Train abnormal running state recognition method based on traction current time-frequency feature fusion

Abstract

A method for identifying abnormal train operation states based on traction current time-frequency feature fusion includes the following steps: constructing and training a train abnormal operation state identification model based on traction current time-frequency feature fusion; inputting the traction current to be identified into the model, and taking the operation state category with the highest probability obtained by the model as the abnormal train operation state; in the identification model, the time-domain feature extraction branch and the frequency-domain feature extraction branch are used to extract the time-domain feature vector and frequency-domain feature vector of the traction current, respectively; a gated fusion module is used to weightedly concatenate the time-domain feature vector and the frequency-domain feature vector to obtain a joint feature vector; a fully connected classifier is used to classify the joint feature vector to obtain the probability of each operation state category. This invention uses traction current as a unified observation, which can reduce the dependence on external sensors, dedicated acquisition equipment, and additional state quantities, reduce system deployment and maintenance costs, and improve the engineering economy of abnormal operation state identification.

Description

Technical Field

[0001] This invention relates to the field of rail transit train condition monitoring technology, and in particular to a method for identifying abnormal train operation conditions based on the fusion of traction current time-frequency characteristics. Background Technology

[0002] With the continuous expansion of the scale and intensity of rail transit operations, the risks of abnormal states encountered during train operation are also increasing, placing higher demands on train operation safety, stability, and condition monitoring capabilities. Traction current, as a key onboard electrical quantity during train operation, reflects the changes in the train's state during current intake, energy transfer, energy conversion, and traction output, and is characterized by convenient data acquisition and strong engineering applicability. Therefore, conducting abnormal operation state identification for metro trains based on traction current is of great significance for improving the train's operational status perception capability and the efficiency of abnormal state identification.

[0003] Existing methods for identifying abnormal train operation states mainly include state monitoring methods based on dedicated sensors, feature recognition methods based on signal processing, and intelligent recognition methods based on deep learning. Among these, state monitoring methods based on dedicated sensors typically rely on specific sensors, dedicated acquisition equipment, or additional state variables, resulting in problems such as inconsistent observations, complex equipment deployment, high maintenance costs, and limited engineering economics. Feature recognition methods based on signal processing rely on human experience for feature construction, and the feature extraction process is easily affected by parameter settings and changes in operating conditions, lacking the ability to uniformly represent multiple types of abnormal states. While intelligent recognition methods based on deep learning can automatically learn signal features, existing methods typically use structures such as one-dimensional convolutional networks, recurrent neural networks, and temporal convolutional networks to directly process the original time-domain signal, or convert the signal into a spectrogram or time-frequency diagram and then use a two-dimensional convolutional network for classification. Some methods also combine signal decomposition methods with neural networks to enhance feature representation. While the aforementioned deep learning methods have improved recognition capabilities to some extent, they typically still focus on a single signal representation. Although some methods extract both time-domain and time-frequency features simultaneously, they often employ direct concatenation or fixed weighting, making it difficult to dynamically adjust the contributions of these two types of features based on their relative importance in different traction current samples. This results in the complementary effect of time-domain dynamic information and time-frequency energy distribution information not being fully utilized. Furthermore, the current recognition methods, observations, and feature representations used for different abnormal operating states vary, lacking uniformity and applicability, making it difficult for the system to quickly and accurately identify multiple abnormal states in complex operating scenarios.

[0004] From the perspective of the traction current signal itself, the electrical response caused by abnormal operating conditions is not merely a simple amplitude change, but is simultaneously reflected in both the time process and frequency structure. In the time domain, the traction current may exhibit dynamic changes such as abrupt amplitude changes, local impacts, damped oscillations, and continuous fluctuations; in the frequency domain, its frequency components and energy distribution also change with the occurrence, development, and recovery of the anomaly. In other words, the abnormal state information in the traction current is simultaneously contained in the waveform evolution process and the time-frequency energy distribution structure. If only time-domain features are extracted, the frequency distribution differences corresponding to the abnormal disturbances are easily overlooked; if only frequency-domain features are extracted, it is difficult to reflect the dynamic evolution process of the abnormal changes on the time axis. Therefore, for the identification of abnormal operating conditions of subway trains based on traction current, simply increasing the network depth or changing the classifier cannot fundamentally solve the problem of incomplete feature representation. The key lies in constructing an identification structure that can extract and adaptively fuse the two types of features separately, taking advantage of the characteristic that the abnormal response of traction current simultaneously possesses both time-series evolution features and time-frequency distribution features, so as to achieve complementary expression of time-domain dynamic features and frequency-domain time-frequency distribution features. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying abnormal train operating states based on the fusion of traction current time-frequency features. The technical solution to achieve this purpose is as follows:

