Explanatable fault diagnosis method and system for nuclear power station system fault, and storage medium
By combining a sliding window and a positive-channel attention-multilayer perceptron model, the problem of insufficient algorithm transparency in nuclear power plant fault diagnosis is solved, achieving high-precision and efficient fault diagnosis and improving the operation and maintenance efficiency and safety of nuclear power plants.
Patent Information
- Application Number
- CN202511161746.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-19
- Publication Date
- 2025-12-23
AI Technical Summary
Existing nuclear power plant fault diagnosis methods suffer from insufficient algorithmic transparency, difficulty in interpreting and solving analytical results, and lack of verification. This is especially true under high-dimensional data conditions, which leads to low diagnostic efficiency and unreliable results.
A fault diagnosis model is constructed using the sliding window method. Combined with the orthogonal channel attention-multilayer perceptron model, channel statistical features are extracted by generating orthogonal basis vectors and channel-independent filters. Fault classification is performed using attention weights, and local and global interpretable analyses are conducted.
It has achieved high-precision fault diagnosis in nuclear power plants, improved the transparency and credibility of diagnosis, increased the efficiency of fault location, adapted to the high-dimensional data scenarios of nuclear power plants, and met the requirements of safety review.
Smart Images
Figure CN121187828A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of nuclear power plant fault diagnosis, and particularly relates to an interpretable fault diagnosis method and system for nuclear power plant system faults and a storage medium. BACKGROUND
[0002] To ensure the safe and stable operation of a nuclear power plant, accurate fault diagnosis of nuclear power plant faults is required. When a nuclear power plant fault occurs, an operator needs to process a large amount of system monitoring data and determine the fault type and fault cause based on experience. This not only has low diagnosis efficiency, but also has the risk of introducing human error. Deep learning technology has the potential to improve the accuracy and efficiency of fault diagnosis in the field of nuclear power plant fault diagnosis due to its ability to mine the implicit relationships in data. However, due to the "black box" nature of deep learning models, it is difficult to provide diagnosis decision-making basis for operators, making it difficult for fault diagnosis methods based on this technology to meet the requirements of nuclear power plant safety review and gain the trust of operators, and facing certain difficulties in practical engineering applications. Therefore, it is necessary to develop an interpretable deep learning nuclear power plant fault diagnosis method with high precision and high transparency. Through interpretable analysis of the fault diagnosis results, key parameters that have a greater impact on various nuclear power plant faults can be identified, which not only helps operators to understand the internal logic of the diagnosis decision and improves the credibility of the fault diagnosis results, but also assists them in locating abnormal data in a large amount of system monitoring data, further improving the efficiency of nuclear power plant fault diagnosis and maintenance.
[0003] Currently, interpretable fault diagnosis methods for nuclear power plants are mainly performed through post-hoc interpretable analysis methods such as LIME and SHAP. The above methods have certain defects when used for interpretable analysis of nuclear power plant fault diagnosis. The nuclear power plant system data has a high dimensionality, and there are many simplifications in the LIME method, which has poor applicability under high-dimensional data conditions. The SHAP method needs to traverse each data dimension combination to determine its impact on the diagnosis result, which is relatively slow in the presence of high-dimensional data, affecting the timeliness of fault diagnosis. Furthermore, the interpretable analysis results given in some existing studies exist independently, and their accuracy cannot be verified through further comparison in practical engineering. In addition to post-hoc interpretable analysis methods, some studies have also used methods such as reducing model complexity or setting attention modules to explain fault diagnosis models from the inside. However, due to the high dimensionality of nuclear power plant data and the simplicity of single-dimensional data features, existing methods have problems such as overfitting and poor generalization performance.
