A rail vehicle fault early warning method, device, equipment and storage medium

Abstract

The application provides a rail vehicle fault early warning method, device, equipment and storage medium, comprising: collecting current and temperature signals of a low-voltage electrical appliance cabinet power supply loop; sending the current high-frequency component into an arc detection channel, performing time-frequency analysis and outputting an arc fault probability vector through a first neural network model; sending the current low-frequency component and the temperature signal into an overheating detection channel, extracting time sequence features and outputting an overheating fault probability vector through a second neural network model; calculating a fusion weight based on the correlation of the two types of faults, weighting and fusing the two probability vectors to obtain a final fusion probability vector, and determining a vehicle warning level accordingly. The above scheme ensures the recognition accuracy of a single fault type and improves the stability and accuracy of the overall fault determination.

Description

Technical Field

[0001] This application relates to the field of electrical safety monitoring in rail transit, and more specifically, to a method, device, equipment, and storage medium for early warning of rail vehicle faults. Background Technology

[0002] The low-voltage electrical cabinet of a rail vehicle is the core execution unit of the vehicle's electrical system, undertaking key functions such as power distribution, circuit protection, and on-board electrical control. During long-term operation, the vehicle is affected by continuous vibration, frequent load fluctuations, and contact aging, which can easily lead to early faults such as local overheating and latent arcing in the power supply circuit within the cabinet. If these faults are not identified and warned in time, they can easily cause electrical fires, line damage, and other safety accidents, directly affecting the safe operation and stable functioning of the rail vehicle.

[0003] Currently, fault monitoring of low-voltage electrical cabinets in rail vehicles mostly adopts a single physical quantity threshold detection scheme. This involves deploying temperature sensors and current acquisition devices to collect temperature and current data of the power supply circuit in real time, and then directly comparing the collected values ​​with preset safety thresholds to determine whether there is an abnormal fault state in the circuit.

[0004] This threshold comparison monitoring method can only make simple judgments on physical quantities that exceed limits. It cannot identify the differences in signal characteristics corresponding to different faults, and it does not have the ability to identify early weak fault signals. It is very easy to have false alarms and missed alarms, and it is difficult to meet the actual needs of accurate early warning of vehicle faults. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method, device, equipment and storage medium for early warning of faults in rail vehicles, which not only ensures the identification accuracy of a single fault type, but also improves the stability and accuracy of the overall fault judgment.

[0006] In a first aspect, embodiments of this application provide a method for early warning of faults in rail vehicles, the method comprising: Acquire the current and temperature signals of the power supply circuit inside the low-voltage electrical cabinet of the rail vehicle; The high-frequency component of the current signal is input to the arc detection channel, and the high-frequency component is subjected to time-frequency analysis to generate a time-frequency characterization. The time-frequency characterization is processed by the first neural network model to output a first probability vector characterizing the arc-type fault state. The low-frequency component of the current signal and the temperature signal are input to the overheat detection channel, their time-series feature sequences are extracted, and the time-series feature sequences are processed by the second neural network model to output a second probability vector representing the overheat fault state. The fusion weight is calculated based on the correlation between the arc fault and the overheat fault. The first probability vector and the second probability vector are weighted and fused according to the fusion weight to obtain the final fusion probability vector. The warning level of the rail vehicle is determined based on the final fusion probability vector.

[0007] Optionally, performing time-frequency analysis on the high-frequency components to generate a time-frequency characterization includes: Using Morlet wavelet as the mother wavelet function, a continuous wavelet transform is performed on the high-frequency components to obtain wavelet coefficients under different scale factors and translation factors. The scale axis is divided into three intervals: low frequency, mid frequency, and high frequency. The wavelet coefficient amplitudes of each interval are mapped to the pixel values ​​of the three RGB color channels to generate a size-normalized time-frequency scalogram image as the time-frequency representation.

[0008] Optionally, the first neural network model includes: The initial convolutional layer is used to perform preliminary convolution and downsampling on the input time-frequency representation; At least one multi-scale parallel convolution module is used to extract features from the input feature map by applying convolution kernels of three sizes: 1×1, 3×3, and 5×5 in parallel. Dilated convolutions with different dilation rates are embedded in the branches of the 3×3 and 5×5 convolution kernels respectively. The feature maps output by each branch are concatenated along the channel dimension and then fused by 1×1 convolution. The channel-space dual attention module includes a cascaded channel attention submodule and a spatial attention submodule. The channel attention submodule is used to generate channel weights through global average pooling and fully connected layers to perform channel-wise weighting on the input feature map. The spatial attention submodule is used to generate a spatial attention map through max pooling and average pooling along the channel dimension to perform element-wise weighting on the input feature map. A global average pooling layer, a fully connected layer, and a Softmax output layer are used to output the first probability vector.

[0009] Optionally, the extraction of its temporal feature sequence includes: The low-frequency component and the temperature signal are segmented by a sliding time window. At least two features from the temperature slope, current RMS value, contact point voltage drop and signal entropy are extracted in each window to form an original feature vector sequence. For each feature vector in the original feature vector sequence, its first-order difference and second-order difference are calculated respectively. The original feature value, first-order difference value and second-order difference value are concatenated along the feature dimension to form an enhanced feature vector sequence, which is used as the input of the second neural network model.

[0010] Optionally, the second neural network model includes at least one layer of trend-gated memory units, each including a forget gate. The forget gate generates a forget gate output based on the hidden state of the previous time step and the input of the current time step. The trend-gated memory unit updates its memory unit state in the following manner: Based on the first-order difference mean of the input at the current moment, a trend gating factor is generated through linear transformation and the Sigmoid activation function; The forget gate output is multiplied element-wise by the trend gating factor to obtain the trend-sensitive effective forgetting rate; Substitute the effective forgetting rate, the memory cell state at the previous time step, the input gate output at the current time step, and the candidate memory state into the memory cell state update formula to calculate the memory cell state at the current time step.

[0011] Optionally, the method further includes: The second-order difference mean of the time-series feature sequence is calculated in real time and used as a measure of the trend change rate. When the mean of the second-order difference exceeds the preset gain threshold, the step size of the sliding time window input to the second neural network model is shortened to one-quarter of the baseline step size; When the mean of the second-order difference does not exceed the gain threshold, the step size of the sliding time window is kept as the baseline step size.

[0012] Optionally, the step of calculating the fusion weight based on the correlation between the arc-type fault and the overheat-type fault includes: Calculate the first sum of the arc warning probability and the arc alarm probability in the first probability vector; Calculate the second sum of the overheat warning probability and the overheat alarm probability in the second probability vector; The difference between the first sum and the second sum is multiplied by a preset adjustment coefficient and then added to a preset base weight to obtain the initial fusion weight. The initial fusion weight is restricted to a preset value range to obtain the fusion weight.

