JPH06174503A - Plant abnormality diagnostic device - Google Patents

Abstract

PURPOSE:To automatically generate data for learning of diagnosis mechanism with learning function base on plant operation history data. CONSTITUTION:The title device is provided with a means 21 for setting parameters for learning which sets parameters all and a12 for learning to a parameter storage part 23 for learning, a plant history data storage part 29 for storing plant data which are input via a plant data input means 27, a means 31 for generating data for learning which extracts history data based on a parameter a 11 for learning and generates data a13' for primary learning, and at the same time generates a parameter a14 for diagnosis, a means 33 for generating secondary learning which generates data a13'' for secondary learning based on the parameter a12 for secondary learning from the data a13' for primary learning, and a learning processing means 7 which performs learning processing using the data a 13 for learning consisting of data a13' and a13'' and outputs the learning result as diagnosis mechanism.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a plant abnormality diagnosing device having a learning function for diagnosing a plant abnormality and identifying the cause.

[0002]

2. Description of the Related Art Generally, a plant abnormality diagnosis is to diagnose the cause of an abnormality from plant data indicating the behavior of the plant when an abnormality occurs in the plant. Also,
The plant data corresponds to the cause of plant abnormality. That is, if "plant data is X",
There is a correspondence such as “the corresponding cause of abnormality is Y”. Learning abnormality diagnosis is learning such correspondence.

FIG. 20 shows a conventional learning system of a plant abnormality diagnosing device having such a learning function. As shown in this figure, the learning data a1 given by the human 1
Is input to the learning data input unit 3 and stored in the learning data storage unit 5. The learning processing means 7 reads the learning data a1 from the learning data storage unit 5, uses the learning mechanism a2 of the neural network type stored in the learning mechanism storage unit 9 to perform learning for plant abnormality diagnosis, and performs learning. The diagnostic mechanism a3 obtained by the above is stored in the diagnostic means storage unit 11.

Here, using a back propagation learning rule, which is a form of conventional neural network,
The learning mechanism a2 for plant abnormality diagnosis will be described with reference to FIG. Learning data a1 given by human 1
Is composed of plant data b as learning input data and abnormality cause occurrence degree data c as learning output data, and is given to a learning mechanism a2, that is, a back propagation learning rule neural network in this example. The learning mechanism a2 is composed of an input layer 100, an intermediate layer 110, and an output layer 120. The input layer 100 has three input layer neurons i (i = 1, 2, 3) 101, 102,
From 103, the intermediate layer 110 has three intermediate layer neuro j
From (j = 1, 2, 3) 111, 112, 113, the output layer 120 has two output layer neurons k (k = 1, 2) 12.
It is composed of 1, 122.

Back-probation learning rule Learning about plant abnormality diagnosis by a neural network means plant data b1, b2, b of learning data a1.
When 3 is applied to the input layer neuros 101, 102, and 103, respectively, the input layer 100, the intermediate layer 110, and the output layer 1
The connection between the neuros so that the output data d1 and d2 of the output layer neurons 121 and 122 of the output layer 120 calculated in the order of 20 approach the abnormality cause occurrence degree data c1 and c2 of the given learning data a1 respectively. That is, the coefficients Wij and Wjk are changed. In the figure, reference numerals 131 and 132 are output deviation devices for calculating deviations between the output data d1 and d2 and the corresponding abnormality cause occurrence degree data c1 and c2, and outputting deviation data e1 and e2.

FIG. 22 shows an example of a learning data pattern. In FIG. 22, a pattern P1 has plant data b1, b2 and b3 as learning input data b XP1 (1), XP1 (2), When XP1 (3), the abnormality cause occurrence degree data c1 and c2 as the learning output data c are the occurrence degree YP1 (1) and the abnormality cause B of the abnormality cause A, respectively.
Of the plant data b1, b2, b3 are XP2 (1), XP2 (2), XP2.
In the case of (3), it is indicated that the abnormality cause occurrence degree data c1 and c2 are the occurrence degree YP2 (1) of the abnormality cause A and the occurrence degree YP2 (2) of the abnormality cause B. Also, the pattern P
The horizontal axis of the learning input data b of 1 and P2 shows the sampling time T, and the vertical axis shows the value X of the plant data. The value Y of the abnormality cause occurrence degree data c1 and c2 which is the learning output data c of the patterns P1 and P2 is, for example, 0 ≦ YP1.
(1), YP1 (2), YP2 (1), YP2 (2) ≤ 1, and the closer the value is to 0, the lower the probability of occurrence of the abnormal cause (probability). ) Can be expressed as high.

