A method for fault diagnosis of a supercharged boiler based on small sample learning, and a training method and a testing method for a fault diagnosis model

Abstract

The present application belongs to the field of supercharged boiler fault diagnosis, and relates to a supercharged boiler fault diagnosis method based on small sample learning and a training method and a testing method of a fault diagnosis model, aiming to solve the problem of data scarcity in supercharged boiler fault diagnosis, and the points include obtaining fault samples; dividing the fault samples into a training set and a validation set; randomly drawing fault samples from the training set to construct a batch of positive and negative sample pairs with the same number; inputting the positive and negative sample pairs into the twin deep network for training in batches; constructing a validation set and a validation support set; traversing the fault samples in the validation set; obtaining the fault diagnosis accuracy of the model, and the effect is to realize the supercharged boiler fault diagnosis with data scarcity belonging to small samples and nonlinear classification.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of supercharged boiler fault diagnosis, and relates to a supercharged boiler fault diagnosis method based on small sample learning and a training method and a test method of a fault diagnosis model. BACKGROUND

[0002] As a power source of a steam power system, the supercharged boiler provides high-temperature and high-pressure steam for the entire system. Its safety directly affects the normal operation of the entire system. However, due to the complex structure of the supercharged boiler and the harsh working environment, the failure occurs frequently. If the failure cannot be found and eliminated in time, it may cause the entire steam power system to be paralyzed, and even cause irreparable consequences.

[0003] The early fault diagnosis technology applied to the supercharged boiler system mainly diagnoses based on models and expert knowledge. These methods depend on the complex structure and physical and chemical processes of the boiler system. However, the supercharged boiler system has many coupled parameters between devices and complex and various operating conditions. With the development of automation technology, the complexity of the supercharged boiler is increasing, which increases the difficulty of fault diagnosis of the supercharged boiler. In order to effectively diagnose the fault of the supercharged boiler system, the complex structure and physical and chemical processes of the supercharged boiler must be simplified, which limits the ability of the model to describe the actual operation and fault of the boiler, and cannot effectively diagnose some complex faults. At the same time, when the structure and physical and chemical processes of the supercharged boiler change, the model-based and expert-knowledge-based fault diagnosis method must be rebuilt, otherwise it is difficult to effectively diagnose the fault. With the development of computer science and artificial intelligence, the data-driven fault diagnosis technology reduces the dependence on artificial and improves the efficiency of fault diagnosis, so it has become a hot spot in the field of fault diagnosis and has been successfully applied in many aspects. However, the traditional data-driven fault diagnosis method needs a large number of samples to train the model, but in reality, it is usually difficult to obtain enough supercharged boiler fault samples to train a usable fault diagnosis model, which restricts the development of the data-driven supercharged boiler fault diagnosis technology.

[0004] In recent years, deep neural networks based on small sample learning have made great progress in solving the problem of data scarcity, and have become an active field of machine learning. Small sample learning algorithms require less sample data and can perform excellent classification in relatively limited data samples, which is a powerful tool for solving small sample and nonlinear classification problems. At the same time, the fault diagnosis technology based on small sample learning algorithm is data-driven and does not need to understand the system fault mechanism, while the supercharged boiler fault data samples are usually small and nonlinearly distributed, which is consistent with the advantages of small sample learning algorithm. SUMMARY

[0005] To solve the problem of data scarcity in supercharged boiler fault diagnosis, in the first aspect, according to the supercharged boiler fault diagnosis model training method based on small sample learning, the supercharged boiler fault diagnosis model based on small sample learning includes a twin deep network, and the training method S100 includes

[0006] S101. Obtain fault samples;

[0007] S102. Divide the fault samples into a training set and a validation set;

[0008] S103. Randomly select fault samples from the training set to construct a batch of same number of positive and negative sample pairs;

[0009] S104. Input the positive and negative sample pairs in batches into the twin deep network for training;

[0010] S105. Construct a validation set and a validation support set, the fault samples in the validation set are assigned with fault category labels, and the validation support set includes the same N fault categories as the validation set, and each fault category includes K fault samples;

[0011] S106. Traverse the fault samples in the validation set, respectively pair with the fault samples in the validation support set, and then input the paired fault sample pairs into the twin deep network to obtain the fault categories of the fault samples in the validation set;

[0012] S107. By comparing the fault category labels of the fault samples in the validation set, the fault diagnosis accuracy of the model is obtained;

[0013] S108. Repeat steps S103-S107 until a predetermined number of iterations is reached or the fault diagnosis accuracy of the model reaches a preset value;

[0014] S109. Assign the model parameters obtained in the last iteration to the model.

