A traction converter fault diagnosis method, device, medium and product

Abstract

The present invention discloses a traction converter fault diagnosis method, equipment, medium and product, which relates to the field of intelligent operation and maintenance of electric locomotives. The method comprises: obtaining a data set; the data set is intermediate DC voltage data; decomposing the data set based on variational modal decomposition to obtain a multi-channel standard fault feature set; determining a matching channel weighting layer according to the variational modal decomposition characteristics; the matching channel weighting layer is a PE-Spearman rank correlation coefficient; establishing a matching channel weighting layer-one-dimensional deep separable convolutional neural network model according to the matching channel weighting layer; the one-dimensional deep separable convolutional neural network is a network constructed by applying a deep separable convolutional neural network to one-dimensional long sequence data; inputting a multi-channel standard fault feature set into the matching channel weighting layer-one-dimensional deep separable convolutional neural network model to obtain a fault classification. The present invention can accurately, quickly and stably identify fault features.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of electric locomotives, and in particular to a method, equipment, medium and product for traction converter fault diagnosis. Background Art

[0002] The HXD1 electric locomotive has excellent traction and is suitable for traction tasks of long-distance and heavy-load freight trains. Its core power source is composed of two TGA9 traction inverters. The traction inverter operates continuously for a long time and the working environment is complex and changeable, resulting in a high failure rate. The existing detection of traction inverter faults still has problems such as insufficient accuracy and slow diagnostic rate, which may lead to false alarms and missed alarms, seriously affecting the operation safety, efficiency and maintenance costs of the locomotive; therefore, designing advanced intelligent algorithms to perform fault diagnosis on traction inverters is of great significance to ensure train safety and improve the development of railway intelligence.

[0003] At present, fault diagnosis methods at home and abroad are mainly divided into two categories: fault diagnosis based on mechanism models and intelligent fault diagnosis based on data-driven. Due to the difficulty in obtaining traction converter fault data, the current research on traction converter fault diagnosis is still mainly based on the research on fault diagnosis based on mechanism models. However, the traction converter is a complex system with multiple components working together and highly integrated, and it is difficult to establish a nonlinear fault mechanism model. With the rapid rise of deep learning in the engineering field, its powerful data processing capabilities and automatic feature extraction functions play a powerful role in fault diagnosis. However, deep learning is currently rarely used in traction converter fault diagnosis. Therefore, how to design a new data-driven traction converter fault diagnosis method has become a technical problem that needs to be solved urgently in this field. Summary of the invention

[0004] The purpose of the present invention is to provide a traction converter fault diagnosis method, device, medium and product, which can accurately, quickly and stably identify fault characteristics.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] In a first aspect, the present invention provides a traction converter fault diagnosis method, the traction converter fault diagnosis method comprising:

[0007] Acquire a data set; the data set is intermediate DC voltage data.

[0008] The data set is decomposed based on variational mode decomposition (VMD) to obtain a multi-channel standard fault feature set.

[0009] According to the variational mode decomposition characteristics, a matching channel weighting layer (MCWL) is determined; the matching channel weighting layer is a PE-Spearman rank correlation coefficient.

[0010] According to the supporting channel weighted layer, a supporting channel weighted layer-one-dimensional deep separable convolutional neural network model (One-dimensional Deep Separable Convolutional Neural Networks, 1DSCNN) is established; the one-dimensional deep separable convolutional neural network is a network constructed by applying the deep separable convolutional neural network to one-dimensional long sequence data.

[0011] The multi-channel standard fault feature set is input into the matching channel weighted layer-one-dimensional deep separable convolutional neural network model to obtain fault classification.

[0012] Optionally, the data set is decomposed based on variational mode decomposition to obtain a multi-channel standard fault feature set, specifically including:

[0013] Based on variational mode decomposition, the data set is decomposed into several eigenmode components.

[0014] A plurality of the eigenmode components are synthesized and reconstructed to obtain a plurality of synthesized and reconstructed components.

