A wheel-rail system vibration abnormality separation and identification method and system for complex working conditions

Abstract

The application discloses a wheel-rail system vibration anomaly separation and identification method and system for complex working conditions, and belongs to the technical field of anomaly identification, and comprises the following steps: acquiring multi-source heterogeneous time sequence vibration signals of a wheel-rail system under complex working conditions, and generating a three-dimensional dynamic feature tensor; inputting the three-dimensional dynamic feature tensor into a pre-constructed vibration source physical constraint embedding model to obtain a physical constraint feature vector set output by the model; decoupling vibration components of the physical constraint feature vector set to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution degree atlas; inputting the probability distribution matrix into an anomaly pattern adaptive clustering engine to generate an anomaly pattern label set; and fusing the anomaly pattern label set and the decoupling contribution degree atlas to generate a wheel-rail system vibration anomaly separation and identification report; and the application can adapt to the fine modeling requirements of multi-source heterogeneous time sequence vibration signals, and effectively avoids a significant decline in subsequent vibration component separation precision.

Description

Technical Field

[0001] This invention relates to a method and system for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions, belonging to the field of anomaly identification technology. Background Technology

[0002] Urban rail transit, high-speed rail, and other rail transit systems have become an important part of the modern comprehensive transportation system. As the core carrier of the coupling between rail vehicles and track foundations, the wheel-rail system's operational health directly determines the safety, stability, and service life of rail transit operations. When trains operate under complex actual conditions, they are affected by multiple factors such as rail corrugation, wheel polygons, wheel-rail tread damage, fastener loosening, track bed deterioration, track irregularities, train speed fluctuations, and load changes. These factors cause the wheel-rail system to generate complex temporal vibration signals with multiple sources, multiple couplings, and strong interference. The vibration components generated by various excitation sources are superimposed and modulated with each other, exhibiting typical multi-source heterogeneous, nonlinear, and non-stationary characteristics. However, existing technologies for identifying vibration anomalies in wheel-rail systems mostly remain at the level of traditional signal processing and shallow feature analysis, and generally have obvious technical shortcomings and application limitations. Most existing methods only carry out vibration signal feature mining from a single time domain, frequency domain, or time-frequency dimension, lacking in-depth characterization of the time-space-frequency correlation of vibration signals. They cannot construct high-dimensional feature expressions that can characterize the evolution law of wheel-rail coupled vibration under complex working conditions, and are difficult to adapt to the fine modeling requirements of multi-source heterogeneous time-series vibration signals. Their feature characterization capabilities are insufficient, which leads to a significant decrease in the accuracy of subsequent vibration component separation.

[0003] It is evident that, in order to meet the requirements of refined modeling of multi-source heterogeneous temporal vibration signals and avoid a significant decrease in the accuracy of subsequent vibration component separation, there is an urgent need for a vibration anomaly separation and identification method for wheel-rail systems under complex working conditions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions. This method and system can adapt to the requirements of fine modeling of multi-source heterogeneous time-series vibration signals and effectively avoid a significant decrease in the accuracy of subsequent vibration component separation.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions, comprising:

[0007] Acquire multi-source heterogeneous temporal vibration signals of the wheel-rail system under complex working conditions, perform time-space-frequency three-dimensional feature mapping on the multi-source heterogeneous temporal vibration sequence numbers, and generate a three-dimensional dynamic feature tensor.

[0008] The three-dimensional dynamic feature tensor is input into the pre-constructed vibration source physical constraint embedding model to obtain the physical constraint feature vector set output by the model; The vibration component decoupling of the physical constraint feature vector set yields a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum.

[0009] The probability distribution matrix is ​​input into the abnormal pattern adaptive clustering engine to obtain the abnormal pattern clusters in the independent vibration components output by the engine, and an abnormal pattern label set is generated based on the abnormal pattern clusters.

[0010] By fusing the abnormal pattern label set with the decoupling contribution map, a vibration anomaly separation and identification report for the wheel-rail system is generated.

[0011] Specifically, after generating the wheel-rail system vibration anomaly separation and identification report, the parameters of the vibration source physical constraint embedding model are adjusted based on the wheel-rail system vibration anomaly separation and identification report.

[0012] Furthermore, the method for generating the three-dimensional dynamic feature tensor includes:

[0013] Simultaneously acquire multi-channel vibration acceleration signals, wheel-rail force signals, and operating status signals of the wheel-rail system under complex working conditions including curve sections, turnout areas, and different wheel-rail interface states, forming a multi-source heterogeneous time-series vibration signal stream;

[0014] The multi-source heterogeneous temporal vibration signal stream is subjected to spatiotemporal alignment and scale normalization to generate a synchronous standardized signal matrix.

[0015] For each signal channel in the synchronous normalized signal matrix, recursive quantum coding technology is applied to encode the amplitude changes of the time-series signal into a symbol sequence, and the symbol transition probability is statistically analyzed to generate the time-domain recursive dynamic feature vector of each channel.

[0016] Adaptive wavelet ridge extraction is performed on the synchronous normalized signal matrix to generate a time-frequency joint distribution feature matrix.

[0017] Based on the time-domain recursive dynamic feature vector and the time-frequency joint distribution feature matrix, a three-dimensional dynamic feature tensor is constructed by fusing the spatial location and operating condition labels in the operating status signal.

[0018] Furthermore, the method for generating the physical constraint feature vector set includes: From the three-dimensional dynamic feature vector, the wheel-rail contact geometric constraints and system dynamic boundary conditions under the current working condition are parsed from the operating state signal to generate the physical rule template corresponding to the working condition;

[0019] A physical constraint loss function is constructed that includes vibration energy conservation, momentum transfer and structural modal coupling. The parameters in the physical constraint loss function are instantiated using the physical rule template corresponding to the working condition.

[0020] The three-dimensional dynamic feature tensor is input into a pre-trained variational encoder. The training objective of the variational encoder is to minimize the physical constraint loss function while reconstructing the input features, so that the encoder learns feature representations that satisfy physical laws in the latent space.

[0021] From the bottleneck layer of the variational encoder, feature vectors of the latent space are extracted, and feature vectors with a deviation from the physical constraint loss function calculation below a preset threshold are selected through physical rule compliance test to form a primary physical constraint feature vector set.

[0022] Spectral clustering analysis is performed on the primary physical constraint feature vector set to merge feature clusters generated by the same physical mechanism, and the centroid vector of each feature cluster is extracted to generate a redundancy-free physical constraint feature vector set. Each vector in this set represents a potential vibration source feature that conforms to physical laws.

[0023] Furthermore, the vibration components of the physical constraint feature vector set are decoupled to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum, including:

[0024] The physical constraint feature vector set is input into a pre-constructed and trained generative adversarial network to obtain the synthetic feature vector output by the generative adversarial network.

