Vehicle fault monitoring and early warning method based on multi-data fusion

Abstract

The invention relates to the technical field of vehicle fault monitoring, in particular to a multi-data fusion vehicle fault monitoring and early warning method, which comprises the following steps of: generating fusion traction fault features from data acquired by an inverter circuit sensor, a motor temperature sensor, a motor rotating speed sensor and a gearbox vibration sensor through an extraction model; wherein the extraction model is constructed based on an improved DRSN network architecture and has a multi-input processing branch and a channel attention mechanism for a multi-modal signal; determining the fault type of the traction system through a multi-stage classifier model according to the fused traction fault features; and matching the fault type of the traction system with the fault tree and the early warning list, and determining an early warning emergency disposal strategy. According to the method, the improved DRSN is adapted to the high-noise and strong-coupling environment of the multi-modal signal of the subway train traction system, and the early-stage accurate early warning and quick response of the fault type of the subway traction system are realized through efficient fusion, multi-stage classification optimization and intelligent emergency disposal of the multi-modal data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle fault monitoring, and in particular to a vehicle fault monitoring and early warning method based on multi-data fusion. BACKGROUND

[0002] With the rapid development of urban rail transit, the safety and reliability of metro trains have become the core problem of operation and maintenance. As the power core of metro trains, the running state of the traction system directly affects the stability and fault risk of the train. Traditional fault monitoring methods mainly rely on threshold judgment of single sensor data or diagnosis models based on shallow feature analysis, which are difficult to cope with the fusion processing of multi-source heterogeneous data and the identification of complex fault modes. For example, the fixed threshold alarm method has the problems of high false alarm rate and poor real-time performance. The feature extraction method based on conventional neural networks lacks sufficient feature representation ability under complex working conditions, which easily leads to the loss of key fault features.

[0003] In recent years, deep learning technology has been introduced into the field of fault diagnosis. However, existing models often use simple feature concatenation methods when processing multi-modal sensor signals, without fully considering the temporal and spatial correlation and feature weight differences between different signals of the metro train traction system. For example, traditional convolutional neural networks (CNN) or recurrent neural networks (RNN) lack targeted feature enhancement mechanisms when fusing multi-dimensional data such as inverter circuit and motor speed, resulting in interference from redundant information between signals and affecting classification accuracy.

[0004] Therefore, how to optimize the feature fusion capability of multi-modal signals of the metro train traction system and achieve fine monitoring and early warning of fault types is a technical problem to be solved at present. SUMMARY

[0005] To this end, the present application provides a vehicle fault monitoring and early warning method based on multi-data fusion, which adapts to the high-noise and strong-coupling environment of multi-modal signals of the metro train traction system through an improved DRSN network architecture, and performs efficient fusion of multi-modal data, deep feature enhancement, multi-level classification optimization and intelligent emergency disposal, thereby achieving early and accurate early warning and rapid response of fault types of the metro traction system.

[0006] To achieve the above-mentioned purpose, the present application provides a vehicle fault monitoring and early warning method based on multi-data fusion, a management platform and a communication between the inverter circuit sensor, the motor temperature sensor, the motor speed sensor and the gearbox vibration sensor of the metro train traction system. The vehicle fault monitoring and early warning method applied to the management platform comprises:

[0007] The data collected by the inverter circuit sensor, motor temperature sensor, motor speed sensor and gearbox vibration sensor is fused into traction fault features through a traction fault feature extraction model, wherein the traction fault feature extraction model is constructed based on an improved DRSN network architecture and has a multi-input processing branch for multi-modal signals and a channel attention mechanism for signal feature enhancement;

[0008] The fusion traction fault features are determined by a multi-level classifier model to determine the traction system fault type;

[0009] The traction system fault type is matched with a fault tree to determine the scope of the subway train fault, and the fault range is matched with a pre-warning list to generate a pre-warning emergency disposal strategy.

[0010] Further, the traction fault feature extraction model has an improved residual shrinkage unit for cross-modal signal fusion, the multi-input processing branch includes an electrical signal branch, a vibration signal branch and a temperature thermal energy branch, and the process of generating the fusion traction fault features includes:

[0011] The data collected by the inverter circuit sensor, motor temperature sensor, motor speed sensor and gearbox vibration sensor is preprocessed to generate circuit fault signals, temperature fault signals, speed fault signals and vibration fault signals, respectively;

[0012] The circuit fault signals and the speed fault signals are processed through the electrical signal branch to generate fault time sequence features;

[0013] The vibration fault signals are processed through the vibration signal branch to generate fault time-frequency features, wherein the vibration signal branch has the channel attention mechanism;

[0014] The temperature fault signals are processed through the temperature thermal energy branch to generate temperature spatio-temporal variation features;

[0015] The fault time sequence features, fault time-frequency features and temperature spatio-temporal variation features are processed through the improved residual shrinkage unit to generate the fusion traction fault features.