[0013] Secondly, embodiments of this application provide a rail vehicle fault early warning device, the device comprising: The signal acquisition module is used to acquire the current and temperature signals of the power supply circuit inside the low-voltage electrical cabinet of the rail vehicle. The first probability vector determination module is used to input the high-frequency component in the current signal to the arc detection channel, perform time-frequency analysis on the high-frequency component to generate a time-frequency representation, process the time-frequency representation using a first neural network model, and output a first probability vector representing the arc-type fault state. The second probability vector determination module is used to input the low-frequency component of the current signal and the temperature signal into the overheat detection channel, extract their time-series feature sequence, process the time-series feature sequence using the second neural network model, and output a second probability vector representing the overheat fault state. The fusion probability vector production module is used to calculate the fusion weight based on the correlation between the arc-type fault and the overheat-type fault, and to perform weighted fusion of the first probability vector and the second probability vector according to the fusion weight to obtain the final fusion probability vector. The early warning level determination module is used to determine the early warning level of the rail vehicle based on the final fusion probability vector.

[0014] Optionally, performing time-frequency analysis on the high-frequency components to generate a time-frequency characterization includes: Using Morlet wavelet as the mother wavelet function, a continuous wavelet transform is performed on the high-frequency components to obtain wavelet coefficients under different scale factors and translation factors. The scale axis is divided into three intervals: low frequency, mid frequency, and high frequency. The wavelet coefficient amplitudes of each interval are mapped to the pixel values ​​of the three RGB color channels to generate a size-normalized time-frequency scalogram image as the time-frequency representation.

[0015] Optionally, the first neural network model includes: The initial convolutional layer is used to perform preliminary convolution and downsampling on the input time-frequency representation; At least one multi-scale parallel convolution module is used to extract features from the input feature map by applying convolution kernels of three sizes: 1×1, 3×3, and 5×5 in parallel. Dilated convolutions with different dilation rates are embedded in the branches of the 3×3 and 5×5 convolution kernels respectively. The feature maps output by each branch are concatenated along the channel dimension and then fused by 1×1 convolution. The channel-space dual attention module includes a cascaded channel attention submodule and a spatial attention submodule. The channel attention submodule is used to generate channel weights through global average pooling and fully connected layers to perform channel-wise weighting on the input feature map. The spatial attention submodule is used to generate a spatial attention map through max pooling and average pooling along the channel dimension to perform element-wise weighting on the input feature map. A global average pooling layer, a fully connected layer, and a Softmax output layer are used to output the first probability vector.

[0016] Optionally, the extraction of its temporal feature sequence includes: The low-frequency component and the temperature signal are segmented by a sliding time window. At least two features from the temperature slope, current RMS value, contact point voltage drop and signal entropy are extracted in each window to form an original feature vector sequence. For each feature vector in the original feature vector sequence, its first-order difference and second-order difference are calculated respectively. The original feature value, first-order difference value and second-order difference value are concatenated along the feature dimension to form an enhanced feature vector sequence, which is used as the input of the second neural network model.

[0017] Optionally, the second neural network model includes at least one layer of trend-gated memory units, each including a forget gate. The forget gate generates a forget gate output based on the hidden state of the previous time step and the input of the current time step. The trend-gated memory unit updates its memory unit state in the following manner: Based on the first-order difference mean of the input at the current moment, a trend gating factor is generated through linear transformation and the Sigmoid activation function; The forget gate output is multiplied element-wise by the trend gating factor to obtain the trend-sensitive effective forgetting rate; Substitute the effective forgetting rate, the memory cell state at the previous time step, the input gate output at the current time step, and the candidate memory state into the memory cell state update formula to calculate the memory cell state at the current time step.

[0018] Optionally, the device further includes a sliding time window adjustment module for: The second-order difference mean of the time-series feature sequence is calculated in real time and used as a measure of the trend change rate. When the mean of the second-order difference exceeds the preset gain threshold, the step size of the sliding time window input to the second neural network model is shortened to one-quarter of the baseline step size; When the mean of the second-order difference does not exceed the gain threshold, the step size of the sliding time window is kept as the baseline step size.

[0019] Optionally, the step of calculating the fusion weight based on the correlation between the arc-type fault and the overheat-type fault includes: Calculate the first sum of the arc warning probability and the arc alarm probability in the first probability vector; Calculate the second sum of the overheat warning probability and the overheat alarm probability in the second probability vector; The difference between the first sum and the second sum is multiplied by a preset adjustment coefficient and then added to a preset base weight to obtain the initial fusion weight. The initial fusion weight is restricted to a preset value range to obtain the fusion weight.

[0020] Thirdly, embodiments of this application provide a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the rail vehicle fault early warning method described in any of the optional embodiments of the first aspect are performed.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the rail vehicle fault early warning method described in any of the optional embodiments of the first aspect.

[0022] The technical solution provided in this application includes, but is not limited to, the following beneficial effects: By acquiring current and temperature signals from the power supply circuit within the low-voltage electrical cabinet of a rail vehicle, electrical and thermal characteristic signals reflecting electrical faults can be collected simultaneously. This comprehensively covers the core characterization information of two typical faults: arc faults and overheating faults. It avoids the information loss problem caused by using only a single signal for monitoring and provides comprehensive, accurate, and stable raw data support for subsequent fault feature extraction and status determination, ensuring the basic reliability of fault early warning from the data source.

[0023] The high-frequency components of the current signal are input into the arc detection channel. Time-frequency analysis is used to generate corresponding time-frequency representations. The first probability vector representing the arc-type fault state is output using the first neural network model. This allows for targeted detection that matches the high-frequency transient signal characteristics of arc faults. By fully exploring the fault features hidden in the high-frequency current through time-frequency analysis, and then using the neural network model to complete feature learning and state determination, the quantitative assessment of arc faults can be achieved, effectively improving the precision and accuracy of arc-type fault detection.

[0024] The low-frequency component of the current signal and the temperature signal are input together into the overheat detection channel. After extracting the time-series feature sequence, the second probability vector representing the overheat fault state is output through the second neural network model. This can adapt to the slow and gradual time-series change pattern of overheat faults. By extracting time-series features, the development trend of the fault is captured. Relying on the dedicated neural network model, the overheat state can be accurately quantified and judged. This can effectively identify early hidden overheat faults and improve the sensitivity and reliability of overheat fault detection.

[0025] The fusion weight is calculated based on the inherent correlation between arc faults and overheat faults. The first probability vector and the second probability vector are then weighted and fused according to the weight to obtain the final fusion probability vector. This can fully combine the physical correlation characteristics of the two types of faults to optimize the decision logic, avoid the judgment bias or abnormal results of a single detection channel from dominating the overall conclusion, make the fault judgment result more comprehensive and objective, and greatly reduce the risk of misjudgment and missed judgment caused by single channel error.

[0026] The warning level of the rail vehicle is determined by the final fusion probability vector. Based on the fusion probability of the comprehensive dual-channel judgment results, the corresponding warning level is divided according to the degree of fault risk. This makes the warning results highly matched with the actual fault state of the equipment, and the warning judgment is more reasonable and referential. It can provide accurate and effective status indications for the safe operation and maintenance of rail vehicles.

