Electric locomotive fault early warning method based on machine learning

Abstract

The invention discloses an electric locomotive fault early warning method based on machine learning, which comprises the following steps: acquiring a real-time vibration signal of an electric locomotive bearing, and generating a time-frequency characteristic matrix containing time domain and frequency domain coupling characteristics according to the real-time vibration signal; calling a fault recognition model, and estimating the probability that the electric locomotive bearing has different fault categories according to the time-frequency characteristic matrix; performing time sequence dynamic mode coding processing on the real-time vibration signal to generate a time sequence fault feature coding vector, and generating a fault evolution prediction feature based on the time sequence fault feature coding vector; according to the fault evolution prediction characteristics, deducing the abnormal confidence of different fault types of the electric locomotive bearing; and according to the abnormal confidence coefficient of different fault types of the electric locomotive bearing and the estimated probability of different fault types of the electric locomotive bearing, the actual fault type of the electric locomotive bearing is judged, and fault early warning is generated according to the fault type, so that the accuracy and timeliness of fault early warning are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a machine learning-based electric locomotive fault early warning method. BACKGROUND

[0002] As the core equipment of railway transportation, the running safety and reliability of electric locomotives directly affect the transportation efficiency and operation safety. Bearings, as the key components of the transmission system of electric locomotives, are prone to failure due to wear and fatigue under long-term alternating load and high-speed friction. If not timely warned, it may lead to equipment downtime or even traffic accidents. Therefore, real-time condition monitoring and accurate fault early warning of electric locomotive bearings are important technical means to ensure railway transportation safety.

[0003] In the prior art, the fault early warning of electric locomotive bearings mostly adopts a threshold judgment method based on a single feature. This method usually extracts time domain features (such as peak value, root mean square value) or frequency domain features (such as characteristic frequency amplitude) of the vibration signal, compares the signal features by setting a fixed threshold, and triggers an early warning when the feature value exceeds the threshold. Some improved schemes introduce a simple statistical model to fit historical fault data to optimize the threshold, but still take single-dimensional feature analysis as the core.

[0004] The above prior art solutions have obvious limitations: due to the complex operating conditions of electric locomotives (such as load fluctuation and speed change), the stability of a single feature is poor, and a fixed threshold is difficult to adapt to signal changes under different conditions, which is prone to false positives or false negatives; at the same time, relying on a single feature cannot fully reflect the evolution process of the fault, making it difficult to accurately identify early faults and resulting in insufficient timeliness of the early warning, which cannot provide sufficient time for maintenance decisions. SUMMARY

[0005] To solve the above technical problems, the present application provides a machine learning-based electric locomotive fault early warning method to at least alleviate the above technical problems.

[0006] The technical scheme provided by the embodiments of the present application is as follows: A machine learning-based electric locomotive fault early warning method, the method comprising: Step 1: acquiring real-time vibration signals of electric locomotive bearings, and generating a time-frequency feature matrix containing time domain and frequency domain coupled features according to the real-time vibration signals; Step 2: calling a fault identification model to estimate the probability of different fault categories of the electric locomotive bearings according to the time-frequency feature matrix; Step 3: performing time series dynamic pattern coding processing on the real-time vibration signals to generate a time series fault feature coding vector, and generating a fault evolution prediction feature based on the time series fault feature coding vector; Step 4: According to the fault evolution prediction feature, the abnormal confidence of the electric locomotive bearing appearing different fault categories is derived; Step 5: According to the abnormal confidence of the electric locomotive bearing appearing different fault categories and the estimated probability of the electric locomotive bearing appearing different fault categories, the actual fault category of the electric locomotive bearing is determined and a fault warning is generated accordingly.

[0007] In this application, the fusion representation of multi-dimensional features is realized by generating a time-frequency feature matrix containing time-domain and frequency-domain coupling features, avoiding the limitations of single features under complex working conditions. In addition, the determination is combined with abnormal confidence and fault probability, rather than relying on a fixed threshold, which can dynamically adapt to differences in working conditions such as load fluctuations and speed changes, reducing false positives and false negatives. Furthermore, the application generates fault evolution prediction features through time sequence dynamic pattern coding, which can capture the dynamic change law of faults over time. By deriving abnormal confidence based on evolution features, it can identify feature anomalies in the early stages of faults, and combined with a comprehensive determination mechanism, it realizes the whole process tracking of the fault evolution process, improves the accuracy of early fault identification, and provides sufficient time for maintenance decisions. In addition, intelligent analysis of multi-source features by the fault identification model replaces traditional single-dimensional analysis, making fault determination more in line with actual fault patterns, further improving the reliability and effectiveness of the warning. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 The flowchart of the electric locomotive fault warning method based on machine learning of the embodiments of the present application.

[0009] Figure 2 The structural diagram of the electric locomotive fault warning system based on machine learning of the embodiments of the present application. DETAILED DESCRIPTION

[0010] As Figure 1 shown, the embodiments of the present application provide a machine learning-based electric locomotive fault warning method, which comprises: Step 1: Obtain the real-time vibration signal of the electric locomotive bearing, and generate a time-frequency feature matrix containing time-domain and frequency-domain coupling features according to the real-time vibration signal; Step 2: Call the fault identification model, and estimate the probability of the electric locomotive bearing appearing different fault categories according to the time-frequency feature matrix; Step 3: Perform time sequence dynamic pattern coding processing on the real-time vibration signal to generate a time sequence fault feature coding vector, and generate a fault evolution prediction feature based on the time sequence fault feature coding vector; Step 4: According to the fault evolution prediction feature, the abnormal confidence of the electric locomotive bearing appearing different fault categories is derived; Step 5: According to the abnormal confidence of different fault categories of the electric locomotive bearing and the estimated probability of different fault categories of the electric locomotive bearing, the actual fault category of the electric locomotive bearing is determined, and a fault warning is generated accordingly.

[0011] Optionally, step 1: Obtain the real-time vibration signal of the electric locomotive bearing, and generate a time-frequency feature matrix containing time-domain and frequency-domain coupling features according to the real-time vibration signal, specifically including the following steps: Step 11: Filter out the environmental noise and interference components in the real-time vibration signal to obtain a denoised vibration signal; Step 12: Decompose the denoised vibration signal into a plurality of sub-signals containing different frequency components to obtain multi-band wavelet components; Step 13: Construct the time-frequency distribution features of the multi-band wavelet components to obtain a time-frequency feature matrix containing time-domain and frequency-domain coupling features, wherein the matrix elements in the time-frequency feature matrix represent the signal energy density of the corresponding time point and frequency band.

[0012] Optionally, step 11: Filter out the environmental noise and interference components in the real-time vibration signal to obtain a denoised vibration signal, specifically including the following steps: Step 111: Decompose the real-time vibration signal into a plurality of eigenmodes, and select the eigenmode components dominated by noise by calculating the kurtosis values of the eigenmodes; Step 112: Based on the eigenmode components dominated by noise, calculate the adaptive noise standard deviation estimate value that changes dynamically over time; Step 113: Based on the adaptive noise standard deviation estimate value, recombine the components to obtain a denoised vibration signal.