[0015] According to the supercharged boiler fault diagnosis model training method based on small sample learning, the step S11. Obtain fault samples, specifically includes:

[0016] S01. For a fault sample data file of a fault category, set a correlation coefficient threshold T between the fault diagnosis variables of the data file;

[0017] S02. Calculate the correlation coefficient between all pairs of fault diagnosis variables;

[0018] S03. Randomly select a pair of fault diagnosis variables A and fault diagnosis variables B with a correlation coefficient greater than the threshold T;

[0019] S04. Calculate the average value a and the average value b of the correlation coefficient of the fault diagnosis variable A, the fault diagnosis variable B and other fault diagnosis variables respectively;

[0020] S05. If a > b, delete the fault diagnosis variable A and keep the fault diagnosis variable B; otherwise, delete the fault diagnosis variable B and keep the fault diagnosis variable A;

[0021] S06. Repeat steps S03-S05 until the correlation coefficient between all fault diagnosis variables is lower than a given threshold T, and output the remaining fault diagnosis variables to the set VS n , n corresponds to the fault category, VS n corresponds to the fault diagnosis variable of a certain fault category;

[0022] S07. Implement steps S101-S106 on the fault sample data file of each fault category of all N fault categories respectively, obtain the fault diagnosis variable of each fault category, and the result is N fault diagnosis variable sets VS1, VS2, …, VS N ;

[0023] S08. Merge N fault diagnosis variable sets VS1, VS2, …, VS N into one fault diagnosis variable set VS t , and record the number of times each fault diagnosis variable appears in VS t ;

[0024] S09. According to the number of times the fault diagnosis variable appears in the set VS t , sort the fault diagnosis variables from many to few, select the fault diagnosis variables with the top number of times according to the requirements, and obtain the fault diagnosis general variable;

[0025] S10. Extract the time sequence data of the fault general variable to form the fault sample data of a specified dimension.

[0026] According to the supercharged boiler fault diagnosis model training method based on small sample learning, the input of the twin neural network in step S104 is a positive and negative sample pair with the same or different fault categories , where i represents the i th batch index, j represents the j th sample pair from the i th batch, and the output of the twin neural network is the probability that the two samples of the positive and negative sample pair are the same

[0027]

[0028]

[0029] where sigmoid is an activation function, FC is a dense fully connected layer, and f is a one-dimensional neural network, is the distance between the output vectors of the two one-dimensional neural networks f, the two one-dimensional neural networks f of the Siamese neural network have the same structure and share the weights.

[0030] According to the small sample learning-based supercharged boiler fault diagnosis model training method of some embodiments of the present application, the regularization cross-entropy is used to calculate the loss function in the model training process according to some embodiments of the present application:

[0031]

[0032] where represents a vector including a label with a length of M, when the fault sample belongs to the same fault category, t j = 1, otherwise, t = 0, λ represents a regularization factor, and W represents a network weight. The model uses the Adam algorithm to optimize the gradient descent method of back propagation.

[0033] In the second aspect, the test method of the small sample learning-based supercharged boiler fault diagnosis model according to some embodiments of the present application, the model is obtained by any of the methods, and the test method 200 includes

[0034] S201. Constructing a test set and a test support set, the test support set containing N known supercharged boiler fault categories and an N+1 unknown supercharged boiler fault category, wherein each of the N known supercharged boiler fault categories contained in the test support set contains K supercharged boiler fault samples, and the initial fault sample of the N+1 unknown supercharged boiler fault category is the average fault sample of the N known supercharged boiler fault categories;

[0035] S202. Pairing a test fault sample in the test set for testing a supercharged boiler fault category with a supercharged boiler fault sample in the test support set, and then inputting the paired fault sample pair into the Siamese deep network of the trained model to obtain the probability that a test fault sample in the test set belongs to each supercharged boiler fault category.

[0036] The maximum probability value C in the probability that the test fault sample belongs to each supercharged boiler fault category is C:

[0037] S2021. When the maximum probability value C is not less than a set value, the test fault sample belongs to the supercharged boiler fault category corresponding to the maximum probability value C, and the step S202 is returned.

[0038] S2022. When the maximum probability value C is less than the set value, the test fault sample belongs to the unknown supercharged boiler fault category, the supercharged boiler fault category of the test fault sample is determined, the unknown supercharged boiler fault category is designated as the determined supercharged boiler fault category, the fault sample of the unknown supercharged boiler fault category is designated as the test fault sample, and the unknown supercharged boiler fault category is classified into the known supercharged boiler fault category;

[0039] S203. An unknown supercharged boiler fault category is added, the initial fault sample of the added unknown supercharged boiler fault category is set as the average fault sample of the known supercharged boiler fault category, and steps S202-S203 are repeated until the test of all test fault samples in the test set is completed.

[0040] According to the test method of the supercharged boiler fault diagnosis model based on small sample learning, the test support set contains N known supercharged boiler fault categories, each of which contains K supercharged boiler fault samples (S1,..., SN), and the test method 200 contains the following steps: K ), represents a test fault sample, (x n+1 ,y n ′ +1 ) represents the initial fault sample of the N+1 unknown supercharged boiler fault category;

[0041] The test support set is represented as:

[0042] S={(x1,y1),....,(x n ,y n ),(x n ′ +1 ,y n ′ +1 )} (4)

[0043] The maximum probability value C is represented as:

[0044]

[0045] wherein x ck is a sample in the support set, c is a known sample category in the support set, and k is the number of known sample categories in the support set.

[0046] According to the test method of the supercharged boiler fault diagnosis model based on small sample learning, the set value of the maximum probability value C is 0.6.

[0047] According to the test method of the supercharged boiler fault diagnosis model based on small sample learning, the obtained fault sample data is supercharged boiler simulation fault sample data and / or other system simulation fault sample data.

[0048] According to the test method of the supercharged boiler fault diagnosis model based on small sample learning, the other system simulation fault sample data includes power plant thermal system simulation fault sample data.

