Fault diagnosis method and system for nuclear power circulating water pump based on optimized capsule network

Abstract

The present invention relates to a nuclear power circulating water pump fault diagnosis method and system based on an optimized capsule network, the method comprising: obtaining vibration sensor data of the nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set; preprocessing the first data set to obtain a second data set; extracting features from the second data set to obtain a feature matrix; reconstructing the feature matrix in phase space to obtain training data; constructing a time convolution capsule network; training the time convolution capsule network using the training data to obtain a trained time convolution capsule network; obtaining vibration sensor data of a nuclear power circulating water pump to be detected; and using the trained time convolution capsule network to determine the fault result of the nuclear power circulating water pump to be detected. The present invention can improve the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method and system for diagnosing a nuclear power circulating water pump fault based on an optimized capsule network. Background Art

[0002] The nuclear power system has a complex structure and potential risk of radioactive release, and has extremely high safety requirements. Therefore, the reliability requirements for key equipment in the nuclear power system are very high; at the same time, with the demand for offshore drilling platforms and island power generation, it is impossible to deploy a large number of operating personnel on the relevant platforms, so the automation and intelligence level of nuclear power plant operation is very high, and there is a strong demand for less-manned and unmanned operation. The operating environment of the nuclear power system is harsh, and key equipment is very prone to failure in long-term continuous operation. If a failure occurs and cannot be discovered and repaired in time, it may lead to serious radioactive consequences and endanger the lives of operators and the public.

[0003] The circulating water pump is an indispensable equipment for the long-term operation of nuclear power plants. It can provide driving pressure for the working fluid to remove the heat energy of the reactor and its auxiliary equipment. However, the failure of the bearing in the pump can only be discovered and handled during regular disassembly and maintenance. Therefore, once the bearing fails during operation, it will seriously threaten the safety and economy of the nuclear power plant. Therefore, it is necessary to carry out research on the fault diagnosis method of the circulating water pump bearing. With the continuous development of artificial intelligence technology and big data theory, the accumulation of a large amount of operation data of nuclear power systems and application experience in other fields, the use of some efficient and accurate artificial intelligence technologies to quickly and accurately diagnose faults can effectively improve the operation and maintenance guarantee capabilities of key equipment in nuclear power systems, assist operation and maintenance personnel in decision-making analysis, and improve operation safety and economy.

[0004] In 1967, the U.S. Naval Research Laboratory established a Mechanical Failure Prevention Group, and began research on fault diagnosis technology. Subsequently, the research and application of fault diagnosis technology gradually spread around the world. In the late 1960s, the establishment of the British Machinery Care and Condition Monitoring Association further promoted the development of fault diagnosis technology. Subsequently, European countries also carried out related research on condition monitoring and fault diagnosis technology, and formed their own distinctive diagnostic technology systems. Japan's fault diagnosis technology started in the mid-1970s. By learning from the research of countries around the world and continuously improving and improving, Japan's fault diagnosis technology in civil industries such as steel production, railway operation, and chemical processes is now very mature. China's research on fault diagnosis technology started in the early 1980s, and a relatively complete theoretical system has been formed. In the field of nuclear power, typical research results include the operation decision support system for operators developed by the Argonne National Laboratory in the United States; the operation status monitoring and diagnosis system developed by the European Union Halden Reactor Project; the Korea Advanced Institute of Science and Technology developed an accident diagnosis and consulting system for nuclear power plant fault diagnosis; Tsinghua University researched and developed a 200MW nuclear heating station fault diagnosis system; Harbin Engineering University designed and developed a nuclear power plant operation support system, which includes functions such as status monitoring, alarm analysis, fault diagnosis, and emergency operation guidance.

[0005] In terms of fault diagnosis methods, they can be divided into three categories: methods based on quantitative analytical models, methods based on qualitative empirical knowledge, and methods based on historical data. In terms of fault diagnosis methods based on quantitative analytical models, in order to solve the problem of nonlinear system failures, Wiinnenberg first proposed a fault diagnosis method for nonlinear unknown observers. For nonlinear discrete systems, Julier et al. proposed a fault diagnosis method based on filters, and added a random distribution of sigma points under the characterization input, which further improved the fault diagnosis accuracy of nonlinear filters. The equivalent space method was first proposed by Chow and Willsky in 1984; in 1997, Isermann and Balle made a detailed review of the analytical model-based method and the equivalent space method therein. In the 1990s, the US Air Force used the equivalent space method to realize fault detection and separation of aircraft control systems. However, for nonlinear systems, it is difficult to establish an accurate mathematical model, so the application of this type of method is severely limited.

