Rail transit electric passenger car bogie dynamic performance monitoring method based on multi-sensor fusion

Abstract

The present application relates to the field of rail transit monitoring technology, and discloses a multi-sensor fusion rail transit electric passenger car bogie dynamic performance monitoring method. The method comprises collecting multi-source heterogeneous sensor data such as bogie vibration acceleration time sequence signal, wheel-rail contact force distribution data, bearing temperature gradient data and structure strain field data; extracting a frequency band energy feature vector from the vibration signal analysis, generating a distribution matrix by spatial gridding mapping of the contact force data, and forming a primary fusion feature set by spatio-temporal alignment fusion of the two; positioning the bearing temperature abnormal area through a deep convolution network, and generating a secondary fusion feature set in combination with the structure strain concentration coefficient; inputting the two-level feature sets into a dynamic performance evaluation model to output a performance degradation index set including scores of axle box bearing wear degree, frame fatigue accumulation and wheel set tread damage, thereby generating a maintenance priority ranking and triggering a preventive maintenance instruction, and realizing accurate monitoring and efficient operation, management and control of the bogie dynamic performance.

Description

Technical Field

[0001] This invention relates to the field of rail transit monitoring technology, specifically a method for monitoring the dynamic performance of rail transit electric train bogies using multi-sensor fusion. Background Technology

[0002] As a high-capacity, high-speed public transportation vehicle, the operational safety and stability of rail transit electric multiple units (EMUs) directly affect the safety of passengers' lives and property and the efficient operation of the public transportation network. The bogie, as the core running gear of an EMU, bears the crucial functions of supporting the weight of the car body, transmitting traction and braking forces, and guiding the train along the track. Its dynamic performance directly determines the overall operational quality of the train. In the long-term, complex service environment, the bogie faces multiple effects, including wheel-rail impact, alternating loads, and temperature fluctuations, making it prone to failures such as axle box bearing wear, frame fatigue accumulation, and wheelset tread damage. Failure to monitor and address these issues in a timely manner can lead to serious traffic safety accidents.

[0003] Currently, the monitoring methods for the dynamic performance of bogies in the rail transit industry mostly rely on data from a single type of sensor, resulting in a one-sided monitoring dimension and incomplete feature extraction. Some monitoring schemes only collect acceleration signals through vibration sensors and make fault judgments based solely on frequency domain characteristics, ignoring the comprehensive impact of key information such as wheel-rail contact state, temperature changes, and structural strain on bogie performance, leading to a risk of misjudgment in fault identification. Other schemes attempt to introduce multi-source data, but they have significant deficiencies in data processing and fusion: wheel-rail contact force distribution data is only subjected to simple numerical statistics, failing to fully explore its spatial distribution characteristics; bearing temperature monitoring is mostly limited to overall temperature threshold judgment, unable to accurately locate the specific location of abnormal heat sources; at the same time, the fusion of multi-source data often adopts a direct splicing method, lacking consideration of spatiotemporal correlation, resulting in insufficient effectiveness of fused features and difficulty in comprehensively reflecting the true working status of various bogie components.

[0004] Existing monitoring methods lag behind in performance evaluation and maintenance guidance. Traditional methods primarily focus on post-fault diagnosis, failing to dynamically track and predict the bogie performance degradation process, thus lacking accurate basis for preventative maintenance. Some evaluation models can only output a single fault diagnosis result, unable to quantify the performance degradation degree of each component, making it difficult for maintenance personnel to develop targeted maintenance plans, often resulting in over-maintenance or untimely maintenance, increasing operating costs and reducing train operating efficiency. Therefore, a monitoring method is needed that can integrate multi-source heterogeneous data, achieve deep feature fusion, accurately evaluate bogie dynamic performance, and guide preventative maintenance to address the many shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring the dynamic performance of bogies in rail transit electric multiple units using multi-sensor fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for monitoring the dynamic performance of bogies in rail transit electric multiple units using multi-sensor fusion, the method comprising: Collect multi-source heterogeneous sensing data on the bogie's operating status. The multi-source heterogeneous sensing data includes vibration acceleration time-series signals, wheel-rail contact force distribution data, bearing temperature gradient data, and structural strain field data. Frequency domain energy distribution analysis is performed on the vibration acceleration time series signal to extract the frequency band energy feature vector. Spatial gridded feature mapping is performed on the wheel-rail contact force distribution data to generate the contact force distribution matrix. The frequency band energy feature vector and the contact force distribution matrix are spatiotemporally aligned and fused to form a primary fusion feature set. An abnormal heat source localization process is performed on the bearing temperature gradient data by a deep convolutional feature extraction network to identify the coordinates of the abnormal temperature region, and a secondary fusion feature set is generated by combining the strain concentration coefficient in the structural strain field data. The primary and secondary fusion feature sets are input into the dynamic performance evaluation model for multi-dimensional feature interaction calculation, and the performance degradation index set of each component of the bogie is output. The performance degradation index set includes the axle box bearing wear score, the frame fatigue accumulation score, and the wheelset tread damage score. Based on the set of performance degradation indices, the bogie maintenance priority ranking results are generated, and preventive maintenance instructions for the corresponding components are triggered.

[0007] Preferably, the step of performing frequency domain energy distribution analysis processing on the vibration acceleration time-series signal to extract frequency band energy feature vectors includes: An adaptive window length Fourier transform is used to decompose the vibration acceleration time-series signal into multiple time-frequency energy units; Calculate the energy percentage of each time-frequency energy unit within a preset characteristic frequency band to form an initial frequency band energy sequence; The initial frequency band energy sequence is optimized by wavelet packet reconstruction algorithm to eliminate frequency band aliasing effect. The optimized frequency band energy sequence is grouped and aggregated according to the inherent frequency range of the key bogie components to generate frequency band energy feature vectors.

[0008] Preferably, the step of performing spatial gridded feature mapping processing on the wheel-rail contact force distribution data to generate a contact force distribution matrix includes: Establish a three-dimensional mesh coordinate system for the wheel-rail contact surface and map the contact force sampling points to the corresponding mesh cells; The inverse distance weighted interpolation algorithm is used to calculate the contact force estimate of the unsampled grid cells; The contact force data of the grid cells are normalized to eliminate the differences in dimensions between different sensors; Extract the spatial gradient and symmetry features of the contact force distribution to construct the contact force distribution matrix.

[0009] Preferably, the step of performing spatiotemporal alignment and fusion processing on the frequency band energy feature vector and the contact force distribution matrix to form a primary fusion feature set includes: Establish a mapping table between the physical locations of vibration frequency bands and contact force grids; The timestamps of the frequency band energy feature vectors are aligned and transformed with the spatial coordinates of the contact force distribution matrix using a time-space encoder. A dual-stream attention mechanism is used to calculate the correlation weight matrix between frequency band energy characteristics and contact force characteristics; The frequency band energy feature vector and the contact force distribution matrix are weighted and concatenated according to the correlation weight matrix to generate a primary fusion feature set.