[0015] The energy percentage (Energy Storage Network, ESN) and signal-to-noise ratio (SNR) of each synthesized and reconstructed component are calculated respectively to obtain several energy percentage values ​​and signal-to-noise ratio values.

[0016] According to the energy percentage value and the signal-to-noise ratio value, a multi-channel standard fault feature set is determined by comparative verification.

[0017] Optionally, according to the variational mode decomposition characteristics, the supporting channel weighting layer is determined, specifically including:

[0018] A correlation coefficient is determined according to the data set and the multi-channel standard fault feature set.

[0019] The correlation coefficient is dimensionally expanded to obtain an expanded correlation coefficient.

[0020] The expanded correlation coefficient is constrained using a constraint function to obtain a matching channel weighted layer.

[0021] Optionally, the expression of the supporting channel weighted layer is:

[0022] ;

[0023] in, is the weighting function of the supporting channel weighting layer, To calculate the first The rank difference of the signals; is the length of the input sequence.

[0024] Optionally, the constraint function is expressed as:

[0025] ;

[0026] in, is the number of channels, is the correlation coefficient The exponential function of is the correlation coefficient, is the channel weight, is the constraint function.

[0027] Optionally, the calculation formula of the one-dimensional depth separable convolutional neural network is:

[0028] ;

[0029] ;

[0030] in, is the parameter quantity, is the amount of calculation, is the length of the one-dimensional convolution kernel, is the number of channels of the convolution kernel input, is the number of channels output by the convolution kernel, is the length of the input sequence.

[0031] Optionally, the supporting channel weighted layer-one-dimensional depth separable convolutional neural network model specifically includes:

[0032] The input layer is used to input a multi-channel standard fault feature set.

[0033] A supporting channel weighting layer is used to weight the multi-channel standard fault feature set to obtain a weighted feature set.

[0034] Two one-dimensional depth-separable convolutional layers are used to extract features from the weighted feature set to obtain a feature set after feature extraction.

[0035] The three fully connected layers are used to integrate the feature set after the feature extraction to obtain the fault classification.

[0036] In a second aspect, the present invention provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any one of the above-described traction inverter fault diagnosis methods.

[0037] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any one of the above-described traction inverter fault diagnosis methods.

[0038] In a fourth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements any one of the above-described traction converter fault diagnosis methods.

[0039] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] The present invention provides a traction inverter fault diagnosis method, equipment, medium and product, by acquiring a data set; the data set is intermediate DC voltage data; the data set is decomposed based on variational modal decomposition to obtain a multi-channel standard fault feature set; wherein, variational modal decomposition can provide a higher frequency domain resolution, and even in the presence of noise, it can better separate the different frequency characteristics of the signal; in addition, according to the variational modal decomposition characteristics, a matching channel weighted layer is determined; the matching channel weighted layer is a PE-Spearman rank correlation coefficient; according to the matching channel weighted layer, a matching channel weighted layer-one-dimensional deep separable convolutional neural network model is established; by using the matching channel weighted layer as a "bridge" between the variational modal decomposition and the one-dimensional deep separable convolutional neural network model, the variational modal decomposition and the one-dimensional deep separable convolutional neural network model can be made more consistent, and at the same time, the network's characterization ability of multi-channel features can be effectively enhanced; finally, the multi-channel standard fault feature set is input into the matching channel weighted layer-one-dimensional deep separable convolutional neural network model to obtain fault classification. Compared with the prior art, the present invention can accurately, quickly and stably identify fault characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] 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 work.

[0042] Figure 1 This is a diagram of the application environment of a traction converter fault diagnosis method in one embodiment of the present invention.

[0043] Figure 2 A schematic flow chart of a traction converter fault diagnosis method provided by one embodiment of the present invention.

[0044] Figure 3 A flow chart of MCWL channel weighting provided by an embodiment of the present invention.

[0045] Figure 4 A 1DSCNN network structure diagram provided for an embodiment of the present invention.

