Data analysis method for train operation faults

Abstract

The invention relates to the technical field of train data analysis, in particular to a data analysis method for train operation faults, which comprises the following steps: carrying out data fusion processing on train operation brake signals acquired by a vibration sensor, an acoustic emission sensor and a state detection device to generate multi-dimensional brake time sequence data; brake fusion features are generated from the multi-dimensional brake time sequence data through a cross-modal fusion model, and the cross-modal fusion model is constructed based on a gating feature interaction layer and a convolution feature extraction layer; and the brake fusion features are used to generate brake fault severity through a space-time joint model, the space-time joint model is constructed based on an improved LSTM architecture, and optimization training is carried out through a physics constraint loss function. According to the method, the accuracy and reliability of predictive fault diagnosis of the subway braking system are remarkably improved through the LSTM architecture of multi-modal data fusion, cross-modal feature modeling and physical constraint embedding on the train operation braking data detected beside the track and the detection data of the braking system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of train data analysis, and in particular to a data analysis method for train operation failure. BACKGROUND

[0002] With the rapid development of urban rail transit system, the operation safety of subway train and the reliability of braking system become a key issue. Subway train braking system failure may cause train delay, equipment damage and even safety accidents, so a high-efficiency and accurate predictive fault diagnosis method is needed. Traditional subway braking detection technology mainly relies on a single type of sensor or mechanical signals of the braking system itself, and fault warning is carried out through threshold judgment or simple time sequence analysis.

[0003] However, such methods have the following limitations: the signals of a single sensor are difficult to fully reflect the complex state of mechanical, vibration, acoustic and other multi-physical field coupling in the braking process, resulting in insufficient fault feature extraction and high missed detection rate; there is a lack of deep fusion of multi-source heterogeneous data, and traditional data splicing or weighted fusion methods are difficult to mine the correlation between modalities, limiting the accurate identification of fault patterns; the fault diagnosis model based on traditional LSTM or CNN does not fully consider the spatio-temporal correlation characteristics in the braking process, and the training process lacks physical law constraints, resulting in weak model generalization ability and insufficient prediction stability under complex working conditions.

[0004] At present, for example, the patent document with application number 202411143270.8 discloses a car chassis fault recognition and diagnosis system, which proposes to analyze the brake pad wear state in combination with vibration and temperature signals, but it does not introduce acoustic emission signals, and the fusion model does not optimize the cross-modal feature interaction mechanism.

[0005] In summary, the existing technology still has significant defects in efficient fusion of multi-source data, joint modeling of spatio-temporal features and embedding of physical constraints, which restricts the accuracy and real-time performance of subway train braking failure analysis. SUMMARY

[0006] Therefore, the present application provides a data analysis method for train operation failure, which significantly improves the accuracy and reliability of predictive fault diagnosis of subway braking system through multi-modal data fusion, cross-modal feature modeling and LSTM architecture with physical constraint embedding of trackside detected train operation braking data and braking system detection data.

[0007] To achieve the above purpose, the present application provides a data analysis method for train operation failure, a management platform and a vibration sensor and an acoustic emission sensor set at a trackside subway train brake point, and a state detection device of a subway train braking system are in communication with the management platform, and the data analysis method applied to the management platform comprises:

[0008] The train operation brake signals collected by the vibration sensor, the acoustic emission sensor and the state detection device are subjected to data fusion processing to generate multi-dimensional brake timing data;

[0009] The multi-dimensional brake timing data is subjected to cross-modal fusion model to generate brake fusion features, wherein the cross-modal fusion model is constructed based on a gating feature interaction layer and a convolution feature extraction layer;

[0010] The brake fusion features are subjected to a spatio-temporal joint model to generate brake fault severity, wherein the spatio-temporal joint model is constructed based on an improved LSTM architecture and is subjected to optimization training through a physical constraint loss function embedding a pressure vibration energy constraint term and a temperature acoustic emission constraint term;

[0011] The operation fault maintenance of the subway train is performed according to the brake fault severity.

[0012] Further, the process of generating brake fusion features from the multi-dimensional brake timing data through the cross-modal fusion model includes:

[0013] The multi-dimensional brake timing data is subjected to the feature interaction layer to generate dynamic weights;

[0014] The multi-dimensional brake timing data is subjected to weighting and reorganization according to the dynamic weights to generate reorganized brake timing data;

[0015] The reorganized brake timing data is subjected to the convolution feature extraction layer to generate the brake fusion features.