[0010] Preferably, the step of performing abnormal heat source localization processing on the bearing temperature gradient data through a deep convolutional feature extraction network and identifying the coordinates of the abnormal temperature region includes: Convert the bearing temperature gradient data into a three-dimensional thermogram tensor; Multiscale temperature distribution features of thermogram tensors are extracted using dilated convolutional layers. The non-maximum suppression algorithm was used to locate the center coordinates of the temperature anomaly region; The physical rationality of the coordinates of the abnormal area is verified by combining the bearing structure topology diagram.

[0011] Preferably, the generation of the secondary fusion feature set by combining the strain concentration coefficients in the structural strain field data includes: Calculate the coordinates of the local strain gradient extreme points in the structural strain field data; Extract the strain energy density distribution characteristics of the region surrounding the extreme point; Spatial correlation matching is performed between the coordinates of the temperature anomaly region and the coordinates of the extreme points of the strain gradient. The coordinate region features that are successfully fused and matched are used to generate a secondary fused feature set.

[0012] Preferably, the step of inputting the primary fusion feature set and the secondary fusion feature set into the dynamic performance evaluation model for multi-dimensional feature interaction calculation includes: Construct a cross-feature association graph between the primary fusion feature set and the secondary fusion feature set; The graph neural network aggregates cross-feature associations with multi-hop neighbor features in the graph. Gated recurrent units are used to capture temporal dependencies in the feature interaction process; Output a set of performance degradation indices for each component of the bogie.

[0013] Preferably, the performance degradation index set includes axle box bearing wear score, frame fatigue accumulation score, and wheelset tread damage score, including: The wear score of the bearing box is calculated based on the area of ​​the abnormal bearing temperature region and the characteristics of vibration energy. The fatigue accumulation score of the frame is calculated by combining the characteristics of strain energy density distribution and contact force asymmetry. The wheelset tread damage score is calculated by the number of abrupt changes in the wheel-rail contact force gradient and the peak value of the frequency band energy anomaly.

[0014] Preferably, the step of generating bogie maintenance priority ranking results based on the performance degradation index set includes: Establish a mapping model between component performance degradation index and remaining service life; Calculate the rate of change of the performance degradation index of each component and its distance from the critical threshold. A multi-objective decision-making algorithm is used to generate maintenance priority ranking results that take into account both security and economy.

[0015] Preferably, the preventive maintenance command that triggers the corresponding component includes: Parse the identifier of the highest priority component in the maintenance priority ranking results; Match the standard maintenance procedure parameters corresponding to this component; Generate preventative maintenance instructions that include maintenance time window positioning and resource allocation plans.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By acquiring multi-source heterogeneous sensor data, including vibration acceleration time-series signals, wheel-rail contact force distribution data, bearing temperature gradient data, and structural strain field data, a comprehensive coverage of the bogie's operating status was achieved. Compared to traditional single-sensor monitoring, multi-dimensional data acquisition can capture the bogie's operating characteristics in different dimensions such as mechanics, thermodynamics, and structural mechanics, fully reconstructing the bogie's complex working state. This avoids performance misjudgments caused by incomplete monitoring information and provides rich and comprehensive basic data support for subsequent feature processing and performance evaluation.

[0017] In the feature processing stage, differentiated processing strategies were adopted for different types of data, significantly improving the effectiveness of the features. Frequency domain energy distribution analysis of vibration acceleration time-series signals accurately extracted energy features from different frequency bands, effectively reflecting the frequency characteristics of bogie vibration anomalies. Spatial gridded feature mapping was performed on wheel-rail contact force distribution data, transforming discrete contact force data into a structured distribution matrix and fully exploring its spatial distribution patterns. Spatiotemporal alignment fusion processing organically combined these two types of features, solving the problem of spatiotemporal dimension mismatch in traditional fusion methods. This enabled the primary fusion feature set to simultaneously cover the temporal frequency information of vibration and the spatial distribution information of wheel-rail contact, improving the comprehensiveness and correlation of the features.

[0018] To address the processing of bearing temperature and structural strain data, a deep convolutional feature extraction network was introduced to accurately locate abnormal heat sources. This overcomes the limitations of traditional temperature monitoring, which can only determine whether the overall temperature exceeds the standard, and can pinpoint the specific coordinates of abnormal temperature areas, providing precise guidance for subsequent fault tracing. Simultaneously, a secondary fusion feature set was generated by combining the strain concentration coefficients from the structural strain field data, establishing a correlation between temperature anomalies and structural stress concentration. This allows the features to more comprehensively reflect the structural health status of the bogie, further enriching the feature dimensions and providing more targeted feature support for performance evaluation.

[0019] The construction of a two-level fusion feature set enables deep fusion of multi-source data. Inputting this data into a dynamic performance evaluation model allows for a comprehensive, multi-dimensional assessment of the performance of key bogie components. The output performance degradation index set covers key indicators such as axle box bearing wear, frame fatigue accumulation, and wheelset tread damage, achieving a quantitative characterization of the bogie's performance degradation state. This allows maintenance personnel to clearly understand the health status of each component. Based on the maintenance priority ranking results generated from this performance degradation index set, the urgency of maintenance for different components can be clearly identified, providing precise decision-making guidance for preventative maintenance. This achieves a shift from "post-failure repair" to "pre-failure prevention," rationally allocating maintenance resources, reducing unnecessary maintenance work, extending the service life of bogie components, and ensuring the safe and stable operation of rail transit trains. Attached Figure Description

[0020] Figure 1 A rating chart for the maintenance priority of various components of the bogie of a rail transit electric passenger car; Figure 2 A flowchart for frequency domain energy distribution analysis and frequency band energy feature vector extraction; Figure 3 Flowchart for generating the initial fusion feature set; Figure 4 This is a time-series trend chart of the performance degradation index of key components of the bogie of a rail transit electric passenger car. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides a multi-sensor fusion method for monitoring the dynamic performance of bogies in rail transit electric multiple units (EMUs). The method includes: achieving a comprehensive assessment of the bogie's condition by integrating multiple sensor data. First, vibration acceleration sensors, wheel-rail contact force sensors, bearing temperature sensors, and structural strain sensors are deployed at key parts of the bogie to collect multi-source heterogeneous sensor data during operation. This data includes vibration acceleration time-series signals, wheel-rail contact force distribution data, bearing temperature gradient data, and structural strain field data. The vibration acceleration time-series signals are processed through frequency domain energy distribution analysis to extract frequency band energy feature vectors. The wheel-rail contact force distribution data undergoes spatial gridding feature mapping to generate a contact force distribution matrix. Subsequently, the frequency band energy feature vectors and the contact force distribution matrix are spatiotemporally aligned and fused to form a primary fusion feature set. The bearing temperature gradient data is input into a deep convolutional feature extraction network to perform abnormal heat source localization processing, identify the coordinates of abnormal temperature regions, and combine this with the strain concentration coefficients in the structural strain field data to generate a secondary fusion feature set. The primary and secondary fusion feature sets are input into the dynamic performance evaluation model for multi-dimensional feature interaction calculations, outputting a set of performance degradation indices for each bogie component. This set includes axle box bearing wear scores, frame fatigue accumulation scores, and wheelset tread damage scores. Based on the performance degradation index set, the method generates a bogie maintenance priority ranking result and triggers preventive maintenance commands for the corresponding components, thereby achieving intelligent condition monitoring and maintenance decision-making.