[0016] Further, the circuit fault signals include DC side current and voltage signals, AC side current and voltage signals and switch device driving voltage signals, the electrical signal branch is provided with a first one-dimensional convolution normalization layer, a second one-dimensional convolution normalization layer, an activation function layer and a down-sampling layer, wherein the convolution kernel size of the first one-dimensional convolution normalization layer is smaller than that of the second one-dimensional convolution normalization layer, and the process of generating the fault time sequence features includes:

[0017] The direct current side current voltage signal and the switch device driving voltage signal are inhibited of noise and short time high frequency mutation characteristics are extracted through the first one-dimensional convolution normalization layer, and the corresponding fault time sequence characteristics are generated through the activation function layer and the down sampling layer;

[0018] After the alternating current side current voltage signal is extracted of low frequency harmonic distortion characteristics through the second one-dimensional convolution normalization layer, the corresponding fault time sequence characteristics are generated through the activation function layer and the down sampling layer.

[0019] Further, the vibration signal branch is provided with a first two-dimensional convolution pooling layer, a second two-dimensional convolution pooling layer and the channel attention mechanism, and the process of generating the fault time-frequency characteristics includes:

[0020] The vibration fault signal is inhibited of high frequency electromagnetic noise and fault frequency sub-band characteristics are extracted through the first two-dimensional convolution pooling layer, and a fault concentrated frequency sub-band is generated;

[0021] The fault concentrated frequency sub-band is extracted of fault impact characteristics through the second two-dimensional convolution pooling layer, and a vibration fault feature is generated;

[0022] The vibration fault feature is adjusted of allocation weight of multi-class fault frequency band channels through the channel attention mechanism, and the fault time-frequency characteristics are generated, wherein the multi-class fault frequency band channels include a gear tooth surface wear pitting channel, a gear tooth breakage defect channel, a gear eccentricity channel and a bearing coupling fault channel.

[0023] Further, the temperature thermal energy branch is provided with a two-dimensional convolution layer, a normalization activation layer, a time feature extraction sub-branch, a space feature extraction sub-branch and a feature splicing layer, wherein the time feature extraction sub-branch and the space feature extraction sub-branch are arranged in parallel, and the process of generating the temperature space-time change characteristics includes:

[0024] The temperature fault signal is extracted of spatial local overheating area frame by frame through the two-dimensional convolution layer, and an overheating area feature is generated;

[0025] The overheating area feature is inhibited of sensor error noise through the normalization activation layer, and a highlighted overheating area feature is generated;

[0026] A plurality of highlighted overheating area features of a plurality of frames are extracted of cross-frame transient over-temperature characteristics and continuous heat diffusion characteristics through the time feature extraction sub-branch, and an abnormal heating and cooling feature is generated;

[0027] The highlighted overheating area feature is extracted and a overheating area shape feature is generated through the space feature extraction sub-branch;

[0028] The temperature space-time change feature is generated by splicing the temperature rise and heat dissipation abnormal feature and the overheating area shape feature through a feature splicing layer.

[0029] In the above scheme, through branch customization processing, dynamic feature enhancement and space-time joint modeling of multi-modal signals, high-precision early warning judgment of complex faults of the subway traction system, such as electrical transient anomaly, mechanical wear, thermal runaway, etc., is realized, and the response speed and reliability of fault warning operation and maintenance are improved.

[0030] Further, the temperature thermal energy branch is provided with a two-dimensional convolution layer, a normalization activation layer, a time feature extraction sub-branch, a space feature extraction sub-branch and a feature splicing layer, wherein the time feature extraction sub-branch and the space feature extraction sub-branch are arranged in parallel, and the process of generating the temperature space-time change feature includes:

[0031] The temperature fault signal is extracted frame by frame through the two-dimensional convolution layer to extract a local overheating area, and an overheating area feature is generated;

[0032] The overheating area feature is inhibited through the normalization activation layer to suppress sensor error noise, and a highlighted overheating area feature is generated;

[0033] The highlighted overheating area features of multiple frames are extracted through the time feature extraction sub-branch to extract cross-frame transient over-temperature features and continuous heat diffusion features, and a temperature rise and heat dissipation abnormal feature is generated;

[0034] The highlighted overheating area feature is extracted through the space feature extraction sub-branch to generate an overheating area shape feature;

[0035] The temperature rise and heat dissipation abnormal feature and the overheating area shape feature are spliced through the feature splicing layer to generate a joint fault feature, and the temperature space-time change feature is generated.