[0049] In the third aspect, according to the supercharged boiler fault diagnosis method based on small sample learning, the method is implemented in a supercharged boiler fault diagnosis model based on small sample learning, the model is obtained by any of the methods, and the diagnosis method includes

[0050] S10. Pairing the input fault data with the supercharged boiler fault samples in the support set in the model;

[0051] S20. Inputting the paired fault sample pairs into the twin deep network of the model to obtain the probability that the fault data belongs to each supercharged boiler fault category, wherein the maximum probability value C in the probability that the fault data belongs to each supercharged boiler fault category is:

[0052] S201. When the maximum probability value C is not less than a set value, the fault data belongs to the supercharged boiler fault category corresponding to the maximum probability value C;

[0053] S202. When the maximum probability value C is less than the set value, the fault data belongs to an unknown supercharged boiler fault category, an unknown supercharged boiler fault category is set, the initial fault sample of the unknown supercharged boiler fault category is the average fault sample of the known supercharged boiler fault category, the supercharged boiler fault category of the fault data is determined, the unknown supercharged boiler fault category is specified as the determined supercharged boiler fault category, the fault sample of the unknown supercharged boiler fault category is specified as the fault data, and the unknown supercharged boiler fault category is classified into the known supercharged boiler fault category.

[0054] Advantages:

[0055] In the first aspect, the twin deep network is applied to supercharged boiler fault diagnosis, and after the application, the supercharged boiler fault diagnosis with data scarcity and small sample and non-linear classification is realized.

[0056] In the second aspect, the present application can quickly find the variable dimension meeting the model training by screening the general fault diagnosis variables, and can ensure that the variable dimensions of all fault samples are consistent, facilitating the model to train using fault samples of other thermal systems in the training stage, widening the use range of training data, overcoming the defect of few fault samples of the supercharged boiler fault diagnosis, and according to the effect described in the embodiment of the present application, the obtained model can effectively diagnose the faults of the supercharged boiler.

[0057] In the third aspect, the present application can obtain a plurality of fault diagnosis variable sets of different fault categories by screening the general fault diagnosis variables, and then quickly find a general variable set meeting the requirements of the fault diagnosis model and capable of reflecting all fault characteristics, so as to realize the most effective fault diagnosis with the least variables, reduce the model redundancy under the condition of ensuring the fault diagnosis accuracy, and improve the diagnosis efficiency of the fault diagnosis model.

[0058] In the fourth aspect, the present application has the ability of continuous iteration and updating. When an unknown fault category appears, the support set of the model will add the fault category, and the initial fault sample of the unknown supercharged boiler fault category added is the average fault sample of the known supercharged boiler fault category in the support set. Under this setting, the model prompts manual analysis of the fault type, and assigns the fault category corresponding to the fault data identified by manual analysis to the new boiler fault category. Thus, the model can continuously learn new fault types and assign fault data to the corresponding fault types.

[0059] From the above, the technical scheme of the present application applies small sample learning to the supercharged boiler fault diagnosis through the twin deep network, solves the problem of expanding the training data field in the application process, and also solves the problem of limited diagnosis type in the application process. After adaptive improvement, on the one hand, the small sample learning can be trained and learned by using the fault data of other thermal systems or similar equipment to obtain prior knowledge for distinguishing different fault characteristics, and then combined with the limited supervised information of the supercharged boiler, the supercharged boiler fault can be accurately diagnosed. On the other hand, the supercharged boiler fault diagnosis based on small sample learning can allow the support set to be imperfect at the beginning, but the more it is used, the closer it is to the target object and the more it fits the target object. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical scheme of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creating any creative labor.

[0061] Figure 1A small sample learning-based supercharged boiler fault diagnosis logic structure schematic diagram provided for the embodiments of the present application;

[0062] Figure 2 A small sample learning-based supercharged boiler fault diagnosis method flowchart provided for the embodiments of the present application;

[0063] Figure 3 (a) The trend diagram of the fault diagnosis accuracy of the method of the present application under 30 training samples with the change of the test noise level;

[0064] Figure 3 (b) The trend diagram of the fault diagnosis accuracy of the method of the present application under 30 training samples with the change of the test noise level;

[0065] Figure 4 The trend diagram of the fault diagnosis accuracy of the method of the present application with the change of the number of training sample fault categories. DETAILED DESCRIPTION

[0066] The embodiments of the present application are described in detail below with reference to the accompanying drawings, examples of which are shown in the drawings. The embodiments of the present application are taken as an example of the logical structure shown in the drawings, a small sample learning-based supercharged boiler fault diagnosis method, the method comprising:

[0067] Step 1. Obtain the fault sample data used for model training and verification, and perform outlier rejection, missing value filling, noise reduction and normalization processing on the fault sample data.

[0068] In this step, all fault sample data is obtained by obtaining general fault diagnosis variables and screening the general fault diagnosis variables. Specifically, the fault sample data acquisition method comprises the following steps:

[0069] (1) Given the correlation coefficient threshold T between two variables.

[0070] (2) Calculate the correlation coefficient between all variables.

[0071] (3) Randomly select a pair of variables A and B with a correlation coefficient greater than the threshold T.

[0072] (4) Calculate the average correlation coefficient of A and B with other variables, respectively.

[0073] (5) If α>β, delete variable A, otherwise delete B.

[0074] (6) Repeat steps (3)-(5) until the correlation coefficient between all variables is lower than the given threshold T, and output the final remaining variables to the set VS n , n corresponds to the fault category, VSn a fault diagnosis variable corresponding to a certain fault category.