[0023] Example 1: See Figure 2The vibration acceleration time-series signal originates from an array of acceleration sensors installed at key locations on the bogie. These sensors capture the multidimensional vibration response of the bogie during operation at a high-frequency sampling rate of 5000 Hz to ensure coverage of the natural frequency range of each bogie component. The raw vibration acceleration time-series signal first undergoes a preprocessing stage, including detrending to eliminate linear or slowly changing drift components and bandpass filtering to remove high-frequency noise and extremely low-frequency interference. The passband frequency range of the bandpass filter is set from 10 Hz to 2000 Hz. An adaptive window-length Fourier transform is then used to decompose the preprocessed vibration acceleration time-series signal into multiple time-frequency energy units. The core of the adaptive window-length Fourier transform lies in the dynamic adjustment mechanism of the window function length. The window length adjustment is based on the stability of the instantaneous frequency of the signal, and the instantaneous frequency is calculated by extracting the phase derivative of the analytic signal using a Hilbert transform. When a rapid change in signal frequency components is detected, such as during an impact vibration event, the algorithm automatically selects a shorter Hanning window with a length of 1024 sampling points to provide high temporal resolution and accurately pinpoint the occurrence of the transient event. When the signal is in a relatively stable vibration state, the algorithm automatically switches to a longer Hanning window with a length of 4096 sampling points to obtain high frequency resolution and accurately distinguish densely distributed frequency components. Each time a Fourier transform of a specific window length is applied to a segment of the signal, a time-frequency energy unit is generated. Each time-frequency energy unit contains a time center point, a frequency range, and an energy value representing the signal strength in that time-frequency region.

[0024] The energy proportion of each time-frequency energy unit within a preset characteristic frequency band is calculated to form an initial frequency band energy sequence. The preset characteristic frequency band is a set of frequency intervals predefined based on the known fault characteristic frequencies and inherent vibration modes of key bogie components. For example, the characteristic frequency band used to monitor the bearing condition of the axle box includes the bearing outer ring fault characteristic frequency bandwidth, inner ring fault characteristic frequency bandwidth, and rolling element fault characteristic frequency bandwidth; the characteristic frequency band used to monitor gear meshing condition includes the bandwidth containing the gear meshing frequency and its harmonic components. The energy proportion is calculated by integrating and summing the energy values ​​of all frequency components within the corresponding characteristic frequency band for each time-frequency energy unit, and then dividing this integral sum by the total energy value of that time-frequency energy unit across the entire frequency band to obtain a dimensionless proportion. This proportion quantifies the importance of vibration energy in a specific frequency band within the total vibration energy. As the signal progresses along the time axis, the energy proportion values ​​of each characteristic frequency band calculated for each time point or time window are arranged chronologically to form the initial frequency band energy sequence. This sequence reflects the evolution of vibration energy in different characteristic frequency bands over time. The initial frequency band energy sequence is optimized using a wavelet packet reconstruction algorithm to eliminate band aliasing effects that may be caused by the inherent limitations of the Fourier transform. The wavelet packet reconstruction algorithm employs the Db10 wavelet basis function because the Db10 wavelet has good compact support and regularity, effectively representing the characteristics of vibration signals. The algorithm performs a deeper time-frequency decomposition on the original vibration signal segment corresponding to the initial frequency band energy sequence. Wavelet packet decomposition decomposes the signal at multiple scales, generating a complete binary tree structure of subbands, with each node representing a frequency band. By analyzing the energy distribution of these subbands, the actual boundaries of the characteristic frequency bands can be determined more accurately. For frequency bands with energy leakage or aliasing, the algorithm redefines the frequency band boundaries by selecting the optimal subtree or reconstruction path. For example, a wide, aliased frequency band can be divided into two or more non-overlapping, energy-concentrated subbands, or several adjacent, energy-related narrow frequency bands can be merged into a more representative wideband. This optimization process effectively sharpens the frequency band edges, enabling each characteristic frequency band to more purely reflect a specific physical phenomenon or component state.

[0025] The optimized frequency band energy sequence is grouped and aggregated according to the natural frequency ranges of key bogie components to generate the final frequency band energy feature vector. The grouping and aggregation are based on the bogie dynamics model, which provides the theoretically calculated natural frequency ranges of key components such as frame side beams, frame crossbeams, axle boxes, spring systems, and gearboxes, as well as the natural frequency ranges verified through modal tests in actual operation. For example, the first-order bending mode frequency range of the frame side beams and the vibration dominant frequency range of the axle box swing arm are clearly defined. Each frequency band in the optimized frequency band energy sequence is divided into its corresponding component frequency range based on its center frequency. The aggregation operation extracts features from the energy sequences of all frequency bands belonging to the same component frequency range. The extracted features include time-domain and frequency-domain statistics such as the average energy, energy standard deviation, peak factor, and kurtosis index of each frequency band within the analysis time window. Then, these features extracted from different frequency bands but belonging to the same component are concatenated to form a feature sub-vector representing the vibration state of that component. Finally, the feature sub-vectors of all key components are concatenated in a predetermined order to form a frequency band energy feature vector with fixed length, structured dimensions, and clear physical meaning. This frequency band energy feature vector serves as the basic input for subsequent fusion with features such as wheel-rail contact force, and its quality directly affects the accuracy of the final state assessment.

[0026] Example 2: See Figure 3The spatial gridded feature mapping processing of wheel-rail contact force distribution data begins with establishing a precise three-dimensional grid coordinate system for the wheel-rail contact surface. This three-dimensional grid coordinate system is defined based on the geometric model and kinematic relationships of the bogie wheelset. The origin of the coordinate system is set at the intersection of the wheelset centerline and the center of the wheel-rail contact patch. The longitudinal axis X extends along the rail direction, the transverse axis Y is perpendicular to the rail plane pointing outwards, and the vertical axis Z is perpendicular to the rail plane and pointing upwards. The grid cell size is determined according to the measurement accuracy requirements and sensor spatial resolution: the longitudinal grid size is set to 10 mm, the transverse grid size to 5 mm, and the vertical grid size to 1 mm, forming a three-dimensional grid covering the potential contact area. Contact force sensor arrays arranged at locations such as the wheel back and rail web collect contact force data at discrete points. Each data point contains three-dimensional coordinate information and a contact force vector value. The mapping process converts the three-dimensional coordinates of each sampling point into the grid coordinate system, determines its corresponding grid cell index number through rounding or floor division, and assigns the contact force vector value of that sampling point to the corresponding grid cell. An inverse distance-weighted interpolation algorithm is used to calculate the contact force estimate of grid cells not directly covered by sensors. Based on the principle of spatial correlation, this algorithm assumes that the influence of unsampled points on their neighboring sampled points is inversely proportional to the distance. For each grid cell to be interpolated, the algorithm searches for all valid contact force sampling points within a set search radius (20 mm), centered on its spatial location. The Euclidean distance from the grid cell to be interpolated to each neighboring sampling point is calculated, and the reciprocal of this distance is used as the basis for the weights. The weight calculation function uses an inverse distance-weighted model with a power of 2, meaning the weight of each sampling point is equal to the reciprocal of the square of its distance to the interpolated point. The contact force values ​​of all sampling points within the search radius are multiplied by their respective weights, summed, and then divided by the sum of all weights to obtain the contact force estimate for that grid cell. This process iterates through all unsampled grid cells, ultimately ensuring that every cell in the entire 3D mesh space has a contact force estimate.