[0016] Further, the feature interaction layer is provided with a self-attention mechanism and an activation function, and the process of generating dynamic weights from the multi-dimensional brake timing data through the feature interaction layer includes:

[0017] The multi-dimensional data relationship is extracted from the multi-dimensional brake timing data through the self-attention mechanism to generate attention weights;

[0018] The attention weights are subjected to the activation function to judge the relationship between the multi-dimensional data to generate the dynamic weights.

[0019] Further, the convolution feature extraction layer is constructed based on a multi-scale pyramid architecture, including a first convolution layer, a second convolution layer, a third convolution layer, a cross-layer aggregation layer and a residual connection layer, wherein the convolution kernel size of the first convolution layer is smaller than that of the second convolution layer, and the convolution kernel size of the second convolution layer is smaller than that of the third convolution layer, and the process of generating the brake fusion features from the reorganized brake timing data through the convolution feature extraction layer includes:

[0020] The brake timing data is extracted by the first convolutional layer to generate first extracted features;

[0021] The first extracted features are extracted by the second convolutional layer to generate second extracted features;

[0022] The second extracted features are extracted by the second convolutional layer to generate third extracted features;

[0023] The first extracted features are pooled by the cross-layer aggregation layer, the second extracted features are sampled, and then the third extracted features are sequentially spliced to generate spliced features;

[0024] The spliced features are calculated by the residual block of the residual connection layer, and then spliced with the first extracted features to generate the brake fusion features.

[0025] In the above scheme, the feature interaction layer is constructed by introducing a self-attention mechanism and an activation function, which can extract the correlation between features of different modalities. The convolutional feature extraction layer based on the multi-scale pyramid architecture can simultaneously capture the transient features, time-frequency correlation features and overall trend features of the brake signal. Through the cross-layer aggregation layer, the multi-scale features are pooled, sampled and spliced, and combined with the residual connection to retain the original information, effectively solving the omission problem of long-period trends or high-frequency details in traditional single-scale convolutional models.

[0026] Further, the multi-dimensional brake timing data includes vibration dimension brake timing data, vibration energy and acoustic emission wave timing data. The process of data fusion processing of the train running brake signals collected by the vibration sensor and the acoustic emission sensor to generate multi-dimensional brake timing data includes:

[0027] The train running brake vibration signals collected by the vibration sensor are subjected to fault vibration feature extraction based on frequency cepstrum coefficients to generate the vibration dimension brake timing data and vibration energy;

[0028] The train running acoustic emission wave signals collected by the acoustic emission sensor are subjected to statistical analysis based on data volatility to generate the acoustic emission wave timing data and acoustic emission rate.

[0029] Further, the state detection device is used to collect pressure data, current data, temperature data, displacement data and contact resistance data of the brake system. The multi-dimensional brake timing data further includes brake system state timing data. The process of data fusion processing of the train running brake signals collected by the state detection device to generate multi-dimensional brake timing data includes:

[0030] The pressure data, current data, temperature data, displacement data and contact resistance data are subjected to multi-scale entropy analysis fusion to generate the brake system state time series data.

[0031] Further, a pressure vibration energy constraint term is constructed based on the brake system pressure gradient, vibration energy and brake fault severity;

[0032] A temperature acoustic emission constraint term is constructed based on the acoustic emission rate, brake disc temperature and brake fault severity;

[0033] The physical constraint loss function is constructed based on the pressure vibration energy constraint term and the temperature acoustic emission constraint term.

[0034] Further, the first product of the brake system pressure gradient and the brake fault severity is calculated, and the pressure vibration energy constraint term is constructed based on the difference between the first product and the vibration energy;

[0035] The second product of the brake disc temperature and the brake fault severity is calculated, and the temperature acoustic emission constraint term is constructed based on the difference between the second product and the acoustic emission rate.

[0036] Further, a splicing vector is constructed based on the hidden state and the input vector;

[0037] The product of the splicing vector and the vibration energy is subjected to an activation function to construct the iterative input data of the improved LSTM architecture.

[0038] Further, the brake fault severity includes light fault, moderate fault and severe fault, and a graded operation fault maintenance strategy is generated for the brake system based on the light fault, moderate fault or severe fault.