[0075] (7) repeating steps (1)-(6) on all N fault sample data files of fault categories to obtain N fault diagnosis variable sets VS1, VS2, …, VS N .

[0076] (8) merging N fault diagnosis variable sets VS1, VS2, …, VS N into 1 fault diagnosis variable set VS t , and recording the number of times each variable appears repeatedly in the set VS t .

[0077] (9) sorting the fault diagnosis variables according to the number of times they appear in the set VS t from the most to the least, and selecting the first several fault diagnosis variables with a larger number of appearances to form the final fault general variable.

[0078] (10) extracting the time series data of the fault general variable from the large data file to form the fault sample data of a specified dimension.

[0079] According to the scheme, the fault sample data acquisition step can ensure that the variable dimension of all fault samples remains consistent by screening the general fault diagnosis variable, and can also quickly find a general variable set that meets the fault diagnosis model requirements and can reflect all fault characteristics.

[0080] According to the scheme of the present application, the present application uses the fault general variable as the basis of the fault sample data, can quickly find a dimension that meets the model training, and facilitates the model to use the fault sample of other thermal systems for training in the training stage, and can widen the use range of the training data. As described in Embodiment 2 below, the training data is the simulation fault sample data of a certain power plant thermal system, and this way can expand the data source of the training data set to the fault sample data of similar or similar systems, can overcome the defect of few fault samples of the supercharged boiler fault diagnosis, and according to the effect recorded in the embodiment of the present application, the obtained model can still effectively diagnose the faults of the supercharged boiler.

[0081] And the present application can obtain a plurality of fault diagnosis variable sets of different fault categories through steps (7)-(10), different fault diagnosis variable sets have certain repetition, or in other words, different fault diagnosis variable sets have a certain number of same variables, the present application intends to obtain fault general variables from the same variables, however, in the case of taking intersection, the present application finds that the number of fault general variables obtained by the method is extremely small, which cannot meet the variable quantity requirement of training. Therefore, the present application arranges the repeated variables according to the number of times of occurrence, takes the variables with more repetition times as fault general variables according to the model test results and the diagnosis accuracy and efficiency of the model, and the fault general variables determined by the method of the present application meet the variable quantity requirement of training, and according to the model effect shown in Figure 3 and Figure 4 , the model still has high diagnosis accuracy. It can be understood that the present application can obtain prior knowledge capable of distinguishing different fault characteristics by training and learning the fault data of other thermal systems or similar equipment through the above steps, and then combined with the limited supervision information of the supercharged boiler, the present application can realize relatively accurate diagnosis of the fault of the supercharged boiler.

[0082] Step 2. Divide all fault sample data into a training set and a validation set.

[0083] Step 3. Randomly extract fault samples from the training set to construct a plurality of positive and negative sample pairs with the same number.

[0084] Step 4. Input the training positive and negative sample pairs into the twin network designed for the supercharged boiler in batches for training.

[0085] Step 5. Construct a validation support set of the validation set, the support set contains the same N fault categories as the validation set, each category contains K fault samples, and each fault sample is different from the fault samples in the validation set. In this step, N=2, 3, 4…, K=1, 2, 3…, for each specific fault category, K can take different values within the allowed range.

[0086] In step 5, the fault sample variable group used for validation has the same dimension as the sample variable group used for training, the variable group can be different, but must be the same as the variable group of the corresponding validation support set sample, through the screening of the general fault diagnosis variable above, the general variable set of the validation set and the validation sample meeting the variable dimension and composition requirements can be quickly found.

[0087] Step 6. Traverse the fault samples in the validation set, respectively pair with the samples in the validation support set, and then input the sample pairs into the twin network to judge the fault categories of the samples in the validation set.

[0088] Step 7. Verify the original fault class label of the fault sample in the comparison validation set to obtain the accuracy of the supercharged boiler fault diagnosis model based on small sample learning.

[0089] Step 8. Return to step 3 and repeat steps 3-7 until the predetermined number of iterations is reached or the accuracy reaches the preset value.

[0090] Step 9. Save the supercharged boiler fault diagnosis model parameters with the highest diagnosis accuracy.

[0091] Step 10. Construct a test support set for the supercharged boiler test fault sample, which contains N possible fault classes and an unknown fault class N+1, each fault class contains K fault samples, each fault sample is different from the fault sample in the test set. In this step, N = 2, 3, 4…, K = 1, 2, 3…, for each specific fault class, K can take different values within the allowed range.

[0092] In step 10, the variable composition of the sample used for testing is the same as the variable dimension of the sample used for training, the variable composition can be different, but it must be the same as the variable composition of the test support set sample, and the general variable set of the test set and test sample that meets the variable dimension and composition requirements can be quickly found through the above general fault diagnosis variable screening method.

[0093] During the construction of the test support set of the test sample in step 10, the added unknown fault class N+1 can make the model discover new unknown fault classes, and the results of analyzing new fault classes can be used to improve the test support set, thereby expanding the fault diagnosis range and diagnosis accuracy of the supercharged boiler fault diagnosis method based on small sample learning.