[0027] In summary, this application independently detects different types of faults through separate channels, combines fault correlations for decision fusion, and then provides graded early warnings based on the comprehensive results. This not only ensures the identification accuracy of a single fault type but also improves the stability and accuracy of the overall fault determination, providing efficient and reliable early fault warning protection for the safe operation of low-voltage electrical cabinets in rail vehicles.

[0028] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart of a rail vehicle fault early warning method provided in Embodiment 1 of this application is shown; Figure 2 This diagram illustrates a hardware deployment architecture of a low-voltage electrical cabinet for a rail vehicle, as provided in Embodiment 1 of this application. Figure 3 This paper illustrates an overall architecture diagram of a dual-channel attention fusion network provided in Embodiment 1 of this application; Figure 4 A flowchart of a time-frequency characterization generation method provided in Embodiment 1 of this application is shown; Figure 5 A schematic diagram of an improved CNN network structure provided in Embodiment 1 of this application is shown; Figure 6 A flowchart of a time-series feature extraction method provided in Embodiment 1 of this application is shown; Figure 7 A flowchart of a memory cell state update method provided in Embodiment 1 of this application is shown; Figure 8 A flowchart of a sliding time window adjustment method provided in Embodiment 1 of this application is shown; Figure 9 A schematic diagram of an improved LSTM network structure provided in Embodiment 1 of this application is shown; Figure 10 A flowchart of a fusion weight calculation method provided in Embodiment 1 of this application is shown; Figure 11 A schematic diagram of a decision fusion module logic provided in Embodiment 1 of this application is shown; Figure 12 This paper shows a schematic diagram of the structure of a rail vehicle fault early warning device provided in Embodiment 2 of this application; Figure 13 A schematic diagram of the structure of a computer device provided in Embodiment 3 of this application is shown. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0032] Example 1 To facilitate understanding of this application, the following is combined with... Figure 1 The flowchart of the fault early warning method for rail vehicles provided in Embodiment 1 of this application illustrates the content of Embodiment 1 in detail.

[0033] See Figure 1 As shown, Figure 1 A flowchart of a rail vehicle fault early warning method provided in Embodiment 1 of this application is shown, wherein the method includes steps S101 to S105: S101: Acquire the current and temperature signals of the power supply circuit inside the low-voltage electrical cabinet of the rail vehicle.

[0034] Specifically, by using high-frequency current transformers, power frequency precision current sensors, and distributed PT100 (Platinum Resistance Thermometer 100) temperature sensors deployed in the low-voltage electrical cabinet of the rail vehicle, real-time multi-parameter parallel acquisition of the power supply circuit is performed to obtain the corresponding current and temperature signals.

[0035] High-frequency current transformers with a sampling rate of ≥5MHz are used to acquire high-frequency current signals corresponding to arc faults; power frequency precision current sensors with a sampling rate of ≥10kHz are used to acquire low-frequency current signals corresponding to overheat faults; PT100 resistance temperature sensors are used to acquire temperature signals of conductors and contact points.

[0036] See Figure 2 As shown, Figure 2 This paper illustrates a hardware deployment architecture diagram of a low-voltage electrical cabinet for a rail vehicle according to Embodiment 1 of this application, wherein... Figure 2 The medium- and high-frequency CT (high-frequency current transformer), the power frequency CT (power frequency current sensor), and the PT100 temperature sensor collect signals and connect them to the ECU edge computing unit. The ECU communicates with the train control system through the train gateway. All hardware symbols and definitions are completely consistent.

[0037] S102: Input the high-frequency component in the current signal to the arc detection channel, perform time-frequency analysis on the high-frequency component to generate a time-frequency characterization, process the time-frequency characterization using a first neural network model, and output a first probability vector characterizing the arc-type fault state.

[0038] Specifically, the acquired raw current signal is split according to its frequency characteristics. The high-frequency component automatically enters the arc detection channel, while the low-frequency timing component enters the overheat detection channel, thus realizing the parallel processing of the two types of fault signals.

[0039] A continuous wavelet transform (CWT) is performed on the high-frequency current signal of the arc detection channel to generate a time-frequency representation. The core formula of the transform is: ; in, Continuous wavelet transform coefficients of high-frequency current signals in the arc channel, dedicated to arc time-frequency analysis; The high-frequency current time-domain signal to be transformed in the arc channel is a dedicated arc signal input; The wavelet transform scaling factor of the electric arc signal is inversely proportional to the frequency of the electric arc signal. : Wavelet transform shift factor of electric arc signal, corresponding to the time position of electric arc signal; Morlet mother wavelet basis functions specifically for arc signal analysis; The complex conjugate form of the Morlet mother wavelet basis function for electric arc signals; : A definite integral symbol specifically used for time-frequency analysis of electric arc signals, representing integration over the entire time axis; : A dedicated time variable for electric arc signals, representing the time dimension of the electric arc signal; The absolute value of the scale factor of the electric arc signal ensures that the denominator of the transformation is non-negative; The time differential symbol is dedicated to the electric arc signal and is used in conjunction with the integral to complete the time-frequency transformation.

[0040] After generating the time-frequency representation, it is input into the first neural network model (an improved CNN (Convolutional Neural Network)) for feature extraction and inference, and finally outputs the first probability vector, denoted as: ; in, The first probability vector specific to arc faults; : Probability of normal state specific to arc fault; : Probability of arc fault-specific early warning status; : Probability of arc fault-specific alarm status.

[0041] S103: Input the low-frequency component of the current signal and the temperature signal into the overheat detection channel, extract their time-series feature sequence, process the time-series feature sequence using the second neural network model, and output a second probability vector representing the overheat fault state.

[0042] Specifically, the low-frequency timing components in the current signal are synchronously input into the overheat detection channel along with the temperature signal. The timing feature sequence is first extracted through sliding window segmentation and feature engineering, and then the feature sequence is input into the second neural network model (an improved LSTM (Long Short-Term Memory) network) for inference.

[0043] The second neural network model ultimately outputs a second probability vector, denoted as: ; in, : The second probability vector specific to overheating faults; : Probability of overheating fault occurring under normal conditions; Probability of overheating fault-specific warning status; Probability of overheating fault-specific alarm status.

[0044] S104: Calculate the fusion weight based on the correlation between the arc fault and the overheat fault, and perform weighted fusion of the first probability vector and the second probability vector according to the fusion weight to obtain the final fusion probability vector.

[0045] Specifically, there is a physical causal coupling relationship between arc faults and overheating faults. Continuous arc discharge will cause the local temperature of the contact point to rise, which will lead to overheating faults. Based on this relationship, a causal-coupling dynamic weighting model (CCDW) is constructed to calculate the fusion weight.