[0036] Further, the improved residual shrinkage unit is sequentially provided with a convolution dimension reduction layer, a channel attention threshold generation block and a residual connection layer, and the process of generating the fusion traction fault feature includes:

[0037] The fault time sequence feature, the fault time-frequency feature and the temperature space-time change feature are respectively reduced and fused through the convolution dimension reduction layer to generate a pre-fusion feature;

[0038] Multiple pre-fusion features are generated through the multi-channel of the channel attention threshold generation block to generate a fusion fault feature;

[0039] The fusion fault feature is generated through the residual connection layer to generate the fusion traction fault feature.

[0040] Further, the channel attention threshold generation block sequentially sets a channel compression excitation layer, a threshold calculation layer, a soft thresholding layer and a residual connection layer, and the process of generating the fused fault feature includes:

[0041] The plurality of pre-fusion features are respectively generated into a plurality of channel weights through the channel compression excitation layer;

[0042] The plurality of channel weights and the pre-fusion features are generated into adaptive thresholds through the threshold calculation layer;

[0043] The adaptive thresholds are generated into filtering thresholds through the soft thresholding layer;

[0044] The filtering thresholds and the pre-fusion features are fused through a residual connection layer to generate the fused fault feature.

[0045] Further, the process of constructing the traction fault feature extraction model includes:

[0046] The branch entropy values of the plurality of input processing branches are calculated based on a softmax function;

[0047] The dynamic weights are calculated based on the plurality of branch entropy values;

[0048] The multi-branch loss term based on cross-entropy loss is constructed;

[0049] The total loss function is calculated based on the dynamic weights and the multi-branch loss term, and the traction fault feature extraction model is optimized and trained through the total loss function.

[0050] Further, the multi-level classifier model includes a first level, a second level and a third level, and the process of determining the traction system fault type includes:

[0051] The fused traction fault feature is discriminated through the first level to determine electrical fault, mechanical fault and thermal fault;

[0052] The fused traction fault feature is positioned through the second level;

[0053] The fused traction fault feature is determined through the third level to determine the traction system fault type.

[0054] Further, the early warning emergency disposal strategy includes determining the protection level of the traction inverter, whether to start the auxiliary inverter and whether to return to the warehouse for maintenance.

[0055] In the above scheme, through multi-modal data deep fusion, adaptive feature enhancement and hierarchical classification, precise, rapid and interpretable monitoring and early warning of the subway traction system fault are realized.

[0056] Compared with the prior art, the application has the beneficial effects that

[0057] 1. The improved DRSN network architecture is adapted to the high-noise and strong-coupling environment of the multi-modal signals of the metro train traction system, and performs efficient fusion of multi-modal data, deep feature enhancement, multi-level classification optimization and intelligent emergency disposal, thereby realizing early and accurate early warning and rapid response of the fault type of the metro traction system.

[0058] 2. Through branch customized processing, dynamic feature enhancement and spatio-temporal joint modeling of multi-modal signals, high-precision early warning and judgment of complex faults of the metro traction system, such as electrical transient anomaly, mechanical wear and tear, thermal runaway, etc., are realized, and the operation and maintenance response speed and reliability of fault warning are improved.

[0059] 3. Through deep fusion of multi-modal data, adaptive feature enhancement and hierarchical classification, accurate, rapid and interpretable monitoring and early warning of the faults of the metro traction system are realized. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 Fig. 1 is a flowchart of the multi-data fusion vehicle fault monitoring and early warning method of the embodiment of the application;

[0061] Figure 2 Fig. 2 is a flowchart of the traction fault feature extraction model of the multi-data fusion vehicle fault monitoring and early warning method of the embodiment of the application;

[0062] Figure 3 Fig. 3 is a flowchart of the vibration signal branch of the multi-data fusion vehicle fault monitoring and early warning method of the embodiment of the application;

[0063] Figure 4 Fig. 4 is a flowchart of the temperature and thermal energy branch of the multi-data fusion vehicle fault monitoring and early warning method of the embodiment of the application. DETAILED DESCRIPTION

[0064] In order to make the purpose and advantages of the application more clear and explicit, the application will be further described below in combination with embodiments; it should be understood that the specific embodiments described herein are only used to explain the application, and do not limit the application.

[0065] The preferred embodiments of the 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 application, and are not intended to limit the protection scope of the application.