[0094] Step 11. Pair the supercharged boiler fault samples used for testing with the samples in the test support set, and input the sample pairs into the trained fault diagnosis model to obtain the probability of each fault class to which the test sample belongs. Find the maximum probability value C, when C ≥ 0.6, select the corresponding fault class, when C < 0.6, directly classify it into the N+1 fault class. Here, 0.6 can be adjusted within the range of [0.5-0.95] according to actual needs. For the case of C < 0.6, human intervention is required for analysis, and after analyzing the fault cause and class, the sample of the fault class is used to replace the original average sample of the N+1 fault class, and is classified as a known fault class. Then take 1~(N+1) fault class sample average sample as the initial sample of the N+2 unknown class, and repeat this process several times to perfect the support set until the N value is stable. This process expands the fault diagnosis range of the invention and improves the diagnosis ability of the invention.

[0095] According to the scheme, the inventors find that in the actual supercharged boiler fault, the existing method can only identify the fault sample that supports the existing fault type. However, the fault sample of the supercharged boiler that has not occurred is difficult to obtain, so it is difficult to provide accurate supercharged boiler fault diagnosis more comprehensively in actual model testing or use, and the diagnosis type has limitations in small sample learning direction.

[0096] Therefore, the present application uses steps 10 and 11, so that the model has the ability of continuous iterative learning. When the fault data without fault sample appears, the model support set will add a new fault category, and the initial fault sample of the new unknown supercharged boiler fault category is the average fault sample of the known supercharged boiler fault category. In this setting, the model prompts manual analysis of the fault type, and assigns the fault category corresponding to the fault data identified by manual analysis to the new boiler fault category. Thus, the model can continuously update and learn new fault types, and the fault data and its corresponding fault data.

[0097] Therefore, according to the above, the technical scheme of the present application applies small sample learning to supercharged boiler fault diagnosis through twin deep network, solves the problem of expanding the training data field in the application process, and also solves the problem of limited diagnosis type in the application process. After adaptive improvement, on the one hand, small sample learning can be applied to the training and learning of other thermal systems or similar equipment fault data to obtain prior knowledge of distinguishing different fault characteristics, and then combined with the limited supervision information of the supercharged boiler, the more accurate diagnosis of the supercharged boiler fault is realized. On the other hand, the supercharged boiler fault diagnosis based on small sample learning can allow the support set to be imperfect at the beginning, but the more it is used, the closer it is to the target object, and the more it fits the target object.

[0098] In summary, the present application can accurately and effectively diagnose the common faults of supercharged boiler, thereby meeting the practical needs of engineering.

[0099] Embodiment 1: As a specific example, the training data is supercharged boiler simulation fault sample data, all fault sample variables are the same, and the method comprises:

[0100] Step 1. Obtain the simulation data of the supercharged boiler fault samples, and perform outlier rejection, missing value filling, noise reduction and normalization processing on the fault sample data. In this step, according to the general fault diagnosis variable screening method, the general variable set of the supercharged boiler fault diagnosis is obtained, and finally 20 parameters are selected: saturated steam flow, drum level, auxiliary steam turbine steam flow, superheater outlet steam pressure, furnace outlet flue gas flow, economizer inlet water pressure, economizer inlet water temperature, auxiliary steam turbine outlet temperature, economizer outlet water temperature, feed water flow, auxiliary steam turbine work, economizer outlet water pressure, furnace outlet flue gas temperature, flue gas turbine outlet temperature, superheated steam flow, drum saturation temperature, superheater inlet flue gas temperature, superheater outlet flue gas temperature, auxiliary steam turbine inlet temperature and economizer outlet flue gas temperature.

[0101] Step 2. 70% of the supercharged boiler fault sample simulation data is divided into training set and validation set according to the ratio of 6:4.

[0102] Step 3. Randomly draw fault samples from the training set to construct a batch of positive and negative sample pairs with the same number.

[0103] Step 4. Input the training positive and negative sample pairs into the supercharged boiler designed twin network in batches for training. In the model training stage, the twin network input is a pair of samples with the same or different categories , where i is the i-th small batch index, and the output is the probability of the same of the two samples The specific calculation process is as follows:

[0104]

[0105] Where f is a 1D-CNN neural network, the two branches of the twin network have the same f structure and share the weights, is the distance between the output vectors of the two branches of f.

[0106]

[0107] Where FC is a dense fully connected layer, and sigmoid is the activation function.

[0108] The regularization cross-entropy is used to calculate the loss function in the training process, and the calculation formula is as follows:

[0109]

[0110] Where is a vector with label length M, and when , t j= 1, otherwise 0, where j is the jth sample pair from the ith mini-batch. λ denotes the regularization factor, W denotes the network weights, and finally the model is optimized using the Adam algorithm for backpropagation-based gradient descent.

[0111] Step 5. Constructing the verification support set S of the verification set veri , the support set contains the same N fault categories as the verification set, each category contains K fault samples, and the fault samples in the support set are different from those in the verification set. In this step, N = 2, 3, 4…, K = 1, 2, 3…

[0112] Step 6. Traverse the fault samples in the verification set, respectively pair with the samples in the support set, and then input the sample pairs into the Siamese network to judge the fault category of the samples in the verification set.

[0113] Step 7. By comparing the original fault labels of the fault samples in the verification set, the accuracy of the supercharged boiler fault diagnosis model based on small sample learning is obtained.

[0114] Step 8. Return to step 3 and repeat steps 3-7 until the predetermined number of iterations is reached or the accuracy reaches the preset value.

[0115] Step 9. Save the parameters of the supercharged boiler fault diagnosis model with the best diagnosis effect.