[0013] Preferably, in step 111, the acquired real-time vibration signal of the electric locomotive bearing is adaptively decomposed using the empirical mode decomposition method to be decomposed into a plurality of eigenmode components with different characteristic scales, which can reflect the local characteristics of the signal in different frequency bands; for each eigenmode component obtained by decomposition, the kurtosis value thereof is calculated, the kurtosis value can reflect the steepness of the signal distribution and the pulse characteristics, and the kurtosis value of the component dominated by noise is usually low and close to 3, while the kurtosis value of the useful component containing fault information is high; by setting a kurtosis value threshold, the eigenmode components with kurtosis values lower than the threshold are selected as the eigenmode components dominated by noise. Preferably, in step 112, for the noise-dominant intrinsic modal component screened out in step 111, the signal amplitude thereof is analyzed, and a sliding time window is adopted to locally statistically analyze the amplitude distribution of the component at different times; in each sliding window, the standard deviation of the amplitude of the noise-dominant component in the window is calculated to serve as an estimation of the noise intensity near the time; with the movement of the sliding window on the time axis, the estimation of the noise standard deviation is continuously updated, thereby obtaining an adaptive noise standard deviation estimation dynamically changing with time, which can reflect the noise level at different times in real time. Preferably, in step 113, all the intrinsic modal components obtained through original decomposition are processed by using the adaptive noise standard deviation estimation dynamically changing with time obtained in step 112, for the intrinsic modal component determined to be noise-dominant, the amplitude thereof is weighted and suppressed according to the noise standard deviation estimation corresponding thereto, so as to reduce the contribution of the component in the overall signal, while for the intrinsic modal component containing useful information with a kurtosis value higher than the threshold, the amplitude thereof remains unchanged; all the intrinsic modal components (including the suppressed noise-dominant components and the useful components not processed) after the above processing are linearly superimposed according to the original time scale thereof, the component recombination is completed, and finally the noise-reduced vibration signal with most of the environmental noise and interference components removed is obtained.

[0014] Optionally, step 12: decomposing the noise-reduced vibration signal into a plurality of sub-signals containing different frequency components to obtain multi-band wavelet components, specifically comprising the following steps: Step 121: adapting the frequency characteristics of the noise-reduced vibration signal, and dynamically selecting an optimal base function from preset wavelet bases to calculate the decomposition level through signal dominant frequency analysis; Step 122: performing variable-scale wavelet decomposition on the noise-reduced vibration signal based on the decomposition level to obtain initial wavelet components; Step 123: purifying the initial wavelet components to retain effective components to obtain multi-band wavelet components.

[0015] Preferably, in step 121, the noise-reduced vibration signal is converted to the frequency domain through fast Fourier transform, the frequency energy distribution is analyzed to determine the dominant frequency component of the signal, the wavelet base function with the highest matching degree to the frequency characteristics of the signal is selected from the preset Daubechies, Symlet, Coiflet, etc. wavelet base function library according to the characteristics of the dominant frequency component, and the optimal decomposition level is calculated in combination with the sampling frequency and the frequency band width of the signal, so as to ensure that the subsequent wavelet decomposition can effectively capture the characteristic information of the signal. Preferably, in step 122, based on the number of decomposition layers determined in step 121, a Mallat algorithm is applied to the denoised vibration signal to perform a variable-scale wavelet decomposition, which performs a multi-resolution analysis of the signal at different scales and frequencies, and each layer of decomposition generates an approximation coefficient and a detail coefficient, and through multi-layer decomposition, initial wavelet components of different frequency bands are obtained, which correspond to the feature representation of the signal at different frequency bands respectively. Preferably, in step 123, threshold processing is performed on the initial wavelet components, the energy entropy value of each component is calculated, the components with energy entropy lower than a set threshold are determined as noise-dominant components and are removed, and at the same time, the frequency band segmentation characteristic of wavelet packet decomposition is applied to perform correlation analysis on the components of adjacent frequency bands, and highly correlated redundant components are merged, and finally, effective components with obvious fault characteristic frequencies are retained to form multi-band wavelet components.

[0016] Optionally, step 13: constructing a time-frequency distribution feature of the multi-band wavelet components to obtain a time-frequency feature matrix containing time-domain and frequency-domain coupling features, specifically including the following steps: Step 131: mapping the multi-band wavelet components to a time-frequency plane, performing instantaneous frequency analysis on each frequency band wavelet component through Hilbert transform to generate a time-frequency energy distribution; Step 132: quantifying the time-frequency energy distribution features based on the time-frequency energy distribution to construct a three-dimensional energy density tensor; Step 133: generating a time-frequency feature matrix containing time-domain and frequency-domain coupling features based on the three-dimensional energy density tensor.

[0017] Preferably, in step 131, the multi-band wavelet components are taken as processing objects, and the wavelet components of each frequency band are mapped to a two-dimensional time-frequency plane according to their corresponding time series and frequency ranges, so that each component forms a specific distribution area on the time-frequency plane; the Hilbert transform is applied to each frequency band wavelet component on the time-frequency plane, and through the transform, the instantaneous frequency value of each component at any time is obtained, which can reflect the instantaneous change speed of the signal at that time; the instantaneous frequency value at each time is multiplied by the amplitude of the corresponding wavelet component to obtain the energy value of each frequency point at each time, and the distribution of these energy values on the time-frequency plane is the time-frequency energy distribution. Preferably, in step 132, the time-frequency energy distribution is taken as the processing object, quantization processing is performed thereon, fixed-length time windows are divided along the time axis, fixed-bandwidth frequency intervals are divided along the frequency axis, the total energy in each time window and frequency interval intersection region is calculated, and an energy density value of the region is obtained; the energy density values are arranged in three dimensions in the order of time window, frequency interval and corresponding energy density value, wherein the first dimension is the time window number, the second dimension is the frequency interval number, and the third dimension is the energy density value of the space-time intersection region, thereby constructing a three-dimensional energy density tensor. Preferably, in step 133, the three-dimensional energy density tensor obtained in step 132 is taken as the processing object, the corresponding relationship of the time dimension, the frequency dimension and the energy density value is retained, the three-dimensional tensor is unfolded along the energy density dimension, the time window order is taken as the row of the matrix, the frequency interval order is taken as the column of the matrix, and the value of each element in the matrix is the energy density value of the corresponding time window and frequency interval. The matrix generated in this arrangement manner is a time-frequency feature matrix containing time domain and frequency domain coupling characteristics, and each element in the matrix can directly reflect the signal energy density of a specific time point and frequency band.

[0018] Optionally, step 2: calling a fault identification model, estimating the probability of different fault categories of the electric locomotive bearing according to the time-frequency feature matrix, specifically including the following steps: Step 21: based on the spatial attention layer in the fault identification model, dynamically assigning weights to the fault sensitive regions in the time-frequency matrix to obtain an attention weighted feature map; Step 22: based on the attention weighted feature map, converting to a fault semantic feature vector through a cross-scale feature fusion network in the fault identification model; Step 23: mapping the fault semantic feature vector to a fault category probability space, and combining a category correlation matrix to obtain the probability of different fault categories of the electric locomotive bearing.

[0019] Optionally, step 21: based on the spatial attention layer in the fault identification model, dynamically assigning weights to the fault sensitive regions in the time-frequency matrix to obtain an attention weighted feature map, specifically including the following steps: Step 211: based on the multi-scale pooling and feature fusion unit in the spatial attention layer, perceiving the spatial distribution features of the time-frequency feature matrix, and extracting time-frequency joint distribution features through multi-scale pooling operation to obtain a spatial feature descriptor; Step 212: based on the fault guided convolution unit in the spatial attention layer, assigning sensitive region weights through a convolution network embedded with fault priori knowledge according to the spatial feature descriptor to obtain a spatial attention weight map; Step 213: Based on the feature enhancement operation unit in the spatial attention layer, the spatial attention weight map is element-wise enhanced with the time-frequency feature matrix to obtain an attention weighted feature map.