[0046] Figure 5 A VMD evaluation index curve diagram provided by an embodiment of the present invention.

[0047] Figure 6 A VMD decomposition diagram provided by an embodiment of the present invention.

[0048] Figure 7 This is a structural diagram of the MCWL-1DSCNN model provided in one embodiment of the present invention.

[0049] Figure 8 A loss rate curve diagram provided for an embodiment of the present invention.

[0050] Fig. 9 This is an accuracy curve diagram provided by an embodiment of the present invention.

[0051] Fig.10 A schematic diagram of the structure of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] 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.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] The purpose of the present invention is to address the problem of fault diagnosis of TGA9 traction converters and propose a traction converter fault diagnosis method based on VMD-MCWL-1DSCNN. The operating state of the traction converter is closely related to the DC side voltage of the rectifier module, which can reflect the power conversion and stability of the entire system. When a traction converter fails, the abnormal change of the intermediate DC voltage is often a significant fault feature. The present invention performs fault diagnosis on the operating state of the TGA9 traction converter by collecting intermediate DC voltage data.

[0055] The traction converter fault diagnosis method provided by the embodiment of the present invention can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the acquired data set to the server 104. After the server 104 receives the acquired data set, for the acquired data set, the server 104 decomposes the data set based on the variational mode decomposition to obtain a multi-channel standard fault feature set; according to the variational mode decomposition characteristics, the supporting channel weighting layer is determined; the supporting channel weighting layer is the PE-Spearman rank correlation coefficient; according to the supporting channel weighting layer, a supporting channel weighting layer-one-dimensional deep separable convolutional neural network model is established; the one-dimensional deep separable convolutional neural network is a network constructed by applying the deep separable convolutional neural network to one-dimensional long sequence data; the multi-channel standard fault feature set is input into the supporting channel weighting layer-one-dimensional deep separable convolutional neural network model to obtain a fault classification. The server 104 can feedback the obtained fault classification to the terminal 102. In addition, in some embodiments, the traction inverter fault diagnosis method can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly perform fault diagnosis on the acquired data set, or the server 104 can obtain the data set from the data storage system and perform fault diagnosis on the acquired data set.

[0056] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones and tablet computers, IoT devices and portable wearable devices. IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0057] In an exemplary embodiment, Figure 2As shown, a traction converter fault diagnosis method is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In an embodiment of the present invention, the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate the method, which includes the following steps S1 to S5. Among them:

[0058] S1: Acquire a data set; the data set is intermediate DC voltage data.

[0059] S2: Decomposing the data set based on variational mode decomposition to obtain a multi-channel standard fault feature set.

[0060] S3: Determine the supporting channel weighting layer according to the variational mode decomposition characteristics; the supporting channel weighting layer is the PE-Spearman rank correlation coefficient.

[0061] S4: According to the supporting channel weighted layer, a supporting channel weighted layer-one-dimensional deep separable convolutional neural network model is established; the one-dimensional deep separable convolutional neural network is a network constructed by applying the deep separable convolutional neural network to one-dimensional long sequence data.

[0062] S5: Input the multi-channel standard fault feature set into the matching channel weighted layer-one-dimensional deep separable convolutional neural network model to obtain fault classification.

[0063] In the implementation of the above steps S1 to S5, first, the intermediate DC voltage data is collected to construct a data set; second, the data set is decomposed by VMD to obtain the intrinsic modal components at different frequencies and construct a multi-channel standard fault feature set; then, MCWL is designed based on the modal component characteristics to weight the network channels and use t -Distributed random neighbor embedding is used to evaluate the weighted effect; finally, MCWL is integrated into 1DSCNN, and the fault diagnosis of traction converter is realized through this model.

[0064] In another exemplary embodiment of the present invention, in step S2, specifically including:

[0065] S21: Based on variational mode decomposition, decompose the data set into a plurality of eigenmode components.