[0025] The synthesized features are subjected to channel-level Softmax normalization to obtain the probability that each potential vibration source feature vector belongs to each separated vibration component, forming a probability distribution matrix;

[0026] Calculate the gradient of the probability value of each separated vibration component in the probability distribution matrix with respect to each local feature in the original three-dimensional dynamic feature tensor, and generate a decoupling contribution map based on the gradient value;

[0027] The generative adversarial network includes a generator network and multiple discriminator networks. Each discriminator network corresponds to a potential vibration component to be separated. The generator network reconstructs a synthetic feature tensor that is consistent with the original three-dimensional dynamic feature tensor in terms of physical constraint loss but is separable in each dimension based on the input physical constraint feature vector. Each discriminator network is used to determine whether the synthetic feature tensor contains its corresponding vibration component, and through adversarial training, the generator network separates vibration features of different physical mechanisms into different dimensions or channels of the synthetic tensor.

[0028] Furthermore, during the stabilization phase of adversarial training, the internal neuron patterns activated by the generator network to reconstruct each physical constraint feature vector are recorded, generating a pattern activation map.

[0029] The physical constraint feature vector set is input into the frozen generator, and a probe layer is inserted into the middle layer of the network to capture the intermediate representation sequence during the feature transformation process. Attention weights are assigned to the intermediate representation sequence, and the importance score of each position in the sequence to each channel of the final output synthetic feature tensor is calculated to generate an attention-guided channel association weight matrix.

[0030] Based on the channel association weight matrix, the synthesized feature tensor output by the generator is weighted, fused, and recombined with channels to obtain the recombined synthesized feature tensor.

[0031] The recombined synthetic feature tensor is subjected to a Softmax operation along the channel dimension to convert the response value of each channel into a probability form, thereby obtaining the probability distribution of each vector in the physical constraint feature vector set corresponding to each separated vibration component, and generating a probability distribution matrix.

[0032] Furthermore, the data processing procedure of the anomaly pattern adaptive clustering engine includes:

[0033] The probability distribution matrix is ​​regarded as a set of points in a high-dimensional feature space. Each point in the set represents the coordinates of a vibration sample in the separated component space. A cluster center self-discovery algorithm based on local density peaks and relative distance is used to identify the cluster center points in the probability distribution matrix.

[0034] The cluster center points are input into a pre-constructed dynamic similarity measurement network to obtain a similarity matrix;

[0035] The similarity is subjected to a spectral clustering algorithm to divide the points in the probability distribution matrix into different clusters, resulting in several vibration mode clusters, each cluster representing a specific vibration mode;

[0036] Several vibration mode clusters are compared with a pre-constructed normal mode baseline distribution model by KL divergence calculation. Vibration mode clusters with divergence values ​​exceeding a preset dynamic threshold are marked as abnormal modes to obtain abnormal mode clusters. The normal mode baseline distribution model is constructed based on the probability distribution of vibration components under historical normal operating conditions.

[0037] An abnormal pattern label set is generated based on the abnormal pattern cluster, including:

[0038] Each abnormal pattern cluster is assigned a unique abnormal pattern label, resulting in an abnormal pattern label set.

[0039] Furthermore, the data processing procedure for the normal pattern baseline distribution model includes:

[0040] Extract the probability distribution of all sample points in each vibration mode cluster to be identified, and calculate the empirical probability distribution of the cluster.

[0041] Calculate the KL divergence between the empirical probability distribution of each vibration mode cluster and the normal mode baseline distribution model, and use the KL divergence as an index of the abnormal deviation of the mode cluster.

[0042] Based on the type and intensity of the current complex working conditions, the dynamic threshold for anomaly determination of KL divergence is determined based on a pre-constructed working condition-threshold mapping table, and pattern clusters whose anomaly deviation index exceeds the dynamic threshold are determined to be abnormal.

[0043] Furthermore, by fusing the abnormal pattern label set with the decoupling contribution map, a wheel-rail system vibration anomaly separation and identification report is generated, including:

[0044] Establish a mapping relationship between the set of anomalous pattern labels and the decoupling contribution map to obtain the spatial contribution region of each anomalous pattern cluster on the decoupling contribution map;

[0045] Based on the spatial contribution regions of each anomaly pattern cluster on the decoupling contribution map, the time-frequency information in the three-dimensional dynamic feature tensor is superimposed to obtain a visualization map with time-space-frequency anomaly localization.

[0046] By combining the anomaly deviation index and the main separation component index in the anomaly pattern label set, the main physical source of the anomaly is analyzed, and the corresponding fault or degradation pattern description text is matched from the pre-built knowledge base.

[0047] Based on the visualized spectrum, main physical sources, matched descriptive text, and key statistics of the original signal, an anomaly separation and identification report is generated according to a preset report template.

[0048] Furthermore, based on the vibration anomaly separation and identification report of the wheel-rail system, the parameters of the vibration source physical constraint embedding model are adjusted, including:

[0049] Extract new anomaly pattern labels and their main physical sources from the vibration anomaly separation and identification report that are not covered by the historical knowledge base, and determine the source feature vector subset of the new anomaly pattern labels based on the physical constraint feature vector set;

[0050] Based on the aforementioned subset of source feature vectors, a new form or parameter range of the physical constraint conditions it satisfies in the vibration source physical constraint embedding model is derived in reverse.

[0051] Using the subset of source feature vectors as positive samples and the normal mode feature vectors as negative samples, an incremental learning dataset is constructed. The variational encoder in the vibration source physical constraint embedding model is fine-tuned using the incremental learning dataset. The new physical constraint conditions derived in reverse are formalized and added as new entries to the physical rule template library. The physical rule template generation logic corresponding to the working condition is also updated.

[0052] Furthermore, the variational encoder in the vibration source physical constraint embedding model is fine-tuned using an incremental learning dataset, including:

[0053] Perform iterative steps until the model's reconstruction error on the new anomalous samples stabilizes and the distance between the clusters of new anomalous feature vectors in the latent space and the clusters of normal feature vectors reaches the preset separation requirement, thus completing the fine-tuning.

[0054] The iterative steps include:

[0055] A batch of vectors is randomly sampled from the source feature vector subset as positive samples, and an equal number of negative sample vectors are sampled from the historical normal pattern feature vector library to form the incremental learning data pairs for the current batch.

[0056] The incremental learning data is used to embed the physical constraints of the input vibration source into the variational encoder of the model, and the reconstruction loss and physical constraint loss are calculated. A contrastive learning loss term is introduced, which makes the feature representations of positive samples in the latent space closer to each other and farther away from the feature representations of negative samples.

[0057] The reconstruction loss, physical constraint loss, and contrastive learning loss are weighted and summed to form the total loss function in the fine-tuning stage. The gradient descent algorithm is then used to update the network parameters of the variational encoder based on the total loss function.

[0058] Secondly, the present invention provides a vibration anomaly separation and identification system for wheel-rail systems under complex operating conditions, comprising:

[0059] The first module is used to acquire multi-source heterogeneous temporal vibration signals of the wheel-rail system under complex working conditions, perform time-space-frequency three-dimensional feature mapping on the multi-source heterogeneous temporal vibration sequence numbers, and generate a three-dimensional dynamic feature tensor.