[0020] Preferably, in step 211, the time-frequency feature matrix is input into a multi-scale pooling and feature fusion unit in the spatial attention layer. The unit includes parallel average pooling and maximum pooling operations. The average pooling operation calculates the average value of the local region of the time-frequency feature matrix, and the maximum pooling operation extracts the maximum value of the local region. The multi-scale distribution features of the time-frequency feature matrix in the time and frequency dimensions are captured through different sizes of the pooling window. The feature maps obtained by different pooling operations are spliced in the channel dimension to form a feature tensor containing multi-scale information. A 1x1 convolution is applied to the feature tensor for dimension reduction processing, which compresses the channel number of the feature tensor while keeping its spatial dimension unchanged, to obtain a spatial feature descriptor capable of representing the spatial distribution characteristics of the time-frequency feature matrix. Preferably, in step 212, the spatial feature descriptor is input into a fault-guided convolution unit. The convolution kernel parameters of the unit are obtained by pre-training with historical fault data. The convolution kernel embeds typical feature patterns of different fault types in the time-frequency domain, such as specific frequency components corresponding to outer ring faults, time-domain impact characteristics of inner ring faults, etc. The spatial feature descriptor is convolved with these convolution kernels embedding prior knowledge of faults to generate multiple feature maps, each corresponding to the response strength of a fault-sensitive pattern. A Sigmoid activation function is applied to these feature maps to map the feature values to the 0-1 interval, obtaining the sensitivity weight of each spatial position to different fault types. These weight values are arranged according to the spatial position to form a spatial attention weight map. The higher the value in the region of the map, the more sensitive it is to faults. Preferably, in step 213, the time-frequency feature matrix and the spatial attention weight map are input into a feature enhancement operation unit. The unit element-wise multiplies the time-frequency feature matrix and the spatial attention weight map. The feature values in the time-frequency feature matrix corresponding to the high weight value regions (i.e. fault sensitive regions) in the spatial attention weight map are amplified, while the feature values in the low weight value regions are suppressed. Batch normalization is performed on the result of the element-wise multiplication to standardize the feature values by calculating the mean and variance of the batch data, making the feature value distribution more stable. A ReLU activation function is then applied to introduce a nonlinear transformation, enhancing the model's ability to express complex fault features. Finally, an attention weighted feature map is obtained after attention enhancement processing.

[0021] Optionally, step 22: based on the attention weighted feature map, a fault semantic feature vector is converted through a cross-scale feature fusion network in the fault recognition model, which specifically includes the following steps: Step 221: Extract multi-scale features of the attention-weighted feature map, extract features of different receptive fields through multiple parallel convolution branches, and obtain multi-scale features; Step 222: Based on the multi-scale features, a gating fusion branch is used for weighted aggregation to generate a fusion feature tensor; Step 223: The fusion feature tensor is subjected to semantic alignment through an attention mechanism embedded with prior fault categories, and is converted into a fault semantic feature vector. Preferably, in step 221, the attention-weighted feature map is input into multiple parallel convolution branches of the cross-scale feature fusion network, each convolution branch is configured with a convolution kernel of different size (such as 3x3, 5x5, 7x7) to obtain features of different receptive fields, wherein the 3x3 convolution kernel captures local detail features, and the 5x5 and 7x7 convolution kernels capture medium-range and large-range context features, respectively; each convolution branch sequentially performs convolution operation, batch normalization processing (stabilizing feature distribution) and ReLU activation function (introducing nonlinear transformation) on the input attention-weighted feature map to generate a feature map of corresponding scale; the feature maps output by the multiple convolution branches are spliced in the channel dimension to obtain multi-scale features containing multi-scale spatial information. Preferably, in step 222, the multi-scale features are input into the gating fusion branch, which first performs channel compression (reduces feature dimension) on the multi-scale features through 1x1 convolution, and then generates a gating weight map (weight value range 0-1) consistent with the size of the multi-scale features through the Sigmoid activation function, wherein the higher the weight value, the more important the feature at the corresponding position; the multi-scale features are multiplied element by element with the gating weight map to strengthen important features and suppress secondary features, and a weighted feature map is obtained; 1x1 convolution is applied to the weighted feature map to integrate channel information, while introducing residual connection (adding the original multi-scale features to the weighted feature map) to avoid feature degradation, and finally a fusion feature tensor that integrates multi-scale key information is output through the ReLU activation function. Preferably, in step 223, the fusion feature tensor is input into the attention mechanism embedded with prior fault categories, which pre-defines prototype feature vectors (typical semantic features extracted based on historical fault data) of fault categories such as inner ring, outer ring and roller; the cosine similarity between the feature vector at each spatial position in the fusion feature tensor and each fault prototype feature vector is calculated to obtain an attention weight matrix (each row represents the correlation degree of the feature at the corresponding position with different fault categories); the fusion feature tensor is weighted and summed through the attention weight matrix to aggregate the feature information in the spatial dimension, and a semantic feature vector corresponding to the number of fault categories is obtained; the semantic feature vectors are subjected to L2 normalization (unified vector scale) to finally convert into fault semantic feature vectors that can represent the semantic information of different fault categories.

[0022] Optionally, step 23: mapping the fault semantic feature vector to a fault category probability space to obtain probabilities of different fault categories of the electric locomotive bearing in combination with the category correlation matrix, specifically comprising the following steps: Step 231: mapping the fault semantic feature vector to an initial probability space to calculate similarity through a fault category embedding vector and generate an initial fault probability distribution accordingly; Step 232: based on the initial fault probability distribution, cross-correction is performed in combination with the dynamic category correlation matrix to obtain an intermediate correction probability; Step 233: uncertainty calibration and bias compensation are performed on the intermediate correction probability to obtain probabilities of different fault categories of the electric locomotive bearing. Preferably, in step 231, the fault semantic feature vector is input to a fully connected layer for linear transformation to map it to an initial probability space with the same number of dimensions as the fault categories, so that each dimension corresponds to a fault category; cosine similarity calculation is performed between the mapped vector and a pre-trained fault category embedding vector (each category corresponds to a unique feature vector) to obtain similarity scores of the fault semantic feature vector with each fault category; the Softmax function is applied to the similarity scores for normalization to convert the scores into a probability distribution form, so that the sum of the probabilities of all fault categories is 1, thereby generating an initial fault probability distribution that reflects the matching degree of the fault semantic feature vector with each fault category. Preferably, in step 232, matrix multiplication is performed between the initial fault probability distribution and the dynamic category correlation matrix, which is a symmetric matrix whose elements represent the degree of association between different fault categories (such as the probability of simultaneous occurrence of inner ring fault and rolling body fault), and the matrix is updated in real time through historical fault data; the matrix multiplication operation allows the probability of each fault category to be adjusted by referring to the information of other related categories, for example, if a category has strong correlation with the current high-probability category, its probability will be correspondingly increased; row normalization is performed on the matrix multiplication result to ensure that the adjusted probability values still maintain the characteristics of probability distribution, thereby obtaining an intermediate correction probability that has been corrected by the correlation between categories. Preferably, in step 233, the intermediate correction probability is input to an uncertainty calibration module, which statistically calibrates the deviation between the prediction confidence and the actual accuracy of each fault category based on historical prediction data, constructs a calibration curve (such as Platt scaling or Isotonic regression), calibrates the intermediate correction probability, reduces the prediction deviation of the model being too confident or conservative; introduce temperature scaling parameters to smooth the calibrated probability, adjust the relative gap between different categories of probability, make the prediction result more reasonable; bias compensation is performed on the calibrated and smoothed probability, and a compensation factor (such as the influence of load, speed and other factors on fault probability) based on the running state of the equipment and environmental parameters is added, and finally the probability of different fault categories of the electric locomotive bearing is obtained considering the uncertainty and actual working condition.

[0023] Optionally, step 3: performing time sequence dynamic mode encoding processing on the real-time vibration signal to generate a time sequence fault feature encoding vector, and generating a fault evolution prediction feature based on the time sequence fault feature encoding vector, specifically including the following steps: Step 31: capturing the time sequence dynamic feature of the real-time vibration signal, extracting the signal change trend at different times through a bidirectional gated recurrent unit to obtain a time sequence dynamic feature sequence; Step 32: based on the time sequence dynamic feature sequence, aggregating key time sequence patterns through a self-attention mechanism to obtain a time sequence fault feature encoding vector; Step 33: based on the time sequence fault feature encoding vector, extracting multi-scale time sequence evolution patterns through a time convolution network to generate a fault evolution prediction feature.