[0006] A method for identifying abnormal train operation states based on traction current time-frequency feature fusion includes the following steps:

[0007] Construct and train a train abnormal operation state identification model based on the fusion of traction current time-frequency features;

[0008] The traction current to be identified is input into the model, and the operating state category with the highest probability obtained by the model is taken as the abnormal operating state of the train.

[0009] The recognition model includes a time-domain feature extraction branch, a frequency-domain feature extraction branch, a gated fusion module, and a fully connected classifier. The time-domain feature extraction branch and the frequency-domain feature extraction branch are used to extract the time-domain feature vector and the frequency-domain feature vector of the traction current, respectively. The gated fusion module is used to weightedly concatenate the time-domain feature vector and the frequency-domain feature vector to obtain a joint feature vector. The fully connected classifier is used to classify the joint feature vector to obtain the probability of each operating state category.

[0010] Preferably, the extraction of the time-domain feature vector of the traction current by the time-domain feature branch includes the following steps:

[0011] S1.1 performs empirical wavelet transform on the traction current to obtain multiple frequency band sub-signals;

[0012] S1.2 Divide multiple frequency band sub-signals into low-frequency component group, mid-frequency component group and high-frequency component group according to frequency from low to high;

[0013] S1.3 The low-frequency component group, mid-frequency component group and high-frequency component group are respectively subjected to multi-scale one-dimensional convolution to obtain low-frequency trend features, mid-frequency fluctuation features and high-frequency transient disturbance features;

[0014] S1.4 combines low-frequency trend features, mid-frequency fluctuation features, and high-frequency transient disturbance features to obtain multi-band fusion features;

[0015] S1.5 uses a bidirectional long short-term memory network to extract forward and backward temporal dependency features from the multi-band fused features, resulting in a temporal feature vector.

[0016] Furthermore, it also includes:

[0017] S1.4' uses a self-attention mechanism to weight the features at each time position in the multi-band fusion features to obtain attention-enhanced features;

[0018] Step S1.5 is replaced by: passing the attention enhancement features through a bidirectional long short-term memory network to extract forward and backward temporal dependency features, thereby obtaining a temporal feature vector.

[0019] Preferably, the frequency domain feature branch extraction of the traction current frequency domain feature vector includes the following steps:

[0020] S2.1 Perform a short-time Fourier transform on the traction current to obtain a one-dimensional traction current sequence;

[0021] S2.2 The amplitude of the one-dimensional traction current sequence is squared and logarithmically transformed to obtain a two-dimensional time-frequency diagram;

[0022] S2.3 The frequency domain feature vector is obtained by passing the two-dimensional time-frequency graph through a two-dimensional convolutional neural network.

[0023] Preferably, the training is based on a train abnormal operation state identification model fused with traction current time-frequency features, and adopts the cross-entropy loss function.

[0024] The beneficial effects of this invention are as follows:

[0025] (1) Compared with state monitoring methods that rely on dedicated sensing information such as images, speed, vibration or internal variables of the control system, the present invention uses traction current as a unified observation, which can reduce the dependence on external sensors, dedicated acquisition equipment and additional state quantities, reduce system deployment and maintenance costs, and improve the engineering economy of abnormal operation state identification.

[0026] (2) Compared with traditional signal processing methods that rely on artificial spectrum, statistics or empirical thresholds, this invention performs adaptive frequency band decomposition of traction current through empirical wavelet transform, and extracts time-domain dynamic features by combining multi-scale convolution, self-attention mechanism and bidirectional long short-term memory network, which can reduce the impact of artificial feature design and fixed parameter settings on recognition effect.

