., b _m represent the input to the neuron or the output of the neuron, respectively. The weight of the connection from the layer 1 to the layer 2 is determined by the weight matrix vector W which is a correlation matrix stored in the cross-correlation matrix storage unit 27, and the weight of the connection from the layer 2 to the layer 1 is determined by the transposed matrix vector W ^T. Represented by

In the bidirectional association (recall operation) process using the correlation matrix stored in the cross-correlation matrix storage unit 27 as an associative memory model, the bipolar information of the symptom / sign pattern output from the vector quantization recall unit 24 is used. The symptom data pattern B as the conversion result is input to each neuron of layer 2 and the output of each neuron of layer 1 is obtained. Then, the operation in the opposite direction, that is, the output obtained in the layer 1 is input to each neuron in the layer 1 as a new input, and the output of each neuron in the layer 2 is obtained. By repeating such an operation, the output of the layer 1 in a stable state, that is, the pattern A is set as a candidate for a cause for a symptom / sign as monitoring data input from the monitoring target plant 11.

Since a plurality of correlation matrices are generally stored in the cross-correlation matrix storage unit 27, the cross-correlation matrix having the minimum stable state energy in the bidirectional associative action using each correlation matrix is used. The pattern A as the output of the layer 1 is diagnosed as the cause of the abnormality, and the display device 12
Is output to This will be described with reference to FIG. The bidirectional association (recall operation) process can be expressed as follows using mathematical expressions.

[0056]

(Equation 2)

Where, vector B: transposed matrix of the output vector of layer 2 vector A: transposed matrix of the output vector of layer 1 vector W: cross-correlation matrix between layer 1 and layer 2 vector W ^T : transposed matrix of vector W F: Activity function (for example, sigmoid function) NET _i : Weighted sum of inputs to neuron i λ: Slope of activity function In the above-described bidirectional association (recall operation) process in step S22 in FIG. C _qz corresponding to cause A is taken as the first input (to layer 1), a bidirectional recall operation is performed, and the output in the stable state is obtained as the pair state (X, Y).

FIG. 7 is a flowchart of the abnormality diagnosis processing performed by the abnormality diagnosis unit 28 based on the distributed associative memory model. When the process is started in the figure, first, in step S41, the value of the read count number q of the correlation matrix vector W _q stored in the cross-correlation matrix storage unit 27 becomes 0.
And the value is incremented in step S42.

Then, in step S43, the input pattern, that is, the pattern B as bipolar information of symptoms / signs
Is given to the layer 2 of FIG. 6, a bidirectional associative action using the correlation matrix vector W _q is performed, and a resonance pattern (X _q , Y _q ) as a stable state is output. Here, X _q corresponds to the cause pattern A, and Y _q corresponds to the symptom pattern B.

Then, the energy E _{q of the} resonance pattern is calculated in step S44, and the energies of the input pattern B and the resonance output pattern X _q are calculated in the following equation in step S45.

[0061]

(Equation 3)

Here, the reason why the above equation corresponds to energy will be described. Physical kinetic energy is represented by (1/2) × mass × velocity × velocity. Most of the energy can be expressed in the form of a coefficient × a certain physical quantity × a certain physical quantity. The above equation corresponds to this, and can be considered as energy.

Subsequently, in step S46, S44 and S45
It is determined whether or not the energy obtained in
If they do not match, it is determined in step S48 whether or not the read count number q of the correlation matrix has reached the number L of matrices stored in the cross-correlation matrix storage unit 27. If not, step S42 and subsequent steps are performed. Are repeated, and the processing after step S43 for outputting the resonance pattern as a result of the bidirectional associative action using the next correlation matrix is performed.

If it is determined in step S46 that the two energies _Eq and _Eq 'match, then step S46 is executed.
At 47, the resonance pattern (X,
Y) corresponds to the cause of the symptom / sign pattern corresponding to the monitoring data input from the monitoring target plant 11, and X = A of this pattern is diagnosed as the cause of the abnormality. In general, the energy is minimized when reaching a stable state from an unstable state. Therefore, the coincidence between _Eq and _Eq 'means that the energy has been minimized.
In this case as well, a series of processes after step S42 is performed by reading the correlation matrix read count q in step S48.
Is performed until the number reaches L, and the processing ends when the number reaches L.

