Railway vehicle traction motor fault monitoring method, device, equipment and storage medium

Abstract

The application provides a rail vehicle traction motor fault monitoring method, device, equipment and storage medium, and relates to the technical field of rail transit equipment fault diagnosis. The method comprises the following steps: obtaining a traction motor current signal by using a sliding window and preprocessing to obtain an input sequence, and then sending the input sequence into a time-frequency joint network for fault classification. The time-frequency joint network processing procedure is as follows: performing S transformation on the input sequence, decoupling the modulus and the phase of the second-order tensor after transformation, and constructing a third-order tensor; performing normalization, convolution and activation on the third-order tensor in sequence to extract time-frequency depth features; constructing a time-frequency joint attention mechanism, calculating attention weights along the time and frequency dimensions respectively, and weighting and focusing the time-frequency depth features; and performing pooling, unfolding and full connection classification on the weighted features in sequence to output fault types. The application has high diagnosis precision, strong anti-interference ability and fast reasoning speed, and is suitable for real-time fault monitoring of a rail vehicle on-board end.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for rail transit equipment, and in particular to a method, device, equipment, and storage medium for monitoring faults in traction motors of rail vehicles. Background Technology

[0002] As the core transportation tool in the rail transit field, the operational safety, reliability, and maintenance efficiency of rail vehicles directly affect the stable operation of the rail transit network. Traction motors, as the core component of the rail vehicle's power system, bear the heavy responsibility of converting electrical energy into mechanical energy. They are key actuators for vehicle traction and braking, and their health directly determines the vehicle's power output performance and driving safety.

[0003] In actual operation, rail vehicle traction motors operate under complex and harsh conditions for extended periods. They must withstand alternating loads from frequent starts, stops, accelerations, and decelerations, as well as multiple external disturbances such as vibration, shock, high temperature and humidity, dust and oil contamination, and power grid fluctuations. This makes them highly susceptible to typical faults such as stator winding short circuits, rotor bar breakage, bearing wear, and eccentricity. If these faults are not monitored, identified, and warned of in a timely and accurate manner, they can lead to anything from reduced vehicle power, abnormal operating noises, and increased energy consumption to motor burnout, vehicle downtime, and even train accidents. Furthermore, they significantly increase vehicle maintenance costs and repair workload.

[0004] Traction motors are gradually upgrading towards higher power, compactness, and higher integration, resulting in increasingly sophisticated internal structures and more complex fault causes. Traditional fault monitoring methods are no longer adequate for the operation and maintenance needs of modern rail vehicles. Current mainstream fault monitoring methods in the industry largely rely on single time-domain signal analysis, conventional frequency-domain transformation, or shallow machine learning models, which present numerous technical bottlenecks. First, single signal analysis can only capture local fault characteristics and cannot take into account both time-series dynamic characteristics and frequency domain energy distribution. It has extremely low sensitivity to the identification of early weak faults and complex faults. Secondly, conventional time-frequency analysis methods suffer from problems such as time-frequency resolution imbalance, poor noise robustness, and incomplete feature decoupling. They are also susceptible to interference from complex electromagnetic noise and load fluctuations in vehicles, leading to distorted feature extraction. Third, existing deep learning models are not optimized for the time-frequency coupling characteristics of traction motor current signals, lack the ability to focus on key fault features, have high model redundancy and slow inference speed, making it difficult to meet the real-time monitoring needs of the vehicle end; Fourth, some methods rely on manual feature extraction and parameter tuning, resulting in weak generalization ability and a significant drop in diagnostic accuracy under unsteady conditions of variable load and variable speed.

[0005] Therefore, overcoming the shortcomings of existing single time-domain analysis, conventional time-frequency transformation and shallow machine learning models in non-steady-state operating conditions, such as time-frequency resolution imbalance, poor noise robustness, incomplete feature decoupling, insufficient ability to focus key fault features, low model inference efficiency and weak generalization ability, is an urgent technical problem that needs to be solved. Summary of the Invention

[0006] This invention provides a method, device, equipment, and storage medium for monitoring faults in traction motors of rail vehicles, aiming to solve the technical problems of insufficient time-frequency analysis accuracy, inadequate feature extraction, poor adaptability to operating conditions, and difficulty in balancing diagnostic accuracy and real-time performance in existing traction motor fault monitoring methods.

[0007] In a first aspect, the present invention provides a method for monitoring faults in a traction motor of a rail vehicle, the method comprising: The traction motor current signal is acquired using a sliding window and preprocessed to obtain the input sequence. The input sequence is fed into a time-frequency joint network to obtain a fault type prediction result. The time-frequency joint network responds to the input sequence by performing an S-transform to obtain a second-order tensor. The magnitude and phase of the second-order tensor are decoupled to construct a third-order tensor. Normalization, convolution, and activation are performed sequentially on the third-order tensor to extract time-frequency depth features. Based on these time-frequency depth features, a time-frequency joint attention mechanism is constructed, calculating attention weights along both the time and frequency dimensions. These attention weights are then used to weight and focus the time-frequency depth features, resulting in weighted time-frequency features. The weighted time-frequency features are then subjected to pooling dimensionality reduction and vector expansion to obtain a one-dimensional feature vector. A fully connected classification is performed on the one-dimensional feature vector, and the fault type prediction result is output after passing through a Softmax activation function.