[0004] In summary, developing an interpretable fault diagnosis method for nuclear power plant system faults is of great significance for achieving high-precision fault diagnosis of nuclear power plants, ensuring efficient operation and maintenance of nuclear power plants, meeting safety review requirements, and gaining the trust of operators. Summary of the Invention
[0005] The purpose of this invention is to solve the problems of insufficient algorithm transparency, difficulty in solving interpretable analysis results, and lack of verification in the fault diagnosis technology of nuclear power plant systems, and to propose an interpretable fault diagnosis method, system and storage medium for nuclear power plant system faults.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] An interpretable fault diagnosis method for nuclear power plant system faults, comprising the following steps:
[0008] Step 1: Collect data files of multiple fault types from the nuclear power plant system, and construct a training set for the fault diagnosis model based on the sliding window method;
[0009] Step 2: Establish and train the orthogonal channel attention-multilayer perceptron model, generate orthogonal basis vectors of input data to construct channel-independent filters, extract channel statistical features, generate orthogonal channel attention weights based on statistical features, and input the attention weights into the multilayer classifier for fault classification training;
[0010] Step 3: Use the trained model to perform fault diagnosis and classification on the unlabeled dataset to be analyzed;
[0011] Step 4: Based on the fault classification results and the attention channel weights of the main traffic lane, perform local interpretability analysis on each data sample;
[0012] Step 5: Take the average value of the local interpretation results for the same fault type to obtain the global interpretable analysis results.
[0013] Further, step 1 includes:
[0014] Step 1.1: Collect data files of multiple fault types from the nuclear power plant system. Select a file and start from the first time step. Divide the file according to the set window size to obtain a data sample with a length equal to the window size and the same dimensions as the fault data.
[0015] Step 1.2: Shift the window along the time step of the dataset with a fixed time step to obtain the next sample;
[0016] Step 1.3: Repeat the sliding window operation to collect data until no more samples can be generated from the data, then discard the remaining data;
[0017] Step 1.4: Traverse all fault type data files and perform the above operations;
[0018] Step 1.5: Randomly select a specified number of data samples from the data samples of each fault type as the training set.
[0019] Further, step 2 includes:
[0020] Step 2.1: Process the input data Generate a set of random initialization vectors Generate orthogonal vectors recursively:
[0021]
[0022] In the formula, X represents the input data, C represents the dimension of the input data sample, L represents the length of the input data sample, and u i For the generated orthogonal vectors, k i The normalized orthogonal vectors are used to obtain the orthogonal filter K = {k1, k2, ..., k} in the orthogonal channel attention module. C},K T K = 1;
[0023] Step 2.2: Multiply the orthogonal filter by the input data X to obtain the filtered result Z:
[0024]
[0025] Step 2.3: Perform global average pooling on Z along the channel dimension to obtain the global statistics s for each channel:
[0026]
[0027] In the formula, s i The elements in the pooling result s, s = {s1, s2, ..., s...} C};Z n,i These are elements in the filtered result Z;
[0028] Step 2.4: Perform two fully connected processing steps, dimensionality reduction and dimensionality increase, on s to generate channel weights, and multiply them with X to obtain the attention weights of the module input;
[0029] Step 2.5: Input the attention weights sequentially into the multilayer perceptron module. Add regularization methods such as Dropout to the multilayer perceptron module to improve the model's generalization ability. The output h of the i-th layer multilayer perceptron module... i for:
[0030] h i =Dropout p (ReLU(W i ·h i-1 +b i ))
[0031] In the formula, Dropout pThis indicates that the dropout rate is p, the layer is a Dropout layer, ReLU is the activation function, and W is a dropout rate. i b represents the weights of the nonlinear links in the multilayer perceptron module. i For bias;
[0032] Step 2.6: Set up a SoftMax layer after the multilayer perceptron module to output the fault classification results;
[0033] Step 2.7: Train the established model using the training dataset, and save the trained model and the channel weights of the positive traffic channel attention module.
[0034] Furthermore, the multilayer perceptron module includes three fully connected layers.
[0035] Furthermore, in step 4, based on the fault classification results and the attention channel weights of the main traffic lane, a local interpretable analysis is performed on each data sample. Each data sample to be analyzed is multiplied by the saved channel weights of the main traffic lane attention module to obtain its attention weight, which is the local pre-interpretive interpretable analysis of this data sample. The interpretable analysis results are saved based on the fault type labels obtained from the diagnostic results.
[0036] A computer device / equipment / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of an interpretable fault diagnosis method for nuclear power plant system faults.
[0037] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of an interpretable fault diagnosis method for nuclear power plant system faults.
[0038] A computer program product includes a computer program / instructions that, when executed by a processor, implement steps of an interpretable fault diagnosis method for nuclear power plant system faults.