[0060] The second module is used to input the three-dimensional dynamic feature tensor into the pre-constructed vibration source physical constraint embedding model to obtain the physical constraint feature vector set output by the model.

[0061] The third module is used to decouple the vibration components of the physical constraint feature vector set to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum.

[0062] The fourth module is used to input the probability distribution matrix into the anomaly pattern adaptive clustering engine to obtain the anomaly pattern clusters in the independent vibration components output by the engine, and to generate an anomaly pattern label set based on the anomaly pattern clusters.

[0063] The fifth module is used to fuse the abnormal pattern label set with the decoupling contribution map to generate a wheel-rail system vibration anomaly separation and identification report;

[0064] Specifically, after generating the wheel-rail system vibration anomaly separation and identification report, the parameters of the vibration source physical constraint embedding model are adjusted based on the wheel-rail system vibration anomaly separation and identification report.

[0065] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0066] Fourthly, the present invention provides a computer device, comprising:

[0067] Memory, used to store computer programs / instructions;

[0068] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0069] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0070] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0071] This invention provides a method and system for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions. By acquiring multi-source heterogeneous temporal vibration signals of wheel-rail systems under complex operating conditions, performing time-space-frequency three-dimensional feature mapping on them, and generating a three-dimensional dynamic feature tensor, this step breaks through the limitations of single-dimensional feature analysis, comprehensively acquires multi-source heterogeneous temporal vibration signals under complex operating conditions, and completely captures the vibration response of wheel-rail systems at different times, different spatial locations, and different frequencies.

[0072] This invention provides a method and system for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions. Through a three-dimensional feature mapping of time, space, and frequency, it integrates dispersed single-dimensional features into a three-dimensional dynamic feature tensor, clearly presenting the correlation between vibration signals in terms of temporal evolution, spatial distribution, and frequency distribution. This comprehensively characterizes the complex evolutionary laws of wheel-rail coupled vibration, overcoming the shortcomings of incomplete single-dimensional feature representation. The generated three-dimensional dynamic feature tensor provides high-dimensional and comprehensive feature support for subsequent feature extraction of potential vibration sources and decoupling of vibration components, effectively improving feature representation capabilities and fundamentally solving the problem of insufficient feature characterization in existing methods.

[0073] This invention provides a method and system for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions. By constructing a physical constraint embedding model of vibration sources, feature vectors of potential vibration sources that meet the physical constraints of the wheel-rail system are extracted from the three-dimensional dynamic feature tensor. This step of constructing the physical constraint embedding model of vibration sources integrates the physical constraints of the wheel-rail system into the feature extraction process, which can accurately screen out the feature vectors of potential vibration sources that conform to physical laws from the three-dimensional dynamic feature tensor, eliminate redundant and interfering features, and ensure that the extracted features have clear physical meaning.

[0074] This invention provides a method and system for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions. By generating a set of physically constrained feature vectors, it focuses on the core vibration source features of the wheel-rail system, strengthening the correlation between features and vibration sources. This solves the problems of existing methods, such as lack of physical rationality in features and excessive redundant interference, thus improving the targeting and effectiveness of feature extraction. Based on the set of physically constrained feature vectors, an interpretable adversarial separation network is driven to decouple vibration components, generating a probability distribution matrix containing multiple independent vibration components and corresponding decoupling contribution maps. This step relies on the set of physically constrained feature vectors to drive the interpretable adversarial separation network, achieving accurate decoupling of multi-source coupled vibration components through adversarial training. It decomposes the aliased vibration signal into multiple independent vibration components, each corresponding to a specific vibration source.

[0075] This invention provides a method and system for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions. The generated probability distribution matrix clearly presents the distribution patterns of each independent vibration component, and the decoupling contribution spectrum intuitively reflects the degree of contribution of each vibration component to the overall vibration. At the same time, the interpretability of the network can clarify the physical meaning of each vibration component. This solves the problems of incomplete decoupling of vibration components and inability to distinguish the contribution of vibration sources in existing methods, thereby improving the accuracy and interpretability of vibration anomaly separation.

[0076] This invention provides a method and system for vibration anomaly separation and identification in wheel-rail systems under complex operating conditions. By inputting a probability distribution matrix into an adaptive clustering engine for anomaly patterns, the method identifies and classifies anomalous pattern clusters within the separated independent vibration components. This step utilizes an adaptive clustering engine that requires no manual intervention. It adaptively adjusts clustering parameters based on the characteristic differences of each independent vibration component in the probability distribution matrix, accurately identifying anomalous pattern clusters and systematically classifying different types of anomalous patterns to generate a standardized set of anomalous pattern labels. This process achieves automated identification and classification of anomalous patterns, overcoming the limitations of manual interpretation and fixed threshold judgments. It solves the problems of low efficiency, poor adaptability, and unsystematic classification in existing methods, clearly defining the anomaly types within each independent vibration component and improving the efficiency and accuracy of anomaly identification.

[0077] This invention provides a method and system for identifying and separating vibration anomalies in wheel-rail systems under complex operating conditions. By fusing anomaly pattern label sets and decoupling contribution maps, a vibration anomaly identification report is generated. This report is then fed back to the vibration source physical constraint embedding model to optimize physical constraints. This step, fusing the anomaly pattern label set and decoupling contribution map, generates a vibration anomaly identification report that clearly presents the anomaly pattern type, the contribution degree of each anomalous vibration component, and the anomaly correlation, providing a comprehensive and systematic reference for maintenance personnel to investigate the causes of anomalies and formulate maintenance strategies. Simultaneously, feeding the report back to the vibration source physical constraint embedding model allows for optimization of physical constraints based on the anomaly identification results, correcting deviations in the initial feature extraction process, and improving the accuracy of subsequent feature extraction and vibration decoupling. This step solves the problems of existing methods such as missing closed-loop mechanisms, unsystematic reporting, and the inability to continuously improve identification accuracy. It achieves refined, systematic, and continuously optimized wheel-rail vibration anomaly identification, ensuring the safe and stable operation of rail transit systems. Attached Figure Description

[0079] Figure 1 This is a schematic diagram of the steps in the method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to the present invention.

[0080] Figure 2 for Figure 1 A detailed flowchart of step S1;

[0081] Figure 3 Example of a multi-source heterogeneous time-domain vibration signal stream;

[0082] Figure 4 For synchronization and normalization signal matrix;

[0083] Figure 5 For time-domain recursive dynamic feature vectors;

[0084] Figure 6 This is the time-frequency distribution characteristic matrix;

[0085] Figure 7 It is a three-dimensional dynamic feature tensor matrix;

[0086] Figure 8 It is a set of physical constraint feature vectors;

[0087] Figure 9 Contribution map for decoupling from multiple sources;

[0088] Figure 10 Results of identifying abnormal vibration sources in wheel-rail systems. Detailed Implementation

[0090] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0091] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0092] Example 1:

[0093] Figure 1 This is a flowchart of a method for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions, according to Embodiment 1 of the present invention. The method for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions provided in this embodiment can be applied to a terminal and can be executed by a system for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions. This system can be implemented in software and / or hardware and can be integrated into the terminal, such as any smartphone, tablet, or computer device with communication capabilities. See also... Figure 1 The method implemented in this way specifically includes the following steps:

[0094] Acquire multi-source heterogeneous temporal vibration signals of the wheel-rail system under complex working conditions, perform time-space-frequency three-dimensional feature mapping on the multi-source heterogeneous temporal vibration sequence numbers, and generate a three-dimensional dynamic feature tensor.