[0066] 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.

[0067] 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 fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or 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.

[0068] As shown in Figures 1 to 4 The present application provides a vehicle fault monitoring and early warning method of multi-data fusion, which is adapted to the high-noise, strong-coupling environment of the multi-modal signal of the metro train traction system through the improved DRSN network architecture, and performs efficient fusion of multi-modal data, deep feature enhancement, multi-level classification optimization and intelligent emergency disposal, realizing early and accurate early warning and rapid response of the fault type of the metro traction system.

[0069] As shown in Figures 1 to 4 The present embodiment provides a vehicle fault monitoring and early warning method of multi-data fusion, the management platform communicates with the inverter circuit sensor, motor temperature sensor, motor speed sensor and gearbox vibration sensor of the metro train traction system, the vehicle fault monitoring and early warning method applied to the management platform comprises:

[0070] The data collected by the inverter circuit sensor, motor temperature sensor, motor speed sensor and gearbox vibration sensor are fused through a traction fault feature extraction model, wherein the traction fault feature extraction model is constructed based on an improved DRSN network architecture, and has a multi-input processing branch for multi-modal signal and a channel attention mechanism for signal feature enhancement;

[0071] The fusion traction fault feature is determined by a multi-level classifier model to determine the traction system fault type;

[0072] The traction system fault type is matched with the fault tree to determine the fault range of the metro train, and the fault range is matched with the early warning list to generate an early warning emergency disposal strategy.

[0073] It can be understood that the improved DRSN (Deep Residual Shrinkage Network) is an improved residual network optimized for high-noise industrial scenes. The network architecture can enhance the anti-interference ability of the model to multi-modal signals through residual structure and adaptive threshold shrinkage mechanism. The multi-input branch design is aimed at the spatio-temporal characteristics of different sensor signals, such as the transient characteristics of circuit signals and the frequency domain characteristics of vibration signals, to realize differentiated feature extraction. This architecture can still maintain stable feature representation performance under complex working conditions and is suitable for high-noise and strong-coupling environments of subway traction systems.

[0074] As shown in Figure 2 Further, the traction fault feature extraction model has an improved residual shrinkage unit for cross-modal signal fusion based on an improved DRSN network architecture. The multi-input processing branch includes an electrical signal branch, a vibration signal branch, and a temperature thermal energy branch. The process of generating the fusion traction fault features includes:

[0075] The data collected by the inverter circuit sensor, motor temperature sensor, motor speed sensor, and gearbox vibration sensor are preprocessed to generate circuit fault signals, temperature fault signals, speed fault signals, and vibration fault signals, respectively.

[0076] The circuit fault signals and the speed fault signals are processed through the electrical signal branch to generate fault time sequence features.

[0077] The vibration fault signals are processed through the vibration signal branch to generate fault time-frequency features, wherein the vibration signal branch has the channel attention mechanism.

[0078] The temperature fault signals are processed through the temperature thermal energy branch to generate temperature spatio-temporal change features.

[0079] The fault time sequence features, fault time-frequency features, and temperature spatio-temporal change features are processed through the improved residual shrinkage unit to generate the fusion traction fault features.

[0080] Specifically, the process of generating circuit fault signals includes synchronously processing the three-phase current signals and voltage signals of the inverter circuit and the switch device (IGBT) drive signals at 100 kHz and 6 channels, and preprocessing them through sliding average filtering and Teager energy operator filtering algorithm.

[0081] Specifically, the temperature sequence of the motor stator winding and bearing seat is encoded and mapped to generate a temperature fault signal representing the temperature distribution. The data collected by the motor speed sensor is processed through differential calculation to generate a speed fault signal. The three-axis acceleration time domain data collected by the gearbox vibration sensor is processed through wavelet packet decomposition algorithm to generate a vibration fault signal.

[0082] Therefore, the modal specificity of different physical signals is reserved by the multi-branch input setting of the electrical signal branch, the vibration signal branch and the temperature thermal energy branch, the electrical signal branch, the vibration signal branch and the temperature thermal energy branch are preferably arranged in parallel, and efficient preliminary feature extraction of multi-source heterogeneous data is realized.