[0116] Step 10. Randomly select a number of samples from the remaining 30% of supercharged boiler fault sample simulation data to construct the test support set of test samples, which contains N known fault categories that the test samples may belong to and the N+1 unknown fault category, each fault category contains K fault samples, and each fault sample is different from the fault sample in the test set. In this step, N = 2, 3, 4…, K = 1, 2, 3…, and for each specific fault category, K can take different values within the allowed range.

[0117] Step 11. Randomly select a number of samples from the remaining 30% of supercharged boiler fault sample simulation data as test samples, then pair the test samples with the samples in the test support set, and input the sample pairs into the trained fault diagnosis model to judge the fault category of the unknown fault samples of the supercharged boiler, and realize the fault diagnosis of the supercharged boiler. In this step, the model testing process can be divided into (N+1)-way 1-shot and (N+1)-way K-shot testing, (N+1)-way K-shot is K times repetition based on (N+1)-way 1-shot.

[0118] In the (N+1)-way 1-shot test task, 1 test sample Classification is performed, and a test support set S is given, which contains N samples of known possible categories and a sample of the (N+1)th unknown category, the initial sample of the unknown category is obtained by averaging the samples of the categories 1~N, and is represented by (x n+1 ,y n ′ +1 ).

[0119] S={(x1,y1),...,(x n ,y n ),(x n ′ +1 ,y n ′ +1 )} (4)

[0120] Then, the test sample is classified according to the most similar sample in the test support set, and the maximum probability of the category to which the test sample belongs is obtained

[0121]

[0122] When , the corresponding fault category is selected, when , it is directly classified into the (N+1)th fault category, and after artificial intervention analysis, the fault reason and category are analyzed, and after the fault reason and category are analyzed, it is classified into a known fault category, and the original average sample is replaced, and then the average sample of 1~(N+1) fault categories is taken as the initial sample of the (N+2)th unknown category, and the process is repeated to continuously enrich the support set and strengthen the fault diagnosis range and diagnosis ability of the model. After multiple diagnoses and perfecting the support set, the N value is finally stable at 10, which is consistent with the simulated supercharged boiler fault category number.

[0123] In the (N+1)-way K-shot test, the model is given a support set composed of N known possible categories and the (N+1)th unknown category, and each category has K samples (S1, …, S K ). Then the model determines the maximum probability of the category to which the test sample belongs

[0124]

[0125] When , the corresponding fault category is selected, when , it is directly classified into the (N+1)th fault category, and the remaining process is the same as the above (N+1)-way 1-shot test task.

[0126] By implementing Example 1, the present application compares the influence of different noise levels in the test environment on the fault diagnosis accuracy of the 1D-CNN algorithm, the small sample learning algorithm One-shot and Five-shot under the conditions that the training sample size is 30 and 60, as shown in Figure 3 The smaller the signal-to-noise ratio in the figure indicates the greater the noise, Figure 3 Overall, the small sample learning method is less affected by the noise environment, and the fault diagnosis accuracy fluctuates less within a certain noise range and is higher than the traditional 1D-CNN method. When the training sample size is 30, compared with the benchmark 1D-CNN algorithm, the accuracy of the small sample learning method is increased by 10.05%, 6.53%, 5.26% and 4.14% at noise levels of -4dB, 2dB, 8dB and without noise; indicating that the accuracy of the small sample learning method is very significant in the test environment with small sample size and high noise. For the problem of fault diagnosis of the supercharged boiler with limited data in reality, the small sample learning method can better meet the actual application requirements.

[0127] Example 2: As a specific example, the training data is the fault sample data of a certain power plant thermal system simulation, all fault sample variables are the same, and the method comprises:

[0128] Step 1. Obtain the fault sample data of a certain power plant thermal system simulation for training and verification of the supercharged boiler fault diagnosis model based on small sample learning, and perform outlier rejection, missing value filling, noise reduction and normalization processing on the fault sample data. In this step, by using the general fault diagnosis variable screening method, a set of general variables of the fault sample of the certain power plant thermal system simulation is obtained, the first 20 variables are taken, and the data of the 20 variables are separated from the original data as the sample for training and verification of the supercharged boiler fault diagnosis model.

[0129] Step 2. Divide the fault sample data of the certain power plant thermal system simulation into a training set and a verification set according to a ratio of 6:4.

[0130] Step 3. Randomly extract fault samples from the training set to construct a batch of positive and negative sample pairs with the same number.

[0131] Step 4. Input the training positive and negative sample pairs into the well-designed twin network for the supercharged boiler in batches for training. In the model training stage, the input of the twin network is a pair of samples with the same or different categories, and the output is the probability of the two samples being the same. During the training process, the regularization cross-entropy is used to calculate the loss function, and the Adam algorithm is used to optimize the gradient descent method of back propagation.

[0132] Step 5. Construct the validation support set S veri, support set contains the same N fault categories as the validation set, each category contains K fault samples, and the fault samples in the support set are different from the fault samples in the validation set, in this step, N = 2, 3, 4…, K = 1, 2, 3…

[0133] Step 6. Traverse the fault samples in the validation set, respectively pair with the samples in the support set, and then input the sample pair into the twin network to judge the fault category of the sample in the validation set.

[0134] Step 7. By comparing the original fault label of the fault sample in the validation set, the accuracy of the supercharged boiler fault diagnosis model based on small sample learning is obtained.

[0135] Step 8. Return to step 3 and repeat steps 3-7 until the predetermined number of iterations is reached or the accuracy reaches the preset value.