[0066] S22: synthesizing and reconstructing a plurality of the eigenmode components to obtain a plurality of synthesized and reconstructed components.

[0067] S23: Calculate the energy percentage and signal-to-noise ratio of each synthesized and reconstructed component respectively to obtain a plurality of energy percentage values ​​and signal-to-noise ratio values.

[0068] S24: Determine a multi-channel standard fault feature set by comparative verification according to the energy percentage value and the signal-to-noise ratio value.

[0069] For the voltage signal of the traction converter, VMD can effectively extract the local oscillation components in the signal, reveal the intrinsic characteristics of the signal, and suppress the noise in the signal. In the VMD decomposition process, if the parameters are not set properly, under-decomposition and over-decomposition problems may occur. In order to avoid this situation, the IMFs components are synthesized and reconstructed, and two indicators are calculated: energy percentage and signal-to-noise ratio. The optimal number of decomposition layers is found through comparative verification. K value.

[0070] In another exemplary embodiment of the present invention, in step S3, the following steps are specifically included:

[0071] S31: Determine a correlation coefficient according to the data set and the multi-channel standard fault feature set.

[0072] S32: Dimensionally expand the correlation coefficient to obtain an expanded correlation coefficient.

[0073] S33: Use a constraint function to constrain the expanded correlation coefficient to obtain a matching channel weighted layer.

[0074] When weighting the network input channel, the channel attention mechanism is commonly used, such as SENet. The channel weighting layer learns the attention weight of each channel based on the functions of the fully connected layer and the convolutional layer, which usually requires a large amount of calculation; and the channel attention compresses and excites the global information of the input signal channel, and cannot effectively weight the channel according to the VMD decomposition characteristics. Therefore, the present invention designs a matching channel weighting layer. According to the VMD characteristics, the Spearman rank correlation coefficient is selected as the weighting function. The function is a statistic used to measure the degree of nonlinear correlation between two variables and is often used to analyze the relationship between two variables; but the Spearman rank correlation examines the strength of the monotonic relationship between the two, and has limitations when dealing with periodic relationships or data mutations. Permutation Entropy (PE) has the randomness and dynamic mutation of time series, which can effectively overcome this problem; therefore, the present invention introduces PE to calculate the arrangement of each IMFs rank, and establishes a PE-Spearman rank correlation coefficient mathematical model as the weighting function of MCWL to reflect the relationship between each component and the original signal.

[0075] First, the sorting entropy of the original signal and IMFs of each fault curve is calculated. and original signal Refactoring to introduce embedding dimension and time delay , we can reconstruct the matrix .right Internal variables are sorted in ascending order. If two values ​​are equal, they are sorted by subscript. Appears after being mapped There are permutations, and the probability of each permutation is ,in ,get PE The calculation formula is:

[0076] (1);

[0077] Secondly, the permutation entropy of each component is obtained by the above formula and the permutation entropy of the original signal Then, the rank of each component is evaluated according to the sort entropy. and the rank of the original signal , , where Indicates that the values ​​are sorted in ascending order, and the result returned is the rank after sorting and . The ranking is assigned in ascending order.

[0078] Then, by calculating and , calculate the rank difference , the Spearman rank correlation coefficient based on PE can be calculated:

[0079] (2);

[0080] in, is the weighting function of the supporting channel weighting layer, To calculate the first The rank difference of the signals; is the length of the input sequence.

[0081] Finally, in order to prevent the probability of each channel weight from differing too much and avoid the gradient disappearance or explosion of the model, the constraint function Correlation coefficient Constraints are performed and the channel weights are finally obtained , to ensure the stability and effectiveness of the model. The calculation formula is:

[0082] (3);

[0083] in, is the number of channels, is the correlation coefficient The exponential function of is the correlation coefficient, is the channel weight, is the constraint function.

[0084] In VMD decomposition, the larger the correlation coefficient between the IMF and the original signal, the more information the component contains, and the better it can express some characteristics of the original signal. The number of components of the original signal decomposed by VMD is the number of channels input to the model. The PE-Spearman rank correlation coefficient function is used as the weighting function of the channel weighting layer, and the correlation coefficient between each IMF and the original signal is the weighted weight of each channel.