[0006] In the research of fault diagnosis based on qualitative empirical knowledge, they do not need to establish a systematic analytical model, and the diagnosis results are easy to understand and have good robustness; however, it is difficult to obtain expert knowledge; when there are many rules, there are problems such as matching conflicts and combinatorial explosion in the reasoning process. As early as 1980, expert systems were applied to fault diagnosis. This was the first time that humans transformed the experience learned in the past into an evaluation system for fault diagnosis. Pang et al. proposed a distributed expert system that can distribute the functions of the expert system to multiple processors to work in parallel, thereby improving the processing efficiency of the system. BO et al. proposed an object-oriented knowledge representation method to address the dual problems of low versatility and low scalability of various expert systems currently available, so that the fault rules of specific machines can be solved with general rules. Due to the limited number of measurement points, the acquired fault phenomenon will show fuzziness. The introduction of fuzzy fault method is conducive to solving the problems of imprecise information, uncertainty and noise encountered in detection and diagnosis. Liu et al. proposed to combine fuzzy measurement with fuzzy integral to analyze mechanical fault data, which performed well in bearing and motor fault diagnosis.

[0007] In terms of fault diagnosis based on historical data, its advantage over the above two methods is that it does not require the establishment of an accurate analytical model of the core, and can directly process data or signals. Therefore, this type of method is widely used in both linear and nonlinear systems. The methods based on historical data mainly include fault diagnosis methods based on multivariate statistical methods, signal analysis, and artificial intelligence and pattern recognition:

[0008] (1) Multivariate statistical methods such as principal component analysis (PCA), kernel principal component analysis, and independent component analysis developed rapidly at the end of the last century. Misra et al. proposed the application of PCA and its improved methods in actual industrial process fault detection. The proposed improved method MSPCA significantly reduced the false alarm rate compared with the traditional PCA-based method; however, this type of method is mainly used for fault detection and has poor results in identifying and classifying fault causes.

[0009] (2) Fault diagnosis methods based on signal analysis only began to emerge in the 1980s. Such methods mainly include wavelet transform, Hilbert-Huang transform, S transform, etc. Wavelet transform is currently the most commonly used method for processing signals and is also the most reliable method. Leung et al. reviewed the application of wavelet transform in chemical analysis, which is used for noise elimination and data compression in different fields of analytical chemistry. In actual industrial processes, the signals obtained generally contain various forms of noise. Signal analysis-based methods can be used to decompose useful signals containing noise to achieve the purpose of distinguishing useful signals from noise. Therefore, signal analysis-based methods are mainly used in data denoising, preprocessing, etc. Since the signal analysis method itself does not have the ability of pattern recognition and classification, signal analysis-based methods are often used in combination with pattern recognition methods.

[0010] (3) Methods based on artificial intelligence and pattern recognition. As early as 1988, some scholars applied neural networks to the fault diagnosis of rotating machinery. The main types of neural networks currently used for fault detection and diagnosis are: adaptive networks, radial basis function networks (RBF networks), back propagation algorithms (BP networks), etc. Venkata subramanian et al. first proposed the application of BP networks in process fault diagnosis. Gome et al. used Gaussian radial basis function neural networks to analyze pressurized water reactor power plant accidents. Sinuhe used a strategy based on artificial neural networks to detect core component blockage failures in sodium-cooled fast reactors. A "jump" type multi-layer neural network is proposed, using two neural networks to dynamically identify and verify the results of identification. In addition to shallow neural networks, many scholars have also used various models such as logistic regression, support vector machine, and decision tree to study fault diagnosis technology. However, these machine learning methods need to combine artificial experience to select feature parameters, the network training stability is poor, and the accuracy cannot be further improved, so it is difficult to adapt to the needs of intelligent fault diagnosis. With the rapid development of artificial intelligence technology, deep learning research has achieved great success in image recognition, speech recognition, natural language, language translation and other fields. At present, fault identification and diagnosis research based on deep learning algorithms is generally still in the initial exploration stage. Tamilselvan et al. proposed a multi-sensor health diagnosis method based on deep belief network. Lu Chunyan et al. realized the effective diagnosis of refining air compressor faults based on deep belief network, and the results also showed that the diagnostic accuracy and stability of this method are better than those of traditional shallow neural networks.

[0011] The present invention selects artificial intelligence and pattern recognition technology from the method based on historical data to realize the intelligent fault diagnosis method. Since the deep learning method can avoid the manual selection of feature parameters and the stability and accuracy of the diagnosis results are better, the present invention adopts deep learning technology for intelligent fault diagnosis. The temporal convolutional network is a special deep neural network. Its working principle is to construct multiple filters to perform feature extraction on the input samples layer by layer through convolution and pooling calculation, and to mine the hidden information in the data layer by layer.