[0083] Further, the circuit fault signal includes a direct current side current voltage signal, an alternating current side current voltage signal and a switching device driving voltage signal, the electrical signal branch is provided with a first one-dimensional convolution normalization layer, a second one-dimensional convolution normalization layer, an activation function layer and a down-sampling layer, wherein the convolution kernel size of the first one-dimensional convolution normalization layer is smaller than the convolution kernel size of the second one-dimensional convolution normalization layer, and the process of generating the fault time sequence feature includes:

[0084] After the direct current side current voltage signal and the switching device driving voltage signal are passed through the first one-dimensional convolution normalization layer to suppress the noise of the switching device driving voltage signal and extract the short-time high-frequency mutation feature, the corresponding fault time sequence feature is generated through the activation function layer and the down-sampling layer;

[0085] After the alternating current side current voltage signal is passed through the second one-dimensional convolution normalization layer to extract the low-frequency harmonic distortion feature, the corresponding fault time sequence feature is generated through the activation function layer and the down-sampling layer.

[0086] Specifically, the first one-dimensional convolution normalization layer adopts a small convolution kernel with a size of 3x3, a length of 7 and a corresponding window of 0.7ms, and a batch normalization operation. The second one-dimensional convolution normalization layer adopts a large convolution kernel with a size of 5x5 and a dilation rate of 5, and a batch normalization operation. The activation function layer adopts a ReLU activation function to introduce a nonlinear transformation and enhance the expression ability of the model to complex fault patterns. The down-sampling layer sequentially performs maximum pooling (MaxPooling1D), average pooling (AvgPooling1D) and step convolution (Strided Conv1D) to realize feature abstraction and extraction.

[0087] As shown in Figure 3 Further, the vibration signal branch is provided with a first two-dimensional convolution pooling layer, a second two-dimensional convolution pooling layer and the channel attention mechanism, and the process of generating the fault time-frequency feature includes:

[0088] The vibration fault signal is passed through the first two-dimensional convolution pooling layer to suppress high-frequency electromagnetic noise and extract fault frequency sub-band features to generate a fault concentrated frequency sub-band;

[0089] The fault concentrated frequency sub-band is passed through the second two-dimensional convolution pooling layer to extract fault impact features to generate a vibration fault feature;

[0090] The vibration fault feature is adjusted by the channel attention mechanism to adjust the allocation weight of the multi-class fault frequency band channel, to generate the fault time-frequency feature, wherein the multi-class fault frequency band channel includes a gear surface wear pitting channel, a gear tooth missing channel, a gear eccentricity channel and a bearing coupling fault channel.

[0091] Specifically, the first two-dimensional convolution pooling layer adopts a 3x3 size convolution kernel and a maximum pooling (MaxPooling2D) to retain the most prominent features in the window and suppress random noise. The second two-dimensional convolution pooling layer adopts a 5x5 size convolution kernel and an average pooling (AvgPooling2D) to smooth the continuous vibration energy.

[0092] Specifically, the channel attention mechanism sequentially sets a global average pooling (Global Average Pooling2D), a dimension reduction fully connected layer, a channel weight generation layer and a feature rescaling layer, wherein the channel weight generation layer is used to generate the allocation weight of the multi-class fault frequency band channel, and the feature rescaling layer is used to multiply the vibration fault feature by the allocation weight to adjust the multi-class fault frequency band channel.

[0093] It can be understood that the channel attention mechanism can learn the gear surface wear pitting, gear tooth missing, gear eccentricity and bearing coupling fault feature frequency (GMF) through training data and test data, which can be expanded through simulation based on experimental acquisition.

[0094] As shown in Figure 4 Further, the temperature thermal energy branch is provided with a two-dimensional convolution layer, a normalization activation layer, a time feature extraction sub-branch, a spatial feature extraction sub-branch and a feature concatenation layer, wherein the time feature extraction sub-branch and the spatial feature extraction sub-branch are arranged in parallel, and the process of generating the temperature space-time change feature includes:

[0095] The temperature fault signal is extracted frame by frame by the two-dimensional convolution layer to generate an overheating area feature;

[0096] The overheating area feature is inhibited by the normalization activation layer to suppress sensor error noise and generate a highlighted overheating area feature;

[0097] A plurality of highlighted overheating area features of a plurality of frames are extracted by the time feature extraction sub-branch to extract cross-frame transient over-temperature features and continuous heat diffusion features, to generate a temperature rise and heat dissipation abnormal feature;

[0098] The highlighted overheating area feature is extracted by the spatial feature extraction sub-branch to generate an overheating area shape feature;

[0099] The temperature spatio-temporal change feature can realize, for example, failure monitoring of a heat dissipation fan.

[0100] Specifically, the motor temperature sensor is a temperature sensor array arranged at the stator winding and bearing seat of the motor, so that the temperature distribution can be determined. The two-dimensional convolution layer extracts spatial features frame by frame for two-dimensional data of height and width, maintains time independence, and sets 32 convolution kernels of 3x3 size. The normalization activation layer adopts a ReLU activation function to solve the distribution difference of temperature data across time steps, and improve the training stability and convergence speed.