[0008] In one possible design, a sliding window is used to acquire the traction motor current signal and preprocess it to obtain the input sequence, including: Set the sampling frequency to The motor current data is collected in real time using sensors, and the sliding window width is set to... Get the previous time. Up to the current moment The original time series ,in , M This represents the total number of current sampling points contained within a single sliding window. x t Let be the original sampled value of the traction motor current at the t-th sampling time within the sliding window. t This is the time-series index of the sampling points within the window; The original time series is analyzed using the following formula. Preprocessing: in, This is the mean of the original time series; The standard deviation of the original time series; It is a constant; This represents the normalized current value at the t-th sampling point within the preprocessed window. Obtain the preprocessed input sequence .

[0009] In one possible design, the input sequence is subjected to an S-transform to obtain a second-order tensor. The magnitude and phase of the second-order tensor are then decoupled to construct a third-order tensor, including: The input sequence is processed using the following formula. Perform S-transform: in, The transformed second-order tensor The imaginary unit, For time indexing, For frequency index, e It is a natural constant; Decouple the magnitude and phase of the second-order tensor to construct a third-order tensor. : in, Indicates modulo, Indicates taking the phase. X 1[ m , n ,0] represents The transformed modulus components, X 1[ m , n ,1] indicates The transformed phase component.

[0010] In one possible design, the time-frequency depth features are extracted by sequentially performing normalization, convolution calculation, and activation processing on the third-order tensor, including: Based on the third-order tensor The normalized third-order tensor is obtained by normalizing it using the following formula. : in, , These represent taking the minimum and maximum values, respectively. X 2[ m, n ,0] represents X 1[ m , n The normalized result of [,0] X 2[ m , n ,1] indicates X 1[ m , n The normalized result of [,0]; Convolution and activation are performed on the normalized third-order tensor to extract time-frequency depth features.

[0011] In one possible design, convolution and activation are performed on the normalized third-order tensor to extract time-frequency depth features, specifically including: Constructing convolution kernels According to the normalized third-order tensor Perform convolution calculations to obtain the third-order tensor after convolution. The formula for calculating convolution is as follows: in, Indicates output The bias term, ; k To output the feature channel index, c For input feature channel index, p For row index, q For column indexes, X 3[ m , n , k ]express X 2. Convolutional feature values ​​are calculated at time index m, frequency index n, and output channel k after convolution. C 2[ p , q , n , k ] represents the convolution kernel C 2. Weight parameters at row index p, column index q, input channel c, and output channel k; X 2[ m + p -1, n + q -1, c ] represents a third-order tensor X 2. At the convolution window (m,n), the corresponding kernel position (p,q) and the input feature value of input channel c; The Leaky ReLU function is used to process the convolved third-order tensor. Activation is performed to obtain time-frequency depth features.X 4.

[0012] In one possible design, based on the time-frequency depth features, a time-frequency joint attention mechanism is constructed. Attention weights are calculated along both the time and frequency dimensions, and the time-frequency depth features are weighted and focused using these attention weights to obtain weighted time-frequency features, including: Based on the time-frequency depth features X 4. Take the average along both the time and frequency dimensions to obtain the time-dimensional average result. X time Average results of frequency dimension X freq : in, N For the total number of frequency indices, M The total number of time indices. X 4[:, n ,:]for X 4. Fixed nth frequency, all time and all channels corresponding to feature slices. X 4[ m ,:,:]for X 4. Fixed feature slices corresponding to the m-th time, all frequencies, and all channels; Convolutional kernels are constructed separately for the time dimension and the frequency dimension. , The average result over the time dimension X time and the average result of the frequency dimension X freq Dimensionality reduction is performed, and the Sigmoid function is used for activation to obtain the time dimension weight vector. A time,raw and frequency dimension weight vector A freq,raw : Where * is the convolution operator, which means that the convolution kernel and the corresponding feature matrix are convolved. The time dimension weight vector A time,raw Extending to the full frequency dimension, we obtain the attention weight matrix in the time dimension. A time The frequency dimension weight vector A freq,raw Extending to the full time dimension, we obtain the frequency-dimensional attention weight matrix. A freq : in, This means that the single-point weight value will be copied and extended to all dimensions. A time,raw [0, n [,0] represents the original temporal attention scalar weight value corresponding to the frequency index n. A freq,raw [ m [,0,0] represents the original frequency attention scalar weight value corresponding to time index m; Based on the time-dimensional attention weight matrix A time and the frequency dimension attention weight matrix A freq The time-frequency joint attention weight matrix is ​​calculated using the Hadamard product. ; The time-frequency joint attention weight matrix Applied to the time-frequency depth features X 4. Obtain the weighted time-frequency features: in, This represents the Hadamard product.