FIG. 23 is a flow chart showing a processing example of the learning processing means 7 for such learning data. In FIG. 23, the learning processing means 7 is started by the start of learning.
First, a learning mechanism a2 of a neural network type as shown in FIG. 21 is read from the learning mechanism storage unit 9 (20
0). Then, the learning data pattern Pl (l = 1, 2, ..., M, stored in the learning data storage unit 5).
In the example of FIG. 22, m = 2. ), The learning data of the pattern P1 is read (201), the respective values are given to the input layer neuros 101, 102, 103 and the output layer neurones 121, 122 of the learning mechanism a2, and the learning is executed in the following procedure. Yes (202).

(1) Given learning plant data b
For 1, b2 and b3, the input layers 100,
The intermediate layer 110 and the output layer 120 are calculated in this order, and the output data d1 and d2 of the output layer neuros 121 and 122 are calculated.

D1 = F ₁ (Wij, Wjk, b1, b2, b3) d2 = F ₂ (Wij, Wjk, b1, b2, b3) However, F ₁ , F ₂ ; output function of neuro Wij; input layer neuro Coupling coefficient Wjk between i and the intermediate layer neuro j: coupling coefficient between the intermediate layer neuro j and the output layer neuro k (2) Deviation data e1 by the output deviation devices 131 and 132
, E2 is calculated.

E1 = d1−c1 e2 = d2−c2 (3) The deviations e1 and e2 are output to the output layer 120, the intermediate layer 110,
While backpropagating in the order of the input layer 100, the deviations e1 and e2
Wij and Wjk are changed so that approaches 0.

By the above procedure, the coupling coefficients Wij and Wjk between the neurons are corrected.

Similarly (203), patterns P2 and P
3, ..., Pm, learning is performed in the same procedure as in the case of the pattern P1, and coupling coefficients Wij, W between the neuros are combined.
Modify jk sequentially.

The learning for the patterns P1 to Pm is repeated (204), but when either of the two end conditions (205) is satisfied next, the diagnostic mechanism a
3 is output (206) and the learning ends.

Termination condition 1; pattern Pl (l = 1, 2,
The deviation data between the output data d1 and d2 of the output layer neuros 121 and 122 calculated from the learning input data of m) and the learning output data c1 and c2 of the pattern Pl are e1 (Pl) and e2. If (Pl), then | e1 (P1) | ≤ε | e2 (P1) | ≤ε :: | e1 (Pm) | ≤ε | e2 (Pm) | ≤ε are simultaneously established. However, ε is the required diagnostic accuracy and is a preset constant.

Termination condition 2; n ≧ η, where n is the number of learning iterations (204). However, η is a preset allowable learning repetition constant.

When the learning is completed by the end condition 2, it means that the diagnostic accuracy does not satisfy the required tolerance value. Therefore, in this case, η is increased or the number of neurons in the intermediate layer 120 is changed to Re-learn.

[0017]

In the conventional plant abnormality diagnosing device with a learning function, since the learning data is given by a human, there are the following problems.

(1) The above-mentioned learning plant data b1,
Since learning data such as b2 and b3 are created or input to the device, labor is required and the creator or the input person is burdened.

(2) There is a risk of human error in the process of data creation or data input.

(3) Since it takes time to create the learning data or input it to the device, the time required to generate the diagnostic mechanism a3 becomes long.

The present invention has been made in order to solve the conventional problems, and provides a plant abnormality diagnosing device for automatically generating learning data of a diagnostic mechanism having a learning function based on plant operation history data. The purpose is to do.

[0022]

In order to solve the above problems, the present invention performs learning for abnormality diagnosis by using learning data composed of learning input data and learning output data, and performs learning. In the plant abnormality diagnosis device for performing abnormality diagnosis on plant data from the plant by the diagnosis mechanism adjusted by, the plant history data storage means for sequentially storing the plant data from the plant and the learning data generated by the information given by the human Parameter setting means for setting a parameter for setting, and based on the parameters set by the parameter setting means, the plant data is extracted from the plant history data stored in the plant history data storage means to generate learning data. And learning data generating means for .