[0012] Each neuron in the capsule network is a vector, rather than a traditional scalar, which enables the capsule network to extract more detailed features from the input data and reduce the loss of feature information; the capsule network updates the capsule layer parameters through a dynamic routing mechanism, further increasing the coupling coefficient between the child node and the parent node, and making full use of local information to enrich the feature representation ability and information inclusion; the capsule network structure itself has translation invariance, which can extract the relative position relationship of the input features and improve the accuracy of nuclear power equipment fault diagnosis. Therefore, the capsule network is more suitable for processing highly nonlinear data, and the data of nuclear power circulating water pumps meets these characteristics.

[0013] However, the amount of computation required for dynamic routing iteration of the capsule network is relatively large. In addition, the deep learning method uses a deep structure that is several times deeper than the traditional shallow machine learning model, which is far less efficient than the shallow model and has relatively high hardware requirements. Therefore, the present invention proposes to use support vectors to optimize the dynamic routing of the capsule network to effectively reduce the computational difficulty and improve the efficiency of fault diagnosis. Finally, the temporal convolutional capsule network proposed in the present invention can accurately diagnose the circulating water pump of the nuclear power plant, has good stability and versatility, can improve the accuracy of fault diagnosis, and ultimately provide analysis and reference basis for operators, thereby improving the safety and reliability of the nuclear power plant. Summary of the invention

[0014] The purpose of the present invention is to provide a nuclear power circulating water pump fault diagnosis method and system based on optimized capsule network, which can improve the accuracy of fault diagnosis.

[0015] To achieve the above object, the present invention provides the following solutions:

[0016] A nuclear power circulating water pump fault diagnosis method based on optimized capsule network, comprising:

[0017] Acquire vibration sensor data of a nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set;

[0018] Preprocessing the first data set to obtain a second data set;

[0019] Performing feature extraction on the second data set to obtain a feature matrix;

[0020] Reconstructing the characteristic matrix in phase space to obtain training data;

[0021] Construct a temporal convolutional capsule network;

[0022] Using the training data to train the temporal convolutional capsule network to obtain a trained temporal convolutional capsule network;

[0023] Obtain vibration sensor data of the nuclear power circulating water pump to be tested;

[0024] The trained temporal convolution capsule network is used to determine the fault result of the nuclear power circulating water pump to be detected.

[0025] Optionally, preprocessing the first data set to obtain the second data set includes:

[0026] The first data set is subjected to noise reduction processing by wavelet packet transform, and the low-frequency part and the high-frequency part of the first data set are orthogonally decomposed at the same time to obtain the second data set.

[0027] Optionally, after the step of “obtaining vibration sensor data of the nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set” and before the step of “preprocessing the first data set to obtain a second data set”, the step further includes:

[0028] Different labels are set for the first data set according to the degree of failure.

[0029] Optionally, the constructing a temporal convolutional capsule network includes:

[0030] Construct a convolutional neural network to extract the nonlinear features of the detection data;

[0031] Constructing a temporal convolution kernel at the output end of the convolutional neural network to extract deep temporal features;

[0032] Constructing a capsule network at the output end of the temporal convolution kernel to extract vector features;

[0033] A dynamic routing algorithm is set in the capsule network to iterate and update the vector features.

[0034] Optionally, the temporal convolutional capsule network adopts a cross entropy loss function as a loss function.

[0035] Optionally, the temporal convolutional capsule network is trained using an SGD optimization algorithm.

[0036] Optionally, after the step of “obtaining vibration sensor data of the nuclear power circulating water pump to be detected” and before the step of “using the trained temporal convolution capsule network to determine the fault result of the nuclear power circulating water pump to be detected”, it also includes:

[0037] The vibration sensing data of the nuclear power circulating water pump to be detected is preprocessed.

[0038] Optionally, the preprocessing of the vibration sensing data of the nuclear power circulating water pump to be detected includes:

[0039] The vibration sensor data of the nuclear power circulating water pump to be detected is subjected to noise reduction processing by wavelet packet transformation, and the low-frequency part and the high-frequency part of the vibration sensor data of the nuclear power circulating water pump to be detected are orthogonally decomposed at the same time.

[0040] Optionally, after the step of “constructing a time convolution capsule network” and before the step of “training the time convolution capsule network using the training data to obtain a trained time convolution capsule network”, the method further includes:

[0041] Set hyperparameters for the temporal convolutional capsule network.