[0027] To eliminate dimensional inconsistencies caused by differences in sensor models or calibration parameters, the contact force data of the grid cells is normalized using a minimum-maximum scaling method. This method linearly transforms the original contact force data to a closed interval between zero and one. The minimum and maximum values ​​of the contact force in all grid cells of the entire grid dataset are calculated. For each contact force value in the grid dataset, the minimum value is subtracted, and the result is divided by the difference between the maximum and minimum values ​​to obtain the normalized contact force value. Spatial gradient and symmetry features of the contact force distribution are extracted. Spatial gradient features reflect the rate of change of contact force in three-dimensional space. The Sobel operator is used to calculate the gradient components of the contact force in the longitudinal, transverse, and vertical directions. The gradient magnitude of each grid cell is calculated by taking the square root of the sum of the squares of its three gradient components. Symmetry feature analysis is performed to determine the mirror relationship between the left and right wheel-rail contact force distributions. The left wheel-rail contact force distribution matrix is ​​mirrored along the longitudinal center plane and then compared cell by cell with the right wheel-rail contact force distribution matrix. The correlation coefficient and mean absolute error are calculated as symmetry metrics. The contact force distribution matrix is ​​constructed to integrate the normalized contact force data, spatial gradient magnitude data, and symmetry index into a multi-dimensional data structure. The first channel of the contact force distribution matrix stores the normalized contact force value of each grid cell, the second channel stores the spatial gradient magnitude of each grid cell, and the third channel stores the difference map data generated by the symmetry analysis. The row, column, and layer indices of the matrix correspond to the longitudinal, transverse, and vertical coordinates of the grid coordinate system, respectively.

[0028] The spatiotemporal alignment and fusion of frequency band energy feature vectors and contact force distribution matrices requires establishing a physical location mapping table between vibration frequency bands and contact force grids. This mapping table is constructed based on a 3D computer-aided design model of the bogie and a sensor layout diagram. The model clearly marks the installation location and sensitive direction of each vibration acceleration sensor, as well as the area covered by the wheel-rail contact force sensor array. The mapping table records the main vibration source components corresponding to each vibration feature frequency band, and the area range where these components may affect the wheel-rail contact surface. For example, the vibration feature frequency band of the axle box bearing is mainly related to the axle box and its connected components, and its influence area on the wheel-rail contact surface is concentrated in the inner region of the wheel-rail contact patch. The mapping table is stored in the form of a lookup table, where the key is the dimension index of the frequency band energy feature vector, and the value is the set of spatial grid cell indices corresponding to the contact force distribution matrix. A spatiotemporal encoder is used to align and transform the timestamps of the frequency band energy feature vectors with the spatial coordinates of the contact force distribution matrix. The spatiotemporal encoder is a data processing unit containing a time interpolation module and a spatial transformation module. The time interpolation module addresses the asynchrony between the frequency band energy feature vector and the contact force distribution matrix at different acquisition times. Since the acquisition frequency of contact force distribution data is typically lower than that of vibration acceleration data, the time interpolation module uses a linear interpolation method on the frequency band energy feature vector to generate an equally spaced frequency band energy feature sequence perfectly aligned with the timestamps of the contact force distribution matrix. The spatial transformation module transforms the contact force distribution matrix from its own three-dimensional grid coordinate system to a unified coordinate system referenced to the bogie body. This transformation involves coordinate rotation and translation, with the rotation matrix and translation vectors determined through calibration experiments to ensure that the vibration sensor position and the wheel-rail contact surface grid have a consistent spatial reference frame.

[0029] A two-stream attention mechanism is employed to calculate the correlation weight matrix between frequency band energy features and contact force features. This mechanism comprises two independent feature extraction streams: one stream processes the frequency band energy feature vector, and the other processes local region features related to the vibration frequency band extracted from the contact force distribution matrix. Each stream first projects the original features into a latent space of the same dimension using a fully connected neural network to obtain key and value vectors. The frequency band energy feature stream also generates a query vector. The attention weights are calculated using a scaled dot product attention model. The query vector generated by the frequency band energy feature stream is dot-producted with the key vector generated by the contact force feature stream. The dot product result is scaled by the square root of the key vector dimension and then normalized using a softmax function to obtain the attention weight matrix. Each element of the attention weight matrix represents the correlation strength between a specific frequency band energy feature and a specific contact force grid region feature; a larger value indicates a stronger correlation in the physical mechanism. Based on the correlation weight matrix, the local features of the frequency band energy feature vector and the contact force distribution matrix are weighted and concatenated to generate a primary fusion feature set. The weighting operation uses an attention weight matrix to sum the value vectors of the contact force feature flow, resulting in a set of weighted contact force context features. The concatenation operation connects the weighted contact force context features with the original frequency band energy feature vector along the feature dimension, forming a fused high-dimensional feature vector. This fused feature vector simultaneously contains the frequency domain energy information of the vibration signal and the spatial distribution information of the wheel-rail contact force, and highlights the physically related parts between the two through an attention mechanism. The initial fused feature set serves as the basis for further fusion with features such as bearing temperature and structural strain, providing richer input information for a comprehensive evaluation of the bogie's dynamic performance.

[0030] Example 3: The acquisition of bearing temperature gradient data relies on a high-precision temperature sensor array embedded in the outer ring, inner ring, and roller cage of the bearing housing. The sensor array is distributed in a grid pattern, with each sensor recording the real-time temperature value at its mounting point. The sampling frequency is set to 1 Hz to capture relatively slow temperature changes during bearing operation. Preprocessing of the bearing temperature gradient data includes defect removal and noise filtering. Defect removal is based on a statistical threshold, removing abnormal readings that significantly deviate from the group average. Noise filtering uses a moving average filter with a window length of 5 sampling points to smooth random fluctuations. The preprocessed temperature data is organized according to the physical location of the sensors on the bearing, forming a three-dimensional temperature data cube. The three dimensions of the data cube correspond to the axial, circumferential, and radial positions of the bearing, respectively. Converting the bearing temperature gradient data into a three-dimensional thermogram tensor is a prerequisite for locating abnormal heat sources. The conversion process is achieved through spatial interpolation. Each temperature sensor has a defined spatial coordinate within the three-dimensional data cube. For grid points in the data cube not directly covered by the sensors, their temperature values ​​are estimated using a trilinear interpolation algorithm. The trilinear interpolation algorithm calculates the temperature at any point inside a cube by linearly weighting the known temperature values ​​at its eight vertices. After interpolation filling, the entire bearing volume is discretized into a regular three-dimensional mesh, with each mesh vertex storing a temperature value, thus forming a three-dimensional heatmap tensor. The dimension of the heatmap tensor is determined by the mesh resolution; for example, 32 elements along the axial direction, 64 elements along the circumference, and 8 elements along the radial direction. Each element of the tensor represents the temperature of the corresponding tiny volume cell.