Further, according to the present invention, in the above-mentioned plant abnormality diagnosing apparatus, there is provided a secondary learning data generating means for further generating learning data from the learning data using the extracted data from the plant history data by the interpolation calculation. It is characterized by

[0024]

In the above structure, the plant history data storage means sequentially stores the plant data from the plant and stores it as plant history data. The learning parameter setting means inputs and sets a learning parameter given by a human. The learning data generation means extracts the necessary learning plant history data from the plant history data storage means based on the learning parameter learning parameters set by the learning parameter setting means, and uses it as learning data for learning. Stored in the data storage means.

Thus, the learning data of the diagnostic mechanism can be automatically generated from the plant history data and stored in the storage device. Therefore, the labor for creating or inputting the learning data is not required, the burden on the operator is reduced, and human error is reduced.

Further, by providing the secondary learning data generating means for further generating the learning data from the learning data composed of the extracted plant history data by the interpolative calculation, more learning data is automatically generated. The learning ability can be generated, and the diagnostic ability can be further enhanced by learning using this larger amount of learning data.

[0027]

Embodiments of the present invention will now be described in detail with reference to the drawings. Note that the same portions are denoted by the same reference symbols throughout the drawings, and overlapping description will be omitted.

FIG. 1 shows an embodiment of the plant abnormality diagnosing device of the present invention. This plant abnormality diagnosing device has a learning parameter setting means 21 for setting a learning parameter a11 and a secondary learning parameter a12 given by a human 1, and a learning parameter for storing a learning parameter a11 and a secondary learning parameter a12. A parameter storage unit 23,
A plant data input unit 27 for inputting plant data from the plant 25, a plant history data storage unit 29 for storing the plant data input from the plant 25,
Learning data generating means 31 for extracting plant history data based on the learning parameter a11 to generate primary learning data a13 'and a diagnostic parameter a14.
And a larger amount of learning data a13 from the generated primary learning data a13 'based on the secondary learning parameter a12.
And a learning data storage unit 35 for storing the generated learning data a13.
And a diagnostic parameter storage unit 3 for storing the diagnostic parameter a14 generated by the learning data generating means 31.
7, and a learning mechanism storage unit 9 for storing a learning mechanism a2 in a neural network format as shown in FIG. 21,
Learning processing means 7 that applies the learning data a13 stored in the learning data storage section 35 to the learning mechanism a2 to execute the learning process, and the diagnostic mechanism storage means 11 that stores the diagnostic mechanism a3 obtained by learning. Based on the learning parameter a11 stored in the learning parameter storage unit 23, the diagnostic parameter a14 stored in the diagnostic parameter storage unit 37, and the current plant data input by the plant data input means 27. The diagnosis mechanism a3 is used to perform abnormality diagnosis and to present the result to the plant operator 39.

Next, the operation of the plant abnormality diagnosing device having the above-mentioned configuration will be described by taking as an example the case where the neural network shown in FIG. 21 is used as the learning mechanism a2. First, prior to the plant abnormality diagnosis, learning data for generating the diagnosis mechanism a3 for performing the plant abnormality diagnosis by learning is created as follows.

When the human 1 gives the learning parameter a11 and the secondary learning parameter a12, the learning parameter input means 21 receives them and stores them in the learning parameter storage unit 23. Learning parameter storage unit 23
2 and 3 show examples of data structures of the learning parameter a11 and the secondary learning parameter a12 stored in FIG. As shown in FIG. 2, as the learning parameter a11,
For each learning data pattern to be created, the pattern name, plant data name, plant data sampling start time T _o , sampling time interval Δt, number of plant data, plant abnormality cause name and abnormality occurrence degree Y are set. It Further, as the secondary learning parameter a12, a pattern name and an abnormality degree stage number N are set as shown in FIG.

On the other hand, the plant data input means 27 constantly inputs the plant data from the plant 25 and sequentially stores the plant data in the plant history data storage unit 29 together with the input time. FIG. 4 shows a configuration example of the plant history data stored in the plant history data storage unit 29.