[0027] (3) Compared with deep learning recognition methods that only use a single time-domain network or a single frequency-domain network, this invention simultaneously constructs a time-domain feature extraction branch and a frequency-domain feature extraction branch, which can respectively characterize the waveform evolution characteristics and time-frequency energy distribution characteristics of the traction current, thereby improving the comprehensive expression ability of non-stationary, multi-scale, and time-varying signals.

[0028] (4) Compared with simple splicing or fixed weighted fusion, the present invention adopts a gating fusion mechanism to perform dimension-wise adaptive weighting of time domain features and frequency domain features, so that the model can dynamically adjust the contribution of the two types of features according to different input samples, thereby improving the ability to distinguish between similar operating states. Attached Figure Description

[0029] Figure 1 The overall structure diagram of the time-frequency dual-branch abnormal operation state identification model is shown in the embodiment.

[0030] Figure 2 The diagram shows the branch structure of temporal feature extraction in this embodiment.

[0031] Figure 3 The diagram shows the frequency domain feature extraction branch structure for this embodiment.

[0032] Figure 4 The traction current time-frequency diagrams for different operating states in the embodiment are shown.

[0033] Figure 5 This is a comparison chart of convergence curves for an example and different models.

[0034] Figure 6 This is a comparison chart of the recognition performance metrics of the example and different models.

[0035] Figure 7 This is a comparison chart of confusion matrices between the example and different model test sets.

[0036] Figure 8 This is a visualization of the feature distribution of different models in the example.

[0037] Figure 9 This is a visualization of the feature extraction process of the time-frequency dual-branch model in the example. Detailed Implementation

[0038] This invention provides a train abnormal operation state identification method based on traction current time-frequency feature fusion, addressing the problems of inconsistent observations, strong dependence on external sensors or additional state variables, inadequate representation of non-stationary changes in traction current by single time-domain or single-frequency-domain features, and insufficient ability to distinguish between similar operation states in existing abnormal operation state identification methods. This method uses traction current as a unified input, extracts features from both time-domain waveform evolution and time-frequency energy distribution perspectives, and achieves adaptive fusion of these two types of features through a gated fusion mechanism, thereby enabling the identification of multiple types of abnormal operation states.

[0039] The key technical concept of this invention lies in the following: Addressing the characteristic that abnormal traction current responses simultaneously exist in both time-series waveforms and time-frequency energy distributions, the traction current identification process is divided into two parallel branches: time-domain dynamic feature extraction and frequency-domain time-frequency feature extraction. Matching feature extraction networks are constructed based on the data structure characteristics of different feature representation forms. Specifically, the traction current and its frequency band sub-signals after empirical wavelet decomposition are essentially one-dimensional time-series signals, with abnormal information evolving over time. Therefore, the time-domain branch obtains multiple sub-bands matching the frequency structure of the traction current through empirical wavelet transform, and extracts frequency band waveform features and their temporal evolution features using multi-scale one-dimensional convolution, self-attention mechanisms, and bidirectional long short-term memory networks. The time-frequency map obtained from short-time Fourier transform is a two-dimensional matrix that can represent the energy distribution of frequency components over time, possessing local texture and spatial distribution features. Therefore, the frequency-domain branch uses a two-dimensional convolutional neural network to extract time-frequency energy distribution features. Finally, a gating fusion mechanism dynamically adjusts the contributions of time-domain and frequency-domain features in the feature dimensions, thereby achieving unified identification of multiple abnormal operating states.

[0040] The technical solution of the present invention is as follows:

[0041] A method for identifying abnormal train operating states based on traction current time-frequency feature fusion includes:

[0042] The traction current sample to be identified is input into the abnormal operating state identification model, which includes a time-domain feature extraction branch, a frequency-domain feature extraction branch, a gating fusion module, and a classification output module.

[0043] The time-domain feature extraction branch performs empirical wavelet transform on the traction current samples, decomposing the traction current signal into multiple frequency band sub-signals. The frequency band components obtained from the empirical wavelet transform are represented as follows:

[0044]

[0045] In the formula, This represents the inverse Fourier transform. Frequency domain representation of the traction current signal. Indicates the first An empirical wavelet filter, Indicates the first Each frequency band component.