[0013] In one possible design, the weighted time-frequency features are sequentially pooled for dimensionality reduction and expanded to obtain a one-dimensional feature vector. This one-dimensional feature vector is then subjected to fully connected classification, and a fault type prediction result is output via a Softmax activation function, including: Max pooling is performed on the weighted time-frequency features to obtain the pooled third-order tensor. : in, , ; P The total time dimension length of the pooled feature X6 is given. Q This represents the total length of the frequency dimension of the pooled feature X6. a This is a local offset index in the time direction. b For local offset index in the frequency direction; X 6[p,q,k] represents the pooling feature values ​​of the third-order tensor X6 at row index p, column index q, and channel k after pooling. X 5[2 p + a ,2 q + b , k [] represents the original input feature value of channel k within the 2×2 pooling window in tensor X5; The pooled third-order tensor is expanded into a one-dimensional vector. The expanded mapping relationship is as follows: in, l For index mapping relationships, ; According to the one-dimensional vector The following formula is used for classification and prediction to obtain the prediction vector Y: in, For size The weight matrix, , For size The bias vector, d The feature length of the one-dimensional vector X7; The predicted vector Y is normalized using the Softmax activation function, and the index of the maximum value in the normalized vector is taken as the fault category.

[0014] Secondly, the present invention provides a fault monitoring device for a rail vehicle traction motor, the device comprising: The data preprocessing module is configured to use a sliding window to acquire the traction motor current signal and perform preprocessing to obtain the input sequence; The fault prediction module is configured to feed the input sequence into a time-frequency joint network to obtain a fault type prediction result. The time-frequency joint network responds to the input sequence by performing an S-transform to obtain a second-order tensor. The magnitude and phase of the second-order tensor are decoupled to construct a third-order tensor. Normalization, convolution, and activation are performed sequentially on the third-order tensor to extract time-frequency depth features. A time-frequency joint attention mechanism is constructed based on these depth features, calculating attention weights along both the time and frequency dimensions. These attention weights are then used to weight and focus the time-frequency depth features, resulting in weighted time-frequency features. The weighted time-frequency features are then subjected to pooling dimensionality reduction and vector expansion to obtain a one-dimensional feature vector. A fully connected classification is performed on the one-dimensional feature vector, and the fault type prediction result is output after passing through a Softmax activation function.

[0015] Thirdly, embodiments of the present invention provide an electronic device, comprising: at least one processor and a memory; the memory storing computer execution instructions; the at least one processor executing the computer execution instructions stored in the memory, causing the at least one processor to perform the rail vehicle traction motor fault monitoring method as described in the first aspect and various possible designs of the first aspect.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the rail vehicle traction motor fault monitoring method described in the first aspect and various possible designs of the first aspect.

[0017] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the rail vehicle traction motor fault monitoring method as described in the first aspect and various possible designs of the first aspect.

[0018] The method, apparatus, equipment, and storage medium for monitoring traction motor faults in rail vehicles provided by this invention have at least the following beneficial effects: (1) This invention performs an S-transform on the input sequence. Due to the high frequency resolution in the low-frequency band and the high time resolution in the high-frequency band, the S-transform can meet the time-frequency analysis requirements of the unsteady current signal of the traction motor. On this basis, this invention further decouples the magnitude and phase of the transformed second-order tensor to construct a third-order tensor, thereby completely preserving the amplitude and phase information of the signal. Through the above technical means, the time-frequency characteristics of early weak faults and complex faults can be effectively captured, which is beneficial to improving the sensitivity of fault identification and overcoming the defects of information loss and insufficient identification sensitivity caused by single signal analysis or conventional time-frequency transformation in the prior art.

[0019] (2) Based on the time-frequency depth features extracted by convolution, this invention performs average pooling along the frequency and time dimensions, reduces the dimensionality of the features by a dimensionality-reducing convolution kernel, activates them with a Sigmoid function, and then expands them to all dimensions to obtain attention weight matrices for the time and frequency dimensions. A joint time-frequency attention weight matrix is ​​then constructed using the Hadamard product, and the time-frequency depth features are weighted accordingly. Since this joint time-frequency attention mechanism can autonomously calculate weights along both the time and frequency dimensions, it can enhance effective fault features, suppress noise and redundant information interference, and maintain stable feature extraction capabilities under complex electromagnetic environments and variable load conditions in vehicles.

[0020] (3) This invention constructs a convolution kernel that matches the third-order tensor structure of the S-transform's magnitude-phase dual-channel for convolution calculation, avoiding the extra computational cost caused by redundant channels; and reduces the dimensionality of time-frequency features by max pooling, thereby reducing the number of parameters in the subsequent fully connected layers. With the above structural settings, while maintaining high diagnostic accuracy, the model has a small number of parameters and fast inference speed, which can meet the deployment requirements of real-time fault monitoring on the rail vehicle's onboard end.