[0085] The specific process of MCWL weighting network channels is as follows: Figure 3 As shown. First, initialize the correlation coefficient matrix , then input the original signal matrix and the eigencomponent matrix ;in and Represent the sample index and channel index respectively, and the initial value is set to 0. Then, through a nested loop, the channel correlation coefficient of each sample is calculated separately , that is, the PE-Spearman rank correlation coefficient; when The maximum number of cycles is greater than the sample batch size (batch_size=32), j When the maximum number of loops is greater than the channel size (channels=3), the loop is exited; and the channel correlation coefficient is obtained Assign to Finally, Expand one dimension so that the correlation coefficient matrix dimension can match the input dimension of the network model; and constrain the weight probability through the constraint function to output the final channel weight matrix .

[0086] It should be noted that: 1. The characteristic component matrix is ​​the selected IMF component matrix. The original signal matrix and the characteristic component matrix are used as input together in order to calculate the rank difference of each component by sorting with the original signal. d , and use this to calculate the correlation coefficient.

[0087] 2. Inputting the original signal and the characteristic component matrix, as well as initializing the correlation coefficient matrix are necessary preparations before calculation, and the two do not affect each other. Inputting the original signal and the characteristic component matrix is ​​to calculate the correlation coefficient; first establish an initialized correlation coefficient matrix to facilitate the assignment after calculating the correlation coefficient.

[0088] 3. The input of the 1DSCNN network only includes the characteristic component matrix and does not include the original signal; while the MCWL designed in the present invention requires the assistance of the original signal to calculate the correlation coefficient of each component, so it is necessary to input two quantities, the original signal and the characteristic component; therefore, the output after the input passes through MCWL is a weighted coefficient matrix, and the weighted coefficient matrix and the characteristic component matrix are given as input to other layers of the 1DSCNN network, where the original signal matrix is ​​removed by indexing and does not participate in subsequent operations.

[0089] In another exemplary embodiment of the present invention, in step S4, the supporting channel weighted layer-one-dimensional depth separable convolutional neural network model specifically includes:

[0090] The input layer is used to input a multi-channel standard fault feature set.

[0091] A supporting channel weighting layer is used to weight the multi-channel standard fault feature set to obtain a weighted feature set.

[0092] Two one-dimensional depth-separable convolutional layers are used to extract features from the weighted feature set to obtain a feature set after feature extraction.

[0093] The three fully connected layers are used to integrate the feature set after the feature extraction to obtain the fault classification.

[0094] At present, DSCNN is mainly used in two-dimensional image processing and is rarely used in one-dimensional time series data. The present invention applies DSCNN to one-dimensional long sequence data to construct a 1DSCNN network model, which can realize the rapid identification of long sequence fault data. 1DSCNN is mainly constructed by connecting multiple depthwise convolutions (DC) and pointwise convolutions (PC) in series. DC performs a time series convolution operation on the input feature sequence to extract the time series information; then PC is used to perform a channel convolution operation on the features at each position to extract the channel features. DC has multiple independent filters, each filter only acts on the corresponding input channel; PC is a The convolution kernel is used after DC. This convolution can fuse and mix information between different channels. The operation of one-dimensional depth-separable convolution is as follows: Figure 4 As shown (for the convenience of demonstration, the number of channels of the filter, convolution kernel, and output sequence in the figure is only 3).

[0095] The calculation formula of one-dimensional depth separable convolutional neural network is:

[0096] (4);

[0097] (5);

[0098] in, is the parameter quantity, is the amount of calculation, is the length of the one-dimensional convolution kernel, is the number of channels of the convolution kernel input, is the number of channels output by the convolution kernel, is the length of the input sequence.