[0046] Before weighted fusion, the first and second probability vectors need to be calibrated for temperature-shortened reliability. The calibration formula is as follows: ; in, Confidence calibration exclusive first Class state probability; : The natural exponential function for probability calibration; Probability calibration exclusive first Class state logit vector; Probabilistic calibration of dedicated temperature scaling parameters to meet [specific requirements]. >0; : A special summation symbol for probability calibration, representing the summation of probabilities over all fault categories; Probability calibration exclusive first Class state logit vector; : Probabilistic calibration-specific category number, representing the current calibration category; The probability calibration-specific category traversal index represents all calibration categories.

[0047] Specifically, the calibration probability of the arc channel is obtained after calibration. Overheating path calibration probability Then, the final fusion probability vector is obtained through the weighted fusion formula: ; in, : Decision fusion-specific final probability vector; Decision Integration Exclusive Dynamically integrate weights in real time; : Decision fusion uses a dedicated time variable to represent the time point for weight calculation; : Decision fusion of probability vector after calibration of dedicated arc channel; : Decision fusion dedicated overheating channel weighting coefficient; : Probability vector after calibration of the dedicated overheating channel for decision fusion; Dynamic fusion weights The calculation formula is: ; in, Weight calculation exclusive Initial fusion weights at time step; The weight is a dedicated base weight for weight calculation, with a default value of 0.5. : Decision-making integration specific adjustment coefficient; Weighted calculation of the probability of risk in the dedicated arc channel; Weighted calculation of the probability of overheating risk in dedicated channels.

[0048] The calculated initial weights need to be truncated to the [0.3, 0.7] interval to avoid extreme weighting in a single channel, thus obtaining the final fused weights.

[0049] S105: Determine the warning level of the rail vehicle based on the final fusion probability vector.

[0050] Specifically, based on the tiered early warning logic, and combined with the final fusion probability, the duration of the temperature slope, and the absolute temperature limit, the early warning level is divided into three levels: normal, early warning, and alarm.

[0051] Normal state determination: and The system only monitors and records.

[0052] in, : The probability of a final fusion warning specific to the graded early warning system; : Specific warning probability thresholds for tiered early warning systems; : The probability of a final fusion alarm for graded early warning; : Specific alarm probability thresholds for graded early warning.

[0053] Warning status determination: Or the temperature slope (the slope of the linear fit of temperature over time within the sliding window) is continuous. The number of windows is positive and exceeds The system issued a warning.

[0054] Alarm status determination: or an absolute temperature exceeding 120°C The system triggers an audible and visual alarm and protection linkage.

[0055] in, : Early warning determination of the probability of final fusion warning; : Specific warning probability thresholds for early warning determination; The warning is based on the duration of the temperature slope within a specific window, and is only used for trend warnings. The warning threshold is a specific temperature slope, and is only used for trend warnings.

[0056] Alarm determination is the probability of a final, integrated alarm.

[0057] See Figure 3 As shown, Figure 3 This paper illustrates the overall architecture of a dual-channel attention fusion network provided in Embodiment 1 of this application, wherein... Figure 3 The overall architecture of the early warning system for electrical fires in low-voltage electrical cabinets of rail vehicles is as follows: The system collects high-frequency current signals of arc faults through high-frequency current transformers, low-frequency current signals of overheating faults through power frequency current sensors, and contact point temperature signals through PT100 resistance temperature sensors. The raw signals are split into two paths according to frequency after signal splitting. The high-frequency component enters the arc channel, and after continuous wavelet transform (CWT) to generate a time-frequency representation, it is input into an improved CNN, which outputs three probability vectors for arc faults: normal, early warning, and alarm. The low-frequency time-series component and the temperature signal enter the overheating channel, and after time-series feature extraction, they are input into an improved LSTM, which outputs three probability vectors for overheating faults: normal, early warning, and alarm. The two probability vectors are input into the decision fusion module, and weighted fusion is completed through causal coupling dynamic weight CCDW and confidence calibration. Finally, the graded early warning module outputs three-level judgment results: normal, early warning, and alarm.

[0058] In an optional implementation, see Figure 4As shown, Figure 4 The flowchart of a time-frequency characterization generation method provided in Embodiment 1 of this application is shown, wherein the step of performing time-frequency analysis on the high-frequency components to generate a time-frequency characterization includes steps S401-S402: S401: Using Morlet wavelet as the mother wavelet function, perform continuous wavelet transform on the high-frequency components to obtain wavelet coefficients under different scale factors and translation factors.

[0059] Specifically, the continuous wavelet transform uses the Morlet wavelet as the mother wavelet, which is adapted to the wide frequency band of 5kHz to 5MHz and the microsecond-level transient characteristics of the arc signal, and accurately extracts the time and frequency features.

[0060] The formula for generating a dedicated continuous wavelet transform for time-frequency representation is as follows: ; in, : Generate custom wavelet transform coefficients from time-frequency graphs; The time-frequency graph generates a dedicated high-frequency current time-domain signal, which is used only for time-frequency characterization input. The time-frequency plot generates a proprietary scaling factor, which is used only for frequency band division of Scalogram images; : A dedicated translation factor is generated for the time-frequency plot, which is only used for time axis positioning in Scalogram images; : Generate a dedicated Morlet mother wavelet basis function from the time-frequency plot; : Generating a dedicated mother wavelet complex conjugate from the time-frequency graph; : Generates a dedicated time variable from the time-frequency graph; : Generating the absolute value of the exclusive scale factor for time-frequency plots : Generate custom time differential symbols from time-frequency graphs.

[0061] S402: Divide the scale axis into three intervals: low frequency (corresponding to 0~50kHz), mid frequency (corresponding to 50~500kHz), and high frequency (corresponding to 500kHz~5MHz), and map the wavelet coefficient amplitude of each interval to the pixel values ​​of the three color channels of RGB (Red Green Blue) to generate a size-normalized time-frequency scalogram image as the time-frequency representation.

[0062] Specifically, the scale axis is divided into three frequency bands: low, medium, and high, which correspond to the three RGB channels respectively, and the wavelet coefficient amplitude is mapped to the corresponding channel pixel value.

[0063] The final output is a normalized time-frequency scalogram image with dimensions of 224 (time-frequency plot-specific height pixel value) × 224 (time-frequency plot-specific width pixel value) × 3 (time-frequency plot-specific RGB channel number), which is adapted to the standard input format of the first neural network model.

[0064] In an optional implementation, the first neural network model includes: The initial convolutional layer is used to perform preliminary convolution and downsampling on the input time-frequency representation.

[0065] Specifically, the initial convolutional layer uses a 3 (initial convolutional kernel height) × 3 (initial convolutional kernel width) convolutional kernel, a stride of 2 (initial convolutional zero-padding value), and zero-padding of 1 (initial convolutional stride), and convolves the 224×224×3 time-frequency map.

