CNN-LSTM hybrid network-based small pressurized water reactor fault diagnosis method

Abstract

A small pressurized water reactor fault diagnosis method based on a CNN-LSTM hybrid network comprises the following steps: firstly, acquiring sensor signals in a pressurized water reactor control system under a normal working condition, introducing corresponding faults at different positions according to characteristics of various typical faults, and acquiring corresponding sensor signals under different working conditions; discrete sampling processing is carried out as a training set and a test set; preprocessing the fault training set by adopting a signal-image conversion method, inputting obtained image data into a CNN-LSTM diagnosis model and training, and then verifying the diagnosis accuracy and robustness of the model based on test set data; parameters in the diagnosis model are continuously adjusted, so that an optimal diagnosis model is obtained; according to the technical means of combining the CNN network for extracting image features and the LSTM for extracting time sequence data fault features, the defects of a traditional fault diagnosis method in the aspects of diagnosis efficiency, accuracy, effect and the like are overcome; the invention further comprises a system, equipment and a storage medium for implementing the method.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning fault diagnosis for small pressurized water reactors, and particularly relates to a fault diagnosis method for small pressurized water reactors based on a hybrid network of CNN-LSTM (Convolutional Neural Networks-Long Short-Term Memory). Background Technique

[0002] The structural layout of a small pressurized water reactor (SPWR) is relatively complex, its operating conditions often change, and the potential application environment is relatively harsh. The problems of faults in the control system, especially the internal sensor and actuator components, are relatively prominent, seriously threatening the normal operation of the pressurized water reactor. Therefore, in order to ensure the safe and stable operation of the nuclear power plant, it is necessary to monitor the nuclear power device in real time, so as to perform online diagnosis on the abnormalities occurring during operation, help relevant personnel accurately and quickly understand the details of the faults, and take corresponding measures to avoid the further expansion of the incident.

[0003] With the advent of the era of industrial big data and the continuous development of technologies such as sensors and computer AI, the nuclear power device has gradually realized the transformation from simple analog technology to digital intelligence. Establishing an intelligent fault diagnosis model based on deep learning has become a trend in this field. However, the traditional fault diagnosis methods are no longer able to meet the diagnostic requirements of massive data. The fault diagnosis based on manual experience has limited capabilities in dealing with abnormal operating conditions. In order to ensure the safety of mechanical equipment under fault conditions, an intelligent fault diagnosis model can be proposed for application in nuclear power devices as a reference for relevant practitioners to improve their ability to handle sudden faults.

[0004] Relevant information shows that the main techniques used for the fault diagnosis of nuclear power plants include the graph theory method, data-driven method, analytical model method, and knowledge-based (including deep learning) method. Among them, the graph theory method is difficult to solve complex fault diagnosis problems because of its relatively simple modeling; although the analytical model method has good diagnostic effects, it is limited in use because it is difficult to obtain an accurate mathematical model of the diagnostic object; although the data-driven method is easy to implement, it is difficult to explain the obtained results; compared with the previous fault diagnosis methods, the fault diagnosis method based on deep learning can effectively fit the optimal network weights by accurately extracting signal features and learning the characteristic patterns of different faults, so as to reasonably classify the signal features and achieve efficient and accurate end-to-end fault diagnosis of complex systems. The patent application with the application publication number CN113011248A discloses a fault diagnosis method for nuclear power plants based on LSTM, and this method has good application effects in the diagnosis of fault data with insufficient features. However, this method only uses a pure LSTM layer. Although the LSTM network has a memory function and can effectively learn historical information, it can only learn forward information and cannot make full use of backward information. Therefore, the model fault diagnosis rate of this network for fault data with a large number of feature quantities is not high. In addition, some patent applications only use traditional neural networks to diagnose related nuclear equipment. Since the structure of related nuclear equipment is relatively complex and a large amount of operation data generated has rich features, it cannot meet the diagnostic requirements. Summary of the Invention

[0005] In order to overcome the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a fault diagnosis method for a small pressurized water reactor based on a CNN-LSTM hybrid network, which combines a CNN network capable of fully extracting image features and an LSTM capable of fully extracting fault features of time-series data, and solves the deficiencies of traditional fault diagnosis methods in terms of diagnostic efficiency, accuracy, and effect.