[0039] In the above scheme, the diagnostic error of a single sensor is avoided, multi-dimensional features of vibration, acoustic emission and mechanical state are fused, mechanical, electrical and thermodynamic multi-domain fault modes are covered, the missed detection rate of the brake system is reduced, the physical constraint is embedded into the loss function, the model output is constrained to comply with the physical law, the untrustworthy pure data-driven model is avoided, the false alarm rate is reduced, and the problems of insufficient fault feature extraction, poor model generalization and lagging maintenance response under complex working conditions are solved.

[0040] Compared with the prior art, the beneficial effects of the present application are that,

[0041] 1. Through the multi-modal data fusion, cross-modal feature modeling and LSTM architecture with physical constraint embedding of the train operation brake data and the detection data of the brake system detected on the track side, the precision and reliability of the predictive fault diagnosis of the subway brake system are significantly improved.

[0042] 2、By introducing self-attention mechanism and activation function to construct feature interaction layer, the correlation of features between different modalities can be extracted, and the convolution feature extraction layer based on multi-scale pyramid architecture can capture the transient features, time-frequency correlation features and overall trend features of the brake signal at the same time. Through the cross-layer aggregation layer, the multi-scale features are pooled, sampled and spliced, and combined with the residual connection to retain the original information, effectively solving the omission problem of long-period trend or high-frequency details of the traditional single-scale convolution model.

[0043] 3、Avoiding the diagnostic error of a single sensor, fusing multi-dimensional features of vibration, acoustic emission and mechanical state, covering mechanical, electrical and thermodynamic multi-domain failure modes, reducing the missed detection rate of the brake system, embedding physical constraints into the loss function to constrain the model output to comply with the physical law, avoiding the untrustworthiness of pure data-driven models, reducing the false alarm rate of faults, and solving the problems of insufficient fault feature extraction, poor model generalization and maintenance response lag under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The figure is a flowchart of the data analysis method for train operation faults according to an embodiment of the present application.

[0045] Figure 2 The figure is a flowchart of the generation of brake fusion features in the data analysis method for train operation faults according to an embodiment of the present application.

[0046] Figure 3 The figure is a flowchart of the convolution feature extraction layer in the data analysis method for train operation faults according to an embodiment of the present application.

[0047] Figure 4 The figure is a flowchart of the optimization training process for train operation faults according to an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to make the purpose and advantages of the present application clearer and more apparent, the present application will be further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0049] The preferred embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application and do not limit the protection scope of the present application.

[0050] It should be noted that in the description of the present application, the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicate the direction or positional relationship of the terms based on the direction or positional relationship shown in the drawings, which is only for the convenience of description, and does not indicate or imply that the device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0051] In addition, it should be noted that in the description of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0052] As shown in Figures 1 to 4 The present application provides a data analysis method for train operation failure, which significantly improves the accuracy and reliability of predictive fault diagnosis of subway braking system by fusing the train operation braking data and the detection data of the braking system through a multi-modal data fusion, a cross-modal feature modeling and a LSTM architecture with physical constraint embedding.

[0053] As shown in Figures 1 to 4 The present embodiment provides a data analysis method for train operation failure, the management platform communicates with the vibration sensor and the acoustic emission sensor set at the trackside subway train brake point, and the state detection device of the subway train braking system, the data analysis method applied to the management platform comprises:

[0054] The train operation brake signals collected by the vibration sensor, the acoustic emission sensor and the state detection device are subjected to data fusion processing to generate multi-dimensional brake time series data;

[0055] The multi-dimensional brake time series data is subjected to brake fusion feature generation through a cross-modal fusion model, wherein the cross-modal fusion model is constructed based on a gating feature interaction layer and a convolution feature extraction layer;

[0056] The brake fusion feature is subjected to brake fault severity generation through a spatio-temporal joint model, wherein the spatio-temporal joint model is constructed based on an improved LSTM architecture, and is subjected to optimization training through a physical constraint loss function embedding a pressure vibration energy constraint term and a temperature acoustic emission constraint term;

[0057] According to the brake fault severity, the operation failure maintenance of the subway train is carried out.

[0058] The cross-modal fusion model is preferably optimized and trained using the mean square error (MSE) of the predicted value of the multi-dimensional brake time series data and the true label. The cross-modal fusion model is more efficient in segmented training with the spatio-temporal joint model, and has a lower mean square error (MSE) in fault severity prediction.