[0066] After convolution, batch normalization (BN) and ReLU (Rectified LinearUnit) activation are performed sequentially, outputting a feature map of 112 (height of the initial convolution-specific feature map) × 112 (width of the initial convolution-specific feature map) × 64 (number of channels of the initial convolution-specific feature map), completing the initial feature extraction and downsampling.

[0067] At least one multi-scale parallel convolution module (MSPC) is used to extract features from the input feature map by applying convolution kernels of three sizes: 1×1, 3×3, and 5×5 in parallel. Dilated convolutions with different dilation rates are embedded in the branches of the 3×3 and 5×5 convolution kernels. The feature maps output from each branch are concatenated along the channel dimension and then fused by a 1×1 convolution.

[0068] The input feature map of the multi-scale parallel convolutional module (MSPC) is represented as follows: ; in, : The input feature map is specific to the MSPC module and is used only for multi-scale feature extraction; The MSPC module has its own set of real numbers, representing the elements of the feature map as real numbers. : Feature map height specific to the MSPC module, used only as input for multi-scale convolutions; : The feature map width is specific to the MSPC module and is only used as input for multi-scale convolutions; : The number of channels in the feature map specific to the MSPC module, used only for multi-scale convolution input.

[0069] The 3×3 branch embeds a dilated convolution with a dilation rate of 2 (the dilation rate specific to 3×3 convolution), and the 5×5 branch embeds a dilated convolution with a dilation rate of 4 (the dilation rate specific to 5×5 convolution). The three-branch output is... .

[0070] The formula for feature fusion in the MSPC module is: ; in, : MSPC module-specific output feature map; : 1×1 convolution specific to the MSPC module; MSPC module-specific channel concatenation operator; : 1×1 convolution branch output specific to the MSPC module; MSPC module-specific 3×3 dilated convolution branch output; : MSPC module exclusive 5×5 dilated convolution branch output.

[0071] The Channel-Spatial Dual Attention (CSDA) module includes a cascaded channel attention submodule and a spatial attention submodule. The channel attention submodule is used to generate channel weights by global average pooling and fully connected layers to perform channel-wise weighting on the input feature map. The spatial attention submodule is used to generate a spatial attention map by max pooling and average pooling along the channel dimension to perform element-wise weighting on the input feature map.

[0072] The channel attention submodule first uses global average pooling to obtain the channel description vector. Then, channel weights are generated through compression-excitation.

[0073] in, Channel attention-specific pooling feature output vector, used only for channel weight calculation; Channel attention exclusive A real space of dimension 1, used only for defining the dimension of a vector; : Number of channels in the channel attention-specific feature map, used only for vector dimension calculation.

[0074] Channel weight calculation formula: ; in, Channel attention-specific weight vector; Channel attention-specific Sigmoid activation function; Channel attention-specific weights for the second fully connected layer; Channel attention-specific ReLU activation function; Channel attention-specific weights for the first fully connected layer; Channel attention-specific pooling feature output vector.

[0075] Channel weighted feature formula: .

[0076] in, Channel attention-specific weighted feature map; Channel attention-specific channel weight; Channel attention-specific multiplication operator; Channel attention-specific input feature map.

[0077] Spatial attention map calculation formula: ; in, Spatial attention-specific attention map; Spatial attention-specific Sigmoid activation function; Spatial attention-specific 7×7 convolution; Spatial attention-specific channel splicing symbol; Spatial attention-specific average pooling results; Spatial attention-specific max pooling results; Spatial attention-specific input feature map.

[0078] Spatial weighted feature formula: ; in, Spatial attention-specific weighted feature map; Spatial attention map; : Element-wise multiplication operator specific to spatial attention; Spatial attention-specific input feature map.

[0079] A global average pooling layer, a fully connected layer, and a Softmax (Softmax Activation Function) output layer are used to output the first probability vector.

[0080] Global average pooling compresses a feature map of 14 (height of pooling input feature map) × 14 (width of pooling input feature map) × 512 (number of channels in pooling input feature map) into a feature vector of 1 (height of pooling output feature map) × 1 (width of pooling output feature map) × 512.

[0081] Access drop probability =0.5 (Dropout-specific dropout rate value) Dropout (Dropout Regularization) layer prevents overfitting, and fully connected layer maps 512 (fully connected input-specific feature dimension) dimensions to 256 (fully connected output-specific feature dimension) dimensions.

[0082] in, Dropout-specific dropout probability.

[0083] See Figure 5 As shown, Figure 5 A schematic diagram of an improved CNN network structure provided in Embodiment 1 of this application is shown, wherein, Figure 5 An improved CNN network structure was specifically designed for the arc detection channel. The input is a 224×224×3 CWT time-frequency image. After initial convolutional layers (3×3 convolutions, stride 2, BN+ReLU), a 112×112×64 feature map is output. Subsequently, three MSPC+CSDA modules are cascaded, resulting in output feature map sizes of 56×56×128, 28×28×256, and 14×14×512, respectively. The MSPC module uses 1×1 convolutions, 3×3 dilated convolutions with a dilation rate of 2, and 5×5 dilated convolutions with a dilation rate of 4 in parallel to extract features, which are then concatenated. Feature fusion is achieved with 1×1 convolution; the CSDA module sequentially achieves dual feature weighting through channel attention (global average pooling, fully connected compression activation, sigmoid) and spatial attention (max pooling + average pooling, concatenation, 7×7 convolution, sigmoid); finally, after global average pooling, Dropout (the dropout probability coefficient d=0.5 specific to the arc channel), and a 256-dimensional fully connected layer, the Softmax outputs three types of arc fault probability vectors, with all parameters and symbols matching the specific definition of the arc channel.

[0084] In an optional implementation, see Figure 6 As shown, Figure 6 The flowchart of a time-series feature extraction method provided in Embodiment 1 of this application is shown, wherein the extraction of its time-series feature sequence includes steps S601-S602: S601: The low-frequency component and the temperature signal are segmented by a sliding time window, and at least two features from the temperature slope, the effective value of the current, the voltage drop at the contact point, and the signal entropy are extracted in each window to form an original feature vector sequence.

[0085] Specifically, after the signal is segmented by a sliding window, four types of features are extracted: temperature slope, effective current value, contact point voltage drop, and signal entropy, which are then combined to form a 4-dimensional original feature vector.

[0086] The original feature vector is represented as: The sequence is .

[0087] in, Exclusive to overheating channels Original feature vector at time step; Overheating characteristics are unique to a 4-dimensional real space; : The original feature sequence specific to overheating characteristics; : Overheating characteristic specific time sequence number; Total length of the overheating feature-specific sequence.

[0088] S602: For each feature vector in the original feature vector sequence, calculate its first-order difference and second-order difference respectively, and concatenate the original feature value, first-order difference value and second-order difference value in the feature dimension to form an enhanced feature vector sequence, which is used as the input of the second neural network model.

[0089] First-order difference formula: It represents the instantaneous rate of change of a characteristic.