[0031] A deep convolutional feature extraction network (DCNN) is used to perform anomalous heat source localization on a 3D heatmap tensor. The network structure is specifically designed to handle 3D spatial data. The input layer of the DCNN receives a 3D heatmap tensor with dimensions of 32×64×8×1. The first part of the network contains three consecutive 3D dilated convolutional layers, which expand the receptive field without increasing the number of parameters. The first 3D dilated convolutional layer uses 32 kernels of size 3×3×3 with a dilation rate of 1, producing 32 feature maps. The second 3D dilated convolutional layer uses 64 kernels of size 3×3×3 with a dilation rate of 2. The third 3D dilated convolutional layer uses 128 kernels of size 3×3×3 with a dilation rate of 4. Each 3D dilated convolutional layer is followed by a 3D batch normalization layer and a ReLU activation function layer. The second part of the network is a 3D feature pyramid module for fusing multi-scale features. The feature pyramid module generates feature maps of different scales through max pooling, then uses 3D transposed convolution for upsampling, and concatenates feature maps of the same scale through channel stitching to achieve the fusion of multi-scale features.

[0032] Non-maximum suppression (NMS) is used to locate the center coordinates of temperature anomaly regions. The NMS algorithm is performed on the feature map output by a deep convolutional feature extraction network. The network outputs a saliency map with the same tensor space size as the input heatmap. The value of each pixel in the saliency map represents the probability that the location is an anomalous heat source. The NMS algorithm slides a window of a preset size (5×5×3) across the 3D saliency map. For each point within the window, the algorithm compares its saliency value with the saliency values ​​of all other points within the window. If the saliency value of a point is not the maximum value within the window, its saliency value is suppressed to zero. This process traverses the entire 3D saliency map, retaining only those points with the most prominent saliency within local regions. Then, a probability threshold is set, retaining only those points whose saliency probability values ​​exceed the threshold; these points are initially identified as core points of the temperature anomaly region. For each retained core point, the weighted average of the 3D spatial coordinates of its connected region is taken as the center coordinates of the anomaly region, with the weights being the saliency probability values ​​of each point. Verifying the physical plausibility of anomaly region coordinates using the bearing structural topology diagram is a crucial step in avoiding false alarms. The bearing structural topology diagram is a three-dimensional computer-aided design model that details the geometry, dimensions, and spatial relationships of each component, including the inner and outer rings, rolling elements, cage, and seals. The verification process maps the center coordinates of the anomaly regions initially located by the non-maximum suppression algorithm to the coordinate system of the bearing structural topology diagram. For each candidate anomaly region's center coordinates, the shortest distance from it to the surface of each bearing component is calculated. If the coordinate point is located on the working contact surface of the bearing raceway, near the trajectory of the rolling element, or within the guide area of ​​the cage, its physical plausibility is considered high. Conversely, if the coordinate point is located in the external space of the bearing, inside the seal, or in areas where abnormally high temperatures are unlikely to occur, such as the lubrication channel, the location result is deemed unreasonable and discarded.

[0033] Calculating the strain concentration factor in the structural strain field data requires first locating the coordinates of local strain gradient extrema. The structural strain field data is acquired by resistance strain gauge arrays or fiber optic grating sensor networks attached to key parts of the bogie frame. The data contains three-dimensional strain tensor information for each measurement point. The local strain gradient is calculated using the central difference method. For each point in the grid, the rate of change of each component of its strain tensor in the three spatial directions is calculated, thus obtaining the strain gradient tensor. The strain gradient magnitude is obtained by calculating the Frobenius norm of the strain gradient tensor. Local strain gradient extrema detection is performed in the strain gradient magnitude field, searching for points with the largest strain gradient magnitude in their three-dimensional spatial neighborhood. The neighborhood size is defined as a 5×5×5 cube centered on the current point. The coordinates of the detected extrema point identify the areas of most drastic strain change on the frame; these areas are typically potential locations for stress concentration and fatigue crack initiation.

[0034] The strain energy density distribution characteristics around the extreme point are extracted. Strain energy density is a physical quantity that measures the degree of strain energy stored at a point in a material. For linear elastic materials, strain energy density... The calculations are based on the strain tensor and the elastic constants of the material:

[0035] in: These are components of the Cauchy stress tensor. These are components of the strain tensor. Stress components. By using the generalized Hooke's law, the strain components The calculation requires the material's elastic modulus and Poisson's ratio. For each extreme point of the strain gradient, a fixed-volume neighborhood is extracted centered on its coordinates, with the neighborhood size set to a cube with a side length of 20 mm. The strain energy density values ​​of all points within this neighborhood are calculated, and its characteristic statistics are extracted, including the average strain energy density, maximum strain energy density, standard deviation of strain energy density, and spatial distribution moment of strain energy density.

[0036] Spatial correlation matching is performed between the coordinates of temperature anomaly regions and the coordinates of strain gradient extrema points, within the unified coordinate system of the bogie. Spatial correlation matching calculates the three-dimensional Euclidean distance between the center coordinates of each temperature anomaly region and the coordinates of each strain gradient extrema point. If the distance between a pair of temperature anomaly points and strain extrema points is less than a set correlation threshold (15 mm), the two points are considered spatially correlated, possibly stemming from the same fault mechanism, such as localized overheating due to bearing jamming and stress concentration in the frame caused by additional bending moment. For each successfully matched coordinate pair, its spatial distance is recorded as a measure of correlation strength; the closer the distance, the stronger the correlation. A secondary fusion feature set is generated by fusing the features of the successfully matched coordinate regions. The fusion operation is performed on each pair of spatially matched temperature anomaly regions and strain concentration regions. Features extracted from temperature anomaly regions include the three-dimensional spatial dimensions of the anomaly region, the highest temperature value, the regional average temperature, and the asymmetry index of the temperature distribution. Features extracted from strain concentration regions include the gradient magnitude of the strain gradient extrema points, the neighborhood average strain energy density, the maximum strain energy density, and the spatial gradient of the strain energy density. The fusion process concatenates feature vectors from the temperature anomaly region with feature vectors from the corresponding strain concentration region to form a joint feature vector. Furthermore, the spatial distance between the center points of the two regions is added as an additional feature element to the joint feature vector. The joint feature vectors generated from all successfully matched coordinate pairs together constitute the secondary fusion feature set. This secondary fusion feature set captures the thermo-mechanical coupling effect, reflecting the interaction between the local temperature field and strain field caused by factors such as mechanical constraints, frictional heating, or uneven cooling, providing crucial information for assessing the bearing condition and its impact on the surrounding structure.