[0059] As shown in Figures 1 to 2 Further, the process of generating brake fusion features from the multi-dimensional brake time series data by the cross-modal fusion model includes:

[0060] The multi-dimensional brake time series data is generated into dynamic weights by the feature interaction layer;

[0061] The multi-dimensional brake time series data is weighted and reorganized according to the dynamic weights to generate reorganized brake time series data;

[0062] The reorganized brake time series data is generated into the brake fusion features by the convolution feature extraction layer.

[0063] As shown in Figures 1 to 2 Further, the feature interaction layer is provided with a self-attention mechanism and an activation function, and the process of generating dynamic weights from the multi-dimensional brake time series data by the feature interaction layer includes:

[0064] The multi-dimensional data relationship is extracted by the self-attention mechanism to generate attention weights;

[0065] The relationship between the multi-dimensional data is judged by the activation function to generate the dynamic weights.

[0066] Specifically, the process of generating dynamic weights is:

[0067]

[0068] where g represents the dynamic weight, σ represents the Sigmoid activation function,

[0069] represents the self-attention weight, where Att represents the self-attention mechanism, W g , b g respectively represent the weight matrix and the bias vector belonging to the dynamic weight g, represents the multi-dimensional brake time series data, where h brake , h vib , h ace respectively represent the data after the data fusion processing of the train operation brake signal collected by the state detection device, the vibration sensor and the acoustic emission sensor, represents the concatenation of vector data. Therefore, the input features can be re-encoded and integrated through linear transformation, and higher-level feature representations can be extracted. The weight matrix and bias vector are constantly adjusted during the training process to adapt to the inherent patterns of the data.

[0070] Specifically, the process of generating the brake fusion feature is:

[0071]

[0072] where h fu represents the brake fusion feature, h brake , h vib , h ace respectively represent the data after the train operation brake signal of the state detection device, vibration sensor, and acoustic emission sensor are fused, g represents the dynamic weight, represents the concatenation of vector data, and represents the Hadamard product of vector data. Therefore, the Hadamard product allows selective weighting combination of different features according to the dynamic weight g. When g is close to 1, the brake fusion feature will be more inclined to the former state detection device acquisition; when g is close to 0, the brake fusion feature will be more inclined to the concatenation of the vibration sensor and the acoustic emission sensor, so that different features can be dynamically fused, and the model can adaptively adjust the contribution of the features according to the characteristics of the input data.

[0073] Therefore, through the dynamic weight distribution and feature recombination mechanism, different source feature information is adaptively fused, so that the model can more effectively utilize multi-modal or multi-feature data, and improve its processing capability and generalization performance for complex problems. In fault diagnosis, the importance and combination method of the features can be flexibly adjusted according to different input conditions.

[0074] As shown in Figure 3 Further, the convolution feature extraction layer is constructed based on a multi-scale pyramid architecture, including a first convolution layer, a second convolution layer, a third convolution layer, a cross-layer aggregation layer, and a residual connection layer, wherein the convolution kernel size of the first convolution layer is smaller than that of the second convolution layer, and the convolution kernel size of the second convolution layer is smaller than that of the third convolution layer. The process of generating the brake fusion feature from the reorganized brake time series data through the convolution feature extraction layer includes:

[0075] The reorganized brake time series data is extracted through the first convolution layer to generate first extracted features;

[0076] The first extracted features are extracted through the second convolution layer to generate second extracted features;

[0077] The second extracted features are extracted through the second convolution layer to generate third extracted features;

[0078] pooling the first extracted feature and sampling the second extracted feature through the cross-layer aggregation layer, and sequentially performing feature splicing with the third extracted feature to generate spliced features;

[0079] After the residual block calculation of the spliced features through the residual connection layer, the first extracted feature is spliced to generate the brake fusion feature.

[0080] Specifically, the process of generating the brake fusion feature is:

[0081] H pyr = H1 + ResBlock(MaxPool(H1) + UpSample(H2) + H3)

[0082] In the formula, H pyr represents the brake fusion feature, H1, H2 and H3 represent the first extracted feature, the second extracted feature and the third extracted feature respectively, ResBlock(MaxPool(H1) + UpSample(H2) + H3) represents the spliced feature, wherein ResBlock represents the residual block calculation of the residual connection layer, MaxPool represents the maximum pooling operation, and UpSample represents the up-sampling operation. Therefore, by adding the residual connection in the cross-layer feature aggregation, the aggregated features can not only retain the original feature information, but also fuse new cross-layer feature information.