[0090] in, Exclusive to overheating channels First-order difference features at time intervals; Exclusive to overheating channels Original characteristics of the moment; Exclusive to overheating channels Original features at time -1.

[0091] Second-order difference formula: Characteristic acceleration of change.

[0092] in, Exclusive to overheating channels Second-order difference features at time points; Exclusive to overheating channels First-order difference at time; Exclusive to overheating channels First-order difference at time -1.

[0093] Formula for enhancing eigenvectors: .

[0094] in, Exclusive to overheating channels Time-based enhancement of feature vectors; : Special feature splicing symbol for overheated channels; Exclusive to overheating channels Original characteristics of the moment; Exclusive to overheating channels First-order difference at time; Exclusive to overheating channels Second-order difference at time; Enhanced features are dedicated to a 12-dimensional real space.

[0095] In an optional implementation, the second neural network model includes at least one layer of trend-gated memory cells (TGMCs), each including a forget gate that generates a forget gate output based on the hidden state of the previous time step and the input of the current time step; see also Figure 7 As shown, Figure 7 The flowchart illustrates a memory cell state update method provided in Embodiment 1 of this application, wherein the trend-gated memory cell updates its state according to steps S701-S703: S701: Based on the first-order difference mean input at the current time, a trend gating factor is generated through linear transformation and the Sigmoid activation function.

[0096] Specifically, the first-order difference mean is denoted as This represents the current trend direction of window failure.

[0097] Trend gating factor formula:

[0098] in, TGMC Exclusive The time-trend gating factor is used only for memory cell regulation; : TGMC-specific Sigmoid activation function, used only for gating factor normalization; TGMC-specific trend-gated weights are used only for linear transformation calculations. TGMC Exclusive The first-order difference mean at any time is used only for trend perception. : TGMC-specific trend gating bias term, used only for linear transformation calculations.

[0099] S702: Multiply the forget gate output element-wise with the trend gating factor to obtain the trend-sensitive effective forgetting rate.

[0100] Specifically, the standard LSTM forget gate formula: .

[0101] in, TGMC Exclusive The output of the forget gate at all times; The Sigmoid activation function exclusive to the Forgotten Gate; Forgot Gate-Specific Weight Matrix; : The unique splicing symbol of the Forgotten Gate; TGMC Exclusive Hidden state at time -1; TGMC Exclusive Enhanced features at all times; : A bias item specific to the Forgotten Gate.

[0102] Formula for effective forgetting rate: .

[0103] in, TGMC-specific trend-sensitive effective forgetting rate; TGMC exclusive Forgotten Gate output; : TGMC-specific element-wise multiplication operator; TGMC exclusive trend gating factor.

[0104] S703: Substitute the effective forgetting rate, the memory unit state at the previous moment, the input gate output at the current moment, and the candidate memory state into the memory unit state update formula to calculate the memory unit state at the current moment.

[0105] Specifically, the formula for updating the state of the trend-gated memory unit is as follows:

[0106] in, TGMC Exclusive The state of the memory unit is constantly being remembered; TGMC-specific effective forgetting rate; : Memory update-specific element-wise multiplication operator; TGMC Exclusive -1 is the state of the memory unit at time point 1; TGMC Exclusive Input gate output at all times; TGMC Exclusive Candidate memory states at any given time.

[0107] In an optional implementation, see Figure 8 As shown, Figure 8 The flowchart illustrates a sliding time window adjustment method provided in Embodiment 1 of this application, wherein the method further includes steps S801-S803: S801: Calculate the second-order difference mean of the time-series feature sequence in real time, as a measure of the trend change rate.

[0108] Specifically, the second-order difference mean is denoted as This directly reflects the degree of acceleration of overheating failure.

[0109] in, : Adaptive window exclusive The mean of the second difference at time step.

[0110] S802: When the mean of the second-order difference exceeds the preset gain threshold, the step size of the sliding time window input to the second neural network model is shortened to one-quarter of the baseline step size.

[0111] Specifically, the preset gain threshold is ,when At that time, the malfunction accelerated.

[0112] Step size adjustment formula: .

[0113] in, The step size is adjusted specifically for the adaptive window and is only used during the fault acceleration phase. : Adaptive window-specific baseline step size, used only for step size comparison; The step length adjustment has a dedicated shortening factor of 4.

[0114] S803: When the mean of the second-order difference does not exceed the gain threshold, the step size of the sliding time window is kept as the reference step size.

[0115] Specifically, when At that time, the fault trend was stable, and the step size remained constant. .

[0116] The Adaptive Temporal Window (ATW) mechanism only adjusts the sliding step size (without changing the fixed window length of the model input), independent of the training process, and dynamically adjusts it only during online inference, balancing timeliness and computational efficiency.

[0117] See Figure 9 As shown, Figure 9 A schematic diagram of an improved LSTM network structure provided in Embodiment 1 of this application is shown, wherein, Figure 9 An improved LSTM network structure was specifically designed for the overheat detection channel. The input to this network structure is the original feature sequence (4D × T time steps), which is first fed into the Differential Feature Injection (DFI) layer. After feature augmentation, a 12-dimensional enhanced feature sequence (12D × T) is generated. This enhanced feature sequence is then sequentially fed into TGMC Layer 1 (128-dimensional hidden layers), the Dropout layer, TGMC Layer 2 (64-dimensional hidden layers), and the Dropout layer, where the Dropout layer is used for model regularization. Finally, the last hidden state is retrieved. (64-dimensional), after passing through a fully connected layer (32-dimensional), a Softmax layer outputs the probabilities of three states, i.e. These correspond to three overheating fault states: normal, warning, and alarm. The network structure is configured with an adaptive time window mechanism (ATW), which uses the second-order difference mean as the adjustment criterion. Validity: When valid, window step size When not true, the window step size. By dynamically adjusting the window step size, the input rhythm of the original feature sequence is optimized, enabling adaptive detection of fault development trends and improving the sensitivity and reliability of early fault identification.

[0118] In an optional implementation, see Figure 10 As shown, Figure 10The flowchart of a fusion weight calculation method provided in Embodiment 1 of this application is shown, wherein the step of calculating the fusion weight based on the correlation between the arc fault and the overheat fault includes steps S1001 to S1004: S1001: Calculate the first sum of the arc warning probability and the arc alarm probability in the first probability vector.

[0119] Specifically, the first sum formula: .

[0120] in, Weighted calculation of specific arc risk; Weighted calculation of specific arc warning probability; Weighted calculation of the probability of a dedicated electric arc alarm.

[0121] S1002: Calculate the second sum of the overheat warning probability and the overheat alarm probability in the second probability vector.

[0122] Specifically, the second sum formula: .

[0123] in, Weighted calculation of specific overheating risks; Weighted calculation of overheating warning probability; Weighted calculation of overheating alarm probability.

[0124] S1003: Multiply the difference between the first sum and the second sum by a preset adjustment coefficient, and then add it to the preset base weight to obtain the initial fusion weight.