[0021] (4) This invention uses a sliding window to capture current signals in real time and performs standardized preprocessing using the mean and standard deviation, eliminating the need for manual selection of signal segments or manual adjustment of data preprocessing parameters. The entire process, from signal acquisition, time-frequency transformation, feature extraction, attention weighting to fault classification, is automatically completed by preset steps, without the need for manual intervention in feature selection or parameter debugging. Therefore, this invention can operate based on existing onboard current acquisition equipment in rail vehicles without the need for additional complex sensing devices, which helps reduce system modification costs and on-site deployment difficulties. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0023] Figure 1 This is a flowchart illustrating the overall principle of a fault monitoring method for a rail vehicle traction motor, as provided in an embodiment of the present invention. Figure 2 A flowchart of a method for monitoring faults in a traction motor of a rail vehicle provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the time-frequency joint network provided in an embodiment of the present invention; Figure 4 Original signal diagrams of various states provided in embodiments of the present invention; Figure 5 This is a fault classification result diagram provided in an embodiment of the present invention; Figure 6 This is a structural diagram of a rail vehicle traction motor fault monitoring device provided in an embodiment of the present invention.

[0024] The accompanying drawings have illustrated specific embodiments of the present invention, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0026] The collection, storage, use, processing, transmission, provision, and disclosure of relevant data and information in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0027] It should be noted that in the embodiments of the present invention, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of the present invention. However, they do not mean that the inventor has used or necessarily used the solution.

[0028] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0029] Example 1: This invention provides a method for monitoring faults in traction motors of rail vehicles, aiming to solve the technical problems of insufficient time-frequency analysis accuracy, inadequate feature extraction, poor adaptability to operating conditions, and difficulty in balancing diagnostic accuracy and real-time performance in existing traction motor fault monitoring methods. Figure 1As shown, this method first uses a sliding window to capture the real-time current signal collected by the traction motor. Signal standardization preprocessing is then performed through mean and standard deviation optimization to remove noise interference and normalize the signal distribution, resulting in a standardized input sequence. Next, the input sequence is fed into a hierarchical time-frequency joint network for sequential processing. This network includes a sequentially connected S-transform layer, normalization layer, time-frequency joint attention mechanism, pooling layer, unfolding layer, fully connected layer, Softmax layer, and output layer. The S-transform layer first performs time-frequency conversion, decoupling the signal magnitude and phase features to construct a third-order tensor. Then, the normalization layer normalizes the feature values, and convolution calculation and activation processing are combined to extract time-frequency depth features. Finally, the time-frequency joint attention mechanism is used to further process the signal. The method employs a weighted optimization approach for deep features. The time-frequency joint attention mechanism takes the input tensor as its starting point, dividing the features into two paths and performing averaging along the frequency and time dimensions respectively. After averaging, each feature undergoes dimensionality reduction and sigmoid activation, then they are fused to generate a time-frequency joint attention weight matrix. Finally, the time-frequency joint attention is applied to weight and focus the time and frequency dimensions of the features, strengthening the weights of key fault features. The attention-weighted features are then passed through a pooling layer for max pooling dimensionality reduction, an unfolding layer for one-dimensional vector expansion, and finally a fully connected layer to map the high-dimensional features to the category space. A softmax layer then outputs the probability results for various fault types, and finally, the output layer outputs the final fault type. This method, through time-frequency joint feature decoupling and attention weighting, effectively improves the sensitivity of identifying weak and complex faults under varying loads and strong interference conditions. It combines high diagnostic accuracy, strong robustness, and fast inference capabilities, adapting to the real-time fault monitoring needs of rail vehicles. It can accurately achieve rapid identification and early warning of various typical traction motor faults, ensuring the safe and stable operation of rail vehicles.

[0030] Specifically, such as Figure 2 As shown, the fault monitoring method for the traction motor of the rail vehicle can be implemented through the following steps S10-S20.

[0031] S10: Use a sliding window to acquire the traction motor current signal and perform preprocessing to obtain the input sequence.

[0032] In some embodiments, step S10 specifically includes the following steps S11-S12.

[0033] S11: Set the sampling frequency to... The motor current data is collected in real time using sensors, and the sliding window width is set to... Get the previous time. Up to the current moment The original time series ,in , MThis represents the total number of current sampling points contained within a single sliding window. x t Let be the original sampled value of the traction motor current at the t-th sampling time within the sliding window. t This is the timing index of the sampling points within the window.

[0034] S12: For the original time series... Preprocessing is performed using the following formula: in, This is the mean of the original time series; The standard deviation of the original time series; Let be a constant, taken as 1e-6. The normalized current value at the t-th sampling point within the preprocessed window is used to obtain the preprocessed input sequence. .

[0035] S20: Input the input sequence into the time-frequency joint network to obtain the fault type prediction result.

[0036] like Figure 3 The diagram shows a schematic representation of the time-frequency joint network provided in an embodiment of the present invention. (a) is a schematic diagram of the overall structure of the time-frequency joint network, and (b) is a schematic diagram of the time-frequency joint attention mechanism.