[0006] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0007] A fault diagnosis method for a small pressurized water reactor based on a CNN-LSTM hybrid network includes the following steps:

[0008] S1. Collect multi-dimensional time-series data of the steam pressure sensor and the feed water valve of the primary loop nuclear power control system of the small pressurized water reactor under different types and severity levels of fault conditions, as well as under normal conditions, and divide the multi-dimensional time-series data into a training set and a test set;

[0009] S2. Preprocess the multi-dimensional time-series data of the training set and the test set obtained in step S1;

[0010] S3. Construct a fault diagnosis model for a small pressurized water reactor based on CNN-LSTM;

[0011] S4. Use the training set obtained in step S2 to train the small PWR fault diagnosis model based on CNN-LSTM constructed in step S3, and input the test set data obtained in step S2 into the trained small PWR fault diagnosis model based on CNN-LSTM for verification to achieve the diagnosis of faults at different positions, types, and severity levels in the PWR.

[0012] In the said step S1, the multi-dimensional time series data is the physical thermal data of the small PWR, including: the reactor set power and the actual output power, the cold leg and hot leg temperatures, the feed water flow rate of the steam generator, the steam flow rate, the steam pressure, the control rod speed, the feed water valve opening, the feed water valve pressure drop, the steam bypass valve opening, and the steam turbine inlet steam flow rate.

[0013] The division of the training set and the test set in the said step S1 is specifically as follows: After using the m function program for automatically collecting data to collect the training set data, change the time when a fault occurs during the operation of the fault simulation platform to obtain the test set, and the total number of operating conditions of the training set and the test set is the same.

[0014] The preprocessing in the said step S2 is specifically operated as follows: First, perform denoising and reconstruction operations on the multi-dimensional time series fault signals to obtain the whitened time series; Second, re-encode the fault time series based on the gray-scale image conversion method, cut the whitened time series into multiple samples and convert them into a base matrix, and finally, use the normalization method to obtain the fault signal values of the matrix and encode them into RGB pixel values of three channels of red, green, and blue to obtain the image training data.

[0015] The small PWR fault diagnosis model of CNN-LSTM in the said step S3 includes an image input layer, a convolutional layer, a normalization layer, an average pooling layer, an activation function layer, a flattening layer, an lstm layer, a dropout layer, a fully connected layer, a softmax layer, and a classification output layer of a two-dimensional CNN-LSTM network; among which:

[0016] Image input layer: Receive the image training data after preprocessing of the multi-dimensional time series data;

[0017] Convolutional layer: A multi-layer CNN network, whose function is to extract the low-level features of the image in the processing of the image training data, including edges, lines, and corners, and extract more complex features through iterative processing of multiple layers of networks;

[0018] Average pooling layer, placed after the convolutional layer, performs downsampling on the feature map input by the convolutional layer, selects the average value of each pooling window as the output of each pooling window, thereby reducing the spatial size of the feature map while retaining the feature information, and can effectively alleviate the overfitting phenomenon;

[0019] Normalization layer: The core function of this layer is to first normalize the image input data processed by the average pooling layer, and then scale and translate the processed data to make the input distribution of each layer more stable. Among them, the normalization process is to subtract the mean of all eigenvalues of this layer from each eigenvalue, and then divide by the standard deviation of the eigenvalues of this layer;

[0020] Activation function layer: Activate the neurons in the neural network established above, and then transfer the activation information to the next layer of the network;

[0021] Flattening layer: Used to flatten the image feature map output by the above layer into multi-dimensional time series data;

[0022] LSTM layer: Used to extract the features of multi-dimensional time series data processed by the flattening layer;

[0023] Dropout layer: Randomly set the input data elements processed by the LSTM layer to zero with a given probability. The dropout rate of the dropout layer is set to 50%, ensuring the robustness of model training and preventing network overfitting;

[0024] Fully connected layer: Take the data features obtained by the dropout layer as input values, map the distributed feature representations learned by the network to the category number codes, and the output value is a one-dimensional vector that can represent the characteristics of accident types;

[0025] Softmax layer: Use the Softmax function to compress all values of a one-dimensional vector of any size processed by the fully connected layer, and the size of the processed vector remains unchanged; all element values in the Softmax layer are compressed to between [0, 1], and the sum of all elements is 1;

[0026] Classification output layer: Its function is to calculate the cross-entropy loss and perform classification, and output the final prediction result of the model.