[0042] A nuclear power circulating water pump fault diagnosis system based on optimized capsule network, comprising:

[0043] A first data acquisition module is used to acquire vibration sensor data of the nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set;

[0044] A preprocessing module, used for preprocessing the first data set to obtain a second data set;

[0045] A feature extraction module, used to extract features from the second data set to obtain a feature matrix;

[0046] A phase space reconstruction module is used to perform phase space reconstruction on the feature matrix to obtain training data;

[0047] Network building module for building temporal convolutional capsule networks;

[0048] A training module, used to train the temporal convolutional capsule network using the training data to obtain a trained temporal convolutional capsule network;

[0049] A second data acquisition module is used to acquire vibration sensor data of the nuclear power circulating water pump to be tested;

[0050] The detection module is used to use the trained time convolution capsule network to determine the fault result of the nuclear power circulating water pump to be detected.

[0051] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0052] The present invention selects artificial intelligence and pattern recognition technology from the method based on historical data to realize the intelligent fault diagnosis method. Since the deep learning method can avoid the manual selection of feature parameters and the stability and accuracy of the diagnosis results are better, the present invention adopts deep learning technology for intelligent fault diagnosis. The temporal convolutional network is a special deep neural network. Its working principle is to construct multiple filters to perform feature extraction on the input samples layer by layer through convolution and pooling calculation, and to mine the hidden information in the data layer by layer.

[0053] Each neuron in the capsule network is a vector, rather than a traditional scalar, which enables the capsule network to extract more detailed features from the input data and reduce the loss of feature information; the capsule network updates the capsule layer parameters through a dynamic routing mechanism, further increasing the coupling coefficient between the child node and the parent node, and making full use of local information to enrich the feature representation ability and information inclusion; the capsule network structure itself has translation invariance, which can extract the relative position relationship of the input features and improve the accuracy of nuclear power equipment fault diagnosis. Therefore, the capsule network is more suitable for processing highly nonlinear data, and the data of nuclear power circulating water pumps meets these characteristics.

[0054] However, the amount of computation required for dynamic routing iteration of the capsule network is relatively large. In addition, the deep learning method uses a deep structure that is several times deeper than the traditional shallow machine learning model, which is far less efficient than the shallow model and has relatively high hardware requirements. Therefore, the present invention proposes to use support vectors to optimize the dynamic routing of the capsule network to effectively reduce the computational difficulty and improve the efficiency of fault diagnosis. Finally, the temporal convolutional capsule network proposed in the present invention can accurately diagnose the circulating water pump of the nuclear power plant, has good stability and versatility, can improve the accuracy of fault diagnosis, and ultimately provide analysis and reference basis for operators, thereby improving the safety and reliability of the nuclear power plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0056] Figure 1 The present invention is a flow chart of a method for diagnosing a fault of a nuclear power circulating water pump based on an optimized capsule network;

[0057] Figure 2This is a module diagram of the nuclear power circulating water pump fault diagnosis system based on the optimized capsule network of the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] The purpose of the present invention is to provide a nuclear power circulating water pump fault diagnosis method and system based on optimized capsule network, which can improve the accuracy of fault diagnosis.

[0060] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1 The figure shows a flow chart of a method for diagnosing a fault of a nuclear power circulating water pump based on an optimized capsule network of the present invention. The basic steps of the present invention include:

[0062] Step 1: Collect and store the vibration sensor data of the nuclear power circulating water pump during operation and the vibration sensor data of the nuclear power circulating water pump under various faults as the first data set.

[0063] Step 2: The sensor data collected in the first data set are managed in different categories in the computer according to the subsystem to which the sensor belongs. At the same time, different labels can be set for the normal state, fault state and different fault degrees of the sensor data to facilitate subsequent training of the method described in the present invention.

[0064] Step 3: All the data in step 2 are processed by wavelet packet transform to reduce the noise of the collected raw data, and the low-frequency and high-frequency parts of the data are orthogonally decomposed to improve its time domain resolution, and then the wavelet packet is reconstructed. The formula is The wavelet packet of the signal is reconstructed by using the low-frequency coefficients after orthogonal decomposition of the wavelet packet and the high-frequency coefficients obtained after threshold quantization.

[0065] Among them, k is the transformation parameter, h l-2k is the low-pass filter coefficient, g l-2k is the high-pass filter coefficient, and is the decomposition coefficient obtained by orthogonal decomposition of the original signal.