[0039] The beneficial effects of this invention are as follows:
[0040] Compared with existing technologies, this invention has significant advantages in the interpretable diagnosis of nuclear power plant system faults: It does not employ post-hoc analysis methods, but instead improves the model structure. While ensuring generalization performance, it uses channel weights from an attention module to evaluate the importance of each dimension of data. Furthermore, it replaces traditional modules with a positive channel attention module, ensuring the independence of each data channel during training, making the interpretable analysis results more reliable. Compared to post-hoc methods such as SHAP, its local interpretable analysis does not require a large amount of background data, the process is simple and fast, and it is more suitable for scenarios where large amounts of engineering data for nuclear power plant faults are difficult to obtain. This method uses a sliding window technique to transform system data into massive samples, processes them through a positive channel attention module to generate channel weights (ensuring channel independence), and then uses a multilayer perceptron to deepen the model to improve diagnostic accuracy and generalization ability. During training, the module channel weights are saved and combined with test data to obtain pre-hoc interpretable analysis results, achieving interpretable diagnosis of nuclear power plant system faults with good overall diagnostic performance and accurate analysis results. Attached Figure Description
[0041] Figure 1 Overall calculation process diagram;
[0042] Figure 2 A diagram of a sliding window;
[0043] Figure 3 Schematic diagram of the training process of the positive traffic flow attention-multilayer perceptron model;
[0044] Figure 4 Schematic diagram of confusion matrix of fault diagnosis test results of positive traffic channel attention-multilayer perceptron model;
[0045] Figure 5 Comparison chart of interpretable analysis results under normal operating conditions;
[0046] Figure 6 Comparison of interpretable analysis results under water loss accident conditions;
[0047] Figure 7 Comparison of interpretable analysis results under the condition of a steam generator heat transfer tube rupture accident;
[0048] Figure 8 Comparison of interpretable analysis results under spring rod accident conditions;
[0049] Figure 9 Comparison of interpretable analysis results under the main pump rotor jamming accident condition;
[0050] Figure 10 Comparison of interpretable analysis results under the condition of main steam pipeline rupture accident. Detailed Implementation
[0051] The present invention will now be further described with reference to the accompanying drawings.
[0052] This invention relates to interpretable fault diagnosis of nuclear power plant system faults. The invention includes: constructing a training set for training a fault diagnosis model; establishing a forward-channel attention-multilayer perceptron model, training it for fault diagnosis, and saving the trained model and forward-channel attention channel weights; performing fault diagnosis on an unlabeled dataset to be analyzed; based on the fault diagnosis classification results of the dataset to be analyzed and the saved forward-channel attention channel weights, performing local interpretable analysis on each data sample; and averaging the local interpretable analysis results of all samples with the same fault type to obtain the global interpretable analysis result for the dataset to be analyzed.
[0053] according to Figure 1 The present invention is implemented using the following technical solutions:
[0054] The program code for this invention is implemented using Python programming language and PyCharm Community Edition 2024.3x64 as the integrated development platform. Its main function is to achieve interpretable fault diagnosis of nuclear power plant system faults. The specific steps are as follows:
[0055] Step 1: Construct a training set for the fault diagnosis model based on the sliding window method, such as... Figure 2 As shown:
[0056] 1) The collected nuclear power plant system fault data contains multiple data files of different fault types. Select one file and start from the first time step, and divide it according to the set window size to obtain a data sample with the same length as the window size and the same dimensions as the fault data.
[0057] 2) Shift the window along the time step of the dataset according to the set time step to obtain the next sample;
[0058] 3) Repeat the sliding window operation to collect data until no more samples can be generated from the data, then discard the remaining data;
[0059] 4) Perform sliding window and sample acquisition on the remaining data files;
[0060] 5) Randomly select a specified number of data samples from the data samples of each fault type as the training set.
[0061] Step 2: Establishing and training the positive directional attention-multilayer perceptron model, as follows: Figure 3 As shown:
[0062] 1) For input data First, generate a set of random initialization vectors. And recursively generate orthogonal vectors based on random initial vectors:
[0063]
[0064] In the formula, X represents the input data, C represents the dimension of the input data sample, L represents the length of the input data sample, and u i For the generated orthogonal vectors, k i These are normalized orthogonal vectors. The orthogonal filter K = {k1, k2, ..., k} in the orthogonal channel attention module can be obtained from the generated normalized orthogonal vectors. C},K T K = 1;
[0065] 2) Multiply the generated orthogonal filter by the module input data X to obtain the filtering result Z:
[0066]
[0067] 3) Perform global average pooling on the Gram-Schmidt filter output Z along the channel dimension, and the resulting s is the global statistic for each channel:
[0068]
[0069] In the formula, s i The elements in the pooling result s, s = {s1, s2, ..., s...} C};Z n,i These are elements in the filtered result Z.