[0037] In the overall structure of the time-frequency joint network shown in (a), the data flows sequentially through the input sequence, S-transform layer, normalization layer, time-frequency joint attention mechanism, pooling layer, expansion layer, fully connected layer, Softmax layer, and output layer. The input sequence serves as the network's input entry point. It first enters the S-transform layer to complete the conversion from the time domain signal to the time-frequency domain. The output time-frequency features are sent to the normalization layer to complete the standardization and regularization of the numerical range. The normalized features are input to the time-frequency joint attention mechanism to complete the weighted enhancement of key fault features. The attention-weighted features are then passed through the pooling layer for dimensionality reduction and compression, and through the expansion layer to be converted into a one-dimensional feature vector. Subsequently, they are input to the fully connected layer to complete the linear mapping from the high-dimensional features to the fault category space. The mapping result is converted into the probability distribution of various faults through the Softmax layer. Finally, the output layer outputs the final fault type prediction result.

[0038] In the time-frequency joint attention mechanism structure shown in (b), the input tensor is used as the starting point for processing. The input tensor is divided into two parallel processing paths: one path performs an averaging operation on the features along the frequency dimension. The averaged features are then subjected to dimensionality reduction and Sigmoid activation to generate attention weights in the frequency dimension. The other path performs an averaging operation on the features along the time dimension. The averaged features are then subjected to dimensionality reduction and Sigmoid activation to generate attention weights in the time dimension. The attention weights of the two dimensions are fused to generate a time-frequency joint attention weight matrix. Finally, the time-frequency joint attention is applied to perform element-wise weighting of the input tensor, thereby achieving focused enhancement of key fault features and suppression of noise redundancy features.

[0039] In some embodiments, step S20 specifically includes the following steps S21-S27.

[0040] S21: For the input sequence The S-transform is performed as follows: in, The transformed second-order tensor The imaginary unit, For time indexing, Frequency index; Will Decouple the modulus and phase to construct a third-order tensor. Its value is calculated using the following formula: in, Indicates modulo, Indicates taking the phase. X 1[ m , n ,0] represents The transformed modulus components, X 1[ m , n ,1] indicates The transformed phase component.

[0041] S22: Yes The normalized third-order tensor is obtained after processing. The formula is as follows: in, , These represent taking the minimum and maximum values, respectively. X 2[ m , n ,0] represents X 1[ m ,n The normalized result of [,0] X 2[ m , n ,1] indicates X 1[ m , n The normalized result of ,0].

[0042] S23: Construction dimensions are A convolutional kernel with 2 input channels and 2 output channels. Perform convolution calculations to obtain the third-order tensor after convolution. The formula is as follows: in, Indicates output The bias term, ; k To output the feature channel index, c For input feature channel index, p For row index, q For column indexes, X 3[ m , n , k ]express X 2. Convolutional feature values ​​are calculated at time index m, frequency index n, and output channel k after convolution. C 2[ p , q , n , k ] represents the convolution kernel C 2. Weight parameters at row index p, column index q, input channel c, and output channel k; X 2[ m + p -1, n + q -1, c ] represents a third-order tensor X 2. At the convolution window (m,n), the input feature value corresponding to the kernel position (p,q) and input channel c.

[0043] Activation was performed using the Leaky ReLU function to obtain the time-frequency depth features. .

[0044] S24: Construct a time-frequency joint attention mechanism for... Take the average along the frequency dimension and the time dimension respectively: in, X time The average result over time.X freq This is the average result along the frequency dimension. N For the total number of frequency indices, M The total number of time indices. X 4[:, n ,:]for X 4. Fixed nth frequency, all time and all channels corresponding to feature slices. X 4[ m ,:,:]for X 4. Fixed m-th time, all frequencies and all channels corresponding to feature slices.

[0045] Construct 1×1 convolutional kernels for the time dimension and the frequency dimension respectively. , Dimensionality reduction is performed, and the Sigmoid function is used for activation, as shown in the following formula: in, A time,raw The time-dimension weight vector, A freq,raw The vector represents the frequency dimension weights, and * represents the convolution operator, which means that the convolution kernel performs a convolution operation with the corresponding feature matrix.

[0046] Time-based weight vector A time,raw and frequency dimension weight vector A freq,raw Further extending to the complete frequency and time dimensions, the formula is as follows: in, This means that the single-point weight value will be copied and extended to all dimensions. A time This is the attention weight matrix along the time dimension. A freq For the frequency dimension attention weight matrix, A time,raw [0, n [,0] represents the original temporal attention scalar weight value corresponding to the frequency index n. A freq,raw [ m [,0,0] represents the original frequency attention scalar weight value corresponding to the time index m.

[0047] Calculate the time-frequency joint attention weight matrix The formula is as follows: in, Represents the Hadamard product; The formula for applying time-frequency joint attention is as follows: in, X 5 represents the weighted time-frequency features.