[0037] Example 4: The core of the dynamic performance evaluation model is the construction of a cross-feature correlation graph between the primary and secondary fusion feature sets. The cross-feature correlation graph is a graph data structure where nodes represent feature elements and edges represent relationships between them. Each node corresponds to a specific feature value, such as an energy value from a vibration frequency band energy vector in the primary fusion feature set, or the highest temperature value from a temperature anomaly region in the secondary fusion feature set. The total number of nodes equals the sum of the dimensions of the primary and secondary fusion feature sets. Edge construction is based on the physical coupling and statistical correlation between feature elements. The physical coupling is determined according to the bogie's multibody dynamics model. For example, an edge is established between a temperature anomaly feature node in the axle box bearing area and a strain concentration feature node in an adjacent part of the frame. The edge weight can be initialized based on the geometric distance and connection stiffness between the two components. Statistical correlation is determined by analyzing the Pearson correlation coefficient of different feature sequences in historical data. Connection edges are established between feature pairs whose correlation coefficients exceed a set threshold. Cross-feature association graphs can be represented as adjacency matrices or adjacency lists and stored in the memory of the computing device for subsequent graph traversal and feature aggregation operations.

[0038] By aggregating cross-feature associations in a graph neural network (Graph Neural Network), multi-hop neighbor features in the graph are linked. The Graph Neural Network employs a message-passing mechanism to update node features. This mechanism comprises two key steps: message generation and feature update. For each node in the graph, the message generation step collects feature information from its direct neighbors (one-hop neighbors). Each neighbor's feature is transformed through a trainable fully connected layer (message network) to generate a message. These messages from neighbors are then merged using an aggregation function, which can be average pooling, max pooling, or weighted sum. After aggregating the one-hop neighbor information, the feature update step combines the node's current features with the aggregated neighbor messages. This combination can be achieved by concatenating the features and passing them through a fully connected layer (update network), or by using a gating mechanism such as a GRU unit for fusion, thus generating the updated feature representation for that node. To capture broader dependencies, this process iterates multiple times, each iteration extending the depth of neighbor information propagation by one hop. After K iterations, the final feature representation of each node effectively incorporates information from all its neighbors within its K-hop range. This multi-hop aggregation mechanism enables the model to capture the complex interactions between the states of remote components in the bogie system.

[0039] Gated recurrent units (GRUs) are employed to capture temporal dependencies in feature interactions. These GRUs are integrated into the feature update step of a graph neural network (Graph Neural Network) or act as a standalone temporal processing layer on sequential graph features. When processing monitoring data across consecutive time steps, a cross-feature association graph is generated at each time step. The GRU maintains a hidden state, which serves as a summary of the feature interaction history from past time steps. At each new time step, the GRU receives the Graph Neural Network output features of the current time step as input. The internal update and reset gates of the GRU control how the historical hidden state and the current input information are fused to generate a new hidden state. The update gate determines how much historical information needs to be retained, while the reset gate controls the degree of influence of historical information on the current candidate state. Through this mechanism, the GRU can learn the dynamic evolution of feature interaction patterns over time; for example, how the correlation strength between a gradually increasing vibration feature and a slowly rising temperature feature changes as the equipment deteriorates.

[0040] The output of the performance degradation index set is based on a high-level feature representation processed by a graph neural network and a gated recurrent unit. The calculation of the axle box bearing wear score integrates bearing temperature anomaly region features with associated vibration energy features. Temperature anomaly region features include the volume of the anomaly region, the difference between the highest temperature and the average operating temperature, and the rate of temperature rise. Vibration energy features primarily focus on narrowband energy related to bearing failure frequencies, such as the energy amplitude of the bearing outer ring passing frequency, inner ring passing frequency, and their harmonic components. These features are input into a multilayer perceptron regression model, which outputs a wear score between 0 and 1, with higher scores indicating more severe wear. The calculation of the frame fatigue accumulation score combines strain energy density distribution features with wheel-rail contact force asymmetry features. Strain energy density distribution features are extracted from the secondary fusion feature set, including the maximum strain energy density value and the volume of the high strain energy density region. Contact force asymmetry is calculated by comparing the differences in the wheel-rail contact force distribution matrices on the left and right sides, for example, by calculating the ratio of the sum of the contact forces on both sides or the force correlation coefficient of the corresponding mesh elements. These features are input into another specialized multilayer perceptron model, which outputs a fatigue accumulation score for the chassis, the score of which is proportional to the degree of fatigue damage to the material. The calculation of the wheelset tread damage score depends on the number of abrupt changes in the wheel-rail contact force gradient and the abnormal peak values ​​of vibration band energy. Abrupt changes in the contact force gradient are detected by analyzing the gradient amplitude of the contact force distribution matrix in space. A gradient threshold is set, and the number of grid cells whose gradient values ​​exceed the threshold is counted as the number of abrupt changes. Abnormal peak values ​​of vibration band energy refer to energy peaks that occur in the relevant characteristic frequency bands of the wheelset (e.g., 200-800Hz) and are significantly higher than the historical baseline level. After these indices are normalized, the final tread damage score is synthesized through a regression model. See Table 1, which shows the mapping relationship between each score in the performance degradation index set and the key input features.

[0041] Table 1: Mapping Relationship between Scores and Key Input Features in the Performance Degradation Index Set Axle box bearing wear rating Volume of bearing temperature anomaly area Vibration energy at bearing failure frequency Multilayer perceptron regression model Structural fatigue cumulative score Maximum strain energy density value Lateral asymmetry of wheel-rail contact force Multilayer perceptron regression model Wheel tread damage score Number of abrupt changes in contact force gradient Wheelset characteristic frequency band energy peak Multilayer perceptron regression model Training the dynamic performance evaluation model requires a large amount of historical monitoring data and corresponding equipment disassembly and maintenance records as labels. The training process aims to minimize the difference between the model's predicted performance degradation score and the actual degradation state determined based on actual maintenance records. The loss function typically employs mean squared error loss or cross-entropy loss. The optimization algorithm uses an adaptive moment estimation optimizer to update the parameters of the graph neural network, gated recurrent units, and the output layer regression model. The trained dynamic performance evaluation model can receive new primary and secondary fusion feature sets. Through a series of internal feature interactions and time-series analysis calculations, it ultimately generates a comprehensive and quantified set of performance degradation indices, providing accurate data support for predictive maintenance of bogies.