[0024] Optionally, step 31: capturing the time sequence dynamic feature of the real-time vibration signal, extracting the signal change trend at different times through a bidirectional gated recurrent unit to obtain a time sequence dynamic feature sequence, specifically including the following steps: Step 311: aligning the time sequence length of the real-time vibration signal, using a dynamic time warping algorithm to eliminate time sequence offset to obtain an equal-length time sequence signal; Step 312: based on the equal-length time sequence signal, extracting time sequence dependent features at different times through a bidirectional gated recurrent unit to obtain bidirectional time sequence hidden states; Step 313: fusing the bidirectional time sequence hidden states to obtain a time sequence dynamic feature sequence.

[0025] Preferably, in step 311, the real-time vibration signal is subjected to timing length analysis to determine the differences between the real-time vibration signal and the preset reference timing length, which are mainly caused by the timing offset due to non-synchronous sampling; for the real-time vibration signal with differences, a dynamic time warping algorithm is used to calculate the similarity distance between the signal and the reference timing signal, the time axis of the signal is adaptively adjusted through an iterative optimization process to align the key feature points of the real-time vibration signal with the corresponding feature points of the reference timing signal; the real-time vibration signal after the time axis adjustment is truncated or completed according to the reference timing length to eliminate the timing misalignment problem caused by different sampling periods, and finally an equal-length timing signal consistent with the reference timing length is obtained. Preferably, in step 312, the equal-length timing signal is input into a bidirectional gate recurrent unit, which includes two independent gate recurrent sub-networks, a forward sub-network and a backward sub-network, the forward sub-network processes the signal from the starting time to the ending time step by step, and the backward sub-network processes the signal from the ending time to the starting time in reverse; in the processing of the forward sub-network, the key feature information of the historical time is selectively retained through the gating mechanism (including the reset gate and the update gate), the transmission of noise interference such as burst pulse is inhibited, and the forward hidden state reflecting the historical change trend of the signal is generated; at the same time, the backward sub-network also processes the feature association of the future time through the gating mechanism, and generates the backward hidden state reflecting the future change trend of the signal; the forward hidden state and the backward hidden state are correspondingly combined according to the time step to form a bidirectional timing hidden state containing the historical and future association information of each time, which can fully capture the timing dependence features of the equal-length timing signal at different times. Preferably, in step 313, the forward hidden state and the backward hidden state in the bidirectional timing hidden state are subjected to element-level addition operation to preliminarily fuse the feature information of the two directions, and obtain a preliminarily fused hidden state containing bidirectional feature superimposed information; the preliminarily fused hidden state is input into a preset fully connected layer for dimension mapping, the dimension of the hidden state is compressed to 1 / 2 of the original dimension, the feature redundancy is reduced and the key dynamic features are highlighted, and a dimension-compressed hidden state is obtained; the residual connection is introduced into the dimension-compressed hidden state to superimpose the features of the corresponding dimensions in the preliminarily fused hidden state and the dimension-compressed hidden state, so as to avoid the feature degradation problem in the deep network, and then the ReLU activation function is used to enhance the nonlinear expression ability of the features, and finally a timing dynamic feature sequence reflecting the dynamic change law of the signal is obtained.

[0026] Optionally, step 32: based on the timing dynamic feature sequence, a key timing pattern is aggregated through a self-attention mechanism to obtain a timing fault feature encoding vector, specifically including the following steps: Step 321: associate the features at different time points in the time sequence of dynamic features, calculate the correlation weight by scaling the dot product attention, and generate a time sequence attention weight matrix; Step 322: based on the time sequence attention weight matrix, weight and aggregate the time sequence of dynamic features to obtain an attention aggregated feature; Step 323: map the attention aggregated feature to a fixed dimensional space to obtain a time sequence fault feature encoding vector.

[0027] Preferably, in step 321, the time sequence of dynamic features is linearly transformed to convert it into three vector sequences of the same dimension, which are used as query vector sequence, key vector sequence and value vector sequence respectively. The query vector is used to locate the information that needs to be focused on at the current time point, the key vector is used to describe the attributes of the features at each time point, and the value vector is used to provide specific feature content. The query vector sequence and the key vector sequence are subjected to matrix multiplication operation to obtain the original similarity score between the features at different time points, which reflects the correlation degree of the features at each time point with the current time point. The original similarity score is scaled (divided by the square root of the vector dimension) to avoid the problem of too large score value caused by too high vector dimension, and the scaled similarity score is obtained. The scaled similarity score is subjected to Softmax function normalization to convert the score into a probability distribution form, so that the sum of the weights of all time points corresponding to each time point is 1, and a time sequence attention weight matrix is generated, which can represent the correlation strength of the features at different time points. Preferably, in step 322, the time sequence attention weight matrix is subjected to matrix multiplication operation with the value vector sequence obtained in step 321, so that the feature vector at each time point aggregates the feature information of other time points according to the corresponding weight in the weight matrix, and a preliminary aggregated feature sequence is obtained, in which each feature is fused with the feature content of the key related time point. The time sequence attention weight matrix and the value vector sequence are divided into multiple subsets, and matrix multiplication operation is performed in a multi-head parallel calculation manner to obtain multiple aggregated feature subsequences, and then these subsequences are spliced in the channel dimension to capture the correlation patterns in different feature subspaces, and a multi-head aggregated feature is obtained. Linear transformation is applied to the multi-head aggregated feature to integrate the channel information, and residual connection is introduced (the linear transformation result is added to the original time sequence of dynamic features) to avoid feature information loss, and finally layer normalization processing is performed to stabilize the feature distribution, and an attention aggregated feature containing multi-scale time sequence correlation information is obtained. Preferably, in step 323, the attention aggregation feature is input into a preset fully connected layer, and a linear transformation is performed to map the feature dimension to a preset fixed dimension (such as 256 dimensions), so as to reduce feature redundancy and unify the feature scale, and obtain a low-dimensional feature vector; the low-dimensional feature vector is subjected to layer normalization processing, the feature values are standardized to a similar distribution range by calculating the mean and variance of each dimension of the vector, the stability and generalization ability of the feature are improved, and a normalized feature vector is obtained; the normalized feature vector is applied with a ReLU activation function, the influence of negative value dimensions in the vector is suppressed, the non-linear expression ability of the feature is enhanced, the feature dimensions with greater contribution to fault identification are retained, and finally a time sequence fault feature encoding vector capable of condensing key time sequence mode information is obtained.

[0028] Optionally, step 33: based on the time sequence fault feature encoding vector, a multi-scale time sequence evolution mode is extracted through a time convolution network to generate a fault evolution prediction feature, specifically including the following steps: Step 331: extending the time context of the time sequence fault feature encoding vector, the encoding vectors of consecutive time points are spliced through a sliding window to obtain a time context feature matrix; Step 332: based on the time context feature matrix, a multi-scale time sequence evolution mode is extracted through multiple parallel convolution branches of the time convolution network to obtain multi-scale time sequence evolution features; Step 333: fusing the multi-scale time sequence evolution features to generate a fault evolution prediction feature.