[0070] 4) After performing two fully connected layer operations on s (dimensionality reduction and dimension increase), the channel weights W can be obtained. c Multiplying the obtained channel weights by X yields the attention weights W of the module input. a ;
[0071] 5) The attention W generated by the traffic lane attention module a The data is sequentially input into three multilayer perceptron modules. Regularization methods such as Dropout are added to the multilayer perceptron modules to improve the model's generalization ability. The output h of the i-th layer multilayer perceptron module... i for:
[0072] h i =Dropout p (ReLU(W i ·h i-1 +b i ))
[0073] In the formula, Dropout p This indicates that the dropout rate is p, the layer is a Dropout layer, ReLU is the activation function, and W is a dropout rate.i b represents the weights of the nonlinear links in the multilayer perceptron module. i For bias;
[0074] 6) Add a SoftMax layer after the multilayer perceptron module to implement fault classification;
[0075] 7) Train the established model using the training dataset, and save the trained model and the channel weights of the positive traffic channel attention module.
[0076] Step 3: Perform fault diagnosis using the unlabeled dataset to be analyzed and the saved model:
[0077] By inputting the unlabeled dataset to be analyzed into the trained normal traffic lane attention-multilayer perceptron model, fault diagnosis results for the unlabeled dataset can be obtained. Test results on the PCTRAN dataset show that the diagnostic accuracy of this invention on the test dataset reaches 99.40%. Figure 4 ).
[0078] Figure 4 OCA-MLP is used to refer to the traffic lane attention-multilayer perceptron model. The confusion matrix in the figure shows the diagnostic results of samples for each fault type. The diagonal part represents correctly diagnosed samples, and the samples located outside the diagonal in the figure are misdiagnosed samples.
[0079] Step 4: Perform local interpretability analysis on the data to be analyzed, and assign each data sample X to the analysis. i 'With the saved positive traffic channel attention module channel weight W c Multiplying them together yields the attention weight W. a This refers to the local pre-interpretable analysis of this data sample, and the interpretable analysis results are saved based on the fault type labels obtained from the diagnostic results.
[0080] PHI i =W a =X i '·W c
[0081] In the formula, PHIi represents sample X. i The results of the local interpretable analysis.
[0082] Step 5: Perform global interpretable analysis on the dataset X' to be analyzed based on the saved interpretable analysis results. Take the average of the local pre-analytical results with the same fault type to obtain the global interpretable analysis results for each fault type.
[0083]
[0084] In the formula, PHI represents the globally interpretable analysis result of the dataset to be analyzed, and n' represents the number of data samples in the dataset to be analyzed.
[0085] Example:
[0086] The trained orthogonal channel attention-multilayer perceptron model and the stored channel weights of the orthogonal channel attention module are used to perform interpretable analysis on nuclear power plant system fault data. The interpretation results are compared using cosine similarity, which is calculated as follows:
[0087]
[0088] In the formula, S1 and S2 are two vectors to be solved; s 1i s 2i These are the elements of two vectors.
[0089] The interpretable analysis results of this invention are generally consistent with the analysis results of Kernel SHAP under conditions of sufficient background data, and the cosine similarity of the global analysis results for each fault type is greater than 97%. Figures 5-10 ).
[0090] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the interpretable fault diagnosis method for nuclear power plant system faults described in any of the above embodiments.
[0091] In some other preferred embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program / instruction is stored, wherein when the computer program is executed by a processor, the steps of the interpretable fault diagnosis method for nuclear power plant system faults described in any of the above embodiments are implemented.
[0092] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-described embodiments of the interpretable fault diagnosis method for nuclear power plant system faults, which will not be repeated here.
[0093] Computer-readable storage media encompass a variety of types, including persistent and non-persistent, portable and fixed. These media store information using different technologies, and the content can be machine instructions, data structures, program modules, or other types of data. Some typical examples of computer storage media include: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), various types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory and other storage technologies, optical storage media such as CD-ROM and digital video disc (DVD), magnetic storage devices such as magnetic tape and disks, and other non-transferable media used to store information accessible to computing devices. It is important to note that the computer-readable media described herein do not include temporary storage media, such as modulated data signals and carrier waves.