[0101] Specifically, the time feature extraction sub-branch is provided with a Reshape layer, a one-dimensional convolution layer and a global average pooling layer (GlobalMaxPooling1D), which can effectively model and extract the time evolution law of the temperature field. The spatial feature extraction sub-branch is stacked with three two-dimensional convolution layers to realize multi-level spatial feature extraction from edges, textures to semantics. The feature concatenation layer realizes spatio-temporal joint modeling through feature concatenation (Concatenate).

[0102] Therefore, the temperature spatio-temporal change feature can realize, for example, failure monitoring of a heat dissipation fan.

[0103] In the above scheme, through branch customization processing, dynamic feature enhancement and spatio-temporal joint modeling of multi-modal signals, high-precision early warning judgment is realized for complex faults of the subway traction system, such as electrical transient anomaly, mechanical wear, thermal runaway, etc., and the response speed and reliability of fault warning operation and maintenance are improved.

[0104] Further, the improved residual shrinkage unit is sequentially provided with a convolution dimension reduction layer, a channel attention threshold generation block and a residual connection layer, and the process of generating the fusion traction fault feature includes:

[0105] The fault time sequence feature, the fault time-frequency feature and the temperature spatio-temporal change feature are respectively reduced and fused by the convolution dimension reduction layer to generate a pre-fusion feature;

[0106] Multiple pre-fusion features are generated through multiple channels of the channel attention threshold generation block to generate adaptive thresholds;

[0107] The adaptive thresholds and the fusion fault feature are fused through the residual connection layer to generate the fusion traction fault feature.

[0108] Specifically, the convolution dimension reduction layer sets a 1x1 size convolution for dimension reduction and a 3x3 size causal convolution for feature extraction, thereby improving the fault sensitivity of the model to, for example, current harmonics and vibration impacts.

[0109] Further, the channel attention threshold generation block sequentially sets a channel compression excitation layer, a threshold calculation layer and a soft thresholding layer, and the process of generating the adaptive threshold value comprises:

[0110] The plurality of pre-fusion features are respectively generated into a plurality of channel weights through the channel compression excitation layer;

[0111] The plurality of channel weights and the pre-fusion features are generated into a fusion threshold value through the threshold calculation layer;

[0112] The fusion threshold value is generated into the adaptive threshold value through the soft thresholding layer.

[0113] Specifically, the channel compression excitation layer comprises channel compression (Squeeze) and excitation (Excitation), the channel compression sets global average pooling (GAP), and the excitation sets a plurality of fully connected layers. The soft thresholding layer performs nonlinear filtering on the features to suppress noise.

[0114] Specifically, the residual connection layer can be represented as:

[0115] y = Conv1D(shortcut) + y0

[0116] In the formula, Conv1D(shortcut) represents 1*1 size one-dimensional convolution alignment dimension on the fusion fault feature shortcut of the skip channel attention threshold generation block, and y and y0 respectively represent the fusion traction fault feature and the adaptive threshold value.

[0117] Further, the construction process of the traction fault feature extraction model comprises:

[0118] Based on the softmax function, the branch entropy values of the plurality of input processing branches are calculated;

[0119] Based on the plurality of branch entropy values, a dynamic weight is calculated;

[0120] A multi-branch loss term based on cross-entropy loss is constructed;

[0121] Based on the dynamic weight and the multi-branch loss term, a total loss function is calculated, and the traction fault feature extraction model is optimized and trained through the total loss function.

[0122] Specifically, the construction process of the total loss function is:

[0123]

[0124] In the formula, H k represents the branch entropy value, σ represents the softmax function, (1 k ) coriginal output value of the kth branch belonging to the traction system fault type c (wherein l is lower L), ε represents the avoidance of σ (1 k ) c +ε takes the value of zero, w k represents a dynamic weight, T represents a sharpness control coefficient, preferably takes the value 0.3, L k represents a multi-branch loss term, which is constructed based on a cross-entropy loss form, y c represents a predicted value belonging to the traction system fault type c, and Loss represents a total loss function.

[0125] Further, the multi-level classifier model comprises a first level, a second level and a third level, and the process of determining the traction system fault type comprises:

[0126] discriminating electrical faults, mechanical faults and thermal faults through the first level based on the fusion traction fault features;

[0127] performing fault positioning through the second level based on the fusion traction fault features;

[0128] determining the traction system fault type through the third level based on the fusion traction fault features.