[0029] Preferably, in step 331, the time sequence fault feature encoding vector is taken as a processing object, a sliding time window containing multiple consecutive time encoding vectors is set, the window size is determined according to the typical period of fault evolution, and it is ensured that a complete feature change period can be covered; the sliding window is moved along the time axis by time point, and after each movement, the time sequence fault feature encoding vectors in the window are transversely spliced in time sequence to form a vector containing multi-time feature correlation information; with the continuous movement of the window, a series of spliced vectors arranged in time sequence are obtained, and the vectors are stacked in rows to form a two-dimensional matrix, which is a time context feature matrix with extended time context information, and each row of the matrix corresponds to a time point and a feature set of the associated time points before and after the time point. Preferably, in step 332, the time context feature matrix is input into multiple parallel convolution branches of a time convolution network, each convolution branch is equipped with a one-dimensional convolution kernel of different length (such as 3, 5, 7 time steps), and the convolution kernel of different length can capture the evolution pattern of different time span, the short convolution kernel extracts local instantaneous change feature, and the long convolution kernel captures long-period trend feature; each convolution branch performs convolution operation on the time context feature matrix, calculates the weighted sum of each position and adjacent time features through the sliding convolution kernel, and generates a feature map reflecting the evolution pattern of the corresponding scale of the convolution branch; the ReLU activation function is applied to the feature map output by each convolution branch to introduce nonlinear transformation, and then the time dimension is compressed through the max-pooling operation to retain the key feature points, and the scale features of each branch are obtained; the scale features of multiple convolution branches are spliced in the channel dimension to form a multi-scale time evolution feature containing multi-time scale evolution information. Preferably, in step 333, the multi-scale time evolution feature is input into a 1x1 convolution layer, the feature information in different channels is integrated through the convolution layer, and the feature dimension is reduced to reduce the calculation amount, and the multi-scale feature after dimension compression is obtained; the attention mechanism is applied to the multi-scale feature after dimension compression, the importance weight (the weight value is determined by the energy size of the feature) of each scale feature is calculated, the weight is multiplied with the corresponding scale feature element by element, the expression of important scale feature is strengthened, and the weighted multi-scale feature is obtained; the residual connection is introduced, the weighted multi-scale feature is added with the original multi-scale time evolution feature, the feature information is avoided to be lost in the fusion process, and then the time dimension is compressed into a single value through the global average pooling operation, and the global feature of each channel is obtained; the global feature is mapped to a preset dimension through a fully connected layer, and a fault evolution prediction feature capable of comprehensively reflecting the fault evolution law of different time scales is generated.

[0030] Optionally, step 4: according to the fault evolution prediction feature, the abnormal confidence of the electric locomotive bearing in different fault categories is derived, and the specific steps include the following steps. Step 41: quantizing the dynamic anomaly index of the fault evolution prediction feature, and screening out the fault sensitive feature index; Step 42: based on the fault sensitive feature index, the correlation degree analysis is performed with the historical fault feature template, and the feature matching degree of each fault category is obtained; Step 43: the feature matching degrees of each fault category are fused to obtain the abnormal confidence of the electric locomotive bearing in different fault categories.

[0031] According to the method of claim 14, the step 41: quantizing the dynamic anomaly index of the fault evolution prediction feature, and screening out the fault sensitive feature index, specifically includes the following steps. Step 411: Extract the dynamic anomaly indicators of the fault evolution prediction features, and perform time series segmentation calculation to obtain the fault dynamic features; Step 412: Perform correlation analysis and coefficient of variation calculation on the fault dynamic features to eliminate redundant features in the fault dynamic features and obtain the fault sensitive feature indicators accordingly.

[0032] Preferably, in step 411, the fault evolution prediction features are taken as the processing object, time series analysis is performed thereon, the trend slope reflecting the feature change speed and the fluctuation amplitude reflecting the feature stability are extracted as the dynamic anomaly indicators; the inflection point detection of the time series is performed on the extracted dynamic anomaly indicators (trend slope and fluctuation amplitude), the feature change rate is divided into multiple continuous feature segments according to the significant turning points, and the dynamic anomaly indicators in each feature segment have a relatively stable change mode; for each feature segment, the least square method is used to fit the change straight line of the trend slope in the segment to obtain the average change rate of the feature in the segment; at the same time, the range of the fluctuation amplitude in the segment is calculated through the 3σ criterion to determine the stability degree of the feature in the segment; the average change rate and the stability degree range of each feature segment are integrated to form a parameter set that can represent the dynamic characteristics of the features in each segment, and the parameter sets of all feature segments are arranged in time sequence to form a complete fault dynamic feature that can reflect the dynamic change law of the fault evolution prediction features over time. Preferably, in step 412, the fault dynamic features obtained in step 411 are taken as the processing object, the correlation analysis is performed thereon with the feature parameters corresponding to the known fault types in the historical fault samples, the dynamic features having significant correlation with the historical fault features are selected by calculating the correlation degree therebetween; the coefficient of variation (i.e. the ratio of the standard deviation to the mean of the feature parameters) is calculated for the dynamic features selected through the correlation analysis, the coefficient can quantify the dispersion degree of the feature parameters, and the smaller the coefficient of variation is, the more gentle the feature change is; according to the size of the coefficient of variation, the feature parameters having a smaller coefficient of variation (i.e. the features changing gently and having a low sensitivity to faults) are eliminated, and the feature parameters having a larger coefficient of variation and being able to sensitively reflect the fault evolution are retained; the feature parameters retained after the correlation analysis and the coefficient of variation screening are integrated to form the fault sensitive feature indicators that can effectively represent the key characteristics of the fault evolution.

[0033] Optionally, step 42: based on the fault sensitive feature indicators, the correlation degree analysis is performed with the historical fault feature templates to obtain the feature matching degrees of each fault category, which specifically includes the following steps: Step 421: construct a historical fault feature template library, extract the sensitive feature parameter ranges of typical faults at different evolution stages to form a historical fault feature template set; Step 422: Based on the fault sensitive feature index and the historical fault feature template set, the feature matching degree is calculated by multi-dimensional weighted cosine similarity, and the feature matching degree of each fault category is obtained.

[0034] Preferably, in step 421, historical fault data of electric locomotive bearings are collected, and the data are classified and arranged according to fault categories (such as inner ring fault, outer ring fault, roller fault, etc.), to ensure that each category of fault data contains complete evolution process records; for each fault, different evolution stages (such as early stage, middle stage, late stage, etc.) are divided according to the severity of the feature change, and the time nodes or feature thresholds of each stage are determined; for the data of each fault category in each evolution stage, sensitive feature parameters (such as trend slope, fluctuation amplitude, etc. determined in step 41) are extracted, and the value range (such as the interval covered by the mean value ± 2 times the standard deviation) of these parameters in the stage is calculated by statistical analysis, to obtain the feature parameter range of the fault category in the stage; each fault category, each evolution stage and the corresponding feature parameter range are combined to form a feature template, all templates are stored in order according to fault categories and evolution stages, and a historical fault feature template set is constructed; at the same time, a template dynamic updating mechanism is established, and every time a preset number (such as 100) of new fault data are accumulated, the feature parameter range of the corresponding fault category and stage is recalculated, and the template set is updated to ensure its timeliness. Preferably, in step 422, the fault sensitive feature index obtained in step 412 and the historical fault feature template set constructed in step 421 are taken as processing objects, and for each fault category and its evolution stage template, the similarity between the fault sensitive feature index and the corresponding feature parameter in the template is calculated; in multiple feature dimensions such as trend slope and fluctuation amplitude, the cosine similarity (measuring the direction closeness of two vectors, the larger the value, the more consistent the direction) of the parameter vector of the current fault sensitive feature index and the parameter range vector of the template in the dimension is calculated; according to the evolution stage (judged preliminarily by the absolute value of the feature parameter) in which the fault may currently be, dynamic weights (such as higher weight for trend slope in early stage and higher weight for fluctuation amplitude in late stage) are assigned to different feature dimensions, and the weight values are determined based on the contribution of each dimension to fault recognition in historical data; the cosine similarities of the dimensions are multiplied by the corresponding weights and summed up to obtain the weighted matching degree of the fault category in the evolution stage; the weighted matching degrees of the same fault category in different evolution stages are taken as the maximum value or the average value, as the comprehensive feature matching degree of the fault category, and finally the feature matching degrees of all fault categories are obtained.