[0027] The present invention also includes a system, including a diagnosis module that can run the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network.

[0028] The present invention also includes a device, including:

[0029] Memory: Used to store a computer program for implementing the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network;

[0030] Processor: Used to implement the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network when executing the computer program.

[0031] The present invention also includes a computer-readable storage medium storing a computer program, which when executed by a processor implements the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network.

[0032] Compared with the prior art, the advantages of the present invention are as follows:

[0033] (1) In step S1, multi-dimensional time-series data physical thermal parameters of the sensors and actuators of the small PWR during a fault are obtained, and the changes of these parameters after a fault can be used to identify faults at different positions, types, and severity levels.

[0034] (2) In step S2, in the fault signal conversion method proposed by the present invention, the image size is the only hyperparameter that needs to be adjusted in signal conversion, which greatly reduces the influence of expert bias intervention compared with other conversion methods; in addition, this method can effectively extract the fault signal characteristics of the small PWR, realize the conversion from low-dimensional fault characteristics to high-dimensional fault characteristics, and improve the accuracy of system fault diagnosis.

[0035] (3) In step S3, using a multi-layer CNN can further improve the feature extraction ability and capture higher-level semantic information of the image; the CNN has a certain robustness to changes in the size and position of the image, thus avoiding the influence on the diagnosis result when the image undergoes minor changes; the CNN has the characteristics of parameter sharing and sparse connection structure, greatly reducing the number of parameters in the training network and reducing the computational amount during the training process. Using LSTM helps to maintain the gradient flow and can solve the problem of gradient disappearance that occurs in the traditional RNN network during the training process; LSTM can learn and remember long-term dependence relationships through the gating mechanism, solving the problem that traditional neural networks can hardly capture long sequences, and thus performing excellently in complex sequence prediction tasks including fault diagnosis and speech recognition. Therefore, first extract the spatial correlation of multi-sensor data through CNN, and then learn the propagation law of fault patterns in the time dimension through LSTM to achieve deep fusion of spatio-temporal features, which can further improve the diagnosis accuracy of the fault model.

[0036] In summary, the present invention realizes the accurate diagnosis of faults in the sensors and actuators of the small PWR control system by constructing a hybrid neural network architecture based on CNN-LSTM, providing an important reference for researchers and engineering and technical personnel in related fields. Description of the Drawings

[0037] Figure 1 It is a fault simulation platform diagram based on Simulink.

[0038] Figure 2 It is a logic block diagram of the present invention.

[0039] Figure 3 It is the architecture diagram of the diagnostic model based on the CNN-LSTM network.

[0040] Figure 4 It is the schematic diagram of the structure of LSTM.

[0041] Figure 5 It is the diagram of the setting of the working condition types and the fault data information of the steam pressure sensor and the feed water valve actuator of the small pressurized water reactor.

[0042] Figure 6 It is the model training accuracy curve and the model training cross-entropy loss curve of the fault training set and the test set during the training process of the diagnostic model.

[0043] Figure 7 It is the diagnostic accuracy diagram of the diagnostic model for the fault test set after training.

[0044] Figure 8 It is the result diagram of the diagnostic model diagnosing three types of typical faults. Among them, Figure 8 in (a) is the real-time diagnostic result of the steam pressure sensor having a constant deviation +0.7MPa fault simultaneously at 98s, Figure 8 in (b) is the real-time diagnostic result of the steam pressure sensor having a constant gain 1.2 times fault at 70s, Figure 8 in (c) is the real-time diagnostic result of the steam pressure sensor having a stuck fault at 145s, Figure 8 in (d) is the real-time diagnostic result of the feed water valve having a constant deviation +12% fault at 98s, Figure 8 in (e) is the real-time diagnostic result of the feed water valve having a constant deviation -8% fault at 205s, Figure 8 in (f) is the real-time diagnostic result of the feed water valve having a constant gain 1.3 times fault at 175s. Specific implementation manner

[0045] The present invention will be described in more detail below with reference to the accompanying drawings.