[0129] It can be understood that the first level, the second level and the third level can correspond to the levels of the fault tree currently carried by the management platform in sequence, so that the traction system fault type can be matched with the fault tree.

[0130] Further, the early warning emergency disposal strategy comprises determining a protection level of the traction inverter, whether to start an auxiliary inverter and whether to return to the warehouse for maintenance.

[0131] Specifically, whether to return to the warehouse for maintenance comprises verifying whether the vehicle has a serious fault causing the vehicle to be unable to start, such as traction invalidation or brake failure, and it is suggested to exit operation and return to the warehouse for maintenance. Whether to start the auxiliary inverter comprises extracting the response characteristics of the charging contactor based on the process data of the closing and opening of the charging contactor, and judging whether the response characteristic value is over the threshold. Determining the protection level of the traction inverter comprises reporting a warning when the pantograph catenary voltage exceeds a certain value.

[0132] In the above scheme, through multi-modal data deep fusion, self-adaptive feature enhancement and hierarchical classification, precise, rapid and interpretable monitoring and early warning of the subway traction system fault are realized.

[0133] In this embodiment, the improved DRSN network architecture is adapted to the high-noise and strong-coupling environment of the multi-modal signals of the metro train traction system, and performs efficient fusion of multi-modal data, deep feature enhancement, multi-level classification optimization and intelligent emergency disposal, realizing early and accurate early warning and rapid response of the fault type of the metro traction system. Through branch customization processing, dynamic feature enhancement and spatio-temporal joint modeling of multi-modal signals, high-precision early warning and judgment of complex faults of the metro traction system, such as electrical transient anomalies, mechanical wear and tear, thermal runaway, etc. are realized, and the response speed and reliability of fault warning operation and maintenance are improved. Through deep fusion of multi-modal data, adaptive feature enhancement and hierarchical classification, accurate, rapid and interpretable monitoring and early warning of the faults of the metro traction system are realized.

[0134] So far, the technical solutions of the present application have been described in connection with the preferred embodiments shown in the drawings, but those skilled in the art will readily 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 such changes or replacements will fall within the protection scope of the present application.

[0135] The above description is only the preferred embodiments of the present application and is not intended to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. A vehicle fault monitoring and early warning method based on multi-data fusion, characterized in that: The management platform communicates with the inverter circuit sensor, motor temperature sensor, motor speed sensor, and gearbox vibration sensor of the subway train traction system. The vehicle fault monitoring and early warning method applied to the management platform includes: The data collected by the inverter circuit sensor, the motor temperature sensor, the motor speed sensor, and the gearbox vibration sensor are used in a traction fault feature extraction model to generate a fused traction fault feature, wherein the traction fault feature extraction model is constructed based on an improved DRSN network architecture and has a multi-input processing branch for multimodal signals and a channel attention mechanism for signal feature enhancement; Determine the type of traction system fault by using the fused traction fault features through a multi-level classifier model; The traction system fault type is matched with the fault tree to determine the fault range of the subway train, and the fault range is matched with the early warning list to generate an early warning emergency response strategy.

2. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 1 is characterized in that: The traction fault feature extraction model has an improved residual shrinkage unit for cross-modal signal fusion. The multi-input processing branch includes an electrical signal branch, a vibration signal branch, and a temperature and thermal energy branch. The process of generating the fused traction fault feature includes: Preprocessing the data collected by the inverter circuit sensor, the motor temperature sensor, the motor speed sensor and the gearbox vibration sensor to generate a circuit fault signal, a temperature fault signal, a speed fault signal and a vibration fault signal respectively; Generate a fault timing feature by using the circuit fault signal and the speed fault signal through an electrical signal branch; Generate fault time-frequency features from the vibration fault signal through a vibration signal branch, wherein the vibration signal branch has the channel attention mechanism; The temperature fault signal is used to generate a temperature spatiotemporal variation feature through a temperature thermal energy branch; The fault time series features, fault time-frequency features and temperature spatiotemporal variation features are used to generate the fused traction fault features through an improved residual shrinkage unit.

3. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 2 is characterized in that: The circuit fault signal includes a DC side current and voltage signal, an AC side current and voltage signal, and a switching device drive voltage signal. The electrical signal branch is provided with a first one-dimensional convolution normalization layer, a second one-dimensional convolution normalization layer, an activation function layer, and a downsampling layer. The convolution kernel size of the first one-dimensional convolution normalization layer is smaller than the convolution kernel size of the second one-dimensional convolution normalization layer. The process of generating the fault time series feature includes: After the DC side current and voltage signals and the switching device driving voltage signals are passed through the first one-dimensional convolution normalization layer to suppress the noise of the switching device driving voltage signals and extract the short-term high-frequency mutation features, the corresponding fault time series features are generated through the activation function layer and the downsampling layer; After the low-frequency harmonic distortion features of the AC side current and voltage signals are extracted through the second one-dimensional convolution normalization layer, the corresponding fault time series features are generated through the activation function layer and the downsampling layer.

4. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 2 is characterized in that: The vibration signal branch is provided with a first two-dimensional convolution pooling layer, a second two-dimensional convolution pooling layer, and the channel attention mechanism. The process of generating the fault time-frequency feature includes: Passing the vibration fault signal through the first two-dimensional convolutional pooling layer to suppress high-frequency electromagnetic noise and extract fault frequency sub-band features to generate a fault concentrated frequency sub-band; Extracting fault impact features from the fault concentrated frequency sub-band through the second two-dimensional convolutional pooling layer to generate vibration fault features; The vibration fault feature is used to adjust the allocation weights of multiple fault frequency band channels through the channel attention mechanism to generate the fault time-frequency feature, where the multiple fault frequency band channels include tooth surface wear and pitting channel, broken tooth defect channel, gear eccentricity channel and bearing coupling fault channel.

5. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 2 is characterized in that: The temperature and thermal energy branch is provided with a two-dimensional convolution layer, a normalized activation layer, a time feature extraction sub-branch, a spatial feature extraction sub-branch, and a feature splicing layer, wherein the time feature extraction sub-branch and the spatial feature extraction sub-branch are provided in parallel. The process of generating the temperature spatiotemporal variation feature includes: Extracting spatial local overheating regions frame by frame from the temperature fault signal through the two-dimensional convolution layer to generate overheating region features; Passing the overheating region feature through the normalized activation layer to suppress sensor error noise and generate a prominent overheating region feature; The multiple prominent overheating area features of the multiple frames are extracted through the time feature extraction sub-branch to extract the transient overtemperature features and continuous heat diffusion features across the frames, and generate the temperature rise and heat dissipation abnormality features; Extracting the prominent overheating area features through the spatial feature extraction sub-branch and generating overheating area shape features; The abnormal characteristics of temperature rise and heat dissipation and the shape characteristics of the overheating area are generated through a feature splicing layer to generate a joint fault judgment feature to generate the temperature spatiotemporal change feature.

6. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 2 is characterized in that: The improved residual shrinkage unit sequentially sets a convolutional dimension reduction layer, a channel attention threshold generation block, and a residual connection layer to generate the fused traction fault feature, including: The fault time series features, fault time-frequency features and temperature spatiotemporal variation features are respectively dimensionality-reduced and fused through the convolutional dimensionality reduction layer to generate pre-fusion features; Passing the plurality of pre-fused features through the multi-channel of the channel attention threshold generation block to generate an adaptive threshold; The adaptive threshold and the fused fault feature are fused through the residual connection layer to generate the fused traction fault feature.

7. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 6 is characterized in that: The channel attention threshold generation block sequentially sets a channel compression excitation layer, a threshold calculation layer, and a soft thresholding layer residual connection layer. The process of generating the adaptive threshold includes: Passing the plurality of pre-fused features through the channel compression excitation layer to generate a plurality of channel weights; Passing the plurality of channel weights and the pre-fusion features through the threshold calculation layer to generate a fusion threshold; The fusion threshold is passed through the soft thresholding layer to generate the adaptive threshold.

8. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 1 is characterized in that: The construction process of the traction fault feature extraction model includes: Calculate the branch entropy value of each multi-input processing branch based on the softmax function; Calculating a dynamic weight based on a plurality of said branch entropy values; Construct a multi-branch loss term based on cross entropy loss; A total loss function is calculated based on the dynamic weight and the multi-branch loss term, and the traction fault feature extraction model is optimized and trained using the total loss function.

9. The vehicle fault monitoring and early warning method based on multi-data fusion according to claim 1 is characterized in that: The multi-level classifier model includes a first level, a second level, and a third level. The process of determining the type of traction system fault includes: Using the fused traction fault feature to identify electrical fault, mechanical fault, and thermal fault at the first level; Perform fault location on the fused traction fault feature through the second level; The fused traction fault features are used to determine the traction system fault type through the third level.

10. The vehicle fault monitoring and early warning method based on multi-data fusion according to any one of claims 1 to 9, characterized in that: The early warning emergency response strategy includes determining the protection level of the traction inverter, whether to start the auxiliary inverter, and whether to return to the depot for maintenance.

Images