[0035] Optionally, step 43: fuse the multi-dimensional feature matching degrees to obtain abnormal confidence of different fault categories of electric locomotive bearings, which specifically includes the following steps: Step 431: Based on the historical occurrence probability of each fault category, the feature matching degree of each fault category is weighted and corrected to obtain a weighted matching degree. Step 432: The weighted matching degree is normalized and filtered to obtain the abnormal confidence of the electric locomotive bearing appearing different fault categories.

[0036] Preferably, in step 431, the feature matching degrees of each fault category obtained in step 422 are taken as the processing object, the occurrence times of each fault category in the historical fault data are collected, the occurrence probability of each fault category in the historical data (i.e. the proportion of the number of occurrences of this category of fault to the total number of faults) is calculated, and the historical occurrence probability of each fault category is obtained; according to the historical occurrence probability, a correction coefficient is assigned to the feature matching degree of each fault category, and the higher the historical occurrence probability of a fault category, the greater the correction coefficient, so as to enhance the weight of the feature matching degree of the high-incidence fault; the feature matching degree of each fault category is multiplied by the corresponding correction coefficient to obtain the matching degree of the fault category after the historical occurrence probability weighting adjustment; the weighted and adjusted matching degrees of all fault categories are preliminarily normalized (to ensure that the sum of all matching degrees is 1) to obtain a weighted matching degree that can reflect the influence of the historical occurrence probability, which retains both the feature similarity information and the prior probability of fault occurrence. Preferably, in step 432, the weighted matching degree obtained in step 431 is input into a Sigmoid function for nonlinear transformation, the value of the weighted matching degree is mapped to the interval of 0-1, the output value is more in line with the characteristics of probability distribution, and the discrimination degree of the matching degrees between different fault categories is enhanced; a confidence lower limit threshold (such as 0.2) is set, the filtered matching degrees are adjusted to 0 if they are lower than the threshold, so as to eliminate low-confidence results caused by noise interference; the filtered matching degrees are normalized again to make the sum of all non-zero matching degrees be 1 again, so as to ensure that the final result meets the probability distribution requirements; a scaling factor is dynamically adjusted according to the overall discrete degree of the feature matching degree (i.e. the difference between each matching degree value), the normalized matching degrees are fine-tuned, the confidence results under different working conditions are made comparable, and finally the abnormal confidence that can accurately reflect the possibility of the electric locomotive bearing appearing different fault categories is obtained.

[0037] Optionally, step 5: according to the abnormal confidence of the electric locomotive bearing appearing different fault categories and the estimated probability of the electric locomotive bearing appearing different fault categories, the actual fault category of the electric locomotive bearing is determined and a fault warning is generated accordingly, which specifically includes the following steps: Step 51: data fusion is performed on the abnormal confidence of the electric locomotive bearing appearing different fault categories and the estimated probability of the electric locomotive bearing appearing different fault categories to construct a fault comprehensive evaluation index and obtain a multi-dimensional evaluation vector; Step 52: determining the fault category actually occurring in the electric locomotive bearing through a hierarchical decision mechanism based on the multi-dimensional evaluation vector and the preset decision threshold; Step 53: generating fault warning information with differentiated response strategies according to the determined fault category actually occurring in the electric locomotive bearing and its evolution stage.

[0038] Optionally, step 51: data fusion is performed on the abnormal confidence of different fault categories occurring in the electric locomotive bearing and the estimated probability of different fault categories occurring in the electric locomotive bearing to construct a fault comprehensive evaluation index, thereby obtaining a multi-dimensional evaluation vector, which specifically includes the following steps: Step 511: standardization preprocessing is performed on the abnormal confidence of different fault categories occurring in the electric locomotive bearing and the estimated probability of different fault categories occurring in the electric locomotive bearing to eliminate dimensional differences and filter out invalid data, thereby obtaining a standardized feature matrix; Step 512: the standardized feature matrix dynamically allocates fusion weights of abnormal confidence and fault probability based on the contribution degree of each index in the historical fault sample; Step 513: data fusion is performed on the abnormal confidence of different fault categories occurring in the electric locomotive bearing and the estimated probability of different fault categories occurring in the electric locomotive bearing through nonlinear fusion mapping based on the allocated fusion weights of abnormal confidence and fault probability, thereby obtaining a fault comprehensive evaluation index; Step 514: multi-dimensional feature expansion processing is performed on the fault comprehensive evaluation index by introducing a fault evolution rate and a characteristic variation coefficient, thereby generating a multi-dimensional evaluation vector.

[0039] Preferably, in step 511, the abnormal confidence of different fault categories occurring in the electric locomotive bearing and the estimated fault probability are taken as processing objects, and Z-score standardization processing is respectively performed on the two types of data. The data is converted into standardized values with a mean of 0 and a standard deviation of 1 by calculating the difference between each data point and the mean value of the same type of data and dividing by the standard deviation, thereby eliminating the differences caused by different dimensions. A lower limit (such as 0.3) of abnormal confidence and a lower limit (such as 0.1) of fault probability are set to filter out invalid data points below the lower limit and retain valid data with practical significance. The filtered standardized abnormal confidence and standardized fault probability are arranged in order of fault categories, with the abnormal confidence as the first column of the matrix and the fault probability as the second column of the matrix, thereby forming a standardized feature matrix containing two types of standardized data. Each row of the matrix corresponds to the standardized features of a fault category. Preferably, in step 512, the standardized feature matrix obtained in step 511 is taken as the processing object, the contribution of the abnormal confidence and the fault probability in the historical fault sample to fault discrimination (determined by analyzing the correlation strength of the two types of indicators in the historical data with the actual fault result) is extracted, and the higher the contribution, the greater the weight in fusion; the basic weight is dynamically adjusted based on the fluctuation degree of the two types of data in the current standardized feature matrix (the greater the fluctuation, the more critical the indicator in the current data); the adjusted weight is multiplied by the corresponding column in the standardized feature matrix to obtain the weighted feature value of each fault category in the abnormal confidence and the fault probability, and a standardized feature matrix with completed weight distribution is formed, which not only retains the original data features, but also reflects the dynamic importance of the indicators. Preferably, in step 513, the standardized feature matrix with completed weight distribution obtained in step 512 is taken as the processing object, and the weighted geometric mean method is used for nonlinear fusion of the weighted abnormal confidence and the weighted fault probability of each fault category, that is, the square root of the product of the two (the weight is the exponent), which strengthens the influence of high-value indicators and suppresses the interference of low-value indicators; the fusion results are arranged in the order of fault categories, and each value corresponds to a comprehensive evaluation score of a fault category; linear scaling is performed on all comprehensive evaluation scores to make them fall within the interval of 0-100, which is convenient for intuitively reflecting the evaluation intensity, and a fault comprehensive evaluation indicator that can comprehensively reflect the synergistic effect of abnormal confidence and fault probability is obtained, and each value of the indicator corresponds to the overall evaluation level of a fault category. Preferably, in step 514, the fault comprehensive evaluation indicator obtained in step 513 is taken as the processing object, the fault evolution rate (obtained by dividing the difference between adjacent time comprehensive evaluation indicators by the time interval, reflecting the fault development speed) of each fault category is calculated, and the characteristic variation coefficient (the ratio of the standard deviation to the mean of the comprehensive evaluation indicator, reflecting the stability of the indicator) is calculated; the fault evolution rate is taken as the third dimension, the characteristic variation coefficient is taken as the fourth dimension, and the original fault comprehensive evaluation indicator (first dimension) and the two types of data in the standardized feature matrix (second dimension, integrated into a comprehensive value) are combined in the order of fault categories; the four-dimensional values of each fault category are arranged in order to form a four-dimensional vector, and the four-dimensional vectors of all fault categories are stacked in order to generate a multi-dimensional evaluation vector that can describe the fault state from multiple angles.