[0066] Step 4: Perform feature extraction on the data after denoising in step 3, reflect the time series mutation of signal components through permutation entropy, and reflect the degree of order of data through envelope entropy, adjust and align according to the dimension of the obtained data, and then combine the two to form a hybrid enhanced feature matrix.

[0067] Step 5: Since the input data of the capsule network is at least three-dimensional data, the first dimension represents the total amount of data, the second dimension represents the length of a single data, and the third dimension represents the width of a single data; and the data in step 4 is a two-dimensional array, the first dimension represents the total amount of data, and the second dimension represents the dimension of the feature. In order to enable the data of the nuclear power circulating water pump to be input into the convolutional neural network for effective fault diagnosis, the present invention performs phase space reconstruction on the data in step 4, wherein the interval time is set to 1s, the sliding window length is set to 20s, and finally the two-dimensional data (N×D dimension) in step 4 is converted into a three-dimensional stacked data block of (N-num_steps+1)×(num_steps×D), wherein N is the total amount of data, D is the dimension of the feature parameter, num_steps is the length of the sliding window, and since there is overlap between the data during each sliding process, the total data input length for the algorithm of the present invention is (N-num_steps+1).

[0068] Step 6: The present invention first establishes a convolutional neural network layer to preliminarily extract the nonlinear characteristics of the measurement data, which is composed of an input layer (pre-processed training data) and a pooling layer.

[0069] The convolution layer uses the formula described in equation (1) to extract features. After the convolution operation, the feature map needs to be fed forward to the pooling layer through the activation function, where k is the convolution kernel, b is the bias parameter, and x is j is the output of the convolutional layer, y j is the output of the convolutional layer, M j The feature map formed for the data obtained in step 5.

[0070] The present invention adopts the Leaky ReLU activation function, which can avoid dead nodes on the basis of the ReLU activation function and better reflect the nonlinear characteristics in the data; the calculation of the pooling layer is calculated using equation (2), where x j is the output of the pooling layer, y j is the input of the pooling layer, down is the pooling function, β is the network multiplicative bias of the pooling layer l, and b is the bias; the present invention adopts the maximum pooling calculation, and the pooling operation can downsample the training data to prevent the occurrence of model overfitting.

[0071]

[0072] x j =βj down(y j )+b j (2)

[0073] Step 7: Use the time convolution kernel to extract deep temporal features, expand the receptive field by dilation convolution, and input the sequence X{x 1 ,x 2 ,…,x n}, where x 1 ~x n is the quantity in the sequence X, and the dilated convolution is:

[0074]

[0075] Where d represents the dilation factor, k represents the filter size, sd*i represents the past direction, f(i) represents the convolution kernel function, and F(s) is the output of the dilated convolution. And residual convolution is used to avoid the gradient vanishing problem:

[0076] o=Activation(X+F(X)) (4)

[0077] Where X is the result of the dilated convolution in formula (3), which is used as the input of the residual convolution, o is the output, F(X) represents the output of the last hidden layer of the residual network, and Activation is the activation function. The present invention adopts a two-layer temporal convolution structure.

[0078] Step 8: Input feature information into the capsule network to extract vector features and reduce the loss of key feature information. Multiply the features output by the temporal convolutional layer by the weight matrix to obtain the prediction vector:

[0079] U j,i =U i W j,i (5)

[0080] Where W j,i is the weight of the main capsule layer, U i is the feature output of the temporal convolutional layer, U j,i Represents the vector generated by the input feature prediction. The prediction vector is transmitted to the digital capsule layer:

[0081]

[0082] S j =∑ i U j,i C i,j (7)

[0083]

[0084] b i,j =bi,j +V j U j,i (9)

[0085] C in the formula i,j and b i,j represents the coupling and bias coefficient, S j is the total input vector.

[0086] Step 9: In order to solve the problem that the dynamic routing iteration calculation of equation (6) to equation (9) in the capsule network in step 8 is large, the present invention proposes a dynamic routing algorithm optimized by support vector for the first time. j,i After inputting into the support vector machine for training, a set of support vectors sv can be obtained. 1 ,sv 2 ,…,sv Q With the Lagrangian factor a 1 ,a 2 , …, a N is the Lagrangian vector a of the element 1×N The original training samples are reconstructed by the feature extraction formula shown in formula (10):

[0087]

[0088] Where sv i is the support vector, a i is the Lagrangian factor corresponding to the support vector, is the support vector corresponding to the label, and b is the bias.

[0089] The number of reconstructed samples remains the same as the initial training sample set, which is N, and the dimension of each sample changes from the initial sample dimension M to Q. i ∈U j,i , formula (10) for the input x i After processing, a new set of reconstructed samples can be obtained as shown in formula (11):

[0090]

[0091] Then the reconstructed h i As the new U j,i Substitute it into the dynamic routing for update iteration.