[0048] S25: Yes Perform 2x2 max pooling to obtain the pooled third-order tensor. The formula is as follows: Among them, index , ; P The total time dimension length of the pooled feature X6 is given. Q This represents the total length of the frequency dimension of the pooled feature X6. a This is a local offset index in the time direction. b For local offset index in the frequency direction; X 6[p,q,k] represents the pooling feature values ​​of the third-order tensor X6 at row index p, column index q, and channel k after pooling. X 5[2 p + a ,2 q + b , k ] represents the original input feature value of channel k within the 2×2 pooling window in tensor X5.

[0049] Will Expand into a one-dimensional vector : The index mapping relationship is as follows: .

[0050] S26: Yes The formula for classification prediction is as follows: in, The size is The weight matrix, , The size is The bias vector, The size is ; d Let X be the feature length of the one-dimensional vector X7.

[0051] S27: Calculate the fault category The formula is as follows: in, This represents the Softmax activation function. This indicates the index position of the maximum value in the vector.

[0052] Example 2: This embodiment takes a train traction motor (rated power 250kW, rated speed 1800rpm) as an example and combines specific experimental data to explain in detail the feasibility and progress of the method proposed in Embodiment 1.

[0053] In this embodiment, the sampling frequency is set to... =10kHz, sliding window width is =2048 sampling points, with a sampling time range from t-0.2048s to time t (the current time), resulting in the original time series. =[x(t-2047),x(t-2046),...,x(t)]. The original sequence was calculated to have a mean of μ=15.23A and a standard deviation of σ=3.45A. A constant ε=1e-6 was set to obtain the preprocessed input sequence. .

[0054] like Figure 4 The diagram shown is a representation of the original signal diagrams for various states provided in an embodiment of the present invention, wherein the parameters of the embodiment are... =10kHz N =2048, Figure 4 The horizontal axis represents time (ms), and the vertical axis represents the traction motor current signal value (A). From top to bottom, the current time-domain waveforms are displayed under five operating conditions: normal state, bearing outer ring fault, rolling element fault, bearing inner ring fault, and winding fault. Figure 4 It can be seen that the current signals corresponding to different fault types have distinguishable differences in fluctuation patterns, amplitude characteristics, and harmonic components.

[0055] Based on the above preprocessed input sequence The input time-frequency joint network is used to realize the fault monitoring of the traction motor of the rail vehicle, specifically as follows: steps S21-S27. It should be noted that the specific data processing method of steps S21-S27 is described in the above embodiment 1. In the following S21-S27, only some exemplary parameter settings and corresponding data results are described, and the specific data processing method will not be repeated.

[0056] In step S21, the input sequence is... Perform an S-transform to obtain a second-order tensor. S ( m , n Time-domain index m =0,1,...,2047, Frequency Domain Index n=0,1,...,1023, constructing a third-order tensor ∈R 2048×1024×2 .

[0057] In step S22, calculate The minimum value is 0.0012, the maximum value is 245.678, and the output is a normalized third-order tensor. ∈R 2048×1024×2 .

[0058] In step S23, the convolution kernel is constructed. ∈R (3×3×2×2) Bias term b =[0.1,-0.05], convolution computation and LeakyReLU activation ( α =0.01) .

[0059] In step S24 (time-frequency joint attention), the average result of the time dimension is calculated. X time Average results of frequency dimension X freq Applying 1×1 convolution kernels in both frequency and time dimensions , Perform dimensionality reduction and dimensionality expansion, and calculate the time-frequency joint attention weight matrix. After application, the weighted time-frequency features are obtained. .

[0060] In step S25, 2×2 max pooling is used: pool_size=(2,2), stride=2, pooling output. Expand into a one-dimensional vector .

[0061] In steps S26 and S27, set =5, meaning there are 5 categories, corresponding to the following fault types: label=0: normal state; label=1: bearing outer ring fault; label=2: bearing rolling element fault; label=3: bearing inner ring fault; label=4: winding fault; Classification layer weight matrix ∈ R 5×1048576 Bias vector: ∈ R 5 .

[0062] The probability output of fault classification is as follows: Figure 5 The predicted probabilities for each category are shown below. Figure 5The vertical axis represents the predicted probability, and the horizontal axis corresponds to five categories: normal state, bearing outer ring fault, rolling element fault, bearing inner ring fault, and winding fault. The predicted probability for the normal state is 6.92%, for the bearing outer ring fault it is 5.72%, for the rolling element fault it is 5.55%, for the bearing inner ring fault it is 74.83%, and for the winding fault it is 6.99%. The predicted probability for the bearing inner ring fault is significantly higher than the other categories. Therefore, the output result in this embodiment is label = 3, indicating a bearing inner ring fault.