[0040] Optionally, step 52: based on the multi-dimensional evaluation vector and the preset decision threshold, the actual fault category of the electric locomotive bearing is determined through a hierarchical decision mechanism, specifically including the following steps: Step 521: dynamic decision thresholds are established for different fault categories, and the thresholds are determined based on the ROC curve optimization of historical fault data and are corrected in real time according to the operating conditions of the electric locomotive; Step 522: Construct a multi-dimensional discriminant rule considering the three dimensions of evaluation score, change trend and duration, and determine the fault category and severity based on the preset threshold; Step 523: When multiple fault categories simultaneously meet the determination condition, determine the final fault category through the priority ranking mechanism of evaluation score priority and change trend slope priority.

[0041] Preferably, in step 521, for different fault categories such as inner ring fault, outer ring fault, roller fault, etc., collect the corresponding historical fault data and the corresponding multi-dimensional evaluation vector, determine the critical point that makes the sensitivity and specificity reach the optimal balance as the initial decision threshold, including the early warning threshold, the alarm threshold and the emergency shutdown threshold, by drawing the ROC curve of each fault category (taking the evaluation score in the evaluation vector as the horizontal axis and the actual fault occurrence or not as the vertical axis); Real-time acquisition of the current operating condition parameters of the electric locomotive (such as load size, running speed, environmental temperature), based on the preset operating condition influence coefficient table (obtained by historical data training, reflecting the influence degree of different operating conditions on fault features), dynamically correct the initial decision threshold, for example, appropriately lower the early warning threshold in high load condition to early warning, form a dynamic decision threshold matrix for different fault categories and operating conditions, each element in the matrix corresponds to a certain level threshold of a specific fault category under the current operating condition. Preferably, in step 522, the dynamic decision threshold matrix obtained in step 521 and the multi-dimensional evaluation vector generated in step 514 are used as the processing object, and a discriminant rule containing three dimensions of evaluation score, change trend and duration is constructed; Extract the evaluation score of each fault category (first dimension) from the multi-dimensional evaluation vector, calculate the change trend of the evaluation score (second dimension, obtained by dividing the difference of adjacent time evaluation scores by the time interval, positive value indicates that the score rises, i.e. the fault may worsen), and calculate the duration of the evaluation score exceeding the corresponding dynamic decision threshold (third dimension); Compare the parameters of the three dimensions with the corresponding threshold in the dynamic decision threshold matrix, if the evaluation score exceeds the early warning threshold, the change trend is positive and the duration reaches the early warning time, it is determined as a potential fault; If it exceeds the alarm threshold, the change trend is positive and the duration reaches the alarm time, it is determined as a serious fault; If it exceeds the emergency shutdown threshold, the change trend is positive and the duration reaches the shutdown time, it is determined as an emergency fault, so as to complete the determination of each fault category and severity. Preferably, in step 523, when multiple fault categories in step 522 satisfy the respective discrimination rules, the evaluation scores of these fault categories are used as the primary comparison basis to screen out the fault category with the highest evaluation score as the candidate category; if there are candidate categories with the same evaluation score, the change trend slope (the absolute value of the change trend, reflecting the fault deterioration speed) of these candidate categories is extracted, and the fault category with a larger slope (i.e., the fault with a faster deterioration speed) is selected as the final candidate; the final candidate category is checked again to see if it satisfies the discrimination rule of the corresponding level (to avoid misjudgment due to priority sorting), and after confirmation, it is determined as the fault category actually occurring in the electric locomotive bearing, ensuring that the most concerned fault can be identified preferentially when multiple faults coexist.

[0042] Optionally, step 53: generating fault warning information with differentiated response strategies according to the determined fault category actually occurring in the electric locomotive bearing and its evolution stage, specifically including the following steps: Step 531: mapping the determined fault category and its evolution stage to a five-level warning system to match the warning level, different warning levels in the five-level warning system correspond to different fault emergency levels and disposal time limit recommendations, the warning level is determined based on the influence of the fault on the running safety of the electric locomotive, and is dynamically adjusted with reference to historical fault handling experience; Step 532: based on the determined fault category actually occurring in the electric locomotive bearing, matching the corresponding structured disposal scheme from the pre-set disposal strategy library, the structured disposal scheme includes the maintenance site, spare parts list and process specification, and is associated with historical similar maintenance cases for reference; the structured disposal scheme is generated through the association rules of fault characteristics and disposal measures, ensuring the pertinence and operability; Step 533: selecting the distribution channel of the structured disposal scheme according to the warning level, including pushing through the intranet, notifying through short message and email, starting the sound and light alarm, and sending instructions to the vehicle-mounted system.

[0043] Preferably, in step 531, the fault category actually occurring in the electric locomotive bearing determined in step 523 and its evolution stage (such as early potential fault, mid-stage serious fault, etc.) are associated with the pre-set five-level warning system (green level I to black level V), and according to the typical hazard degree of the fault category and the urgency of the evolution stage, a preliminary warning level is obtained; the influence of the fault category on the running safety of the electric locomotive (such as the inner ring fault usually having a more serious influence) and the historical fault handling experience (such as the typical response of similar faults at similar stages) are combined for dynamic adjustment, for example, if the historical data shows that a fault deteriorates faster than the average level at the current stage, the warning level is appropriately increased, and finally the warning level matching the actual situation of the fault is determined, which clearly shows the emergency level of the fault and the corresponding disposal time limit recommendation. Preferably, in step 532, the actual fault category of the electric locomotive bearing determined in step 523 is taken as the processing object, a preset treatment strategy library (which is stored in categories according to fault categories, containing standard treatment schemes for various faults such as inner ring, outer ring, and roller) is retrieved, and a structured treatment scheme containing the maintenance site (such as the inner ring of the specific bearing), the required spare parts list (such as the inner ring assembly and lubricating grease corresponding to the model), and the specific process specification (such as the disassembly torque and installation precision requirement) is matched out; the fault feature parameters in the structured treatment scheme are subjected to similarity calculation with the case features in the historical maintenance case library (by comparing the fault evolution trend, feature parameter value, etc.), and 3-5 most similar historical cases (containing the treatment process and effect) are associated as references; the structured treatment scheme is dynamically generated through the association rules of fault features and treatment measures (generated through historical data mining), ensuring the pertinence of the scheme to the current fault and the operability of the on-site operation. Preferably, in step 533, the early warning level obtained in step 531 and the structured treatment scheme obtained in step 532 are taken as the processing objects, the corresponding information distribution channel is determined according to the emergency degree of the early warning level, the higher the emergency degree, the more diversified and direct the distribution channel; the early warning information and the structured treatment scheme of different levels are distributed through the corresponding channel: the I-II level early warning is pushed to the equipment management system through the enterprise intranet for routine checking by the maintenance personnel; the III level early warning is additionally triggered to send short message and email to the responsible engineer to remind him to pay attention and arrange inspection while being pushed to the system; the IV-V level early warning further starts the on-site sound and light alarm device (such as the control room warning light and alarm bell) in addition to the above channels, and sends an emergency instruction (such as suggesting to reduce speed or stop) containing the treatment suggestion to the vehicle-mounted control system through a special communication protocol, ensuring that the early warning information and the treatment scheme efficiently reach the relevant personnel or system.

[0044] As shown in Figure 2 The application also provides a machine learning-based electric locomotive fault early warning system, which comprises: A signal processing module for acquiring real-time vibration signals of the electric locomotive bearing and generating a time-frequency feature matrix containing time-domain and frequency-domain coupled features according to the real-time vibration signals; A probability estimation module for calling a fault recognition model and estimating the probabilities of different fault categories of the electric locomotive bearing according to the time-frequency feature matrix; A feature evolution module for performing time-series dynamic pattern coding processing on the real-time vibration signals, generating a time-series fault feature coding vector, and generating a fault evolution prediction feature based on the time-series fault feature coding vector; A confidence derivation module for deriving abnormal confidence of different fault categories of the electric locomotive bearing according to the fault evolution prediction feature; The comprehensive judgment module is configured to judge the actual fault category of the electric locomotive bearing according to the abnormal confidence of the electric locomotive bearing in different fault categories and the estimated probability of the electric locomotive bearing in different fault categories, and generate a fault warning according to the actual fault category.