[0092] Step 10: Sort out the hyperparameters in the temporal convolutional capsule network; Steps 5, 6, 7, and 8 involve a large number of hyperparameters in setting the structure of the temporal convolutional capsule network, including the number of intermediate hidden layers of the convolutional neural network, the convolution kernel size of the convolution layer, the step size of the convolution process, and the number of feature maps; the convolution kernel size of the temporal convolution kernel, the step size and sparsity rate of the convolution process; the number of input capsules, input vector dimension, number of output capsules, output vector dimension, and number of routing iterations of the capsule network; the penalty coefficient and activation function of the support vector machine, etc.

[0093] Step 11: Define loss function and parameter optimization; the present invention adopts cross entropy loss function as loss function. In order to optimize the weights and biases in the above-mentioned temporal convolution capsule network, the SGD optimization algorithm is used to solve the network during the training process so that the loss function value is as small as possible.

[0094] At the same time, during the training process of the temporal convolution capsule network, all data are split into multiple batches of training samples, each batch has 32 groups of data, and the processed data are randomly shuffled to reduce uncertainty and prevent overfitting. As the number of training rounds increases, the training error gradually decreases, indicating that the temporal convolution capsule network model can continuously approach the parameter change characteristics under actual faults.

[0095] Step 12: During the actual fault diagnosis process, the data after the abnormality occurs is preprocessed as described in steps 1-5 to ensure that the data processing method is completely consistent with the training data.

[0096] Step 13: Use the time convolution capsule network model optimized in step 11 to diagnose typical faults of circulating water pumps in nuclear power plants and obtain classification results.

[0097] In addition, the present invention can also evaluate the fault diagnosis results of the model. The present invention uses confusion matrix and fault diagnosis accuracy as indicators to evaluate the accuracy and effectiveness of the model of the present invention. The relevant results can be used as a reference for operation and decision-making personnel to take relevant measures in a timely manner, thereby ensuring safety and improving economic efficiency.

[0098] It should be noted that the fault diagnosis of the circulating water pump of a nuclear power plant can also be performed through machine learning such as support vector machines, back-propagation neural networks, and ordinary convolutional neural network technologies. However, the method described in the present invention can effectively extract the time series characteristics and deep vector characteristics of the input data by combining the time convolution network and the capsule network algorithm, and can obtain more accurate fault diagnosis results for the highly linear characteristics of the operating data of the circulating water pump of the nuclear power plant. The circulating water pump operating data is denoised by wavelet packet transform to reduce the impact of noise on the data; the denoised data is feature extracted by permutation entropy and envelope entropy to obtain hybrid enhanced features; the high-dimensional time series characteristics of the data obtained by feature extraction are mined by the time convolution kernel; the vector features and the relative position relationship between the data are mined by the capsule network, and the dynamic routing of the capsule network is optimized by the support vector to improve the efficiency of the capsule network. Finally, the method described in the present invention can adaptively, accurately and quickly diagnose the potential causes of faults in the circulating water pump of a nuclear power plant, and provide analysis and reference basis for operators.

[0099] Based on the above method, the present invention also provides a nuclear power circulating water pump fault diagnosis system based on optimized capsule network, such as Figure 2 As shown, including:

[0100] The first data acquisition module 201 is used to acquire vibration sensor data of the nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set;

[0101] A preprocessing module 202, configured to preprocess the first data set to obtain a second data set;

[0102] A feature extraction module 203 is used to extract features from the second data set to obtain a feature matrix;

[0103] A phase space reconstruction module 204 is used to perform phase space reconstruction on the feature matrix to obtain training data;

[0104] A network construction module 205 is used to construct a temporal convolutional capsule network;

[0105] A training module 206 is used to train the temporal convolutional capsule network using the training data to obtain a trained temporal convolutional capsule network;

[0106] The second data acquisition module 207 is used to acquire vibration sensor data of the nuclear power circulating water pump to be tested;

[0107] The detection module 208 is used to use the trained temporal convolution capsule network to determine the fault result of the nuclear power circulating water pump to be detected.

[0108] Based on the above content, the present invention also discloses the following technical effects:

[0109] Compared with the prior art, the method of the present invention has a higher fault diagnosis accuracy than other methods, and can also achieve a better convergence effect when the training data set is small, and can provide a fault diagnosis model with higher accuracy and better convergence effect.

[0110] The reason for the high accuracy is the overall implementation of step 3, step 4, step 5, step 7, step 8, step 9, and step 10.