[0042] See Figure 4 This graph, with the monitoring period on the horizontal axis and the performance degradation index on the vertical axis, visually presents the performance degradation trends of the three key bogie components: axle box bearing wear, frame fatigue accumulation, and wheelset tread damage, while also marking critical thresholds. The curves show that axle box bearing wear increases continuously with the monitoring period, exhibiting the fastest rate of increase; wheelset tread damage degradation follows; and frame fatigue accumulation degradation is relatively gradual. The critical threshold is a warning line set based on the dynamic performance evaluation model. When the component performance degradation index exceeds this value, preventative maintenance strategies should be closely monitored. This graph is a visualization of the dynamic performance evaluation model's temporal dependency analysis of multi-source fusion features. It accurately quantifies the performance degradation process of each component by aggregating multi-hop neighbor features through graph neural networks and capturing temporal correlations through gated cyclic units. It addresses the shortcomings of traditional monitoring methods, which focus on post-fault diagnosis and neglect process prediction. It enables dynamic tracking of bogie component performance degradation, providing maintenance personnel with a clear health status evolution map. Based on the time nodes and degradation rates of each component crossing critical thresholds, targeted preventive maintenance plans can be formulated. While ensuring train operation safety, it optimizes the allocation of maintenance resources, demonstrating the technical value of multi-sensor fusion and deep feature interaction in the condition monitoring of rail transit equipment.

[0043] Example 5: Establishing a mapping model between component performance degradation index and remaining service life is the foundation for prioritization. The construction of this mapping model relies on a large amount of historical operating data and maintenance records. The historical data warehouse stores the performance degradation index sequence periodically collected for the same model of bogie throughout its entire lifecycle, as well as the actual remaining service life of components determined through disassembly measurement or condition monitoring. The mapping model employs a combination of physics-based degradation models and data-driven models. Physics-based models, such as Paris's law, describe fatigue crack propagation, and the Archard wear model simulates sliding wear. These models treat the performance degradation index as an internal state variable, describing the evolution of the state with operating time or load cycles through differential equations, extrapolating to the failure threshold to obtain the theoretical remaining service life. The data-driven model uses random forest regression or support vector regression algorithms to directly learn the complex nonlinear mapping from the performance degradation index patterns (including current values, trends, and fluctuation characteristics) extracted from historical data to the actual remaining service life. The model output is a probability distribution of the remaining service life for each key component of the bogie, typically represented by a Weibull distribution or a log-normal distribution, with its distribution parameters determined by fitting historical data. Calculating the rate of change of the performance degradation index of each component and its distance from the critical threshold provides a basis for dynamic risk assessment. The rate of change is calculated using time series analysis, applying a sliding window linear fit to the performance degradation index sequence of each component. The window length is set based on the data acquisition frequency and the urgency of the degradation process, for example, using data points from the past 30 monitoring periods. The slope obtained from the linear fit is the average rate of change within that time window; a positive slope indicates accelerated degradation, while a negative slope indicates mitigation. Calculating the distance to the critical threshold requires pre-defining a safe critical threshold for the performance degradation index of each component. These critical thresholds are derived from technical specifications, safety standards, and reliability engineering analysis. The distance is calculated as the absolute difference between the current performance degradation index value and the critical threshold. Warning thresholds are also set for the predicted remaining useful life; for example, a warning is triggered when the predicted remaining useful life is less than 3 months. There is also a time distance between the current predicted remaining useful life and the warning threshold. The rate of change and the distance to the critical threshold together characterize the urgency of component deterioration; a component with a high rate of change and a small distance to the critical threshold requires close attention.

[0044] A multi-objective decision-making algorithm is employed to generate maintenance priority ranking results that balance safety and economy. This algorithm handles conflicting optimization objectives. The safety objectives for maintenance decisions include minimizing system failure probability and maximizing operational safety margin; the economic objectives include minimizing direct maintenance costs, minimizing downtime losses, and maximizing component utilization. The Analytic Hierarchy Process (AHP) is a commonly used algorithm. Its first step is to construct a hierarchical decision structure. The top layer is the objective layer, which determines the optimal maintenance priority. The middle layer is the criterion layer, containing safety, economy, and other criteria. The bottom layer is the scheme layer, which comprises the various bogie components to be evaluated. The second step involves constructing a judgment matrix using expert scoring, comparing the relative importance of safety and economy criteria pairwise, and comparing the performance of each component under different criteria pairwise. The third step calculates the weight vectors of each judgment matrix and performs a consistency check, finally synthesizing a comprehensive weight score for each component relative to the overall objective. A technically improved sequential preference algorithm is another option. This method requires determining the ideal and negative ideal solutions for each decision criterion. The ideal solution is the set of optimal values ​​under each criterion, and the negative ideal solution is the set of worst values ​​under each criterion. Calculate the Euclidean distance from each alternative component to the ideal solution and the negative ideal solution, and sort them by relative proximity. Regardless of the algorithm used, the final output is a list of components, arranged from highest to lowest maintenance urgency.

[0045] Parsing the highest priority component identifier in the maintenance priority ranking is the first step in performing maintenance operations. The ranking result is an ordered sequence, with the first element being the component most in need of maintenance. The component identifier parsing process maps the abstract component codes in the ranking result to specific, unambiguous physical entity descriptions. Component identifier information is stored in the bogie asset information model, which includes attributes such as a unique code, name, model, installation location, production batch, and historical maintenance records for each component. The parsing process queries the asset information model based on the component code in the ranking result to obtain the complete identifier description of the component, such as "the left axle box bearing of bogie 01, serial number SN20240521001". This precise identifier ensures that subsequent maintenance operations can accurately target the component. Matching the standard maintenance procedure parameters corresponding to the highest priority component relies on the maintenance knowledge base. The maintenance knowledge base is a structured database that stores standard maintenance procedures for various components under various typical defect modes. The knowledge base is indexed by component type and failure mode. Each record details the work content, required tools and equipment, spare parts list, estimated working hours, skill requirements, safety precautions, and quality acceptance standards. The matching process uses the highest priority component's identifier and its main performance degradation type (such as wear, fatigue, or damage) as input to retrieve the corresponding standard maintenance procedure parameters from the maintenance knowledge base. For example, for a component identified as "axle box bearing wear," the matched procedure parameters might include a series of standardized operating instructions and their key parameter values, such as "bearing disassembly procedures, cleaning and inspection standards, new bearing installation torque values, grease application amounts, and no-load test run specifications."