[0094] Those skilled in the art will further recognize that the operation of the module can be achieved using existing technical protocols or programs, without relying on new computer programs themselves. The units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0095] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An interpretable fault diagnosis method for nuclear power plant system faults, characterized in that: The specific steps include: Step 1: Collect data files of multiple fault types from the nuclear power plant system, and construct a training set for the fault diagnosis model based on the sliding window method; Step 2: Establish and train the orthogonal channel attention-multilayer perceptron model, generate orthogonal basis vectors of input data to construct channel-independent filters, extract channel statistical features, generate orthogonal channel attention weights based on statistical features, and input the attention weights into the multilayer classifier for fault classification training; Step 3: Use the trained model to perform fault diagnosis and classification on the unlabeled dataset to be analyzed; Step 4: Based on the fault classification results and the attention channel weights of the main traffic lane, perform local interpretability analysis on each data sample; Step 5: Take the average value of the local interpretation results for the same fault type to obtain the global interpretable analysis results.
2. The interpretable fault diagnosis method for nuclear power plant system faults according to claim 1, characterized in that: Step 1 includes: Step 1.1: Collect data files of multiple fault types from the nuclear power plant system. Select a file and start from the first time step. Divide the file according to the set window size to obtain a data sample with a length equal to the window size and the same dimensions as the fault data. Step 1.2: Shift the window along the time step of the dataset with a fixed time step to obtain the next sample; Step 1.3: Repeat the sliding window operation to collect data until no more samples can be generated from the data, then discard the remaining data; Step 1.4: Traverse all fault type data files and perform the above operations; Step 1.5: Randomly select a specified number of data samples from the data samples of each fault type as the training set.
3. The interpretable fault diagnosis method for nuclear power plant system faults according to claim 1, characterized in that: Step 2 includes: Step 2.1: Process the input data Generate a set of random initialization vectors {v1, v2, ..., v C }, Generate orthogonal vectors recursively: In the formula, X represents the input data, C represents the dimension of the input data sample, L represents the length of the input data sample, and u i For the generated orthogonal vectors, k i These are normalized orthogonal vectors; the orthogonal filter K = {k1, k2, ..., k} in the orthogonal channel attention module is obtained from the generated normalized orthogonal vectors. C },K T K = 1; Step 2.2: Multiply the orthogonal filter by the input data X to obtain the filtered result Z: Step 2.3: Perform global average pooling on Z along the channel dimension to obtain the global statistics s for each channel: In the formula, s i The elements in the pooling result s, s = {s1, s2, ..., s...} C };Z n,i These are elements in the filtered result Z; Step 2.4: Perform two fully connected processing steps, dimensionality reduction and dimensionality increase, on s to generate channel weights, and multiply them with X to obtain the attention weights of the module input; Step 2.5: Input the attention weights sequentially into the multilayer perceptron module. Add regularization methods such as Dropout to the multilayer perceptron module to improve the model's generalization ability. The output h of the i-th layer multilayer perceptron module... i for: h i =Dropout p (ReLU(W i ·h i-1 +b i )) In the formula, Dropout p This indicates that the dropout rate is p, the layer is a Dropout layer, ReLU is the activation function, and W is a dropout rate. i b represents the weights of the nonlinear links in the multilayer perceptron module. i For bias; Step 2.6: Set up a SoftMax layer after the multilayer perceptron module to output the fault classification results; Step 2.7: Train the established model using the training dataset, and save the trained model and the channel weights of the positive traffic channel attention module.
4. The interpretable fault diagnosis method for a nuclear power plant system fault according to claim 3, characterized in that: The multilayer perceptron module includes three fully connected layers.
5. The interpretable fault diagnosis method for a nuclear power plant system fault according to claim 1, characterized in that: Step 4 involves performing local interpretability analysis on each data sample based on the fault classification results and the attention channel weights of the main traffic lane. Each data sample to be analyzed is multiplied by the saved channel weights of the main traffic lane attention module to obtain its attention weight, which is the local pre-interpretive interpretability analysis of this data sample. The interpretability analysis results are saved based on the fault type labels obtained from the diagnostic results.
6. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.
7. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 5.
8. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 5.