[0099] The number of parameters for a standard one-dimensional convolution operation is and the amount of calculation for:

[0100] (6);

[0101] (7);

[0102] The ratio P of the computational cost of a one-dimensional depthwise separable convolution to a standard one-dimensional convolution is:

[0103] (8);

[0104] When actually processing one-dimensional sequences, Usually greater than 1, one-dimensional convolution kernel Usually a vector of length 3 is taken. Therefore, by comparing the computational costs, it can be found that the one-dimensional depth-separable convolution greatly reduces the computational cost while ensuring high accuracy of the classification task.

[0105] The present invention is implemented based on the intermediate DC voltage signal data of the TGA9 traction converter in the locomotive monitoring system, which includes 4 common fault conditions. The voltage signals under 5 states of normal operation of the traction converter, intermediate voltage undervoltage, intermediate voltage overvoltage, four-quadrant input overcurrent, and main circuit grounding are respectively used. The sampling frequency is 10kHz, the length of each sample is 1000, the number of samples under each state is 100 groups, and the total number of samples under 5 states is 500 groups. The specific parameters of the data samples are shown in Table 1.

[0106] Table 1 TGA9 traction converter data set

[0107]

[0108] First, the original data set is decomposed by VMD. In order to constrain the number of decomposition layers, comparative verification is used to find the optimal K Value. The hyperparameter quadratic penalty factor Set to 2000, the range of decomposition levels is set to The curves of the two indicators of VMD at different decomposition levels are as follows: Figure 5 As shown.K When 9 is taken, the decomposition effect is best; when K When the value is less than 9, there is an under-decomposition situation, some information is lost, resulting in a decrease in the energy percentage, and the separated signal and noise components are aliased, resulting in a decrease in the signal-to-noise ratio; K When the value is greater than 9, there is an over-decomposition situation, the original signal is wrongly decomposed into multiple modes, the increase of noise components leads to a decrease in signal-to-noise ratio, and the energy percentage is no longer improved. Therefore, the VMD decomposition layer is increased. K Set to 9. Take a set of data from the intermediate voltage undervoltage fault sample as an example. The characteristic components of the fault signal are extracted by the VMD algorithm. The visualization results are as follows: Figure 6 As shown, the first curve is the original signal, and the other nine curves are IMF components at different frequencies.

[0109] In order to reduce the computational cost of the model, the IMFs that retain the original signal feature information are screened. By calculating the variance contribution rate of each IMF, the IMF with a low variance contribution rate is the redundant component and should be removed. The variance contribution rate of each IMF component is shown in Table 2. From the table, it can be seen that the variance contribution rates of IMF1, IMF2, and IMF3 are relatively high. These three components are synthesized to reconstruct the signal. u , the variance contribution rate of the reconstructed signal reaches 0.995, and almost all the characteristic information of the original signal is retained. Therefore, IMF1, IMF2, and IMF3 are selected as the feature inputs of the three channels of the model to construct a multi-channel feature data set.

[0110] Table 2 Variance contribution of each IMF, reconstructed signal and original signal

[0111]

[0112] Secondly, the multi-channel feature data set is weighted by SENet and MCWL designed by the present invention, and the results of channel weighting are analyzed by comparing SENet and MCWL. As shown in Table 3, the mutual information and variance contribution rate of each IMF component can be seen (where IMF1, IMF2, and IMF3 correspond to the input feature sequences of channel 1, channel 2, and channel 3, respectively), and the component of channel 1 retains more feature information, while channel 2 and channel 3 retain relatively less. The weighted weight results are shown in Table 4. The weighted weight channel 3 output by SENet has the largest weight, while the feature information retained by channel 3 is the least; while the weighted weight output by MCWL can be weighted accordingly by the feature information retained by each channel, strengthening the feature representation of a specific channel and enhancing the overall representation ability of the network.