[0095] The three-dimensional dynamic feature tensor is input into the pre-constructed vibration source physical constraint embedding model to obtain the physical constraint feature vector set output by the model;

[0096] The vibration components of the physical constraint feature vector set are decoupled to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum.

[0097] The probability distribution matrix is ​​input into the abnormal pattern adaptive clustering engine to obtain the abnormal pattern clusters in the independent vibration components output by the engine, and an abnormal pattern label set is generated based on the abnormal pattern clusters.

[0098] By fusing the abnormal pattern label set with the decoupling contribution map, a vibration anomaly separation and identification report for the wheel-rail system is generated.

[0099] Specifically, after generating the wheel-rail system vibration anomaly separation and identification report, the parameters of the vibration source physical constraint embedding model are adjusted based on the wheel-rail system vibration anomaly separation and identification report.

[0100] Furthermore, the method for generating the three-dimensional dynamic feature tensor includes:

[0101] Simultaneously acquire multi-channel vibration acceleration signals, wheel-rail force signals, and operating status signals of the wheel-rail system under complex working conditions including curve sections, turnout areas, and different wheel-rail interface states, forming a multi-source heterogeneous time-series vibration signal stream;

[0102] The multi-source heterogeneous temporal vibration signal stream is subjected to spatiotemporal alignment and scale normalization to generate a synchronous standardized signal matrix.

[0103] For each signal channel in the synchronous normalized signal matrix, recursive quantum coding technology is applied to encode the amplitude changes of the time-series signal into a symbol sequence, and the symbol transition probability is statistically analyzed to generate the time-domain recursive dynamic feature vector of each channel.

[0104] Adaptive wavelet ridge extraction is performed on the synchronous normalized signal matrix to generate a time-frequency joint distribution feature matrix.

[0105] Based on the time-domain recursive dynamic feature vector and the time-frequency joint distribution feature matrix, a three-dimensional dynamic feature tensor is constructed by fusing the spatial location and operating condition labels in the operating status signal.

[0106] Furthermore, the method for generating the physical constraint feature vector set includes:

[0107] From the three-dimensional dynamic feature vector, the wheel-rail contact geometric constraints and system dynamic boundary conditions under the current working condition are parsed according to the operating state signal, and the physical rule template corresponding to the working condition is generated.

[0108] A physical constraint loss function is constructed that includes vibration energy conservation, momentum transfer and structural modal coupling. The parameters in the physical constraint loss function are instantiated using the physical rule template corresponding to the working condition.

[0109] The three-dimensional dynamic feature tensor is input into a pre-trained variational encoder. The training objective of the variational encoder is to minimize the physical constraint loss function while reconstructing the input features, so that the encoder learns feature representations that satisfy physical laws in the latent space.

[0110] From the bottleneck layer of the variational encoder, feature vectors of the latent space are extracted, and feature vectors with a deviation from the physical constraint loss function calculation below a preset threshold are selected through physical rule compliance test to form a primary physical constraint feature vector set.

[0111] Spectral clustering analysis is performed on the primary physical constraint feature vector set to merge feature clusters generated by the same physical mechanism, and the centroid vector of each feature cluster is extracted to generate a redundancy-free physical constraint feature vector set. Each vector in this set represents a potential vibration source feature that conforms to physical laws.

[0112] Furthermore, the vibration components of the physical constraint feature vector set are decoupled to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum, including:

[0113] The physical constraint feature vector set is input into a pre-constructed and trained generative adversarial network to obtain the synthetic feature vector output by the generative adversarial network.

[0114] The synthesized features are subjected to channel-level Softmax normalization to obtain the probability that each potential vibration source feature vector belongs to each separated vibration component, forming a probability distribution matrix;

[0115] Calculate the gradient of the probability value of each separated vibration component in the probability distribution matrix with respect to each local feature in the original three-dimensional dynamic feature tensor, and generate a decoupling contribution map based on the gradient value;

[0116] The generative adversarial network includes a generator network and multiple discriminator networks. Each discriminator network corresponds to a potential vibration component to be separated. The generator network reconstructs a synthetic feature tensor that is consistent with the original three-dimensional dynamic feature tensor in terms of physical constraint loss but is separable in each dimension based on the input physical constraint feature vector. Each discriminator network is used to determine whether the synthetic feature tensor contains its corresponding vibration component, and through adversarial training, the generator network separates vibration features of different physical mechanisms into different dimensions or channels of the synthetic tensor.

[0117] Furthermore, during the stable phase of adversarial training, the internal neuron patterns activated by the generator network to reconstruct each physical constraint feature vector are recorded, generating a pattern activation map.

[0118] The physical constraint feature vector set is input into the frozen generator, and a probe layer is inserted into the middle layer of the network to capture the intermediate representation sequence during the feature transformation process. Attention weights are assigned to the intermediate representation sequence, and the importance score of each position in the sequence to each channel of the final output synthetic feature tensor is calculated to generate an attention-guided channel association weight matrix.

[0119] Based on the channel association weight matrix, the synthesized feature tensor output by the generator is weighted, fused, and recombined with channels to obtain the recombined synthesized feature tensor.

[0120] The recombined synthetic feature tensor is subjected to a Softmax operation along the channel dimension to convert the response value of each channel into a probability form, thereby obtaining the probability distribution of each vector in the physical constraint feature vector set corresponding to each separated vibration component, and generating a probability distribution matrix.

[0121] Furthermore, the data processing procedure of the anomaly pattern adaptive clustering engine includes:

[0122] The probability distribution matrix is ​​regarded as a set of points in a high-dimensional feature space. Each point in the set represents the coordinates of a vibration sample in the separated component space. A cluster center self-discovery algorithm based on local density peaks and relative distance is used to identify the cluster center points in the probability distribution matrix.

[0123] The cluster center points are input into a pre-constructed dynamic similarity measurement network to obtain a similarity matrix;

[0124] The similarity is subjected to a spectral clustering algorithm to divide the points in the probability distribution matrix into different clusters, resulting in several vibration mode clusters, each cluster representing a specific vibration mode;

[0125] Several vibration mode clusters are compared with a pre-constructed normal mode baseline distribution model by KL divergence calculation. Vibration mode clusters with divergence values ​​exceeding a preset dynamic threshold are marked as abnormal modes to obtain abnormal mode clusters. The normal mode baseline distribution model is constructed based on the probability distribution of vibration components under historical normal operating conditions.