[0045] The exemplary technical description of each module is described above Figure 1 The embodiments are described.

[0046] Specifically, the exemplary description of the system realized by the hardware carrier is as follows: The industrial control computer is used as the core, and a vibration signal acquisition front end and a communication module are matched to form a complete hardware architecture. The vibration signal acquisition front end includes a piezoelectric acceleration sensor (used for acquiring real-time vibration signals) installed on the bearing seat of the electric locomotive, a signal conditioner (used for filtering and amplifying the original signals), and a data acquisition card (used for converting analog signals into digital signals). The industrial control computer is connected with the industrial bus (such as PCIe) to complete the acquisition and preprocessing of real-time vibration signals.

[0047] The industrial control computer is built-in with a multi-core processor (such as Intel Core i7 series) and a high-performance GPU (such as NVIDIA Tesla series), which undertakes the core computing task: the processor runs the software program of the signal processing module, and performs time-frequency analysis on the collected digital signals to generate a time-frequency feature matrix; the GPU deploys a trained fault recognition model (such as a neural network model based on the deep learning framework TensorFlow / PyTorch), and completes the fault probability calculation through the probability estimation module, and simultaneously uses the parallel computing capability of the GPU to accelerate the time sequence coding processing of the feature evolution module and the evolution feature analysis of the confidence derivation module.

[0048] In addition, the system is configured with an Ethernet communication module (supporting the IEEE 802.3 standard) and a vehicle-mounted CAN bus interface. The fault warning information generated by the comprehensive judgment module is sent to the ground monitoring center through the Ethernet and transmitted to the vehicle-mounted display terminal through the CAN bus, so as to realize the multi-terminal synchronization of the warning information. The hardware device further includes a redundant power module (to ensure continuous power supply) and a storage unit (such as an SSD hard disk, used for buffering real-time data and historical fault templates), which ensures the stable operation of the system under complex working conditions.

[0049] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A machine learning-based method for early warning of electric locomotive faults, characterized in that, The method includes: Step 1: Obtain the real-time vibration signal of the electric locomotive bearing, and generate a time-frequency feature matrix containing time-domain and frequency-domain coupled features based on the real-time vibration signal; Step 2: Call the fault identification model and estimate the probability of different fault categories occurring in the electric locomotive bearings based on the time-frequency feature matrix; Step 3: Perform time-series dynamic pattern encoding processing on the real-time vibration signal to generate a time-series fault feature encoding vector, and generate fault evolution prediction features based on the time-series fault feature encoding vector; Step 4: Based on the fault evolution prediction characteristics, deduce the abnormal confidence levels for different fault categories in power locomotive bearings; Step 5: Based on the abnormal confidence levels of different fault categories of electric locomotive bearings and the estimated probability of different fault categories of electric locomotive bearings, determine the actual fault category of the electric locomotive bearings and generate a fault warning accordingly.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Step 11: Filter out environmental noise and interference components from the real-time vibration signal to obtain a noise-reduced vibration signal; Step 12: Decompose the noise-reduced vibration signal into multiple sub-signals containing different frequency components to obtain multi-band wavelet components; Step 13: Construct the time-frequency distribution features of multi-band wavelet components to obtain a time-frequency feature matrix containing time-domain and frequency-domain coupling features. In the time-frequency feature matrix, the matrix elements represent the signal energy density of the corresponding time point and frequency band.

3. The method according to claim 2, characterized in that, Step 11 specifically includes: Step 111: Decompose the real-time vibration signal into multiple intrinsic modes, and filter out the intrinsic mode components dominated by noise by calculating the kurtosis value of each intrinsic mode; Step 112: Based on the intrinsic mode components dominated by noise, calculate the adaptive noise standard deviation estimate that changes dynamically over time; Step 113: Based on the adaptive noise standard deviation estimate, perform component reconstruction to obtain the noise-reduced vibration signal.

4. The method according to claim 1, characterized in that, Step 2 specifically includes: Step 21: Based on the spatial attention layer in the fault identification model, dynamically assign weights to the fault-sensitive regions in the time-frequency matrix to obtain the attention-weighted feature map; Step 22: Based on the attention-weighted feature map, convert it into a fault semantic feature vector through the cross-scale feature fusion network in the fault identification model; Step 23: Map the fault semantic feature vector to the fault category probability space, and combine it with the category correlation matrix to obtain the probability of different fault categories occurring in the electric locomotive bearing.

5. The method according to claim 4, characterized in that, Step 21 specifically includes: Step 211: Based on the multi-scale pooling and feature fusion unit in the spatial attention layer, perceive the spatial distribution features of the time-frequency feature matrix, and align and extract the time-frequency joint distribution features through multi-scale pooling operation to obtain the spatial feature descriptor; Step 212: Based on the fault-guided convolutional unit in the spatial attention layer, according to the spatial feature descriptor, the weights of sensitive regions are assigned by the convolutional network embedding fault prior knowledge to obtain the spatial attention weight map. Step 213: Based on the feature enhancement operation unit in the spatial attention layer, perform element-wise enhancement operation on the spatial attention weight map and the time-frequency feature matrix to obtain the attention weighted feature map.

6. The method according to claim 1, characterized in that, Step 3 specifically includes: Step 31: Capture the temporal dynamic characteristics of the real-time vibration signal, and extract the signal change trend at different times through a bidirectional gated loop unit to obtain the temporal dynamic characteristic sequence; Step 32: Based on the temporal dynamic feature sequence, key temporal patterns are aggregated through a self-attention mechanism to obtain the temporal fault feature encoding vector; Step 33: Based on the temporal fault feature encoding vector, extract multi-scale temporal evolution patterns through a temporal convolutional network to generate fault evolution prediction features.

7. The method according to claim 6, characterized in that, Step 31 specifically includes: Step 311: Align the time sequence length of the real-time vibration signal, and use the dynamic time warping algorithm to eliminate the time sequence offset to obtain a time sequence signal of equal length; Step 312: Based on the equal-length time series signal, extract the time series dependency features at different times through a bidirectional gated cyclic unit to obtain the bidirectional time series hidden state; Step 313: Fuse the bidirectional temporal hidden states to obtain the temporal dynamic feature sequence.

8. The method according to claim 6, characterized in that, Step 32 specifically includes: Step 321: Associate features at different times in the temporal dynamic feature sequence, calculate the correlation weights by scaling dot product attention, and generate a temporal attention weight matrix; Step 322: Based on the temporal attention weight matrix, perform weighted aggregation on the temporal dynamic feature sequence to obtain the attention aggregated features; Step 323: Map the attention aggregation features to a fixed-dimensional space to obtain the temporal fault feature encoding vector.

9. The method according to claim 1, characterized in that, Step 4 specifically includes: Step 41: Quantify the dynamic variation indicators of fault evolution prediction characteristics and screen out fault-sensitive characteristic indicators; Step 42: Based on the fault sensitivity feature index, perform correlation analysis with historical fault feature templates to obtain the feature matching degree of each fault category; Step 43: Combine the feature matching degree of each fault category to obtain the abnormal confidence degree of different fault categories of electric locomotive bearings.

10. The method according to claim 1, characterized in that, Step 5 specifically includes: Step 51: The abnormal confidence scores of different fault categories of electric locomotive bearings and the estimated probabilities of different fault categories of electric locomotive bearings are fused together to construct a comprehensive fault assessment index and obtain a multi-dimensional assessment vector. Step 52: Based on the multi-dimensional evaluation vector and the preset decision threshold, determine the actual fault category of the electric locomotive bearing through a hierarchical decision-making mechanism; Step 53: Based on the determined fault type and evolution stage of the electric locomotive bearing, generate fault warning information with differentiated response strategies.

Images