[0046] The multiple frequency band sub-signals obtained by empirical wavelet transform are divided into low-frequency component groups, mid-frequency component groups, and high-frequency component groups according to their frequency from low to high. These components are then input into three multi-scale one-dimensional convolutional branches for feature extraction. Each component group consists of one or more frequency band sub-signals with continuous frequency. The three multi-scale one-dimensional convolutional branches are used to extract low-frequency trend features, mid-frequency fluctuation features, and high-frequency transient disturbance features, respectively.

[0047] Let the convolution outputs corresponding to the low-frequency component group, the mid-frequency component group, and the high-frequency component group be respectively... , and The multi-band fusion characteristics are then expressed as:

[0048]

[0049] In the formula, This indicates a splicing operation in the direction of the channel.

[0050] The multi-band fused features are input into a bidirectional long short-term memory network to extract the forward and backward time-series dependency features of the traction current signal, resulting in a time-domain feature vector. .

[0051] The optimized approach introduces a self-attention mechanism after the multi-band fusion features. By calculating the correlation between different time positions, the features at each time position in the multi-band fusion features are weighted to obtain attention-enhanced features.

[0052] Let the input features be The query matrix is ​​obtained through linear mapping. Key matrix Sum matrix :

[0053]

[0054] In the formula, , , The weight matrix is ​​learnable. The self-attention output is:

[0055]

[0056] In the formula, The dimension of the key vector.

[0057] The attention-enhanced features are input into a bidirectional long short-term memory network to extract the forward and backward temporal dependency features of the traction current signal, resulting in a time-domain feature vector. .

[0058] The frequency domain feature extraction branch performs a short-time Fourier transform on the traction current samples, converting the one-dimensional traction current sequence into a two-dimensional time-frequency graph. The short-time Fourier transform is expressed as:

[0059]

[0060] In the formula, For input traction current signal, For window functions, For time location, For frequency.

[0061] The short-time Fourier transform result is squared and then logarithmically transformed to obtain the time-frequency diagram:

[0062]

[0063] In the formula, A constant used to prevent numerical underflow.

[0064] The time-frequency graph is input into a two-dimensional convolutional neural network. Convolution, pooling, and global average pooling are used to extract the energy distribution features of the traction current in the time-frequency plane, resulting in a frequency domain feature vector. .

[0065] The gated fusion module will convert the temporal feature vector and frequency domain eigenvectors The features are concatenated and then input into a fully connected network to generate fusion weights for the two types of features across various dimensions.

[0066]

[0067] And satisfy:

[0068]

[0069] In the formula, Indicates the first Gating weights corresponding to 3D time-domain features Indicates the first The gating weights corresponding to the frequency domain features and These are the weight matrix and the bias, respectively. The time-domain and frequency-domain features are weighted sequentially according to the gated weights:

[0070]

[0071] In the formula, This indicates element-wise multiplication.

[0072] The weighted features from the two classes are concatenated to obtain a joint feature vector:

[0073]

[0074] The classification output module will combine the feature vectors. Input the fully connected classifier to obtain the probability distribution of each running state category:

[0075]

[0076] In the formula, , This is the weight matrix of the fully connected layer. , For bias terms, This represents the class probability distribution output by the model.

[0077] The category with the highest probability is used as the result of identifying abnormal operating conditions of subway trains:

[0078]

[0079] In the formula, For the final category identification, For the first Predicted probability of class running state.

[0080] During model training, the cross-entropy loss function is used to optimize the model parameters. The loss function is:

[0081]

[0082] In the formula, For the number of categories, For real labels, Predict probabilities for the model.

[0083] By updating the parameters of the time-domain feature extraction branch, the frequency-domain feature extraction branch, the gating fusion module, and the classification output module through backpropagation, the trained abnormal operation state recognition model is obtained.

[0084] Example:

[0085] The traction current sample set is used as the model input object, and the sample is a fixed-length one-dimensional traction current sequence. The traction current sample to be identified is input as follows: Figure 1 The model shown is a time-frequency dual-branch abnormal operating state identification model. The model mainly consists of a time-domain feature extraction branch, a frequency-domain feature extraction branch, a gating fusion module, and a classification output module, which are used to realize feature extraction, feature fusion, and operating state classification of traction current samples.