[0136] Step 9. Save the supercharged boiler fault diagnosis model parameters with the best diagnosis effect.

[0137] Step 10. Obtain the supercharged boiler fault simulation data, obtain the supercharged boiler fault diagnosis general variable set according to the general fault diagnosis variable screening method, and finally select 20 parameters: saturated steam flow, drum level, auxiliary steam turbine steam flow, superheater outlet steam pressure, furnace outlet flue gas flow, economizer inlet water pressure, economizer inlet water temperature, auxiliary steam turbine outlet temperature, economizer outlet water temperature, feed water flow, auxiliary steam turbine work, economizer outlet water pressure, furnace outlet flue gas temperature, flue gas turbine outlet temperature, superheated steam flow, drum saturation temperature, superheater inlet flue gas temperature, superheater outlet flue gas temperature, auxiliary steam turbine inlet temperature and economizer outlet flue gas temperature. Then a number of samples are extracted from the test support set to construct the test sample, which contains N known fault categories and the N+1 unknown fault category that the test sample may belong to, each fault category contains K fault samples, and each fault sample is different from the fault sample in the test set, in this step, N = 2, 3, 4…, K = 1, 2, 3…, for each specific fault category, K can take different values within the allowed range.

[0138] Step 11. Randomly extract a part of the sample from the obtained supercharged boiler fault simulation data as the test sample, and pair the test sample with the sample in the test support set, then input the sample pair into the trained twin network to judge the fault category of the unknown fault sample of the supercharged boiler, and realize the fault diagnosis of the supercharged boiler. The subsequent model test process of this step is the same as that of embodiment 1.

[0139] Through the above steps, the present application selects 5, 7, 9, 11, 13, 15 different types of faults from a certain power plant thermal system respectively, trains the model by selecting 5 fault samples for each fault type, constructs One-shot support set or Five-shot support set by randomly selecting 1 or 5 samples from the supercharged boiler fault data, and performs fault diagnosis on the supercharged boiler. Such experiments are repeated for 20 times. The final obtained result is shown in Figure 4 As can be seen from the figure, the present application provides a supercharged boiler fault diagnosis method based on small sample learning. The diagnosis accuracy within a certain range is improved with the increase of the training sample categories. When 15 categories of 75 fault samples are used for training, the diagnosis accuracy of small sample learning is more than 80%, which shows that the present application can obtain prior knowledge for distinguishing different fault characteristics by training and learning the fault data of other thermal systems or similar equipment, and then combined with the limited supervision information of the supercharged boiler, the accurate diagnosis of the supercharged boiler fault can be realized.

[0140] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The functions specified in one or more flows and / or blocks.

[0141] These computer program instructions can also be stored in a computer-readable memory that can guide the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The functions specified in one or more flows and / or blocks.

[0142] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The functions specified in one or more flows and / or blocks.

[0143] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for training a supercharged boiler fault diagnosis model based on small sample learning, characterized in that, The supercharged boiler fault diagnosis model based on small sample learning comprises a twin deep network, and the training method S100 comprises S101. acquiring fault samples; S102. dividing the fault samples into a training set and a validation set; S103. randomly extracting fault samples from the training set to construct a batch of positive and negative sample pairs with the same number; S104. inputting the positive and negative sample pairs into the twin deep network for training in batches; S105. constructing a validation set and a validation support set, the fault samples in the validation set being assigned with fault category labels, the validation support set comprising the same N fault categories as the validation set, and each fault category comprising K fault samples; S106. traversing the fault samples in the validation set, pairing each fault sample with a fault sample in the validation support set, and then inputting the paired fault sample pairs into the twin deep network to obtain the fault categories of the fault samples in the validation set; S107. comparing the fault category labels of the fault samples in the validation set with the fault categories of the fault samples in the validation set to obtain the fault diagnosis accuracy of the model; S108. repeating steps S103-S107 until a predetermined iteration number is reached or the fault diagnosis accuracy of the model reaches a preset value; S109. assigning the model parameters obtained in the last iteration to the model.

2. The small sample learning-based supercharged boiler fault diagnosis model training method according to claim 1, characterized in that, The step S11. acquiring fault samples specifically comprises: S01. for a fault sample data file of a fault category, setting a threshold T for the correlation coefficients between the fault diagnosis variables of the data file; S02. calculating the correlation coefficients between all pairs of fault diagnosis variables; S03. randomly selecting a pair of fault diagnosis variables A and fault diagnosis variables B with a correlation coefficient greater than the threshold T; S04. calculating the average values α and β of the correlation coefficients of the fault diagnosis variable A and the fault diagnosis variable B with other fault diagnosis variables, respectively; S05. if α>β, deleting the fault diagnosis variable A and retaining the fault diagnosis variable B; otherwise, deleting the fault diagnosis variable B and retaining the fault diagnosis variable A; S06. Repeat steps S03-S05 until the correlation coefficient between all fault diagnosis variables is lower than a given threshold T, output the reserved fault diagnosis variables to the set VS n , n corresponds to the fault category, VS n corresponds to the fault diagnosis variable of a certain fault category; S07. Steps S101-S106 are implemented on each of the N failure class-specific failure sample data files, respectively, to obtain failure diagnosis variables for each failure class, resulting in N sets of failure diagnosis variables VS1, VS2,..., VS N ; S08. Merge the N sets of fault diagnosis variables VS1, VS2,..., VS N into one set of fault diagnosis variables VS t and record the number of times each fault diagnosis variable appears in VS t . S09. According to the number of times the fault diagnosis variable appears in the set VS t , the fault diagnosis variables are sorted from more to less, and the fault diagnosis variables appearing in the first several are selected according to the requirement to obtain the fault diagnosis general variable; S10. extracting the time series data of the fault general variables to form fault sample data of a specified dimension.