[0063] Example 3: This invention also provides a fault monitoring device for rail vehicle traction motors, used to implement the methods described in any of the above embodiments, such as... Figure 6 As shown, the device includes: The data preprocessing module 601 is configured to use a sliding window to acquire the traction motor current signal and perform preprocessing to obtain the input sequence; The fault prediction module 602 is configured to feed the input sequence into a time-frequency joint network to obtain a fault type prediction result. The time-frequency joint network responds to the input sequence by performing an S-transform to obtain a second-order tensor. The magnitude and phase of the second-order tensor are decoupled to construct a third-order tensor. Normalization, convolution, and activation are performed sequentially on the third-order tensor to extract time-frequency depth features. A time-frequency joint attention mechanism is constructed based on the time-frequency depth features. Attention weights are calculated along both the time and frequency dimensions, and these attention weights are used to weight and focus the time-frequency depth features, resulting in weighted time-frequency features. The weighted time-frequency features are then subjected to pooling dimensionality reduction and vector expansion to obtain a one-dimensional feature vector. A fully connected classification is performed on the one-dimensional feature vector, and the fault type prediction result is output after passing through a Softmax activation function.

[0064] This invention provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; for example, the processor and the memory communicate via a communication bus.

[0065] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0066] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0067] The electronic device provided in this embodiment of the invention can be the terminal device described in the above embodiments.

[0068] This invention also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the above-described method for monitoring faults in a rail vehicle traction motor.

[0069] This invention also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the rail vehicle traction motor fault monitoring method described in the above embodiments.

[0070] In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0071] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0072] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0073] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute certain steps of the methods of the various embodiments of the present invention.

[0074] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0075] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0076] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0077] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0078] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0079] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring faults in a traction motor of a rail vehicle, characterized in that, The method includes: The traction motor current signal is acquired using a sliding window and preprocessed to obtain the input sequence. The input sequence is fed into a time-frequency joint network to obtain a fault type prediction result. The time-frequency joint network responds to the input sequence by performing an S-transform to obtain a second-order tensor. The magnitude and phase of the second-order tensor are decoupled to construct a third-order tensor. Normalization, convolution, and activation are performed sequentially on the third-order tensor to extract time-frequency depth features. Based on these time-frequency depth features, a time-frequency joint attention mechanism is constructed, calculating attention weights along both the time and frequency dimensions. These attention weights are then used to weight and focus the time-frequency depth features, resulting in weighted time-frequency features. The weighted time-frequency features are then subjected to pooling dimensionality reduction and vector expansion to obtain a one-dimensional feature vector. A fully connected classification is performed on the one-dimensional feature vector, and the fault type prediction result is output after passing through a Softmax activation function.

2. The method for monitoring faults in a rail vehicle traction motor according to claim 1, characterized in that, The traction motor current signal is acquired using a sliding window and preprocessed to obtain the input sequence, including: Set the sampling frequency to The motor current data is collected in real time using sensors, and the sliding window width is set to... Get the previous time. Up to the current moment The original time series ,in , M This represents the total number of current sampling points contained within a single sliding window. x t Let be the original sampled value of the traction motor current at the t-th sampling time within the sliding window. t This is the time-series index of the sampling points within the window; The original time series is analyzed using the following formula. Preprocessing: in, This is the mean of the original time series; The standard deviation of the original time series; It is a constant; This represents the normalized current value at the t-th sampling point within the preprocessed window. Obtain the preprocessed input sequence .

3. The method for monitoring faults in a rail vehicle traction motor according to claim 2, characterized in that, The input sequence is subjected to an S-transform to obtain a second-order tensor. The magnitude and phase of the second-order tensor are decoupled to construct a third-order tensor, including: The input sequence is processed using the following formula. Perform S-transform: in, The transformed second-order tensor The imaginary unit, For time indexing, For frequency index, e It is a natural constant; Decouple the magnitude and phase of the second-order tensor to construct a third-order tensor. : in, Indicates modulo, Indicates taking the phase. X 1[ m , n ,0] represents The transformed modulus components, X 1[ m , n ,1] indicates The transformed phase component.

4. The method for monitoring faults in a rail vehicle traction motor according to claim 3, characterized in that, Based on the third-order tensor, normalization, convolution, and activation are performed sequentially to extract time-frequency depth features, including: Based on the third-order tensor The normalized third-order tensor is obtained by normalizing it using the following formula. : in, , These represent taking the minimum and maximum values, respectively. X 2[ m , n ,0] represents X 1[ m , n The normalized result of [,0] X 2[ m , n ,1] indicates X 1[ m , n The normalized result of [,0]; Convolution and activation are performed on the normalized third-order tensor to extract time-frequency depth features.

5. The method for monitoring faults in a rail vehicle traction motor according to claim 4, characterized in that, Convolution and activation are performed on the normalized third-order tensor to extract time-frequency depth features, specifically including: Constructing convolution kernels According to the normalized third-order tensor Perform convolution calculations to obtain the third-order tensor after convolution. The formula for calculating convolution is as follows: in, Indicates output The bias term, ; k To output the feature channel index, c For input feature channel index, p For row index, q For column indexes, X 3[ m , n , k ]express X 2. Convolutional feature values ​​are calculated at time index m, frequency index n, and output channel k after convolution. C 2[ p , q , n , k ] represents the convolution kernel C 2. Weight parameters at row index p, column index q, input channel c, and output channel k; X 2[ m + p -1, n + q -1, c ] represents a third-order tensor X 2. At the convolution window (m,n), the corresponding kernel position (p,q) and the input feature value of input channel c; The Leaky ReLU function is used to process the convolved third-order tensor. Activation is performed to obtain time-frequency depth features. X 4.