[0086] First, the traction current sample input time-domain feature extraction branch is used, and then... Figure 2 The feature extraction process shown involves shape transformation. The time-domain feature extraction branch first uses empirical wavelet transform to decompose the traction current sample into frequency bands, breaking down the original one-dimensional current signal into multiple frequency band sub-signals. To obtain stable frequency band division results, this embodiment determines the frequency band boundaries of the empirical wavelet transform based on the average spectrum of the traction current sample. Let the adjacent peak frequencies in the average power spectrum of the traction current be... and Then the candidate boundaries of adjacent frequency bands can be represented as:

[0087]

[0088] In the formula, For the first Each frequency band boundary. This results in a set of frequency band boundaries arranged sequentially from low to high frequency:

[0089]

[0090] In the formula, The number of frequency band divisions, The sampling frequency.

[0091] In this embodiment, the number of frequency bands for the Empirical Wavelet Transform (EWT) is set to 6 to control the subsequent network size while preserving the characteristics of different frequency ranges of the traction current. The six frequency band sub-signals are arranged from low to high frequency. Since the frequency ranges and time scales corresponding to adjacent sub-bands are relatively close, sub-bands 1 and 2 are grouped into a low-frequency component group, sub-bands 3 and 4 into a mid-frequency component group, and sub-bands 5 and 6 into a high-frequency component group. The three component groups are respectively input to low-frequency, mid-frequency, and high-frequency multi-scale one-dimensional convolutional branches to extract low-frequency trends, mid-frequency fluctuations, and high-frequency transient disturbance features in the traction current. This grouping method can preserve the EWT subdivision frequency band information while avoiding mutual interference between features at different scales caused by mixing all sub-bands, and reducing the structural complexity caused by setting up separate convolutional branches for each sub-band.

[0092] In this embodiment, the number of frequency bands for EWT decomposition is set to 6 to control the size of the subsequent network while preserving the characteristics of different frequency ranges of the traction current. Too few decompositions can easily lead to variations across different time scales being mixed within the same sub-band; too many decompositions increase sub-band redundancy and the computational burden of subsequent feature extraction. The 6 sub-bands are arranged in ascending order of frequency, and each pair of adjacent sub-bands is grouped together because the frequency ranges and corresponding time scales of adjacent sub-bands are relatively close, allowing for feature extraction within the same convolutional branch. Thus, sub-bands 1 and 2 form the low-frequency group, sub-bands 3 and 4 form the mid-frequency group, and sub-bands 5 and 6 form the high-frequency group.

[0093] The three convolutional branches—low-frequency, mid-frequency, and high-frequency—are set up to extract the low-frequency trend, mid-frequency fluctuation, and high-frequency transient disturbance features from the traction current, respectively. If all six sub-bands are input into the same branch, features from different time scales are easily mixed; if each sub-band has its own branch, the network structure and computational load will increase. Therefore, this embodiment adopts a structure of "6 sub-bands—pairwise grouping—3 branches," which retains the subdivided frequency band information while merging similar frequency scales.

[0094] Of course, when the number of subbands obtained by EWT decomposition is different, it can also be divided into three continuous component groups of low frequency, mid frequency and high frequency in order of frequency from low to high, and the number of subbands contained in each group can be different.

[0095] Subsequently, the low-frequency, mid-frequency, and high-frequency component groups are input into a multi-scale one-dimensional convolution branch to extract local waveform features at different frequency bands. Since different frequency bands correspond to different dominant time scales, this embodiment determines the convolution kernel scale based on the sub-band center frequency. Let the... The upper and lower boundaries of each sub-band are respectively and Then the center frequency of the sub-band is:

[0096]

[0097] The corresponding dominant cycle can be represented as:

[0098]

[0099] In order for the convolution kernel to cover the local periodic structure of the corresponding frequency band, the initial size of the convolution kernel is determined according to the following formula:

[0100]

[0101] And on Oddity factorization and upper / lower bound constraints are applied. Through these methods, low-frequency components correspond to larger convolutional receptive fields, while high-frequency components correspond to smaller convolutional receptive fields, enabling the model to capture low-frequency trends, mid-frequency fluctuations, and high-frequency transient perturbation features, respectively.