As shown in the flow chart of FIG. 5, the learning data generating means 31 reads the learning parameter a11 stored in the learning parameter storage section 23 when the learning data generation is started (300). Then, the plant history data stored in the plant history data storage unit 29 is read (301), and necessary plant history data is extracted based on the learning parameter a11 to create learning input data b (302). Then, the abnormality cause occurrence degree data, which is the learning output data c, is created from the learning parameter a11, and the diagnostic parameter a14 is generated.
(303) and outputs the primary learning data a13 'as shown in FIG. 6 to the learning data storage unit 35 (30
4) together with the diagnostic parameter a14 as shown in FIG.
Is output to the diagnostic parameter storage unit 37 (305).

Extraction process of the learning input data b (30
2) will be described in more detail with reference to the flowchart of FIG. First, a learning parameter a as shown in FIG.
For the 11 patterns P1, the set collection time TP1 (h) of the plant data I1 is calculated according to the following equation (3
10). _{TP1 (h) = T o +} Δt * (h-1) ......... (I) where, h = 1,2, ..., M (M is the number of plant data for learning, in the example of FIG. 2 M = It is 3.).

Then, the plant data value X is extracted from the plant history data as shown in FIG. 4 using the plant data name (I1) and the obtained collection time TP1 (h) as search keys (311). In addition, in the plant history data stored in the plant history data storage unit 29, when there is no plant data such that the plant data collection time completely coincides with the time TP1 (h), for example, TP1 (h) is the most Extract close plant data. In addition, when two pieces of data closest to TP1 (h) exist in this way, for example, the earliest collection time is extracted.

The plant data XP1 (h) in the pattern P1 thus extracted is obtained by the learning mechanism A2 as shown in FIG.
In the case of the neural network having the structure shown in 1, h =
1, 2, 3 and learning input data b1 input to neuro 1, neuro 2 and neuro 3 of the input layer 100,
b2 and b3.

Extracted plant data XP1 (1), XP1
(2), XP1 (3) are learning plant data b1, b2,
It is stored in the learning data storage unit 35 as b3 (31
2).

In the same manner, the above processing is repeated for all the patterns set in the learning parameter a11 (313) to create learning input data.

Further, in the process of creating the abnormality cause occurrence degree data (303), the learning data generating means 31
Is the output layer neuron 1 of the primary learning data a13 'as shown in FIG. 6, based on the plant abnormality cause names A and B set in the learning parameter a11 as shown in FIG. 2 and the abnormality occurrence degree Y. The learning abnormality cause occurrence degree data c1 and c2 of 2 are generated, and the diagnostic parameter a14 in which the plant abnormality cause names A and B are associated with the output layer neuros 1 and 2 as shown in FIG.

The learning data generating means 31 creates the primary learning data a13 ', and the learning data storage unit 35
Then, the secondary learning data generating means 33 stores the secondary learning parameter a in the learning parameter storage unit 23.
12 from the learning data storage unit 35 to the primary learning data a
13 ′ is read, secondary learning data a13 ″ is further generated using the complementary method, and the secondary learning data a13 ″ is stored in the learning data storage unit 35. FIG. The example of the structure of the data a13 is shown, and the primary learning data a1 created by the learning data generating means 31.
3'and secondary learning data a13 "created and added by the secondary learning data generating means 33.

FIG. 10 shows secondary learning data generating means 33.
The operation flow of the secondary learning data generation means 33 will be described in detail with reference to FIG.

The secondary learning data generating means 33 starts the generation of the secondary learning data, and then the learning parameter storage section 23.
The secondary learning parameter a12 is read from (400), and the primary learning data a13 'is read from the learning data storage unit 35 (401). Then, based on the secondary learning parameter a12 and the primary learning data a13 ', the secondary learning data a13 "is generated according to the following equation (402).