[0046] The final output of the implementation is to generate preventative maintenance instructions that include maintenance time window positioning and resource allocation plans. Maintenance time window positioning requires comprehensive consideration of train timetables, predicted remaining service life of components, availability of maintenance depots, and the time required for maintenance operations. The time window positioning algorithm uses the predicted component failure time as a baseline and works backwards by a safety buffer period to determine a suggested time interval for performing maintenance operations. This time interval needs to interact with the train operation dispatching system to ensure that the target bogie can be taken out of service and sent to a suitable maintenance site within this time window. The resource allocation plan is formulated based on the resource requirements defined in the standard maintenance operation process parameters, combined with the real-time status in the enterprise resource management system. Resources include human resources, assigning qualified maintenance engineer teams; tool and equipment resources, dispatching necessary hydraulic pullers, torque wrenches, heating devices, etc.; and spare parts and materials resources, ensuring that the required types of bearings, seals, grease, etc., are in stock and can be delivered to the work station in a timely manner. The preventative maintenance instruction ultimately generates a structured electronic work order. This work order clearly specifies the maintenance target, the maintenance tasks, the execution time window, the required resource list, safety and quality requirements, and the feedback process for the execution results. This instruction will be automatically sent to the relevant maintenance execution system and resource management department to initiate substantive maintenance preparation work.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor fusion dynamic performance monitoring method for a rail transit electric passenger car bogie, characterized in that, include: Collect multi-source heterogeneous sensing data on the bogie's operating status. The multi-source heterogeneous sensing data includes vibration acceleration time-series signals, wheel-rail contact force distribution data, bearing temperature gradient data, and structural strain field data. Frequency domain energy distribution analysis is performed on the vibration acceleration time series signal to extract the frequency band energy feature vector. Spatial gridded feature mapping is performed on the wheel-rail contact force distribution data to generate the contact force distribution matrix. The frequency band energy feature vector and the contact force distribution matrix are spatiotemporally aligned and fused to form a primary fusion feature set. An abnormal heat source localization process is performed on the bearing temperature gradient data by a deep convolutional feature extraction network to identify the coordinates of the abnormal temperature region, and a secondary fusion feature set is generated by combining the strain concentration coefficient in the structural strain field data. The primary and secondary fusion feature sets are input into the dynamic performance evaluation model for multi-dimensional feature interaction calculation, and the performance degradation index set of each component of the bogie is output. The performance degradation index set includes the axle box bearing wear score, the frame fatigue accumulation score, and the wheelset tread damage score. Based on the set of performance degradation indices, the bogie maintenance priority ranking results are generated, and the preventive maintenance instructions for the corresponding components are triggered. The step of performing frequency domain energy distribution analysis on the vibration acceleration time-series signal to extract frequency band energy feature vectors includes: An adaptive window length Fourier transform is used to decompose the vibration acceleration time-series signal into multiple time-frequency energy units; Calculate the energy percentage of each time-frequency energy unit within a preset characteristic frequency band to form an initial frequency band energy sequence; The initial frequency band energy sequence is optimized by wavelet packet reconstruction algorithm to eliminate frequency band aliasing effect. The optimized frequency band energy sequence is grouped and aggregated according to the inherent frequency range of key bogie components to generate frequency band energy feature vectors. The process of performing spatial gridded feature mapping on the wheel-rail contact force distribution data to generate a contact force distribution matrix includes: Establish a three-dimensional mesh coordinate system for the wheel-rail contact surface and map the contact force sampling points to the corresponding mesh cells; The inverse distance weighted interpolation algorithm is used to calculate the contact force estimate of the unsampled grid cells; The contact force data of the grid cells are normalized to eliminate the differences in dimensions between different sensors; Extract the spatial gradient and symmetry features of the contact force distribution to construct the contact force distribution matrix.

2. The multi-sensor fusion based dynamic performance monitoring method of railway transit electric passenger car bogie according to claim 1, characterized in that, The process of spatiotemporally aligning and fusing the frequency band energy feature vector with the contact force distribution matrix to form a primary fused feature set includes: Establish a mapping table between the physical locations of vibration frequency bands and contact force grids; The timestamps of the frequency band energy feature vectors are aligned and transformed with the spatial coordinates of the contact force distribution matrix using a time-space encoder. A dual-stream attention mechanism is used to calculate the correlation weight matrix between frequency band energy characteristics and contact force characteristics; The frequency band energy feature vector and the contact force distribution matrix are weighted and concatenated according to the correlation weight matrix to generate a primary fusion feature set.

3. The multi-sensor fusion based dynamic performance monitoring method of railway transit electric passenger car bogie according to claim 1, characterized in that, The step of performing abnormal heat source localization processing on bearing temperature gradient data through a deep convolutional feature extraction network and identifying the coordinates of abnormal temperature regions includes: Convert the bearing temperature gradient data into a three-dimensional thermogram tensor; Multiscale temperature distribution features of thermogram tensors are extracted using dilated convolutional layers. The non-maximum suppression algorithm was used to locate the center coordinates of the temperature anomaly region; The physical rationality of the coordinates of the abnormal area is verified by combining the bearing structure topology diagram.

4. The multi-sensor fusion method for monitoring the dynamic performance of rail transit electric train bogies according to claim 1, characterized in that, The secondary fusion feature set generated by combining the strain concentration coefficients in the structural strain field data includes: Calculate the coordinates of the local strain gradient extreme points in the structural strain field data; Extract the strain energy density distribution characteristics of the region surrounding the extreme point; Spatial correlation matching is performed between the coordinates of the temperature anomaly region and the coordinates of the extreme points of the strain gradient. The coordinate region features that are successfully fused and matched are used to generate a secondary fused feature set.

5. The multi-sensor fusion method for monitoring the dynamic performance of bogies in rail transit electric passenger vehicles according to claim 1, characterized in that, The step of inputting the primary fusion feature set and the secondary fusion feature set into the dynamic performance evaluation model for multi-dimensional feature interaction calculation includes: Construct a cross-feature association graph between the primary fusion feature set and the secondary fusion feature set; The graph neural network aggregates cross-feature associations with multi-hop neighbor features in the graph. Gated recurrent units are used to capture temporal dependencies in the feature interaction process; Output a set of performance degradation indices for each component of the bogie.

6. The multi-sensor fusion method for monitoring the dynamic performance of bogies in rail transit electric passenger vehicles according to claim 5, characterized in that, The performance degradation index set includes axle box bearing wear score, frame fatigue accumulation score, and wheelset tread damage score, including: The wear score of the bearing box is calculated based on the area of ​​the abnormal bearing temperature region and the characteristics of vibration energy. The fatigue accumulation score of the frame is calculated by combining the characteristics of strain energy density distribution and contact force asymmetry. The wheelset tread damage score is calculated by the number of abrupt changes in the wheel-rail contact force gradient and the peak value of the frequency band energy anomaly.

7. The method for monitoring the dynamic performance of rail transit electric train bogies using multi-sensor fusion according to claim 1, characterized in that, The bogie maintenance priority ranking result generated based on the performance degradation index set includes: Establish a mapping model between component performance degradation index and remaining service life; Calculate the rate of change of the performance degradation index of each component and its distance from the critical threshold. A multi-objective decision-making algorithm is used to generate maintenance priority ranking results that take into account both security and economy.

8. The multi-sensor fusion method for monitoring the dynamic performance of bogies in rail transit electric passenger vehicles according to claim 7, characterized in that, The preventative maintenance command that triggers the corresponding component includes: Parse the identifier of the highest priority component in the maintenance priority ranking results; Match the standard maintenance procedure parameters corresponding to this component; Generate preventative maintenance instructions that include maintenance time window positioning and resource allocation plans.

Images