[0102] The features output from each frequency band branch are weighted and concatenated to form a multi-band fused feature. In this embodiment, the low-frequency, mid-frequency, and high-frequency branches output [B, 256, 2000]-dimensional features respectively. After concatenation of the three branches along the channel dimension, a [B, 768, 2000]-dimensional feature is obtained. Subsequently, after average pooling, a [B, 768, 1000]-dimensional feature is obtained, and then adjusted to [B, 1000, 768] through dimensional transformation to meet the input requirements of the subsequent time-series modeling structure.

[0103] The multi-band fused features are then input into a self-attention module. The self-attention module generates attention weights by calculating the correlation between different time positions in the traction current sequence, and weights the multi-band fused features according to the attention weights to enhance the correlation features between time positions and suppress redundant or interfering features.

[0104] To avoid attention enhancement disrupting the original convolutional feature structure, this embodiment employs residual enhancement to combine the attention output with the convolutional fusion features, the expression of which is:

[0105]

[0106] In the formula, Features resulting from enhanced attention. For multi-band fusion features, For learnable weight parameters, This is for self-attention output.

[0107] Then, the attention-enhanced features are input into a bidirectional long short-term memory network, and the dynamic evolution features of the traction current signal are extracted through forward and reverse time-series modeling. The bidirectional long short-term memory network outputs [B, 1000, 512]-dimensional time-series features, which are then concatenated by the last layer of bidirectional hidden states and linearly mapped to finally obtain a 128-dimensional time-domain feature vector.

[0108] Simultaneously, the same traction current sample is input into the frequency domain feature extraction branch, and its shape transformation is as follows: Figure 3 As shown in the figure, the frequency domain feature extraction branch uses short-time Fourier transform to convert the one-dimensional traction current sequence into a two-dimensional time-frequency graph, which is used to characterize the energy distribution characteristics of the traction current frequency components over time. The traction current time-frequency graphs under different operating conditions are shown in the figure. Figure 4 As shown, differences in time-frequency energy distribution exist in different states. Subsequently, the time-frequency map is input into a two-dimensional convolutional neural network, and local texture features and time-frequency energy distribution features are extracted from the time-frequency map through convolution, pooling, and global average pooling.

[0109] In this embodiment, the frequency domain branch input is a [B, 1, 129, 33]-dimensional time-frequency map, which is sequentially processed through three layers of two-dimensional convolutional structures for feature extraction. The first convolutional layer outputs [B, 32, 129, 33]-dimensional features, which are then pooled to obtain [B, 32, 64, 16]-dimensional features; the second convolutional layer outputs [B, 64, 64, 16]-dimensional features, which are then pooled to obtain [B, 64, 32, 8]-dimensional features; the third convolutional layer outputs [B, 128, 32, 8]-dimensional features; finally, global average pooling and fully connected mapping yield a 128-dimensional frequency domain feature vector. Through this branch, the model can extract traction current energy distribution, local texture, and frequency perturbation features from the time-frequency two-dimensional plane.

[0110] After obtaining the time-domain and frequency-domain feature vectors, the two types of features are input into the gating fusion module. The gating fusion module concatenates the time-domain and frequency-domain features and generates fusion weights for each dimension of the two types of features through a fully connected network. Then, based on the fusion weights, the time-domain and frequency-domain features are weighted dimension-by-dimensionally, and the weighted features are concatenated to obtain a 256-dimensional joint feature vector. This joint feature vector is input into the classification output module to obtain the probability distribution of each operating state category, and the category with the highest probability is taken as the final recognition result. Thus, the model completes the entire recognition process from traction current input, time-domain feature extraction, frequency-domain feature extraction, gating fusion, to operating state classification output.

[0111] To verify the effectiveness of this invention, model training and testing were performed on an existing traction current sample set. The data partitioning method, optimizer configuration, learning rate, and training environment are shown in Table 1. The training convergence curves of different models are shown below. Figure 5 As shown. By Figure 5 It can be seen that the loss of this invention decreases rapidly during training and the verification accuracy improves relatively steadily, indicating that the model has good convergence stability.