X (h, L) = (XP2 (h) -XP1 (h)) * (L-
1) / (N-1) + XP1 (h) ... (II) Y (k, L) = (YP2 (k) -YP1 (k)) * (L-1) / (N-
1) + YP1 (k) ... (III) However, L = 2,3, ..., N-1 h = 1,2, ..., M k; Corresponding to the output layer neuro k, k = 1,2 M The number of plant data set in the learning parameter a11, M = 3 N in the example of FIG. 2; the abnormality degree step number set in the secondary learning parameter a12, and in the example of FIG. N = 5 X (h, L); input layer neuro learning plant data b1, b2 generated by the secondary learning data generating means 33
b3 Y (k, L); Output layer neuro learning abnormality cause occurrence degree data c generated by the secondary learning data generating means 33
1, c2 XP1 (h); input layer neuro learning plant data b of the pattern P1 generated by the learning data generating means 31
1, b2, b3 XP2 (h); input layer neuro learning plant data b of the pattern P2 generated by the learning data generating means 31
1, b2, b3 YP1 (k); output layer neurolearning abnormality cause occurrence degree data c1, c2 YP2 (k) of the pattern P1 generated by the learning data generating means 31; generated by the learning data generating means 31 Output pattern neuron learning abnormality cause occurrence degree data c1 and c2 of the pattern P2.

Therefore, as the secondary learning data a13 ″, the following data is the secondary learning data generating means 33.
Is generated by. X (1,2), X (1,3), X (1,4) ... Plant data for input layer neuro-1 learning b1 X (2,2), X (2,3), X (2,4) ... input layer neuro 2 learning plant data b2 X (3,2), X (3,3), X (3,4) ... input layer neuro 3 learning plant data b3 Y (1,2), Y (1 , 3), Y (1,4) ... For output layer neuro 1 learning Abnormal cause occurrence degree data c1 Y (2,2), Y (2,3), Y (2,4) ... For output layer neuro 2 learning Abnormality cause occurrence degree data c2 In FIG. 9, each value of the learning data of the patterns P3, P4, and P5 is calculated by the equations (II) and (III).

The pattern name in the secondary learning parameter a12 in FIG. 3 corresponds to the pattern name in FIG. 6 or FIG. 9, and the abnormality degree step number in the secondary learning parameter a12 in FIG. 5 means generating learning data in five stages from the learning data patterns P1 and P2.

The secondary learning data generation means 33 outputs the secondary learning data a13 ″ generated as described above to the learning data storage section 35 (403). The learning data a13 is stored together with the primary learning data a13 'generated by the training data generating means 31.

The learning processing means 7 includes the learning data storage section 3
Learning is performed using the learning data a13 stored in the table 5. The learning process in the learning processing means 7 is performed according to the flowchart of FIG. 23 as in the case of the conventional example. In FIG. 23, m = 5 when the learning data a13 shown in FIG. 9 is used.

By this learning, the coupling coefficient W between the neuros
The learning mechanism a2 for which ij and Wjk are determined is output to and stored in the diagnosis mechanism storage unit 11 as the diagnosis mechanism a3.

Next, the abnormality diagnosis processing in the diagnosis processing means 41 will be described with reference to the flowchart of FIG.
The diagnosis processing means 41 reads the diagnosis mechanism a3 from the diagnosis mechanism storage unit 11 when the diagnosis is started (500). at the same time,
From the learning parameter storage unit 23, the learning parameter a11
(501), the plant data input by the plant data input means 27 is read (50)
2), based on the learning parameter a11, the plant data is collected for diagnosis and stored in the diagnosis data storage unit in the diagnosis processing means 41 (503). After the collection is completed, the diagnostic plant data stored in the diagnostic data storage section is input to the diagnostic mechanism a3 to execute the diagnostic (504). Then, the diagnostic parameter a14 is read from the diagnostic parameter storage unit 37 (5
05), the output data from the diagnostic mechanism a3 is associated with the diagnostic parameter a14, and the plant abnormality cause data is output to the plant operator 39 as a diagnostic result (506).

The flowchart of FIG. 12 shows the process of collecting and extracting diagnostic plant data (503) in more detail. When the diagnostic data collection is started, the plant data collection time T (h) is calculated according to the following equation. (51
0).

T (h) = T + Δt * (h−1) ... (IV) where h = 1, 2, ..., M M; given by the number of plant data of learning parameter a11 as shown in FIG. The number of diagnostic plant data, in this case, M = 3 T; plant diagnosis start time, here, the diagnosis execution start time. Δt: plant data collection time interval set in the learning parameter a11 Next, at the time T (h) obtained above, the plant data having the plant data name set in the learning parameter a11 is collected (511), Figure 1 in the diagnostic data storage
Save as diagnostic data as shown in 3 (51
2). In FIG. 13, X (T1), X (T2), and X (T3) are plant data values read at times T (1), T (2), and T (3), respectively. Reference numeral 1 of mechanism a3
Input layer neuros 1, 2, 3 indicated by 01, 102, 103
The data f1, f2 and f3 are input to the input terminal.