[0065]

As described in detail above, the following effects are obtained by using the associative plant abnormality diagnosis apparatus of the present invention. First, by dispersing the pair information of the cause and the symptom / sign of the abnormality and storing the information as a plurality of cross-correlation matrices, the stored state is not destroyed in the already stored cross-correlation matrix. It is possible to add new pair information, and it is not necessary to be conscious of the storage capacity of the memory.
A large amount of symptom pair information can be stored, and abnormality diagnosis can be accurately performed.

Secondly, even if data of symptoms / signs where noise is superimposed or deficient is given, the most appropriate cross-correlation matrix among the stored cross-correlation matrices can be obtained by the associative action using a neural network. By selecting the correlation matrix, the cause of the abnormality can be accurately diagnosed.

Third, by bidirectionally evaluating the cause of the abnormality for the symptom / sign and the symptom / sign for the cause of the abnormality, it is possible to suppress an increase in ambiguity that occurs when the evaluation is performed individually. It is possible to diagnose the cause of an abnormality that matches the tendency of symptoms and signs.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the principle configuration of the present invention.

FIG. 2 is a block diagram showing a configuration of an embodiment of an associative plant abnormality diagnosis device of the present invention.

FIG. 3 is a diagram showing a structure of a neural network provided in a vector quantization learning unit.

FIG. 4 is a flowchart of a learning vector quantization process.

FIG. 5 is a processing flowchart of a distributed associative memory model learning unit.

FIG. 6 is a diagram showing a configuration of a bidirectional neural network provided in an abnormality diagnosis unit based on a distributed associative memory model.

FIG. 7 is a processing flowchart of an abnormality diagnosis unit based on a distributed associative memory model.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 symptom / cause data storage means 2 cross-correlation matrix storage means 3 abnormality diagnosis means using associative memory model 10 associative plant abnormality diagnosis device 11 monitored plant 12 display device 21 observation value vector input unit 22 vector quantization learning unit 23 Weight vector storage unit 24 Vector quantization recall unit 25 Quantized data storage unit 26 Distributed associative memory model learning unit 27 Cross-correlation matrix storage unit 28 Abnormality diagnosis unit using distributed associative memory model

 ──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Hirohisa Migita 1 Fujimachi, Hino-shi, Tokyo Fujifacom Control Co., Ltd.

Claims (5)

[Claims] 1. A plant abnormality diagnosis apparatus for diagnosing a cause of an abnormal state of a plant based on observation data input from a monitoring target, wherein the abnormality diagnosis apparatus detects a symptom as an abnormal state corresponding to time series data for an observation item. Symptom / cause data storage means for storing data and data of a known cause of the abnormal state as pair information, and created in accordance with the pair information of the symptom data and cause data stored in the symptom / cause data storage means A cross-correlation matrix storage means for storing the cross-correlation matrix to be performed as an associative memory model; symptom data corresponding to time-series data indicating the current state of the monitoring target; and an associative memory model stored in the cross-correlation matrix storage means. Using a cross-correlation matrix of the abnormality diagnosis means using an associative memory model to estimate the cause of the abnormality as the current state of the monitored object An associative plant abnormality diagnosis device, comprising: 2. In the plant abnormality diagnosis device, symptom data as an abnormal state is calculated by learning vector quantization in accordance with time-series data for the observation item, and the calculation result is stored in the symptom / cause data storage means. 2. The associative plant abnormality diagnosis apparatus according to claim 1, further comprising a vector quantization learning unit that outputs the result to the system. 3. The cross-correlation matrix storage means stores cross-correlation matrices created in accordance with pair information of the symptom data and the cause data each represented as bipolar data. The associative plant abnormality diagnosis device described in the above. 4. A cross-correlation matrix created according to pair information of symptom data and cause data expressed as the bipolar data is added to a cross-correlation matrix already stored in the cross-correlation matrix storage means. Wherein the addition result is stored again in the cross-correlation matrix storage means, or is stored in the cross-correlation matrix storage means as a cross-correlation matrix different from the already stored cross-correlation matrix. The associative plant abnormality diagnosis apparatus according to claim 3. 5. The abnormality diagnosis means using the associative memory model, when indicating the current state of the monitored object from a plurality of cross-correlation matrices as associative memory models stored in the cross-correlation matrix storage means. Repeat the bi-directional associative action using the cross-correlation matrix with respect to the input of the symptom data corresponding to the series data, and when the result of the bi-directional associative action is stable, correspond to the symptom data and the cause data in the stable state. 2. The associative plant abnormality diagnosis apparatus according to claim 1, wherein a cross-correlation matrix having a minimum energy is selected, and a cause corresponding to the cause data is estimated as a cause for the current abnormality of the monitored object. .