6. The method for monitoring faults in a rail vehicle traction motor according to claim 5, characterized in that, Based on the time-frequency depth features, a time-frequency joint attention mechanism is constructed. Attention weights are calculated along both the time and frequency dimensions, and the time-frequency depth features are weighted and focused using these attention weights to obtain weighted time-frequency features, including: Based on the time-frequency depth features X 4. Take the average along both the time and frequency dimensions to obtain the time-dimensional average result. X time Average results of frequency dimension X freq : in, N For the total number of frequency indices, M The total number of time indices. X 4[:, n ,:]for X 4. Fixed nth frequency, all time and all channels corresponding to feature slices. X 4[ m ,:,:]for X 4. Fixed feature slices corresponding to the m-th time, all frequencies, and all channels; Convolutional kernels are constructed separately for the time dimension and the frequency dimension. , The average result over the time dimension X time and the average result of the frequency dimension X freq Dimensionality reduction is performed, and the Sigmoid function is used for activation to obtain the time dimension weight vector. A time,raw and frequency dimension weight vector A freq,raw : Where * is the convolution operator, which means that the convolution kernel and the corresponding feature matrix are convolved. The time dimension weight vector A time,raw Extending to the full frequency dimension, we obtain the attention weight matrix in the time dimension. A time The frequency dimension weight vector A freq,raw Extending to the full time dimension, we obtain the frequency-dimensional attention weight matrix. A freq : in, This means that the single-point weight value will be copied and extended to all dimensions. A time,raw [0, n [,0] represents the original temporal attention scalar weight value corresponding to the frequency index n. A freq,raw [ m [,0,0] represents the original frequency attention scalar weight value corresponding to time index m; Based on the time-dimensional attention weight matrix A time and the frequency dimension attention weight matrix A freq The time-frequency joint attention weight matrix is ​​calculated using the Hadamard product. ; The time-frequency joint attention weight matrix Applied to the time-frequency depth features X 4. Obtain the weighted time-frequency features: in, This represents the Hadamard product.

7. The method according to claim 6, characterized in that, The weighted time-frequency features are sequentially subjected to pooling dimensionality reduction and vector expansion to obtain a one-dimensional feature vector. This one-dimensional feature vector is then subjected to fully connected classification, and the fault type prediction result is output after passing through a Softmax activation function, including: Max pooling is performed on the weighted time-frequency features to obtain the pooled third-order tensor. : in, , ; P The total time dimension length of the pooled feature X6 is given. Q This represents the total length of the frequency dimension of the pooled feature X6. a This is a local offset index in the time direction. b For local offset index in the frequency direction; X 6[p,q,k] represents the pooling feature values ​​of the third-order tensor X6 at row index p, column index q, and channel k after pooling. X 5[2 p + a ,2 q + b , k [] represents the original input feature value of channel k within the 2×2 pooling window in tensor X5; The pooled third-order tensor is expanded into a one-dimensional vector. The expanded mapping relationship is as follows: in, l This is an index mapping relationship. ; According to the one-dimensional vector The following formula is used for classification and prediction to obtain the prediction vector Y: in, For size The weight matrix, , For size The bias vector, d The feature length of the one-dimensional vector X7; The predicted vector Y is normalized using the Softmax activation function, and the index of the maximum value in the normalized vector is taken as the fault category.

8. A fault monitoring device for traction motors of rail vehicles, characterized in that, The device includes: The data preprocessing module is configured to use a sliding window to acquire the traction motor current signal and perform preprocessing to obtain the input sequence; The fault prediction module is configured to feed the input sequence into a time-frequency joint network to obtain a fault type prediction result. The time-frequency joint network responds to the input sequence by performing an S-transform to obtain a second-order tensor. The magnitude and phase of the second-order tensor are decoupled to construct a third-order tensor. Normalization, convolution, and activation are performed sequentially on the third-order tensor to extract time-frequency depth features. A time-frequency joint attention mechanism is constructed based on these depth features, calculating attention weights along both the time and frequency dimensions. These attention weights are then used to weight and focus the time-frequency depth features, resulting in weighted time-frequency features. The weighted time-frequency features are then subjected to pooling dimensionality reduction and vector expansion to obtain a one-dimensional feature vector. A fully connected classification is performed on the one-dimensional feature vector, and the fault type prediction result is output after passing through a Softmax activation function.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes the computer execution instructions stored in the memory to implement the rail vehicle traction motor fault monitoring method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the rail vehicle traction motor fault monitoring method as described in any one of claims 1-7.

Images