Further, in the diagnostic execution process (504), as shown in FIG. 14, the diagnostic data as shown in FIG. 13 is read from the diagnostic data storage unit 41a, and the input layer neurons 1, 2 of the diagnostic mechanism a3 are read. 3 (101, 102,
103) Input data f1, f2, f3 respectively, and calculate in the order of the input layer 100, the intermediate layer 110, and the output layer 120, and obtain the output values g1, g2 of the output layer neuros 1, 2 (121, 122). And outputs it to the diagnostic result output unit 41b as diagnostic abnormality cause occurrence degree data.

In the diagnosis result output unit 41b, the relationship of the output layer neuro 1 abnormality cause name: A output layer neuro 2 abnormality cause name: B is found from the diagnosis parameter a14 of FIG. 7 generated by the learning data generating means 31. Output layer neuro 1 (12
Output value g1 of 1) is Y1, output layer neuro 2 (122)
If the output value g2 of the above is set to Y2, plant abnormality cause data as shown in FIG. 15 can be generated.

The diagnostic processing means 41 includes a diagnostic result output section 41.
The anomaly cause name and the occurrence degree obtained as described above by b are output to the plant operator 26 as a diagnosis result.

As is clear from the above description, according to the above embodiment, the following effects can be obtained.

(A) Since the learning data is automatically generated from the driving history data and the learning of the diagnostic mechanism is performed using the learning data, the burden of creating the learning data can be reduced. It is possible to reduce human error caused by creating learning data and improve the reliability of the diagnostic mechanism.

(B) Since the diagnosis mechanism is learned by using the plant history data, the plant abnormality pattern obtained as a result of the operation is reflected in the plant abnormality diagnosis device as soon as possible, and the diagnostic ability of the plant abnormality diagnosis device is used to operate the plant. It can be improved more quickly according to the progress, and thus the reliability of the plant can be improved.

(C) Generally, the operation history data is limited,
Although the learning data may be insufficient in some cases, the present embodiment makes it possible to generate a larger amount of learning data from the given historical data, and a diagnostic mechanism obtained as a learning result with more learning data. Can improve their abilities.

(D) Since the correspondence data (diagnosis parameter) between the output layer neuron and the plant abnormality cause name is automatically generated in the learning data generation process, the operator's burden can be reduced and the output layer neuron and the abnormality cause name can be reduced. There is no error caused by artificially establishing a relationship with.

In the description of the above embodiment, the formulas (II) and (III) are used as the formulas for generating the secondary learning data in the secondary learning data generating means 33, but other formulas are used. However, it is the subject matter of the claims of the present invention. Formula (II),
In the case of (III), as shown in FIG. 16, the patterns P1 and P
For the increase (decrease) of the plant data value X between 2
The occurrence degree Y of the corresponding abnormality cause is linearly proportional to
This is applied to the plant phenomenon 51 that increases (decreases). However, depending on the plant phenomenon, the relationship between the increase (decrease) of the plant data value X and the corresponding occurrence degree Y of the abnormality cause may be non-linear. In that case, an equation suitable for such a phenomenon can be used as the secondary learning data generation equation of the secondary learning data generation means 33.

For example, instead of equation (III), Y (k, L) = (YP2 (k) -YP1 (k)) (logY (k, L ')-logYP1 (k)) / (logYP2 (k ) −log YP1 (k)) + YP1 (k) ……… (V) However, the symbol used in formula (V) corresponds to formula (III), and Y (k, L ′) = (YP2 (k ) -YP1 (k)) * (L-1) / (N-
1) + YP1 (k) is used, as shown by reference numeral 52 in FIG.
When the plant data pattern changes from the learning data pattern P1 to P2, the corresponding plant abnormality cause occurrence degree increases sharply at first and gradually increases gradually.