[0126] An abnormal pattern label set is generated based on the abnormal pattern cluster, including:

[0127] Each abnormal pattern cluster is assigned a unique abnormal pattern label, resulting in an abnormal pattern label set.

[0128] Furthermore, the data processing procedure for the normal pattern baseline distribution model includes:

[0129] Extract the probability distribution of all sample points in each vibration mode cluster to be identified, and calculate the empirical probability distribution of the cluster.

[0130] Calculate the KL divergence between the empirical probability distribution of each vibration mode cluster and the normal mode baseline distribution model, and use the KL divergence as an index of the abnormal deviation of the mode cluster.

[0131] Based on the type and intensity of the current complex working conditions, the dynamic threshold for anomaly determination of KL divergence is determined based on a pre-constructed working condition-threshold mapping table, and pattern clusters whose anomaly deviation index exceeds the dynamic threshold are determined to be abnormal.

[0132] Furthermore, by fusing the abnormal pattern label set with the decoupling contribution map, a wheel-rail system vibration anomaly separation and identification report is generated, including:

[0133] Establish a mapping relationship between the set of anomalous pattern labels and the decoupling contribution map to obtain the spatial contribution region of each anomalous pattern cluster on the decoupling contribution map;

[0134] Based on the spatial contribution regions of each anomaly pattern cluster on the decoupling contribution map, the time-frequency information in the three-dimensional dynamic feature tensor is superimposed to obtain a visualization map with time-space-frequency anomaly localization.

[0135] By combining the anomaly deviation index and the main separation component index in the anomaly pattern label set, the main physical source of the anomaly is analyzed, and the corresponding fault or degradation pattern description text is matched from the pre-built knowledge base.

[0136] Based on the visualized spectrum, main physical sources, matched descriptive text, and key statistics of the original signal, an anomaly separation and identification report is generated according to a preset report template.

[0137] Furthermore, based on the vibration anomaly separation and identification report of the wheel-rail system, the parameters of the vibration source physical constraint embedding model are adjusted, including:

[0138] Extract new anomaly pattern labels and their main physical sources from the vibration anomaly separation and identification report that are not covered by the historical knowledge base, and determine the source feature vector subset of the new anomaly pattern labels based on the physical constraint feature vector set;

[0139] Based on the aforementioned subset of source feature vectors, a new form or parameter range of the physical constraint conditions it satisfies in the vibration source physical constraint embedding model is derived in reverse.

[0140] Using the subset of source feature vectors as positive samples and the normal mode feature vectors as negative samples, an incremental learning dataset is constructed. The variational encoder in the vibration source physical constraint embedding model is fine-tuned using the incremental learning dataset. The new physical constraint conditions derived in reverse are formalized and added as new entries to the physical rule template library. The physical rule template generation logic corresponding to the working condition is also updated.

[0141] Furthermore, the variational encoder in the vibration source physical constraint embedding model is fine-tuned using an incremental learning dataset, including:

[0142] Perform iterative steps until the model's reconstruction error on the new anomalous samples stabilizes and the distance between the clusters of new anomalous feature vectors in the latent space and the clusters of normal feature vectors reaches the preset separation requirement, thus completing the fine-tuning.

[0143] The iterative steps include:

[0144] A batch of vectors is randomly sampled from the source feature vector subset as positive samples, and an equal number of negative sample vectors are sampled from the historical normal pattern feature vector library to form the current batch of incremental learning data pairs.

[0145] The incremental learning data is used to embed the physical constraints of the input vibration source into the variational encoder of the model, and the reconstruction loss and physical constraint loss are calculated. A contrastive learning loss term is introduced, which makes the feature representations of positive samples in the latent space closer to each other and farther away from the feature representations of negative samples.

[0146] The reconstruction loss, physical constraint loss, and contrastive learning loss are weighted and summed to form the total loss function in the fine-tuning stage. The gradient descent algorithm is then used to update the network parameters of the variational encoder based on the total loss function.

[0147] like Figure 1 As shown in the figure, the vibration anomaly separation and identification method for wheel-rail systems under complex working conditions provided in this embodiment involves the following steps in its application process:

[0148] Step S1: Obtain multi-source heterogeneous temporal vibration signals of the wheel-rail system under complex working conditions, perform time-space-frequency three-dimensional feature mapping on the multi-source heterogeneous temporal vibration signals, and generate a three-dimensional dynamic feature tensor.

[0149] In this embodiment of the invention, a synchronous acquisition device for vibration acceleration, wheel-rail force, and train operation status signals is deployed in actual track sections containing curves, turnouts, and different wheel-rail interface conditions. Vibration acceleration signals are acquired from multiple axle boxes and frame measuring points; wheel-rail force is measured via force-measuring wheelsets or strain gauges; and operation status signals are read from the train control system, including real-time speed, position, and traction / braking status. All sensors are synchronized with a high-precision clock to ensure strict data timing alignment. The acquired multi-source signals are first subjected to spatiotemporal alignment and normalization. Based on the installation location of each sensor and known delays, time-shift compensation is performed on the signals to ensure that the peak force signal corresponds perfectly with the vibration impact on the time axis. Then, the measured values ​​in the local coordinate system are transformed to a unified vehicle coordinate system. Amplitude normalization is performed based on the standard deviation or upper limit of the range of each channel in the entire data stream, eliminating differences in physical dimensions and sensor sensitivity, and generating a synchronously normalized signal matrix. Recursive quantization encoding technology is applied to each signal channel of this matrix. The amplitude of the time-series signal in each channel is divided into several intervals and encoded into a symbol sequence. The transition probabilities between symbols are statistically analyzed to form a time-domain recursive feature vector characterizing the nonlinear dynamic characteristics of the signal. Adaptive wavelet ridge extraction is performed on the synchronously normalized signal matrix. The time-frequency energy distribution of the signal is obtained through continuous wavelet transform, and energy ridges on the time-frequency surface are extracted. These ridges describe the evolution trajectory of non-stationary frequency components, forming a time-frequency joint distribution feature matrix. Finally, the time-domain recursive feature vector, the time-frequency feature matrix, and the spatial location and operating condition labels in the operating status signal are fused and concatenated. The features of all time windows are stacked along the time dimension to form a three-dimensional array structure, namely a three-dimensional dynamic feature tensor. The three dimensions of this tensor correspond to time evolution, feature type, and eigenvalue, respectively. Its internal data simultaneously contains the time-domain dynamics, frequency-domain characteristics of the vibration signal, and its correlation with the operating conditions.