[0113] Table 3 Evaluation index of information retention of each IMF component

[0114]

[0115] Table 4 Comparison of channel weighting results

[0116]

[0117] Then, the MCWL designed by the present invention is integrated into the 1DSCNN network to establish a diagnostic model based on MCWL-1DSCNN. The MCWL-1DSCNN traction converter fault diagnosis model is built based on the PyCharm2020.1.3 programming platform and the Pytorch environment. The structure of the model is as follows Figure 7 As shown in Figure 5, the main body of the model is: (1) 1 channel weighted layer, namely MCWL, which is placed after the input layer and before the convolution layer. It is used to weight the input channels and enhance the ability of the convolution layer to extract features. It also regularizes the network and reduces the risk of overfitting by limiting the weight of the channel; (2) 2 one-dimensional depth-separable convolution layers. In order to reduce the computational complexity of channel convolution feature extraction, the depth convolution kernel size is set to 1×3, the point-by-point convolution kernel size is 1×1, and the pooling layer uses maximum pooling to ensure that the model fully extracts features while improving the computing speed. (3) 3 fully connected layers can integrate the outputs of the 3 channels and output them as a one-dimensional label vector of the diagnostic category. The specific parameters of the model are shown in Table 5.

[0118] Table 5 Network model structure parameters

[0119]

[0120] The model input is a feature sequence with 3 channels and a length of 1×1000. The model hyperparameter learning rate is set to 0.001, the Adam optimizer is used to update the parameters, and the cross entropy loss function is selected as the model evaluation indicator. The number of iterations of model training is 50, and the sample batch size for each training is 32.

[0121] Finally, the processed multi-channel feature set is input into the model for training. When processing long sequence data, common deep convolutional neural networks may have overfitting problems due to the complexity of the model and the large number of parameters, resulting in poor model robustness. However, when processing long sequence data, 1DSCNN reduces the risk of overfitting due to the reduction in the number of parameters and has better stability. The experimental results are shown in Figure 2. Figure 8 and Fig. 9 As shown. Among them, Figure 8 and Fig. 9 The loss function and accuracy curves of the 350 fault training sets that have been processed are continuously iteratively optimized in the 1DSCNN model. Figure 8 and Fig. 9It can be seen that for the voltage feature set of traction converter fault diagnosis, the model has fast convergence speed, strong stability, and high diagnostic accuracy; it ensures that the model has stronger feature extraction ability while using fewer parameters. The diagnostic accuracy of the model of the present invention is 98.3%, and the diagnostic time of a single sample is 0.09s.

[0122] The experiment proves the accuracy and superiority of VMD decomposition and MCWL-1DSCNN by reducing and replacing the signal decomposition method and channel weighting algorithm in the overall diagnosis model. The ablation experiment results are shown in Table 6. As can be seen from the table, as the method is improved step by step, the accuracy rate is getting higher and higher, and the false alarm rate is getting lower and lower. The VMD-MCWL-1DSCNN model has significantly improved the diagnostic effect of the traction converter.

[0123] Table 6 Comparison of ablation experiment results

[0124]

[0125] The method proposed in the present invention is an intelligent diagnosis strategy for traction inverters based on the combination of feature extraction and deep learning. The fault voltage of the TGA9 traction inverter collected by the electric locomotive monitoring system is a nonlinear, long sequence data containing noise components. Variational mode decomposition can provide a higher frequency domain resolution, and even in the presence of noise, it can better separate the different frequency characteristics of the signal; the one-dimensional deep separable convolutional neural network supports rapid detection of multiple types of faults; the present invention also customizes a matching channel weighting layer, MCWL as a "bridge" between the VMD and 1DSCNN models, which can make the VMD and 1DSCNN models more compatible, and at the same time can effectively enhance the network's ability to characterize multi-channel features. Therefore, the present invention proposes a traction inverter fault diagnosis model based on VMD-MCWL-1DSCNN, which enables the system to accurately, quickly and stably identify fault characteristics.