Further, in the above-mentioned embodiment, the type of the plant data of the given learning parameter is described as one point, but a plurality of plant data are also covered by the claims of the present invention. For example, a case where the number of plant data points of the given learning parameter a11 is two as shown in FIG. 17 will be described as an example. The secondary learning parameter a12 is as shown in FIG. However,
The meaning of the learning parameter of No. 7 will be described by using the pattern P1 as a representative. “The data value of the plant data name I1 is the same as that collected from time TI11 (1) to (TI11 (1) + 1 * Δt). It changes like plant history data, and the data value of the plant data name I2 changes like the same plant history data collected from time TI21 (1) to (TI21 (1) + 1 * Δt). Yes, "the abnormality occurrence degree of the plant abnormality cause A is YP1 (1) and the abnormality occurrence degree of the plant abnormality cause B is YP1 (2)".

Here, assuming that the relationship between the pattern state of the plant data and the occurrence degree of the corresponding plant abnormality cause is linear and directly proportional, the following learning is performed in the secondary learning data generating means 33. The secondary learning data can be generated using the training data generation formula. The generated learning data is as shown in FIG. In this example, the learning mechanism a2 has four input layer neurons and two output layer neurons. In addition, FIG.
In, XI11 (1), XI11 (2), XI12 (1), XI12 (2)
Is plant data extracted from the plant history data by the learning data generation means 31 with the learning parameter a11 shown in FIG.

XI1 (h, L) = (XI12 (h) -XI11 (h)) (L-
1) / (N-1) + XI11 (h) ... (VI) XI2 (h, L ') = (XI22 (h) -XI21 (h)) (L'-1) / (N
-1) + XI21 (h) ......... (VII) Y (k, L, L ') = (Y (k, N, L')-Y (k, 1, L ')) * (L-1)
/ (N-1) + Y (k, 1, L ') (VIII) where Y (k, 1, L') = (YP2 (k) -YP1 (k)) (L'-1) / (N
-1) + YP1 (k) Y (k, N, L ') = (YP4 (k) -YP3 (k)) (L'-1) / (N
-1) + YP3 (k) L = 2,3, ..., N-1 L '= 2,3, ..., N-1 h = 1,2, ..., M k; Corresponding to the output layer neuro k, k = 1, 2 M; the number of plant data set in the learning parameter a11, and in the example of FIG. 17, M = 2 N; the number of abnormality degree stages set in the secondary learning parameter a12, In the example of FIG. 18, N = 5 XI1 (h, L); I1 plant data XI11 (h), XI12 (h)
Input layer neuro 1 (h = 1) or input layer neuro 2 (h = 2) learning data XI2 (h, L '); I2 plant data XI21 (h), XI22 (h)
Input layer neuro 3 (h = 1) or input layer neuro 4 (h = 2) learning input data Y (k, L, L ') generated from: pattern P1 set to the learning parameter a11 4, the degree of learning abnormality cause occurrence YP1 (k),
Anomaly cause occurrence degree data for learning of the output layer neuron k (k = 1, 2) generated from YP2 (k), YP3 (k), and YP4 (k). As described above, in the above embodiment, all The learning plant data of is a given learning parameter a
11 and the secondary learning parameter a12.
However, if part of the learning plant data cannot be collected due to a defective plant input point sensor, or if there is no abnormal phenomenon necessary for learning and related learning plant data cannot be collected, the plant history data Unable to generate learning plant data. Corresponding to such a case, it is possible to provide a means by which the human 1 gives a part of the plant data lacking for learning or the human 1 corrects the generated learning plant data. Moreover, what added such a means to the said Example is also the claim scope of this invention.

Further, in the above embodiment, the plant data is collected from the plant on the basis of the start of diagnosis, and the abnormality diagnosis of the present plant is performed based on the collected plant data. The present invention can be applied to a case of diagnosing a past plant state by using the above method. In this case, the extraction of the diagnostic plant data from the plant history data storage unit 29 can be realized, for example, by using the same means as the extraction of the learning plant data from the plant history data storage unit 29.

When a plant abnormality occurs, the plant data before and after the plant abnormality may be used for diagnosis. In such a case, the plant history data storage unit 2 is also used.
The plant data before the time of occurrence can be retrieved from 9 and used for diagnosis.