[0083] Further, the multi-dimensional brake time series data includes vibration dimension brake time series data, vibration energy and acoustic emission wave time series data, and the process of generating multi-dimensional brake time series data by data fusion processing of the train running brake signals collected by the vibration sensor and the acoustic emission sensor includes:

[0084] The train running brake vibration signals collected by the vibration sensor are subjected to fault vibration feature extraction based on frequency cepstrum coefficients to generate the vibration dimension brake time series data and vibration energy;

[0085] The train running acoustic emission wave signals collected by the acoustic emission sensor are subjected to statistical analysis based on data volatility to generate the acoustic emission wave time series data and acoustic emission rate.

[0086] Specifically, the MFCC (Mel frequency cepstrum coefficient) used for vibration signal analysis includes: after pre-processing the input signal, performing fast Fourier transform on each frame of signal to obtain an amplitude spectrum, calculating the total energy of each frame based on the amplitude spectrum, extracting the discrete cosine transform (DCT) through the Mel filter bank, calculating the first order difference of the discrete cosine transform to obtain the vibration energy, and the discrete cosine transform is used as the vibration dimension brake time series data. Therefore, the vibration sensor can determine whether the bearing of the brake system fails.

[0087] Specifically, the root mean square energy of the background noise is calculated for threshold judgment, the number of events exceeding the threshold is taken as a ringing count rate (RCR), a coefficient of variation (CV) is calculated according to the ringing count rate, and the coefficient of variation is taken as the acoustic emission wave timing data.

[0088] Further, the state detection device is used to collect pressure data, current data, temperature data, displacement data and contact resistance data of the brake system, and the multi-dimensional brake timing data further includes brake system state timing data, and the process of data fusion processing of the train operation brake signal collected by the state detection device to generate the multi-dimensional brake timing data includes: performing multi-scale entropy analysis fusion on the pressure data, current data, temperature data, displacement data and contact resistance data to generate the brake system state timing data.

[0089] As shown in Figure 4 Further, a pressure vibration energy constraint term is constructed based on the brake system pressure gradient, vibration energy and brake fault severity;

[0090] A temperature acoustic emission constraint term is constructed based on the acoustic emission rate, brake disc temperature and brake fault severity;

[0091] The physical constraint loss function is constructed based on the pressure vibration energy constraint term and the temperature acoustic emission constraint term.

[0092] As shown in Figure 4 Further, a first product of the brake system pressure gradient and the brake fault severity is calculated, and the pressure vibration energy constraint term is constructed based on the difference between the first product and the vibration energy;

[0093] A second product of the brake disc temperature and the brake fault severity is calculated, and the temperature acoustic emission constraint term is constructed based on the difference between the second product and the acoustic emission rate.

[0094] Specifically, the physical constraint loss function is:

[0095]

[0096] In the formula, Loss represents the physical constraint loss function, respectively represent the pressure vibration energy constraint term and the temperature acoustic emission constraint term, || ||2 represents the norm operation, λ1, λ2 represent two weighting coefficients, y represents the brake fault severity, E vib respectively represent the brake system pressure gradient and the vibration energy, T, τ ace respectively represent the brake disc temperature and the acoustic emission rate.

[0097] Further, a splicing vector is constructed based on the hidden state and the input vector; a product of the splicing vector and the vibration energy is constructed into the iterative input data of the improved LSTM architecture through an activation function.

[0098] Specifically, the improved LSTM architecture of the spacial-temporal joint model is:

[0099]

[0100] wherein, represents the adjusted cell state at the current time t, f t represents the output of the forget gate for determining how much information in the cell state at the last time needs to be forgotten, C t-1 represents the cell state at the last time, i t represents the output of the input gate, tanh(W c [h t-1 ,x t ]·S t represents the iterative input data, wherein tanh is the hyperbolic tangent activation function for compressing the input to between -1 and 1, W c represents the weight matrix belonging to the cell state C, [h t-1 ,x t ] represents the hidden state h t-1 at the last time t-1 spliced with the input x t at the current time to generate a splicing vector, S t represents the vibration energy, and represents the composite operation of the function and represents the multiplication operation of the function. Therefore, the spacial-temporal joint model realizes dynamic adjustment of the memory unit update strength by using the vibration energy representing the vibration spectrum features, so that the model can adaptively adjust the update mode of its internal state when processing related sequence data, thereby improving the processing capability and adaptability of the model for such data.