[0046] As Figure 1As shown in the figure, a fault simulation platform for the nuclear steam supply system used is established. This platform is mainly divided into the following modules: reactor module, steam generator module, pressurizer module, feed water system module, control system module, cold leg module of the primary coolant system, hot leg module of the primary coolant system, pressurizer surge flow calculation module, and sensor module. Among them, the fault simulation platform includes more than a dozen sensors such as power sensors and temperature sensors, as well as multiple actuators such as feed water valves. The present invention mainly conducts fault simulation analysis on steam pressure sensors and feed water valves. The two green-labeled modules in the control system part in the figure are the objects of fault diagnosis.

[0047] Referring to Figure 2 , the present invention proposes a fault diagnosis method for small PWRs based on a CNN-LSTM hybrid network. By collecting the main physical and thermal parameter data of small PWR sensors and actuators under different fault types and different severity levels, as well as under normal conditions, such as nuclear power, main steam pressure, etc., and dividing them into a fault training set and a test set by changing the fault time point. Then, by converting these time series signals into images, the data is preprocessed to generate image data that can be input into the model, and further a fault diagnosis model for small PWR sensors and actuators based on a CNN-LSTM hybrid neural network is constructed. The specific steps are as follows:

[0048] S1. Collect multi-dimensional time series data of the steam pressure sensor and the feed water valve of the primary nuclear power control system of a small PWR under different types and different severity levels of fault conditions, as well as under normal conditions, and divide the data into a training set and a test set;

[0049] The multi-dimensional time series data is the main physical and thermal data of a small PWR, including: reactor set power and actual output power, cold leg and hot leg temperatures, steam generator feed water flow rate, steam flow rate, steam pressure, control rod speed, feed water valve opening, feed water valve pressure drop, steam bypass valve opening, and turbine inlet steam flow rate. Use the data recording module in Simulink and the m function program for automatic data collection to collect and save the fault simulation data. These operation data cover information on different positions, times, types, and severity levels to ensure the diversity of the training set and the test set.

[0050] The division of the training set and the test set in step S1 is specifically: using the m function program for automatic data collection, after collecting the training set data, change the time point of the fault occurrence during the operation of the fault simulation platform to obtain the test set. The total number of working conditions of the training set and the test set is the same, both being 334 groups.

[0051] Taking a steam pressure sensor and a feed water valve as examples, different types and degrees of fault conditions include steam pressure / feed water valve sensor faults occurring in various typical steady-state or transient processes at high (100% FP, 90% FP) / medium (50% FP) / low power (40% FP) levels.

[0052] According to the above fault types, multiple groups of symptom data for each type of fault are collected at different fault occurrence times, and are divided into a fault training set and a test set (both are 334 groups of working conditions). The difference between the two is the selected fault time points. Based on this, the established small PWR fault diagnosis model is trained using the fault training set, and the model is verified using the fault test set, and then the diagnostic effect of the nuclear power plant fault diagnosis model is evaluated.

[0053] S2. Preprocess the response data of the physical thermal parameters collected in step S1, which is the preprocessing of converting multi-dimensional time series data signals into image signals. The specific operation is as follows: First, perform denoising and reconstruction operations on the multi-dimensional time series fault signals to obtain the whitened time series to eliminate possible noise interference during the system operation; Second, re-encode the fault time series based on the gray-scale image conversion method, cut the whitened time series into multiple samples and convert them into a base matrix. Finally, use the normalization method to obtain the fault signal values of the matrix and encode them to obtain the RGB pixel values of the red, green, and blue three channels to obtain the image training data.