[0150] Step S2: Construct a physical constraint embedding model for vibration sources, extract potential vibration source feature vectors that satisfy the physical constraint conditions of the wheel-rail system from the three-dimensional dynamic feature tensor, and generate a set of physical constraint feature vectors;

[0151] In this embodiment of the invention, a vibration source physical constraint embedding model is constructed. The core of this model is a variational autoencoder, whose encoder maps a high-dimensional three-dimensional dynamic feature tensor to a low-dimensional latent space, while the decoder reconstructs the input from the latent space. The key innovation lies in the introduction of a physical constraint loss function into the training objective. First, the current wheel-rail contact geometry and system dynamic boundary conditions, such as curve radius and vehicle speed, are parsed from the working condition information associated with the input tensor. Based on this, the theoretical characteristic frequency range (such as the theoretical excitation frequency of wheel polygons and rail corrugation) and system transmission characteristics are calculated. These calculated values ​​constitute a "working condition-specific physical rule template." The physical constraint loss function is designed to include multiple sub-terms, including a "frequency matching loss" that encourages the frequency estimates corresponding to latent variables to be close to the theoretical frequency template; an "energy ratio loss" that constrains the energy ratio of different frequency bands based on the principle of energy conservation; and a "modal consistency loss" that ensures the vibration mode shape conforms to the theoretical mode shape. When training the variational autoencoder, its total loss function is a weighted sum of the standard reconstruction loss, the KL divergence regularization term, and the aforementioned physical constraint loss. By minimizing this total loss, the model is forced to embed the physical laws of the wheel-rail system while learning the compressed representation of the data. After model training, the feature tensors of all training data are input into the encoder, and low-dimensional latent variables are extracted from the bottleneck layer as preliminary feature vectors. Next, these feature vectors are subjected to a physical rule compliance check, and vectors with small deviations from the physical constraint loss calculation are selected to form a primary set. Finally, spectral clustering analysis is performed on this primary set to merge features generated by the same physical mechanism, and the centroid vector of each feature cluster is extracted to form the final deredundant "physical constraint feature vector set". Each vector in this set represents a potential, physically consistent vibration source feature.

[0152] Step S3: Based on the physical constraint feature vector set, drive the interpretable adversarial separation network to decouple vibrational components, generating a probability distribution matrix containing multiple independent vibrational components and the corresponding decoupling contribution spectrum, such as... Figure 9 As shown;

[0153] In this embodiment of the invention, an interpretable adversarial separation network is constructed by using a set of physical constraint feature vectors as input. This network consists of a generator and multiple discriminators. The generator receives a physical constraint feature vector and outputs a structured synthetic feature vector, which is internally divided into K consecutive data blocks. Each data block is intended to represent an independent vibration component (such as wheel polygons, rail corrugations, or random irregularities). Each discriminator is specifically responsible for determining whether the input feature data block "truly" belongs to its corresponding vibration component. Through adversarial training, the generator is trained to decouple and separate different components from the mixed input features into their respective output blocks, in order to "deceive" all discriminators as much as possible. Simultaneously, the generator's training objectives also include minimizing the reconstruction loss to ensure no loss of input information and minimizing the physical constraint consistency loss to ensure that the decoupled features still conform to physical laws. After the adversarial training stabilizes, the generator is frozen. The entire set of physical constraint feature vectors is then input back into the generator to obtain the corresponding synthetic features. For each synthetic feature's K data blocks, Softmax normalization is performed along the block dimension, converting the "activation intensity" of each block into a probability value. This yields a probability distribution matrix, where each row represents the probability that an original sample belongs to each independent vibration component. To interpret the model's decision-making process, the gradient of each component probability in the probability distribution matrix with respect to the original three-dimensional feature tensor is calculated, and a "decoupling contribution map" is generated based on the gradient magnitude. This map visually shows which time-frequency regions or operating condition features of the original signal contribute most to "separating a specific vibration component," thus providing interpretability for the model.

[0154] Step S4: Input the probability distribution matrix into the anomaly pattern adaptive clustering engine to identify and classify the anomaly pattern clusters in the separated independent vibration components, and generate an anomaly pattern label set;

[0155] In this embodiment of the invention, the probability distribution matrix is ​​treated as a set of points in a high-dimensional feature space and input into an adaptive clustering engine for abnormal patterns. This engine first employs an algorithm based on local density peaks and relative distances to automatically discover potential cluster centers in the data point set, without pre-setting the number of categories. These center points represent different vibration pattern prototypes naturally presented in the data. Based on the discovered cluster centers, a graph neural network is constructed as a dynamic similarity measurement network. This network takes an initial K-nearest neighbor graph and data point features as input, aggregates neighborhood information through graph convolutional layers, and learns to output a point pair similarity that better reflects the inherent manifold structure of the data. Using this learned dynamic similarity matrix, a spectral clustering algorithm is executed to divide all probability distribution points into different clusters, each cluster representing a specific vibration pattern. To determine which clusters are abnormal, the system needs to compare them with historical normal baselines. In the offline phase, the system has collected a large amount of historical normal operating condition data and, through steps S1 to S3, established a "normal pattern probability distribution database." This database is modeled using a Gaussian mixture model to generate a "normal pattern baseline distribution model" that characterizes the multi-peak distribution of normal patterns. For each newly identified vibration mode cluster, the KL divergence between its empirical probability distribution and the baseline model is calculated, serving as the cluster's "abnormal deviation index." This index quantifies the degree of difference between the mode and its historical normal state. The anomaly determination threshold is not fixed but dynamically retrieved from a pre-defined "operating condition-threshold mapping table" based on the specific operating condition and intensity of the current data. The deviation index of each cluster is compared with the dynamic threshold; clusters exceeding the threshold are identified as anomalous mode clusters. A composite label containing anomaly type code, deviation value, and index of major separating components is generated for each anomalous cluster; all labels are aggregated to form an "abnormal mode label set."

[0156] Step S5: Merge the abnormal mode label set with the decoupling contribution map to generate a vibration anomaly separation and identification report for wheel-rail system under complex working conditions, and feed the vibration anomaly separation and identification report of wheel-rail system to the vibration source physical constraint embedding model to optimize the physical constraint conditions.

[0157] In this embodiment of the invention, an abnormal pattern label set and a decoupling contribution map are received. First, a mapping relationship between the labels and the map is established, and the regions with high contribution of the samples associated with each abnormal label are highlighted on the decoupling contribution map. These regions are then superimposed with the time-frequency information reconstructed from the original three-dimensional feature tensor to generate a visual map that integrates abnormal location, time-frequency performance, and operating condition background. Combining the main component index and deviation in the abnormal labels, the physical source of the abnormality is analyzed, and the corresponding mechanism description and maintenance suggestion text are matched from the fault knowledge base. The visual map, physical analysis text, and key statistics of the original signal are integrated to automatically generate a complete vibration abnormality separation and identification report according to a structured template. The report is not only used to guide maintenance, but its core value lies in driving the system's self-evolution. The system checks whether the abnormal pattern in the report is a new type not covered in the knowledge base. If so, it backtracks the physical constraint feature vector corresponding to the abnormality to form a "new abnormality source feature subset". Based on this subset, the new physical constraint forms or parameters that it may satisfy are deduced in reverse. Using this subset as positive samples and historical normal features as negative samples, an incremental learning dataset is constructed. This dataset is then used to fine-tune the variational encoder embedded in the vibration source physical constraint model, introducing a contrastive learning loss to better encode new anomalous features. The inversely derived new physical constraints are formalized, added to the physical rule template library, and the generation logic of the template corresponding to the operating condition is updated. Thus, the entire system achieves continuous learning and autonomous evolution of its ability to identify unknown anomalous patterns through a closed loop of "identification-feedback-learning-update."