[0125] Specifically, the formula for calculating the initial fusion weights is as follows:

[0126] in, Initial fusion weights specific to decision fusion: : Decision fusion exclusive basic weight, default fixed value of 0.5; : Decision-making integration specific adjustment coefficient; Weighted calculation of specific overheating risk probability; : Decision-making incorporates specific time variables.

[0127] S1004: The initial fusion weight is restricted to a preset value range to obtain the fusion weight.

[0128] Specifically, the preset value range is a closed interval [0.3, 0.7], and a numerical truncation rule is used to constrain the boundary of the initial fusion weights: when the initial fusion weights If the value is greater than the upper limit of the interval (0.7), the final fusion weight will be directly assigned a value of 0.7; if the initial fusion weight is greater than the upper limit of the interval (0.7), the final fusion weight will be directly assigned a value of 0.7. When the value is less than the lower limit of the interval (0.3), the final fusion weight is directly assigned a value of 0.3; when the initial fusion weight... When the weight is within the range of [0.3, 0.7], the initial fusion weight is directly retained as the final fusion weight.

[0129] This boundary constraint operation avoids the extreme situation where the weights of the arc channel or the overheating channel tend to approach 0 or 1, prevents a single channel from dominating all decisions, and ensures the balance and reliability of the dual-channel fusion decision.

[0130] See Figure 11 As shown, Figure 11 This diagram illustrates a decision fusion module logic provided in Embodiment 1 of this application, wherein... Figure 11 For decision fusion and hierarchical early warning, the arc channel outputs a dedicated logic process. (Arc channel calibration probability), temperature scaled The processed signal is input into the CCDW module; the output from the overheating channel... (Overheating channel calibration probability), temperature scaled The processed data is then input into the CCDW module. The CCDW module calculates the weights according to the formula. Calculate the initial fusion weights and truncate them to the interval [0.3, 0.7] to obtain the dynamic fusion weights. With 1- Arc channel calibration probability and Multiplication, overheating path calibration probability and 1- After multiplying, the two are added together to obtain the fusion probability. Subsequently based on Implement tiered judgment: When the warning or alarm triggering conditions are not met, it is judged as normal (CLASS_0); when the conditions are met... When the temperature slope condition is met, it is judged as a warning (CLASS_1); when the condition is satisfied... An alarm will be triggered if the absolute temperature is greater than 120℃ (CLASS_2).

[0131] Example 2 See Figure 12 As shown, Figure 12 This illustration shows a structural schematic diagram of a rail vehicle fault early warning device according to Embodiment 2 of this application, wherein the device includes: The signal acquisition module 1201 is used to acquire the current signal and temperature signal of the power supply circuit in the low-voltage electrical cabinet of the rail vehicle. The first probability vector determination module 1202 is used to input the high-frequency component in the current signal to the arc detection channel, perform time-frequency analysis on the high-frequency component to generate a time-frequency representation, process the time-frequency representation using a first neural network model, and output a first probability vector representing the arc-type fault state. The second probability vector determination module 1203 is used to input the low-frequency component in the current signal and the temperature signal into the overheat detection channel, extract their time-series feature sequence, process the time-series feature sequence using a second neural network model, and output a second probability vector representing the overheat fault state. The fusion probability vector production module 1204 is used to calculate the fusion weight based on the correlation between the arc-type fault and the overheat-type fault, and to perform weighted fusion of the first probability vector and the second probability vector according to the fusion weight to obtain the final fusion probability vector. The early warning level determination module 1205 is used to determine the early warning level of the rail vehicle based on the final fusion probability vector.

[0132] In an optional implementation, the time-frequency analysis of the high-frequency components to generate a time-frequency characterization includes: Using Morlet wavelet as the mother wavelet function, a continuous wavelet transform is performed on the high-frequency components to obtain wavelet coefficients under different scale factors and translation factors. The scale axis is divided into three intervals: low frequency, mid frequency, and high frequency. The wavelet coefficient amplitudes of each interval are mapped to the pixel values ​​of the three RGB color channels to generate a size-normalized time-frequency scalogram image as the time-frequency representation.

[0133] In an optional implementation, the first neural network model includes: The initial convolutional layer is used to perform preliminary convolution and downsampling on the input time-frequency representation; At least one multi-scale parallel convolution module is used to extract features from the input feature map by applying convolution kernels of three sizes: 1×1, 3×3, and 5×5 in parallel. Dilated convolutions with different dilation rates are embedded in the branches of the 3×3 and 5×5 convolution kernels respectively. The feature maps output by each branch are concatenated along the channel dimension and then fused by 1×1 convolution. The channel-space dual attention module includes a cascaded channel attention submodule and a spatial attention submodule. The channel attention submodule is used to generate channel weights through global average pooling and fully connected layers to perform channel-wise weighting on the input feature map. The spatial attention submodule is used to generate a spatial attention map through max pooling and average pooling along the channel dimension to perform element-wise weighting on the input feature map. A global average pooling layer, a fully connected layer, and a Softmax output layer are used to output the first probability vector.

[0134] In an optional implementation, the extraction of its temporal feature sequence includes: The low-frequency component and the temperature signal are segmented by a sliding time window. At least two features from the temperature slope, current RMS value, contact point voltage drop and signal entropy are extracted in each window to form an original feature vector sequence. For each feature vector in the original feature vector sequence, its first-order difference and second-order difference are calculated respectively. The original feature value, first-order difference value and second-order difference value are concatenated along the feature dimension to form an enhanced feature vector sequence, which is used as the input of the second neural network model.

[0135] In an optional implementation, the second neural network model includes at least one layer of trend-gated memory units, each including a forget gate, the forget gate generating a forget gate output based on the hidden state of the previous time step and the input of the current time step; the trend-gated memory unit updates its memory unit state in the following manner: Based on the first-order difference mean of the input at the current moment, a trend gating factor is generated through linear transformation and the Sigmoid activation function; The forget gate output is multiplied element-wise by the trend gating factor to obtain the trend-sensitive effective forgetting rate; Substitute the effective forgetting rate, the memory cell state at the previous time step, the input gate output at the current time step, and the candidate memory state into the memory cell state update formula to calculate the memory cell state at the current time step.

[0136] In an optional implementation, the device further includes a sliding time window adjustment module for: The second-order difference mean of the time-series feature sequence is calculated in real time and used as a measure of the trend change rate. When the mean of the second-order difference exceeds the preset gain threshold, the step size of the sliding time window input to the second neural network model is shortened to one-quarter of the baseline step size; When the mean of the second-order difference does not exceed the gain threshold, the step size of the sliding time window is kept as the baseline step size.

[0137] In an optional implementation, calculating the fusion weight based on the correlation between the arc-type fault and the overheat-type fault includes: Calculate the first sum of the arc warning probability and the arc alarm probability in the first probability vector; Calculate the second sum of the overheat warning probability and the overheat alarm probability in the second probability vector; The difference between the first sum and the second sum is multiplied by a preset adjustment coefficient and then added to a preset base weight to obtain the initial fusion weight. The initial fusion weight is restricted to a preset value range to obtain the fusion weight.