[0101] Further, the brake failure severity includes a mild failure, a moderate failure and a serious failure, and a hierarchical operation failure maintenance strategy is generated for the brake system based on the mild failure, the moderate failure or the serious failure.

[0102] In one specific embodiment, the algorithm model library of the tool component service of the management platform carries the above-mentioned process of generating brake fault severity, the modeling tool carries PyTorch, TensorFlow for improved LSTM modeling and custom physical loss function, Librosa for vibration signal analysis of MFCC, PyWavelets for generating acoustic emission sensor data into acoustic emission wave timing data and acoustic emission rate, the monitoring rule library carries the hierarchical operation fault maintenance strategy of the brake system fault rule library corresponding to the brake fault severity, and the process of hierarchical maintenance management of the brake system is realized. Wherein, the modeling tool can import the brake test bench data of the company for training the improved LSTM model. Part of the rules in the hierarchical operation fault maintenance strategy of the brake system fault rule library carried by the monitoring rule library are as follows:

[0103]

[0104] In the above scheme, the diagnostic error of a single sensor is avoided, multi-dimensional features of vibration, acoustic emission and mechanical state are fused, mechanical, electrical and thermodynamic multi-domain fault modes are covered, the missed detection rate of the brake system is reduced, physical constraints are embedded into the loss function to constrain the model output to comply with the physical law, the untrustworthiness of pure data-driven models is avoided, the false alarm rate of faults is reduced, and the problems of insufficient fault feature extraction, poor model generalization and lagging maintenance response under complex working conditions are solved.

[0105] In this embodiment, through the multi-modal data fusion, cross-modal feature modeling and LSTM architecture with physical constraint embedding of the trackside detected train operation brake data and the detection data of the brake system, the accuracy and reliability of the predictive fault diagnosis of the subway brake system are significantly improved. By introducing the self-attention mechanism and the activation function to construct the feature interaction layer, the correlation of features between different modalities can be extracted. The convolution feature extraction layer based on the multi-scale pyramid architecture can capture the transient features, time-frequency correlation features and overall trend features of the brake signal at the same time. Through the cross-layer aggregation layer, the multi-scale features are pooled, sampled and spliced, and the residual connection is combined to retain the original information, effectively solving the omission problem of long-period trends or high-frequency details of traditional single-scale convolution models. Avoiding the diagnostic error of a single sensor, fusing multi-dimensional features of vibration, acoustic emission and mechanical state, covering mechanical, electrical and thermodynamic multi-domain fault modes, reducing the missed detection rate of the brake system, embedding physical constraints into the loss function to constrain the model output to comply with the physical law, avoiding the untrustworthiness of pure data-driven models, reducing the false alarm rate of faults, and solving the problems of insufficient fault feature extraction, poor model generalization and lagging maintenance response under complex working conditions.

[0106] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

[0107] The above only describes the preferred embodiments of the present application and is not intended to limit the present application; the present application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A data analysis method for train operation failure, characterized in that: The management platform communicates with vibration sensors and acoustic emission sensors installed at the subway train braking points along the track, as well as a status detection device for the subway train's braking system. The data analysis method applied to the management platform includes: Performing data fusion processing on the train operation brake signals collected by the vibration sensor, acoustic emission sensor and state detection device to generate multi-dimensional brake time series data; Generate brake fusion features from the multi-dimensional brake time series data through a cross-modal fusion model, wherein the cross-modal fusion model is constructed based on a gated feature interaction layer and a convolutional feature extraction layer; The brake fusion features are used to generate brake fault severity through a spatiotemporal joint model, where the spatiotemporal joint model is built based on an improved LSTM architecture and is optimized and trained using a physics-constrained loss function that embeds pressure vibration energy constraints and temperature acoustic emission constraints; An operation fault maintenance strategy for the subway train is generated according to the severity of the brake fault.

2. The data analysis method for train operation failure according to claim 1, characterized in that: The process of generating brake fusion features from the multi-dimensional brake time series data through a cross-modal fusion model includes: Generate dynamic weights from the multi-dimensional braking time series data through the feature interaction layer; Performing weighted reorganization on the multi-dimensional brake time series data according to the dynamic weight to generate reorganized brake time series data; The reorganized brake time series data is passed through the convolutional feature extraction layer to generate the brake fusion feature.