[0126] The present invention also provides an application scenario, which applies the above-mentioned traction converter fault diagnosis method. Specifically: The traction converter fault diagnosis method provided in this embodiment can be applied in the traction converter fault diagnosis scenario. The traction converter fault diagnosis scenario includes a data acquisition link, a model building link, and a fault classification link; the acquired data is subjected to variational mode decomposition to separate the different frequency characteristics of the signal, and then the signal is input into the established MCWL-1DSCNN model to obtain the fault classification.

[0127] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.10As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store intermediate DC voltage data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a traction inverter fault diagnosis method is implemented.

[0128] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0129] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the above-mentioned method embodiments are implemented when the processor executes the computer program.

[0130] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the above-mentioned method embodiments are implemented.

[0131] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the above-mentioned method embodiments are implemented.

[0132] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0133] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0134] The database involved in each embodiment provided by the present invention may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided by the present invention may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0135] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0136] 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 traction converter fault diagnosis method, characterized in that: The traction converter fault diagnosis method comprises: Acquire a data set; the data set is intermediate DC voltage data; Decomposing the data set based on variational mode decomposition to obtain a multi-channel standard fault feature set; According to the variational mode decomposition characteristics, the supporting channel weighting layer is determined; the supporting channel weighting layer is the PE-Spearman rank correlation coefficient; According to the supporting channel weighted layer, a supporting channel weighted layer-one-dimensional deep separable convolutional neural network model is established; the one-dimensional deep separable convolutional neural network is a network constructed by applying the deep separable convolutional neural network to one-dimensional long sequence data; Inputting the multi-channel standard fault feature set into the matching channel weighted layer-one-dimensional deep separable convolutional neural network model to obtain fault classification; According to the variational mode decomposition characteristics, the supporting channel weighting layer is determined, including: Determining a correlation coefficient according to the data set and the multi-channel standard fault feature set; Performing dimension expansion on the correlation coefficient to obtain an expanded correlation coefficient; Using a constraint function to constrain the expanded correlation coefficient to obtain a matching channel weighted layer; The expression of the supporting channel weighted layer is: ； in, is the weighting function of the supporting channel weighting layer, To calculate the first The rank difference of the signals; is the length of the input sequence; The supporting channel weighted layer-one-dimensional depth separable convolutional neural network model specifically includes: Input layer, used to input multi-channel standard fault feature set; A supporting channel weighting layer, used for weighting the multi-channel standard fault feature set to obtain a weighted feature set; Two one-dimensional depth-separable convolutional layers are used to extract features from the weighted feature set to obtain a feature set after feature extraction; The three fully connected layers are used to integrate the feature set after the feature extraction to obtain the fault classification.

2. The traction converter fault diagnosis method according to claim 1, characterized in that: The data set is decomposed based on variational mode decomposition to obtain a multi-channel standard fault feature set, which specifically includes: Based on variational mode decomposition, decomposing the data set into a number of eigenmode components; Synthetically reconstructing a plurality of the eigenmode components to obtain a plurality of synthesized and reconstructed components; Calculate the energy percentage and signal-to-noise ratio of each synthesized and reconstructed component respectively to obtain several energy percentage values ​​and signal-to-noise ratio values; According to the energy percentage value and the signal-to-noise ratio value, a multi-channel standard fault feature set is determined by comparative verification.

3. The traction converter fault diagnosis method according to claim 1, characterized in that: The expression of the constraint function is: ； in, is the number of channels, is the correlation coefficient The exponential function of is the correlation coefficient, is the channel weight, is the constraint function.

4. The traction converter fault diagnosis method according to claim 1, characterized in that: The calculation formula of the one-dimensional depth separable convolutional neural network is: ； ； in, is the parameter quantity, is the amount of calculation, is the length of the one-dimensional convolution kernel, is the number of channels of the convolution kernel input, is the number of channels output by the convolution kernel, is the length of the input sequence.

5. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the traction converter fault diagnosis method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the traction converter fault diagnosis method according to any one of claims 1 to 4 is implemented.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the traction converter fault diagnosis method according to any one of claims 1 to 4 is implemented.

Images