[0138] Example 3 Based on the same application concept, see [link / reference] Figure 13 As shown, Figure 13 This illustration shows a structural schematic diagram of a computer device provided in Embodiment 3 of this application, wherein, as shown... Figure 13 As shown, the computer device 1300 provided in Embodiment 3 of this application includes: The system includes a processor 1301, a memory 1302, and a bus 1303. The memory 1302 stores machine-readable instructions executable by the processor 1301. When the computer device 1300 is running, the processor 1301 and the memory 1302 communicate via the bus 1303. When the machine-readable instructions are executed by the processor 1301, they perform the steps of the rail vehicle fault early warning method shown in Embodiment 1 above.

[0139] Example 4 Based on the same concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the rail vehicle fault early warning method described in any of the above embodiments.

[0140] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0141] The computer program product for early warning of rail vehicle faults provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0142] The rail vehicle fault early warning device provided in this application embodiment can be specific hardware on the device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this application embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0143] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0144] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0145] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0146] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0147] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0148] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for early warning of faults in rail vehicles, characterized in that, The method includes: Acquire the current and temperature signals of the power supply circuit inside the low-voltage electrical cabinet of the rail vehicle; The high-frequency component of the current signal is input to the arc detection channel, and the high-frequency component is subjected to time-frequency analysis to generate a time-frequency characterization. The time-frequency characterization is processed by the first neural network model to output a first probability vector characterizing the arc-type fault state. The low-frequency component of the current signal and the temperature signal are input to the overheat detection channel, their time-series feature sequences are extracted, and the time-series feature sequences are processed by the second neural network model to output a second probability vector representing the overheat fault state. The fusion weight is calculated based on the correlation between the arc fault and the overheat fault. The first probability vector and the second probability vector are weighted and fused according to the fusion weight to obtain the final fusion probability vector. The warning level of the rail vehicle is determined based on the final fusion probability vector.

2. The method according to claim 1, characterized in that, The step of performing time-frequency analysis on the high-frequency components to generate a time-frequency characterization includes: Using Morlet wavelet as the mother wavelet function, a continuous wavelet transform is performed on the high-frequency components to obtain wavelet coefficients under different scale factors and translation factors. The scale axis is divided into three intervals: low frequency, mid frequency, and high frequency. The wavelet coefficient amplitudes of each interval are mapped to the pixel values ​​of the three RGB color channels to generate a size-normalized time-frequency scalogram image as the time-frequency representation.

3. The method according to claim 1, characterized in that, The first neural network model includes: The initial convolutional layer is used to perform preliminary convolution and downsampling on the input time-frequency representation; At least one multi-scale parallel convolution module is used to extract features from the input feature map by applying convolution kernels of three sizes: 1×1, 3×3, and 5×5 in parallel. Dilated convolutions with different dilation rates are embedded in the branches of the 3×3 and 5×5 convolution kernels respectively. The feature maps output by each branch are concatenated along the channel dimension and then fused by 1×1 convolution. The channel-space dual attention module includes a cascaded channel attention submodule and a spatial attention submodule. The channel attention submodule is used to generate channel weights through global average pooling and fully connected layers to perform channel-wise weighting on the input feature map. The spatial attention submodule is used to generate a spatial attention map through max pooling and average pooling along the channel dimension to perform element-wise weighting on the input feature map. A global average pooling layer, a fully connected layer, and a Softmax output layer are used to output the first probability vector.

4. The method according to claim 1, characterized in that, The extraction of its temporal feature sequence includes: The low-frequency component and the temperature signal are segmented by a sliding time window. At least two features from the temperature slope, current RMS value, contact point voltage drop and signal entropy are extracted in each window to form an original feature vector sequence. For each feature vector in the original feature vector sequence, its first-order difference and second-order difference are calculated respectively. The original feature value, first-order difference value and second-order difference value are concatenated along the feature dimension to form an enhanced feature vector sequence, which is used as the input of the second neural network model.

5. The method according to claim 1, characterized in that, The second neural network model includes at least one layer of trend-gated memory units, each including a forget gate. The forget gate generates a forget gate output based on the hidden state of the previous time step and the input of the current time step. The trend-gated memory unit updates its memory unit state in the following manner: Based on the first-order difference mean of the input at the current moment, a trend gating factor is generated through linear transformation and the Sigmoid activation function; The forget gate output is multiplied element-wise by the trend gating factor to obtain the trend-sensitive effective forgetting rate; Substitute the effective forgetting rate, the memory cell state at the previous time step, the input gate output at the current time step, and the candidate memory state into the memory cell state update formula to calculate the memory cell state at the current time step.

6. The method according to claim 4 or 5, characterized in that, The method further includes: The second-order difference mean of the time-series feature sequence is calculated in real time and used as a measure of the trend change rate. When the mean of the second-order difference exceeds the preset gain threshold, the step size of the sliding time window input to the second neural network model is shortened to one-quarter of the baseline step size; When the mean of the second-order difference does not exceed the gain threshold, the step size of the sliding time window is kept as the baseline step size.

7. The method according to claim 1, characterized in that, The calculation of the fusion weight based on the correlation between the arc-type faults and the overheat-type faults includes: Calculate the first sum of the arc warning probability and the arc alarm probability in the first probability vector; Calculate the second sum of the overheat warning probability and the overheat alarm probability in the second probability vector; The difference between the first sum and the second sum is multiplied by a preset adjustment coefficient and then added to a preset base weight to obtain the initial fusion weight. The initial fusion weight is restricted to a preset value range to obtain the fusion weight.

8. A fault early warning device for rail vehicles, characterized in that, The device includes: The signal acquisition module is used to acquire the current and temperature signals of the power supply circuit inside the low-voltage electrical cabinet of the rail vehicle. The first probability vector determination module is used to input the high-frequency component in the current signal to the arc detection channel, perform time-frequency analysis on the high-frequency component to generate a time-frequency representation, process the time-frequency representation using a first neural network model, and output a first probability vector representing the arc-type fault state. The second probability vector determination module is used to input the low-frequency component of the current signal and the temperature signal into the overheat detection channel, extract their time-series feature sequence, process the time-series feature sequence using the second neural network model, and output a second probability vector representing the overheat fault state. The fusion probability vector production module is used to calculate the fusion weight based on the correlation between the arc-type fault and the overheat-type fault, and to perform weighted fusion of the first probability vector and the second probability vector according to the fusion weight to obtain the final fusion probability vector. The early warning level determination module is used to determine the early warning level of the rail vehicle based on the final fusion probability vector.

9. A computer device, characterized in that, include: The system includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the rail vehicle fault early warning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the rail vehicle fault early warning method as described in any one of claims 1 to 7.

Images