[0111] Among them, steps 3 and 4 reduce the noise of the original data through wavelet packet transform to reduce the impact of environmental noise on the data during the operation of the circulating water pump, and perform feature extraction processing on the denoised data through permutation entropy and envelope entropy to form hybrid enhanced features, providing rich and accurate data support for subsequent fault diagnosis models.

[0112] In step 5, by converting the original two-dimensional data into a three-dimensional data group with time series attributes, it can not only meet the input data format of the time convolution capsule network, but also each input data is no longer just a single instantaneous parameter, but a time series feature, which can better reflect the data feature changes of the fault process.

[0113] Step 7 establishes a temporal convolution network structure formed by stacking temporal convolution kernels. Compared with traditional convolution kernels, it can more effectively extract the temporal information characteristics of the data. At the same time, it has a flexible receptive field that can be flexibly customized according to the different characteristics of different tasks. It also has a more stable gradient, avoiding the gradient disappearance and explosion problems.

[0114] In the capsule network formed in step 8, each neuron is a vector instead of a traditional scalar, which enables the capsule network to extract more detailed features from the input data and reduce the loss of feature information. The capsule network updates the capsule layer parameters through a dynamic routing mechanism, further increasing the coupling coefficient between the child node and the parent node, and making full use of local information to enrich the feature representation capability and information content. The capsule network structure itself has translation invariance, which can extract the relative position relationship of the input features and improve the accuracy of fault diagnosis.

[0115] Step 9: To address the problem that the capsule network dynamic routing has a large amount of computation and the diagnostic efficiency is not high enough, the proposed support vector dynamic routing algorithm can adaptively obtain the key feature points in the input data through the support vector machine, reconstruct the data, enhance the feature resolution of the data, accelerate the convergence of the capsule network model, and improve the efficiency and accuracy of fault diagnosis.

[0116] Step 10 comprehensively sorts out the hyperparameters that need to be manually given in the temporal convolutional capsule network, and selects the optimal parameters for model testing and adjustment to ensure the accuracy of fault diagnosis and the stability of the network.

[0117] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0118] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A fault diagnosis method for nuclear power circulating water pump based on optimized capsule network, It is characterized in that include: Acquire vibration sensor data of a nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set; Preprocessing the first data set to obtain a second data set; Performing feature extraction on the second data set to obtain a feature matrix; Reconstructing the characteristic matrix in phase space to obtain training data; Construct a temporal convolutional capsule network, including: Construct a convolutional neural network to extract the nonlinear features of the detection data; Constructing a temporal convolution layer at the output end of the convolutional neural network to extract deep temporal features; A capsule network is constructed at the output end of the temporal convolutional layer to extract vector features, specifically including: Multiply the features output by the temporal convolutional layer by the weight matrix to get the prediction vector: U j,i =U i W j,i Where W j,i is the weight of the main capsule layer, U i is the feature output of the temporal convolutional layer, U j,i Represents the vector generated by the input feature prediction; the prediction vector is transmitted to the digital capsule layer: S j =∑ i U j,i C i,j b i,j =b i,j +V j U j,i C in the formula i,j and b i,j represents the coupling and bias coefficient, S j is the total input vector; A dynamic routing algorithm is set in the capsule network to iterate the vector features, specifically including: Will U j,i After inputting into the support vector machine for training, a set of support vectors sv can be obtained. 1 ,sv 2 ,…,sv Q With the Lagrange factor a 1 ,a 2 , …, a N is the Lagrangian vector a of the element 1×N； The original training samples are reconstructed using the feature extraction formula shown in the following formula: Where sv i is the support vector, a i is the Lagrangian factor corresponding to the support vector, is the label corresponding to the support vector, and b is the bias; The number of reconstructed samples is consistent with the initial training sample set, which is N, and the dimension of each sample changes from the initial sample dimension M to Q. i ∈U j,i , the above formula is for the input x i After processing, a new set of reconstructed samples can be obtained as shown below: Then the reconstructed h i As the new U j,i Substitute into dynamic routing for update iteration; Using the training data to train the temporal convolutional capsule network to obtain a trained temporal convolutional capsule network; Obtain vibration sensor data of the nuclear power circulating water pump to be tested; The trained temporal convolution capsule network is used to determine the fault result of the nuclear power circulating water pump to be detected.

2. According to the method for fault diagnosis of nuclear power circulating water pump based on optimized capsule network in claim 1, It is characterized in that The first data set is preprocessed to obtain a second data set, including: The first data set is subjected to noise reduction processing by wavelet packet transform, and the low-frequency part and the high-frequency part of the first data set are orthogonally decomposed at the same time to obtain the second data set.