[0112] Table 1 Dataset Configuration and Training Parameter Settings

[0113]

[0114] To illustrate the recognition performance of this invention, the time-frequency dual-branch fusion model used in this invention is compared with models such as 1DCNN, LSTM, ResNet1D, TCN, TFEF, MSFEF, and FDFES. The recognition performance of different models on the test set is shown in Table 2, with corresponding metrics for comparison. Figure 6 As shown. To quantify model performance, this embodiment uses Accuracy, Precision, Recall, and F1-score as evaluation metrics, where Accuracy represents the overall recognition accuracy, which can be written as:

[0115]

[0116] Precision represents the proportion of samples predicted to belong to a certain class that actually do not. It can be written as:

[0117]

[0118] Recall represents the proportion of samples that are correctly identified within a certain class, and can be written as:

[0119]

[0120] The F1 score is used to comprehensively measure precision and recall, and can be written as:

[0121]

[0122] In the formula, , , , These represent true positive examples, true negative examples, false positive examples, and false negative examples, respectively.

[0123] From Table 2 and Figure 6 It can be seen that models such as 1DCNN, LSTM, ResNet1D, and TCN mainly rely on time-domain sequences for feature learning, and their overall recognition performance is lower than that of this invention. This indicates that single time-domain modeling is insufficient to fully express the time-frequency distribution differences during traction current anomalies. Although FDFES utilizes time-frequency maps to extract frequency-domain features, it lacks characterization of the evolution and long-term dependencies of the original current waveform, and its recognition performance is still lower than that of this invention. Although MSFEF enhances the multi-scale frequency band features in the time domain, it lacks the supplementary expression of the dynamic distribution of frequency energy through time-frequency map branches. In contrast, this invention utilizes both time-domain dynamic features and frequency-domain time-frequency distribution features, and achieves adaptive feature synergy through gated fusion, thus achieving superior results in all indicators.

[0124] Table 2 Performance comparison of different models on the abnormal operation state identification task

[0125]

[0126] To further verify the role of each component in the model of this invention, an ablation experiment was conducted on the model, and the results are shown in Table 3. Table 3 shows that A1 achieved an accuracy of 91.90% when using only the time-domain branch, and A2 achieved an accuracy of 90.64% when using only the frequency-domain branch, both lower than the 94.71% of the complete dual-branch model. This indicates that traction current anomaly information does not exist only in a single time domain or a single frequency domain; the time-domain dynamic features and the frequency-domain time-frequency distribution features have complementary effects. A3 achieved an accuracy of 92.52% when using simple splicing, and A4 achieved an accuracy of 93.09% when using weighted fusion, both lower than the gated fusion structure of A5. This indicates that dimension-wise gated fusion can more effectively allocate the contributions of time-domain and frequency-domain features than ordinary fusion methods. A6's accuracy decreased to 92.63% after replacing STFT with FFT, indicating that obtaining only the overall spectrum information is insufficient to characterize the time-varying frequency structure during traction current anomalies. A7's accuracy decreased to 92.33% after replacing EWT with fixed filtering, indicating that adaptive frequency band division helps improve the multi-scale feature expression capability. It is evident that empirical wavelet transform, short-time Fourier transform, and gating fusion mechanism all play an important role in improving recognition performance. The technical effect of this invention does not come from the simple replacement of a single known model module, but is generated by the synergistic effect between traction current adaptive frequency band decomposition, time-frequency graph representation, and dimension-by-dimensional gating fusion.

[0127] Table 3 Comparison of ablation experimental results using the dual-branch model

[0128]

[0129] To verify the separability of the features extracted by the model, a feature distribution visualization method was further used to reduce the dimensionality of the high-dimensional features of different models, and the results are as follows. Figure 8 As shown. By Figure 8 It can be seen that the features extracted by this invention form a clearer intra-class aggregation structure and inter-class separation boundary in the low-dimensional space, indicating that the fusion of the time-domain branch and the frequency-domain branch can form a more discriminative feature representation.

[0130] To illustrate the internal feature extraction process of this invention, the feature extraction results of the time-domain branch and the frequency-domain branch are visualized, as shown below. Figure 9 As shown. By Figure 9 It can be seen that the time-domain branch can extract local waveform changes, transient disturbances and time-dependent features from the traction current signal; the frequency-domain branch can extract energy texture and frequency disturbance distribution features from the time-frequency plot; the two types of features are combined into joint features after gating fusion, which are used for the final classification output.