[0054] S3. Construct a small PWR fault diagnosis model based on the CNN-LSTM network;

[0055] Refer to Figure 3 , the small PWR fault diagnosis model of CNN-LSTM in step S3 includes an image input layer, a convolutional layer, a normalization layer, an average pooling layer, an activation function layer, a flattening layer, an lstm layer, a dropout layer, a fully connected layer, a softmax layer, and a classification output layer of a two-dimensional CNN-LSTM network; in the multi-layer network model, the gradient threshold is set to 1, and the rest of the initial weights are randomly generated and continuously optimized during the training process; among them:

[0056] Image input layer: Receive the image training data after preprocessing of multi-dimensional time series data;

[0057] Convolutional layer: A multi-layer CNN network, whose function is to extract low-level features of the image in the processing of image training data, including edges, lines, and corners, and extract more complex features through multi-layer network iteration;

[0058] The average pooling layer is placed after the convolutional layer. It downsamples the feature map input to the convolutional layer, selects the average value of each pooling window as the output of each pooling window, thereby reducing the spatial size of the feature map while retaining the feature information, and can effectively alleviate the overfitting phenomenon;

[0059] Normalization layer: The core function of this layer is to first normalize the image input data processed by the average pooling layer, and then scale and translate the processed data to make the input distribution of each layer more stable. Among them, the normalization process is to subtract the mean value of all feature values of this layer from each feature value, and then divide by the standard deviation of the feature values of this layer;

[0060] Activation function layer: Activates the neurons in the neural network established above, and then transmits the activation information to the next layer of the network;

[0061] Flattening layer: Used to flatten the image feature map output by the above layer into multi-dimensional time series data;

[0062] LSTM layer: Used to extract the features of the multi-dimensional time series data processed by the flattening layer;

[0063] The hidden layer of the LSTM network layer includes 200 neurons. The LSTM hidden layer includes multiple neurons. If the number of neurons is too large (such as taking 256, 512 or even more), it may lead to overfitting of the model and the verification accuracy of the model may instead decrease, thereby reducing its generalization ability and increasing the training time. For example, when the value is 256, the verification accuracy is 98.28%, and the training time is 36 minutes and 44 seconds; when it is 512, the verification accuracy is 97.97%, and the training time is 47 minutes and 6 seconds; while if the number of neurons is too small (such as taking 128, 64 or even less), it may lead to underfitting, thus affecting the prediction accuracy. For example, when the value is 64, the verification accuracy is 98.22%; when it is 128, the verification accuracy is 98.04%. Therefore, an appropriate number of neurons is selected through experience to balance the generalization ability and prediction accuracy of the model, so as to achieve the best fault diagnosis effect;

[0064] Furthermore, setting the number of neurons in the LSTM network layer to 200 (the diagnostic effect is the best under this model at this value) can effectively extract the features of the fault data, thereby significantly improving the accuracy of the prediction model. The activation function layer introduces non-linear factors to enhance the expression ability of the model; the normalization layer accelerates the network training process; the dropout layer improves the robustness of the model by randomly discarding neuron information. The fully connected layer effectively maps the extracted features to the fault types, thus realizing the efficient diagnosis of multiple fault types.

[0065] Dropout layer: Randomly set the input data elements processed by the LSTM layer to zero with a given probability. The dropout rate of the dropout layer is set to 50%, ensuring the robustness of model training and preventing network overfitting;

[0066] Fully connected layer: Take the data features obtained from the dropout layer as input values, map the distributed feature representations learned by the network to the category number codes, and the output value is a one-dimensional vector that can represent the accident type features;

[0067] Softmax layer: Use the Softmax function to compress all the values of a one-dimensional vector of any size processed by the fully connected layer, and the size of the processed vector remains unchanged; all the values of the elements in the Softmax layer are compressed to between [0, 1], and the sum of all the elements is 1;

[0068] Classification output layer: Its function is to calculate the cross-entropy loss and perform classification, and output the final prediction result of the model.

[0069] Furthermore, in step S3, the actual input of the LSTM cell at time t includes the state h at time t - 1 t-1 and the current input x t . Through 4 fully connected neurons f t , g t , i t and o t , 3 gates are used to complete the function of memorizing or forgetting information. Among them, the forget gate determines how much of the previous information will be passed forward, the input gate controls the aspect of the new input information, and the output gate determines what will be output at this time step. In terms of output, h t is then sent as input to the next time step and can be considered a short-term state, while c t determines the longer-term dependencies. The entire calculation process is as follows:

[0070] f t = σ(W (f) x t + U (f) h t-1 )

[0071] i t = σ(W (i) x t + U (i) h t-1 )

[0072] g t = tanh(W (c) x t + U (c) h t-1 )

[0073] ot = σ(W (c) x t + U (c) h t-1 )

[0074] h t = o t tanh(c t )

[0075] Where: g t is the temporary memory cell; c t is the new memory cell; i t is the input gate; f t is the forget gate; o t is the output gate; W and U are weight matrices.