3. According to claim 1, the nuclear power circulating water pump fault diagnosis method based on optimized capsule network, It is characterized in that After the step of "obtaining vibration sensor data of the nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set", and before the step of "preprocessing the first data set to obtain a second data set", the method further includes: Different labels are set for the first data set according to the degree of failure.

4. According to claim 1, the nuclear power circulating water pump fault diagnosis method based on optimized capsule network, It is characterized in that The temporal convolutional capsule network adopts a cross entropy loss function as a loss function.

5. The nuclear power circulating water pump fault diagnosis method based on optimized capsule network according to claim 1, It is characterized in that The temporal convolutional capsule network is trained using the SGD optimization algorithm.

6. The nuclear power circulating water pump fault diagnosis method based on optimized capsule network according to claim 1, It is characterized in that After the step of "obtaining vibration sensor data of the nuclear power circulating water pump to be detected", and before the step of "using the trained temporal convolution capsule network to determine the fault result of the nuclear power circulating water pump to be detected", it also includes: The vibration sensing data of the nuclear power circulating water pump to be detected is preprocessed.

7. The method for diagnosing faults of nuclear power circulating water pumps based on optimized capsule network according to claim 6, It is characterized in that Preprocessing of vibration sensor data of nuclear power circulating water pump to be tested includes: The vibration sensor data of the nuclear power circulating water pump to be detected is subjected to noise reduction processing by wavelet packet transformation, and the low-frequency part and the high-frequency part of the vibration sensor data of the nuclear power circulating water pump to be detected are orthogonally decomposed at the same time.

8. The method for fault diagnosis of a nuclear power circulating water pump based on an optimized capsule network according to claim 1, It is characterized in that After the step of "constructing a time convolution capsule network", and before the step of "training the time convolution capsule network using the training data to obtain a trained time convolution capsule network", the method further includes: Set hyperparameters for the temporal convolutional capsule network.

9. A nuclear power circulating water pump fault diagnosis system based on optimized capsule network, It is characterized in that include: A first data acquisition module is used to acquire vibration sensor data of the nuclear power circulating water pump during operation and vibration sensor data under various faults to form a first data set; A preprocessing module, used for preprocessing the first data set to obtain a second data set; A feature extraction module, used to extract features from the second data set to obtain a feature matrix; A phase space reconstruction module is used to perform phase space reconstruction on the feature matrix to obtain training data; Network building modules are used to build a temporal convolutional capsule network, including: The nonlinear feature extraction module builds a layer of convolutional neural network to extract the nonlinear features of the detection data; A deep temporal feature extraction module constructs a temporal convolution layer at the output end of the convolutional neural network to extract deep temporal features; The vector feature extraction module constructs a capsule network at the output end of the temporal convolution layer to extract vector features, specifically including: Multiply the features output by the temporal convolutional layer by the weight matrix to get the prediction vector: U j,i =U i W j,i Where W j,i is the weight of the main capsule layer, U i is the feature output of the temporal convolutional layer, U j,i Represents the vector generated by the input feature prediction; the prediction vector is transmitted to the digital capsule layer: S j =∑ i U j,i C i,j b i,j =b i,j +V j U j,i C in the formula i,j and b i,j represents the coupling and bias coefficient, S j is the total input vector; The vector feature updating module sets a dynamic routing algorithm in the capsule network to iterate the vector features, specifically including: Will U j,i After inputting into the support vector machine for training, a set of support vectors sv can be obtained. 1 ,sv 2 ,…,sv Q With the Lagrange factor a 1 ,a 2 , …, a N is the Lagrangian vector a of the element 1×N； The original training samples are reconstructed using the feature extraction formula shown in the following formula: Where sv i is the support vector, a i is the Lagrangian factor corresponding to the support vector, is the label corresponding to the support vector, and b is the bias; The number of reconstructed samples is consistent with the initial training sample set, which is N, and the dimension of each sample changes from the initial sample dimension M to Q. i ∈U j,i , the above formula is for the input x i After processing, a new set of reconstructed samples can be obtained as shown below: Then the reconstructed h i As the new U j,i Substitute into dynamic routing for update iteration; A training module, used to train the temporal convolutional capsule network using the training data to obtain a trained temporal convolutional capsule network; A second data acquisition module is used to acquire vibration sensor data of the nuclear power circulating water pump to be tested; The detection module is used to use the trained time convolution capsule network to determine the fault result of the nuclear power circulating water pump to be detected.

Images