[0158] Furthermore, as an embodiment of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart of step S1 is shown below. In this embodiment, step S1 includes the following steps:

[0159] Step S11: Synchronously acquire multi-channel vibration acceleration signals, wheel-rail force signals, and operating status signals of the wheel-rail system under complex working conditions including curve sections, turnout areas, and different wheel-rail interface states, to form a multi-source heterogeneous time-series vibration signal stream;

[0160] Step S12: Perform spatiotemporal alignment and scale normalization on the multi-source heterogeneous temporal vibration signal stream to eliminate the temporal offset caused by sensor heterogeneity and operating condition switching, and generate a synchronous standardized signal matrix.

[0161] Step S13: For each signal channel in the synchronous normalized signal matrix, apply recursive quantum coding technology to encode the amplitude changes of the time-series signal into a symbol sequence, and calculate the symbol transition probability to generate the time-domain recursive dynamic feature vector of each channel.

[0162] Step S14: Perform adaptive wavelet ridge extraction on the synchronous normalized signal matrix to capture the time-varying law of non-stationary frequency components in the signal and generate a time-frequency joint distribution feature matrix, which characterizes the migration trajectory of vibration energy on the time-frequency surface.

[0163] Step S15: Based on the time-domain recursive dynamic feature vector and the time-frequency joint distribution feature matrix, the spatial location and operating condition label in the operating status signal are fused to construct a three-dimensional dynamic feature tensor. The three dimensions of this tensor represent the time evolution, frequency distribution and spatial-operating condition correlation characteristics, respectively.

[0164] In this embodiment of the invention, a multi-sensor synchronous acquisition array is deployed in typical complex operating sections containing small-radius curves, high-speed turnouts, and significant corrugation and wheel polygonal composite faults. This array includes vertical and lateral vibration acceleration sensors mounted on multiple bogie axle boxes to directly capture the dynamic response generated by wheel-rail interaction; a force-measuring wheelset system mounted on specific wheelsets to synchronously measure wheel-rail vertical and lateral forces; and an operating status monitoring unit integrated into the train control system to acquire real-time train speed, real-time position, track markings, and traction / braking status signals. All sensors and monitoring units achieve microsecond-level time synchronization through a high-precision synchronous clock module, ensuring strict correspondence of different physical quantities on the time axis. The sampling frequency of the acceleration signal is set to 5000 Hz, the force signal sampling frequency is 2500 Hz, and the operating status signal is recorded at a frequency of 100 Hz. The entire acquisition process covers the complete operating cycle from entering a curve from a straight section, passing a turnout, and then entering a straight section again, forming a multi-source heterogeneous temporal vibration signal stream containing time-domain waveforms, force changes, and operating parameters, such as... Figure 3 As shown, in the signal stream, the data from each channel are time-aligned, but there are significant differences in physical dimensions, numerical ranges, and dynamic characteristics. The acquired multi-source heterogeneous time-series vibration signal stream is processed. First, time alignment correction is performed. Using known sensor response delay and communication delay parameters, time shift compensation is applied to all channel data to ensure that the peak force signal and the characteristic impact of the vibration acceleration signal are perfectly aligned on the time axis. Then, spatial alignment is performed. Based on the installation positions of each sensor on the vehicle body and bogie, the measurements in the local coordinate system are transformed to a unified vehicle coordinate system. After completing the spatiotemporal alignment, the amplitude scale of each channel signal is normalized. The synchronized normalized signal matrix after spatiotemporal alignment and scale normalization is shown below. Figure 4 As shown.

[0165] For the vibration acceleration channel, calculate the standard deviation of the entire signal flow in that channel. Then, divide all data points of that channel by the standard deviation. The calculation formula is as follows:

[0166] ;

[0167] In the formula, For the vibration acceleration channel number Normalization result of discrete sampling points For the vibration acceleration channel number The values ​​of each discrete sampling point.

[0168] For wheel-rail force channels, the upper limit of their range is used. Normalization is performed, and the normalization process is represented by the following formula:

[0169] ;

[0170] In the formula, For the first wheel-rail force channel Normalization result of discrete sampling points For the first wheel-rail force channel The values ​​of each discrete sampling point.

[0171] For the speed in the operating status signal, the maximum speed designed for the line is used. Normalization is performed, as expressed by the following formula:

[0172] ;

[0173] In the formula, The first in the running status signal The normalized result of the velocity of each discrete sampling point. The first in the running status signal The speed of each discrete sampling point.

[0174] After the above processing, the effects of sensor sensitivity differences and physical dimensions are eliminated, generating a synchronized normalized signal matrix where all channel data are within a similar numerical range and are strictly synchronized in time and space. The rows of this matrix correspond to time series points, and the columns correspond to different signal channels. For example... Figure 5 As shown, for each single-channel time-series signal in the synchronization normalization signal matrix, recursive quantization coding technique is applied to extract its nonlinear dynamic characteristics. For a length of... time series , For the amplitude corresponding to the i-th sampling point in the time series, firstly, based on its amplitude distribution range, it is divided into... Each interval has an equal width of amplitude, and each interval corresponds to a symbol. The entire time series is converted into the corresponding symbolic sequence. ,in For the symbol sequence The sign corresponding to the amplitude range. Next, construct a... The sign transition probability matrix .matrix elements in The calculation method is to count the number of symbols in the sequence starting from the i-th symbol. Transfer to the j-th symbol The number of times, then divided by all from The total number of transitions from the starting point. Mathematically expressed as:

[0175] ;

[0176] In the formula, For the i-th symbol Transfer to the j-th symbol The transition probability, For the statistical symbol sequence starting from the i-th symbol Transfer to the j-th symbol Number of times, For all from Total number of departures.