[0076] S4. Use the training set obtained in step S2 to train the small PWR fault diagnosis model based on CNN-LSTM constructed in step S3. Input the test set data obtained in step S2 into the trained small PWR fault diagnosis model based on CNN-LSTM to realize the diagnosis of faults at different positions, types, and severities in the PWR.

[0077] Furthermore, during the model training process in step S4, input the fault training set into the small PWR sensor and actuator fault diagnosis model based on the CNN-LSTM network. The deep neural network continuously extracts features from a large number of samples and optimizes the network weight parameters to achieve an ideal classification result.

[0078] Furthermore, in step S4, adopt the Adam optimization algorithm, set the maximum number of iterations in the training process to 3960 times (i.e., 20 training epochs) to ensure the convergence of the loss function. Regarding the setting of the learning rate, adopt a variable learning rate strategy. Because too large a learning rate may lead to overfitting, while too small a learning rate will reduce the network convergence speed and affect the training efficiency. Specifically, the initial learning rate is set to 0.001, and every 792 iterations (i.e., 4 training epochs), the learning rate is adjusted to 50% of the original value. This adjustment strategy ensures that the model can find the optimal weights faster and more smoothly during the backpropagation process, thus avoiding violent oscillations in the loss function.

[0079] Based on Figure 1 the simulation platform as the research object, referring to Figure 5, the sample data collected by the simulation platform include different types and degrees of faults that occurred in the steam pressure sensor and the feed water valve at 50:15:155 s (training set) and 70:15:250 s (test set), involving a total of 27 fault types. The specific faults include constant deviation faults (fault degrees: ±0.3 MPa, ±0.5 MPa, ±0.7 MPa), constant gain faults (fault degrees: 0.7, 0.8, 0.9, 1.1, 1.2, 1.3), and stuck faults of the steam pressure sensor; constant deviation faults (fault degrees: ±12%, ±8%, ±4%), constant gain faults (fault degrees: 0.7, 0.8, 0.9, 1.1, 1.2, 1.3), and stuck faults of the feed water valve, as well as normal operating conditions. According to these fault types, by setting different fault time periods and collecting data on the simulation platform, the training set and the test set are obtained respectively. The training set is used to train the fault diagnosis model, and the test set is used to evaluate the training effect of the model.

[0080] See Figure 6 , Figure (a) shows the trend of the model diagnosis accuracy rate changing with the number of iterations, and Figure (b) shows the change of the cross-entropy loss with the number of iterations. The smaller the cross-entropy loss, the smaller the model error, the smaller the gap between the predicted value and the actual value, and thus the better the classification effect. It can be seen from the figure that as the number of iterations increases, the loss values of the model on the training set and the test set gradually decrease, and the diagnosis accuracy rate continuously improves. Finally, the diagnosis accuracy rate corresponding to the minimum cross-entropy loss is selected and used as the final accuracy rate. The final accuracy rate of the validation set is 98.38%.

[0081] See Figure 7 , the horizontal axis represents the serial number of the validation set working conditions, and the vertical axis is the proportion of the time points accurately diagnosed in the entire sequence length for each validation set working condition, that is, the diagnosis accuracy rate corresponding to each working condition. It can be seen from the figure that the diagnosis accuracy rates of almost all fault types exceed 80%, and the diagnosis accuracy rates of most fault types exceed 95%, indicating that the trained network model has high diagnosis accuracy.