3. The small sample learning-based supercharged boiler fault diagnosis model training method according to claim 1 or 2, characterized in that, The twin neural network input in the step S104 is a fault sample with the same or different fault category Positive and negative sample pairs, where i represents the i th batch index, j represents the j th sample pair from the i th batch, and the twin neural network output is the probability that the two samples of the positive and negative sample pair are the same where sigmoid is an activation function, FC is a dense fully connected layer, and f is a one-dimensional neural network, is the distance between the output vectors of the two one-dimensional neural networks f, the two one-dimensional neural networks f of the siamese neural network have the same structure and share the weights.

4. The small sample learning-based supercharged boiler fault diagnosis model training method according to claim 3, characterized in that, In the model training process, a regularization cross-entropy is used to calculate a loss function: wherein represents a vector including label length M, when the fault sample belongs to the same fault category, t j = 1, otherwise, 0, λ represents a regularization factor, W represents a network weight, and the model is optimized by using an Adam algorithm for a back propagation gradient descent method.

5. A method for testing a supercharged boiler fault diagnosis model based on small sample learning, characterized in that, The model is trained by the method of any one of claims 1-4, and the test method 200 comprises S201. constructing a test set and a test support set, the test support set comprising N known supercharged boiler fault categories and an N+1 unknown supercharged boiler fault category, wherein each of the N known supercharged boiler fault categories in the test support set comprises K supercharged boiler fault samples, and the initial fault sample of the N+1 unknown supercharged boiler fault category is an average fault sample of the N known supercharged boiler fault categories; S202. Pairing a test fault sample in the test set for testing the supercharged boiler fault category with a supercharged boiler fault sample in the test support set respectively, and inputting the paired fault sample pairs into the twin deep network of the trained model to obtain the probability of the test fault sample belonging to each supercharged boiler fault category; The maximum probability value C in the probability of the test fault sample belonging to each supercharged boiler fault category is C: S2021. When the maximum probability value C is not less than the set value, the test fault sample belongs to the supercharged boiler fault category corresponding to the maximum probability value C, and returning to step S202; S2022. When the maximum probability value C is less than the set value, the test fault sample belongs to an unknown supercharged boiler fault category, determining the supercharged boiler fault category of the test fault sample, designating the unknown supercharged boiler fault category as the determined supercharged boiler fault category, designating the fault sample of the unknown supercharged boiler fault category as the test fault sample, and classifying the unknown supercharged boiler fault category into the known supercharged boiler fault category; S203. Adding a new unknown supercharged boiler fault category, setting the initial fault sample of the added unknown supercharged boiler fault category as the average fault sample of the known supercharged boiler fault category, and repeating steps S202-S203 until all test fault samples in the test set are tested.

6. The method of claim 5, wherein the method is based on small sample learning. In the test method 200, each of the N known supercharged boiler fault categories included in the test support set includes K supercharged boiler fault samples (S1,..., SK K ) that are known to be associated with the category. represents a test fault sample, (x n+1 , y n ′ +1 ) represents an initial fault sample of the N+1 unknown supercharged boiler fault category. The test support set is represented as: S = {(xl,yl),..., (xn,yn)} (1) n n n +1 n +1 (4)​​​​​ The maximum probability value C is represented as: where x ck is a sample in the support set, c is the known class of the sample in the support set, and k is the number of samples in the support set of known class.

7. The method of claim 5, wherein the method is based on small sample learning. The set value of the maximum probability value C is 0.

6.

8. The method of claim 5, wherein the method is based on small sample learning. The data of the obtained fault sample is the data of the supercharged boiler simulation fault sample and / or the data of other system simulation fault samples.

9. The method according to claim 8, wherein, The data of the other system simulation fault sample includes the data of the power plant thermal system simulation fault sample.

10. A method for supercharged boiler fault diagnosis based on small sample learning, characterized in that, Implementing a supercharged boiler fault diagnosis model based on small sample learning, the model is trained by the method of any one of claims 1-4, and the diagnosis method comprises S10. Pairing the input fault data with the supercharged boiler fault samples in the support set in the model; S20. Inputting the paired fault sample pairs into the twin deep network of the model to obtain the probability of the fault data belonging to each supercharged boiler fault category, wherein the maximum probability value C in the probability of the fault data belonging to each supercharged boiler fault category is C: S201. When the maximum probability value C is not less than the set value, the fault data belongs to the supercharged boiler fault category corresponding to the maximum probability value C; S202. When the maximum probability value C is less than the set value, the fault data belongs to an unknown supercharged boiler fault category, setting the unknown supercharged boiler fault category, and setting the initial fault sample of the unknown supercharged boiler fault category as the average fault sample of the known supercharged boiler fault category; Determining the supercharged boiler fault category of the fault data, designating the unknown supercharged boiler fault category as the determined supercharged boiler fault category, designating the fault sample of the unknown supercharged boiler fault category as the fault data, and classifying the unknown supercharged boiler fault category into the known supercharged boiler fault category.

Images