[0131] This invention verifies the synergistic effect of its technical solutions through comparative and ablation experiments. Compared with single time-domain models such as 1DCNN, LSTM, ResNet1D, and TCN, this invention improves the ability to express the time-frequency structure changes of traction current. Compared with models that only use time-domain branches or only use frequency-domain branches, this invention can simultaneously utilize waveform evolution features and time-frequency energy distribution features. Compared with simple splicing and fixed-weighted fusion, the gated fusion mechanism of this invention can dynamically adjust feature contributions according to different samples. Experimental results show that the complete model achieves an accuracy of 94.71% and an F1-score of 95.16% on the test set, respectively, which is higher than various comparative and ablation models. This indicates that this invention is not a simple replacement of a single network structure, but rather improves recognition performance through the synergistic effect of time-domain branches, frequency-domain branches, and gated fusion.

[0132] In summary, this invention uses traction current as a unified input and achieves multi-class operating state identification through time-domain feature extraction, frequency-domain feature extraction, and gated fusion. Compared with single-feature identification methods, this invention has higher identification accuracy and better class discrimination capability, while reducing dependence on external sensors, dedicated acquisition equipment, and additional state variables, thus exhibiting good engineering applicability and economy.

Claims

1. A method for identifying abnormal train operating states based on traction current time-frequency feature fusion, characterized in that, Including the following steps: Construct and train a train abnormal operation state identification model based on the fusion of traction current time-frequency features; The traction current to be identified is input into the model, and the operating state category with the highest probability obtained by the model is taken as the abnormal operating state of the train. The recognition model includes a time-domain feature extraction branch, a frequency-domain feature extraction branch, a gated fusion module, and a fully connected classifier. The time-domain feature extraction branch and the frequency-domain feature extraction branch are used to extract the time-domain feature vector and the frequency-domain feature vector of the traction current, respectively. The gated fusion module is used to weightedly concatenate the time-domain feature vector and the frequency-domain feature vector to obtain a joint feature vector. The fully connected classifier is used to classify the joint feature vector to obtain the probability of each operating state category.

2. The identification method as described in claim 1, characterized in that, The extraction of the time-domain feature vector of the traction current by the time-domain feature branch includes the following steps: S1.1 performs empirical wavelet transform on the traction current to obtain multiple frequency band sub-signals; S1.2 Divide multiple frequency band sub-signals into low-frequency component group, mid-frequency component group and high-frequency component group according to frequency from low to high; S1.3 The low-frequency component group, mid-frequency component group and high-frequency component group are respectively subjected to multi-scale one-dimensional convolution to obtain low-frequency trend features, mid-frequency fluctuation features and high-frequency transient disturbance features; S1.4 combines low-frequency trend features, mid-frequency fluctuation features, and high-frequency transient disturbance features to obtain multi-band fusion features; S1.5 uses a bidirectional long short-term memory network to extract forward and backward temporal dependency features from the multi-band fused features, resulting in a temporal feature vector.

3. The identification method as described in claim 2, characterized in that, Also includes: S1.4' uses a self-attention mechanism to weight the features at each time position in the multi-band fusion features to obtain attention-enhanced features; Step S1.5 is replaced by: passing the attention enhancement features through a bidirectional long short-term memory network to extract forward and backward temporal dependency features, thereby obtaining a temporal feature vector.

4. The identification method as described in claim 1, characterized in that, The frequency domain feature branching method extracts the frequency domain feature vector of the traction current, including the following steps: S2.1 Perform a short-time Fourier transform on the traction current to obtain a one-dimensional traction current sequence; S2.2 The amplitude of the one-dimensional traction current sequence is squared and logarithmically transformed to obtain a two-dimensional time-frequency diagram; S2.3 The frequency domain feature vector is obtained by passing the two-dimensional time-frequency graph through a two-dimensional convolutional neural network.

5. The identification method as described in claim 1, characterized in that, The training is based on a train abnormal operation state identification model that fuses traction current time-frequency features, and uses the cross-entropy loss function.

Images