[0082] See Figure 8 , the horizontal axis is time, and the vertical axis is the fault type. The two curves respectively represent the actual fault type and the predicted fault type obtained by diagnosing through the diagnostic network model. When the two curves are closer or coincide, it indicates that the accuracy of the network diagnostic model is higher. The specific analysis is as follows:

[0083] Figure 8 (a) and (b) in are the fault diagnosis results during the steady-state operation of a small pressurized water reactor at 100% FP (Full Power). Among them, Figure 8The fault diagnosis rate of the 2SGCD+0.7 type shown in (a) is 99.93%, because it was misdiagnosed as 2SGCG+1.2 at the moment of 99.0 s; Figure 8 The fault diagnosis rate of the 2SGCG+1.2 type shown in (b) is 99.67%. It was misdiagnosed as 2SGCD+0.7 at some moments during the short period from 67.9 s to 69.6 s. The reason may be that the operation data of the two fault types are the same or similar during the operation in this period.

[0084] Figure 8 In (c), it is the diagnosis result of the 2SGST+0.0 type fault that occurs during the step load reduction operation of the small PWR from 100% FP to 90% FP. It can be seen from the figure that this test condition was misdiagnosed as a jamming fault 6.3 s before the fault occurred. The reason may be that the opening of the feed water valve no longer changes before the jamming fault occurs under this condition, thus causing misdiagnosis. Figure 8 In (d), it is the fault diagnosis result diagram of the 2FVCD+12 fault that occurs during the step load change operation of the small PWR from 50% FP to 40% FP. The fault diagnosis rate is 99.90%.

[0085] Figure 8 In (e) and (f), they are the diagnosis result diagrams during the step load increase operation of the small PWR from 50% FP to 60% FP. Among them Figure 8 In (e), it shows the fault diagnosis result of 2FVCD-8. The fault diagnosis rate is 99.90%. The reason is that it was misdiagnosed as 2FVST+0.0 within the time from 202.9 s to 203.0 s, and then the correct fault type was diagnosed. Figure 8 In (f), it shows the fault diagnosis result of 2FVCG+1.3. The fault diagnosis rate is 99.90%.

[0086] To sum up, the small PWR fault diagnosis method and system based on CNN-LSTM provided by the present invention generate multi-dimensional time series data of the training set and the test set through the Simulink simulation platform, and construct a small PWR fault diagnosis model based on CNN-LSTM. By training the sample data, this method enables the model to effectively learn and extract the operation characteristics of the fault conditions, and realizes the accurate diagnosis of different types of faults. The experimental results verify that this method can accurately diagnose the faults of the sensors and actuators of the small PWR under different types and degrees in the three types of working conditions of steady-state operation, step load reduction, and step load increase.

[0087] The present invention also includes a system, including a diagnosis module, which can run the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network.

[0088] The present invention further includes a device, comprising:

[0089] A memory: for storing a computer program for implementing the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network;

[0090] A processor: for implementing the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network when executing the computer program.

[0091] The present invention further includes a computer-readable storage medium, which stores a computer program, and the computer program implements the above-mentioned small PWR fault diagnosis method based on a CNN-LSTM hybrid network when executed by a processor.

[0092] The above content is only for explaining the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention falls within the protection scope of the claims of the present invention.

Claims

1. A fault diagnosis method for a small pressurized water reactor based on a CNN-LSTM hybrid network, characterized in that It includes the following steps: S1. Collect multi-dimensional time series data of the steam pressure sensor and the feed water valve of the primary loop nuclear power control system of a small PWR under different types and severities of fault conditions, as well as under normal conditions, and divide the data into a training set and a test set; S2. Preprocess the multi-dimensional time series data of the training set and the test set obtained in step S1; S3. Construct a fault diagnosis model for a small PWR based on CNN-LSTM; S4. Use the training set obtained in step S2 to train the fault diagnosis model for a small PWR based on CNN-LSTM constructed in step S3, and input the test set data obtained in step S2 into the trained fault diagnosis model for a small PWR based on CNN-LSTM for verification, so as to realize the diagnosis of faults at different positions, types and severities of the PWR.

2. A small PWR fault diagnosis method based on a CNN-LSTM hybrid network according to claim 1, characterized in that In step S1, the multi-dimensional time series data are the physical thermal data of the small PWR, including: the reactor set power and the actual output power, the cold leg and hot leg temperatures, the feed water flow rate of the steam generator, the steam flow rate, the steam pressure, the control rod speed, the feed water valve opening, the feed water valve pressure drop, the steam bypass valve opening, and the steam inlet flow rate of the steam turbine.