3. The data analysis method for train operation failure according to claim 2, characterized in that: The feature interaction layer sets a self-attention mechanism and an activation function, and the process of generating dynamic weights from the multi-dimensional brake time series data through the feature interaction layer includes: The multi-dimensional brake time series data is subjected to a self-attention mechanism to extract the relationship between the multi-dimensional data and generate an attention weight; The attention weight is used to judge the relationship between multidimensional data through an activation function to generate the dynamic weight.

4. The data analysis method for train operation failure according to claim 2, characterized in that: The convolutional feature extraction layer is constructed based on a multi-scale pyramid architecture, including a first convolutional layer, a second convolutional layer, a third convolutional layer, a cross-layer aggregation layer, and a residual connection layer, wherein the convolution kernel size of the first convolutional layer is smaller than the convolution kernel size of the second convolutional layer, and the convolution kernel size of the second convolutional layer is smaller than the convolution kernel size of the third convolutional layer. The process of generating the brake fusion feature by passing the reorganized brake time series data through the convolutional feature extraction layer includes: The reorganized braking time series data is subjected to the first convolutional layer to extract transient features to generate first extracted features; The first extracted features are subjected to the second convolutional layer to extract time-frequency correlation features to generate second extracted features; The second extracted feature extracts the overall trend feature through the second convolutional layer to generate a third extracted feature; After pooling the first extracted features and sampling the second extracted features through the cross-layer aggregation layer, the first extracted features are sequentially spliced ​​with the third extracted features to generate spliced ​​features; The splicing feature is calculated through the residual block of the residual connection layer, and then spliced ​​with the first extracted feature to generate the brake fusion feature.

5. The data analysis method for train operation failure according to claim 1, characterized in that: The multi-dimensional braking time series data includes vibration dimension braking time series data, vibration energy, acoustic emission rate and acoustic emission wave time series data. The process of fusing the train operation braking signals collected by the vibration sensor and the acoustic emission sensor to generate the multi-dimensional braking time series data includes: Extracting fault vibration features based on frequency cepstrum coefficients of the train running brake vibration signal collected by the vibration sensor to generate the vibration dimension brake time series data and vibration energy; The train operation acoustic emission wave signal collected by the acoustic emission sensor is statistically analyzed based on data volatility to generate the acoustic emission wave time series data and acoustic emission rate.

6. The data analysis method for train operation failure according to claim 5, characterized in that: The state detection device is used to collect pressure data, current data, temperature data, displacement data and contact resistance data of the brake system. The multi-dimensional brake time series data also includes brake system state time series data. The process of performing data fusion processing on the train operation brake signal collected by the state detection device to generate the multi-dimensional brake time series data includes: The pressure data, current data, temperature data, displacement data and contact resistance data are subjected to multi-scale entropy analysis and fusion to generate the brake system state time series data.

7. The data analysis method for train operation failure according to claim 6, characterized in that: The pressure-vibration-energy constraint term is constructed based on the brake system pressure gradient, vibration energy and brake fault severity. Constructing a temperature acoustic emission constraint term based on the acoustic emission rate, the brake disc temperature, and the severity of the brake fault; The physical constraint loss function is constructed based on the pressure vibration energy constraint term and the temperature acoustic emission constraint term.

8. The data analysis method for train operation failure according to claim 7, characterized in that: calculating a first product of the brake system pressure gradient and the brake fault severity, and constructing the pressure vibration energy constraint term based on a difference between the first product and the vibration energy; A second product of the brake disc temperature and the brake fault severity is calculated, and the temperature acoustic emission constraint term is constructed based on a difference between the second product and the acoustic emission rate.

9. The data analysis method for train operation failure according to any one of claims 1 to 8, characterized in that: Construct a concatenation vector based on the hidden state and the input vector; The product of the splicing vector and the vibration energy is passed through an activation function to construct iterative input data of the improved LSTM architecture.

10. The data analysis method for train operation failure according to any one of claims 1 to 8, characterized in that: The brake fault severity includes a mild fault, a moderate fault and a severe fault, and a graded operation fault maintenance strategy is generated for the brake system based on the mild fault, the moderate fault or the severe fault.

Images