[0066]

As described above, according to the present invention,
It is possible to automatically generate learning data for a diagnostic mechanism having a learning function from plant operation history data. Therefore, it is possible to reduce the burden on the operator for creating the learning data and inputting the data into the device, and reduce the human error due to the creation and input of the learning data. Can be improved.

Further, according to the present invention, by providing the secondary learning data generating means, more learning data can be generated from the given history data, and learning is performed with more learning data. As a result, the diagnostic ability can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing a plant abnormality diagnosis apparatus according to an embodiment of the present invention.

FIG. 2 is a diagram showing a configuration example of a learning parameter.

FIG. 3 is a diagram showing a configuration example of a secondary learning parameter.

FIG. 4 is a diagram showing a configuration example of plant history data.

FIG. 5 is a flowchart showing an operation of the learning data generating means 31.

6 is a diagram showing a configuration example of primary learning data a13 'generated by learning data generating means 31. FIG.

FIG. 7 is a diagram showing a configuration example of a diagnostic parameter a14.

FIG. 8 is a flowchart showing details of a learning data extraction process from the plant history data in the learning data generation means 31.

9 is a diagram showing a configuration example of learning data a13 stored in a learning data storage unit 35. FIG.

FIG. 10 is a flowchart showing the operation of the secondary learning data generation means 33.

FIG. 11 is a flowchart showing the operation of the diagnostic processing means 41.

FIG. 12 is a flowchart showing details of a process of extracting diagnostic plant data in diagnostic processing means 41.

FIG. 13 is a diagram showing a configuration example of diagnostic data.

FIG. 14 is a diagram showing a diagnosis mechanism a3.

FIG. 15 is a diagram showing a configuration example of plant abnormality cause data.

FIG. 16 is a graph showing changes in the degree of occurrence of an abnormality cause depending on plant data values.

FIG. 17 is a diagram showing a configuration example of a learning parameter a11 in the case of two plant input points.

18 is a diagram showing a configuration example of a secondary learning parameter a12 corresponding to the learning parameter a11 shown in FIG.

19 is a diagram showing a configuration example of learning data a13 generated by the learning parameter a11 shown in FIG. 17 and the secondary learning parameter a12 shown in FIG.

FIG. 20 is a block diagram showing a conventional example of a learning system of a plant abnormality diagnosis device.

FIG. 21 is a diagram showing an example of a learning mechanism a2 for plant abnormality diagnosis.

FIG. 22 is a diagram showing an example of a learning data pattern.

FIG. 23 is a flowchart showing the operation of the learning processing means 7.

[Explanation of symbols]

 1 ... Human 7 ... Learning processing means 9 ... Learning mechanism memory 11 ... Diagnosis mechanism memory 21 ... Learning parameter setting means 23 ... Learning parameter memory 27 ... Plant data input means 29 ... Plant history data storage section 31 ... Learning data generation means 33 ... Secondary learning data generation means 35 ... Learning data storage section 37 ... Diagnostic parameter storage Part 39 ………… Plant operator 41 ………… Diagnostic processing means 100 ………… Input layer 101 ………… Input layer neuro 1 102 ………… Input layer neuro 2 103 ………… Input layer neuro 3 110 ………… Intermediate Layer 111 ………… Middle layer neuro 1 112 ………… Middle layer neuro 2 113 ………… Middle layer neuro 3 120 ………… Output layer 121 ……… Output layer Neuro 1 122 ………… Output layer Yuro 2 a2 ......... learning mechanism a3 ......... diagnosis mechanism a11 ......... learning parameters a12 ......... secondary learning parameters a13 ......... learning data a14 ......... diagnostic parameters

Claims (2)

[Claims] 1. An abnormality diagnosis learning is performed using learning data composed of learning input data and learning output data,
In a plant abnormality diagnosis device for performing abnormality diagnosis on plant data from a plant by a diagnosis mechanism adjusted by learning, a plant history data storage means for sequentially storing plant data from the plant, and the learning data according to information given by a human Parameter setting means for setting a parameter for generating, and based on the parameters set by this parameter setting means, extract plant data from the plant history data stored in the plant history data storage means, and perform the learning And a learning data generating unit that generates training data. 2. The plant abnormality diagnosing device according to claim 1, further comprising secondary learning data generation means for further generating learning data by interpolation calculation from learning data using the extracted data from the plant history data. A plant abnormality diagnosis device comprising.