3. A small PWR fault diagnosis method based on a CNN-LSTM hybrid network according to claim 1, characterized in that The division of the training set and the test set in step S1 is specifically as follows: after using the m function program for automatic data collection to collect the training set data, change the time when a fault occurs during the operation of the fault simulation platform to obtain the test set, and the total number of working conditions of the training set and the test set is the same.

4. A fault diagnosis method for a small pressurized water reactor based on a CNN-LSTM hybrid network according to claim 1, characterized in that, The preprocessing in step S2 is specifically operated as follows: First, perform denoising and reconstruction operations on the multi-dimensional time series fault signals to obtain the whitened time series; Second, re-encode the fault time series based on the gray image conversion method, cut the whitened time series into multiple samples and convert them into a base matrix. Finally, use the normalization method to obtain the fault signal values of the matrix and encode them into RGB pixel values of three channels of red, green and blue to obtain the image training data.

5. A fault diagnosis method for a small pressurized water reactor based on a CNN-LSTM hybrid network according to claim 1, characterized in that, The fault diagnosis model for a small PWR based on CNN-LSTM in step S3 includes an image input layer, a convolutional layer, a normalization layer, an average pooling layer, an activation function layer, a flattening layer, an lstm layer, a dropout layer, a fully connected layer, a softmax layer and a classification output layer of a two-dimensional CNN-LSTM network; in the multi-layer network model, the gradient threshold is set to 1, and the rest of the initial weights are randomly generated and continuously optimized during the training process; among them: Image input layer: Receive the image training data after preprocessing the multi-dimensional time series data; Convolutional layer: A multi-layer CNN network, whose function is to extract low-level features of the image in the processing of the image training data, including edges, lines, corners, and extract more complex features through multi-layer network iteration; Average pooling layer, placed after the convolutional layer, downsamples the feature map input by the convolutional layer, selects the average value of each pooling window as the output of each pooling window, thereby reducing the spatial size of the feature map while retaining the feature information, which can effectively alleviate the overfitting phenomenon; Normalization layer: The core function of this layer is to first normalize the image input data processed by the average pooling layer, and then scale and translate the processed data to make the input distribution of each layer more stable. Among them, the normalization process is to subtract the mean of all eigenvalues of this layer from each eigenvalue, and then divide by the standard deviation of the eigenvalues of this layer; Activation function layer: Activate the neurons in the neural network established above, and then transfer the activation information to the next layer of the network; Flattening layer: Used to flatten the image feature map output by the above layer into multi-dimensional time series data; LSTM layer: Used to extract the features of the multi-dimensional time series data processed by the flattening layer. The LSTM layer includes 200 hidden neurons; Dropout layer: Randomly set the input data elements processed by the LSTM layer to zero with a given probability. The dropout rate of the dropout layer is set to 50%, ensuring the robustness of model training and preventing network overfitting; Fully connected layer: Take the data features obtained by the dropout layer as input values, map the distributed feature representations learned by the network to the category number codes, and the output value is a one-dimensional vector that can represent the features of the accident type; Softmax layer: Use the Softmax function to compress all values of an arbitrary-sized one-dimensional vector processed by the fully connected layer. The size of the processed vector remains unchanged; all element values in the Softmax layer are compressed to between [0, 1], and the sum of all elements is 1; Classification output layer: Its function is to calculate the cross-entropy loss and perform classification, and output the final prediction result of the model.

6. A system, characterized in that, It includes a diagnosis module that can run a small PWR fault diagnosis method based on the CNN-LSTM hybrid network described in any one of claims 1-5.

7. A device, characterized in that, It includes: Memory: Used to store a computer program for implementing a small PWR fault diagnosis method based on the CNN-LSTM hybrid network described in any one of claims 1-5; Processor: Used to implement a small PWR fault diagnosis method based on the CNN-LSTM hybrid network described in any one of claims 1-5 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements a small PWR fault diagnosis method based on the CNN-LSTM hybrid network described in any one of claims 1-5.

Images