[0177] This transition probability matrix The dynamic pattern of signal amplitude variation was characterized, and the amplitude distribution and temporal dependence of the signal were encoded. Finally, the transition probability matrix was... The upper triangular elements (including the diagonal) are expanded row by row to form a dimension of This is a one-dimensional vector, which is the time-domain recursive dynamic eigenvector of this channel. This process is repeated for all signal channels to obtain a set of time-domain recursive dynamic eigenvectors. Time-frequency analysis is performed on the synchronization-normalized signal matrix to capture non-stationary frequency components. Continuous wavelet transform is used as the analysis tool, and its definition is as follows: Signal at discrete sampling points for:

[0178] ;

[0179] in, To analyze wavelet basis functions, As a scale factor, The translation factor is... For signal In scale factor Translation factor Energy distribution on the constructed time-frequency plane By changing and The energy distribution of the signal in the time-frequency plane is obtained. To adapt to the time-varying frequency components in wheel-rail vibration signals, an adaptive wavelet ridge extraction algorithm is adopted. The algorithm first calculates the time-frequency amplitude spectrum. Then at each time point Above, searching in the scale direction Envoy The points that achieve local maxima form a time-frequency path called a wavelet ridge. For each ridge, at time point... The corresponding instantaneous frequency It can be done To calculate, where It is the wavelet center frequency. This is the sampling interval. For example... Figure 6 As shown, all extracted ridges and their corresponding time-frequency amplitudes together constitute a time-frequency joint distribution feature matrix. This matrix not only contains the energy information of a traditional time-frequency plot, but also highlights the time-varying trajectory of the dominant non-stationary frequency component in the signal through the ridges, i.e., the migration trajectory of vibration energy on the time-frequency surface. The generated time-domain recursive dynamic feature vector and the time-frequency joint distribution feature matrix are fused with the spatial location and operating condition labels in the operating state signal to construct a unified three-dimensional dynamic feature tensor, such as... Figure 7 As shown. First, the time-domain recursive dynamic eigenvectors corresponding to each time window are arranged into a two-dimensional matrix, with rows corresponding to different signal channels and columns corresponding to various dimensions of the eigenvectors. Simultaneously, the time-frequency joint distribution eigenvector matrix corresponding to the time window is also included as a part. A three-dimensional data structure is defined. The first dimension of the tensor is the time index, corresponding to different analysis time windows. The second dimension is the feature type index, which is divided into multiple blocks, storing in sequence: recursive quantized feature blocks for all channels, time-frequency ridge amplitude feature blocks, and numerical label blocks extracted from the operating state signals (such as normalized velocity, one-hot encoding of curvature indicators, switch zone indicators, etc.). The third dimension is the eigenvalue. Specifically, for each time window... The recursive dynamic feature vectors of all corresponding channels in the time domain are concatenated and placed at the beginning of the feature type dimension; the time-frequency joint distribution feature matrix of the time window (reconstructed into a vector) is placed in the middle; and the running state parameter vector is placed at the end. Finally, these concatenated vectors of all time windows are stacked to form a three-dimensional array, namely the three-dimensional dynamic feature tensor. , ,in It is the number of time windows. It is the total dimension of feature types. Typically, it is 1 (representing a scalar eigenvalue). This tensor organically integrates the signal's time-domain dynamics, frequency-domain evolution, and its correlation with spatial location and operating conditions.

[0180] Step S2 includes the following steps:

[0181] Step S21: From the three-dimensional dynamic feature tensor, based on the running state signal, parse out the wheel-rail contact geometric constraints and system dynamic boundary conditions under the current working condition, and generate the physical rule template corresponding to the working condition;

[0182] Step S22: Construct a physical constraint loss function that includes vibration energy conservation, momentum transfer and structural modal coupling relationships, and instantiate the parameters in the physical constraint loss function using the physical rule template corresponding to the working condition;

[0183] Step S23: Input the three-dimensional dynamic feature tensor into the pre-trained variational encoder. The training objective of the variational encoder is to minimize the physical constraint loss function while reconstructing the input features, thereby forcing the encoder to learn feature representations that satisfy physical laws in the latent space.

[0184] Step S24: Extract feature vectors from the bottleneck layer of the variational encoder, and filter out feature vectors whose deviation from the physical constraint loss function calculation is lower than a preset threshold through physical rule compliance test to form a primary physical constraint feature vector set;

[0185] Step S25: Perform spectral clustering analysis on the primary physical constraint feature vector set, merge feature clusters generated by the same physical mechanism, and extract the centroid vector of each feature cluster to finally generate a redundancy-free physical constraint feature vector set. Each vector in this set represents a potential vibration source feature that conforms to physical laws.

[0186] In this embodiment of the invention, the specific operating parameters corresponding to the current data segment are parsed from the running state information block associated with the three-dimensional dynamic feature tensor, and a quantized physical rule template is generated accordingly. First, the instantaneous velocity v is extracted. Then, based on the line position information, it is determined whether it is on a curve segment; if so, the curve radius is extracted. Determine if the area is a turnout zone and identify the turnout type. Based on these parameters, calculate the theoretical characteristic frequency. For example, the theoretical characteristic frequency of polygonal wheel excitation. The calculation formula is: ,in For the order of the polygon of interest, Given the wheel radius. Rail corrugation excitation frequency. The relationship with wavelength is ,in The typical wave-erosion wavelength range of interest is defined. These theoretical frequencies and their bandwidths constitute the frequency constraint rules. Meanwhile, based on a simplified vehicle-track coupled dynamics model, at a given speed... curve radius Primary suspension stiffness Unsprung mass Under the same parameters, the approximate transfer functions of the system in the vertical and horizontal directions can be calculated. And the theoretical response amplitude ratio at certain characteristic frequencies. These are sets of theoretical frequencies instantiated from operating parameters. and theoretical transfer properties , Let i be the excitation frequency of the i-th rail corrugation. The theoretical characteristic frequencies of the polygonal vibration of the j-th wheel together constitute the physical rule template corresponding to the current working condition. This template provides a specific, computable reference standard for subsequent physical constraints. A composite physical constraint loss function is then constructed. This function constrains the model's learning process from multiple physical principles. The loss function consists of three main parts:

[0187] 1) Frequency matching loss : Encourage the dominant frequency components of features learned from data to be close to the theoretical feature frequencies in the physical rule template. If the frequencies estimated from the features are Then we have:

[0188] ;

[0189] In the formula, For the i-th estimated frequency, This is the template for the j-th physical rule.

[0190] 2) Energy proportion loss Based on the law of conservation of energy, within a specific frequency band, the energy contributed by different vibration sources (such as wheels and rails) should satisfy a certain proportional relationship. This proportion can be derived from the transfer function in the physical rule template. It was estimated.

[0191] 3) Modal mode consistency loss For the overall modal analysis of the bogie or car body, the phase relationship of vibrations at different measuring points should conform to a specific mode shape. The physical constraint loss function is defined as a weighted sum of these terms:

[0192] ;

[0193] in, The physical constraint loss function, for , and Corresponding weighting coefficients, weighting coefficients Based on the prior confidence levels of each constraint, the generated physical rule templates corresponding to the operating conditions are used. ,right Physical rule templates in Parameters such as theoretical energy ratio and theoretical mode shape are assigned specific values, thereby transforming general physical constraints into instantiated physical constraint loss functions specific to the current working condition. A variational autoencoder is constructed as the core of the vibration source physical constraint embedding model. High-dimensional three-dimensional dynamic feature tensor Mapped to a low-dimensional latent space, the mean of the output latent vectors is given. and variance Decoder Sampling from potential space , For potential vectors, Given a diagonal covariance matrix, reconstruct the input tensor. .

[0194] The standard training objective of VAE is to maximize the lower bound of evidence ELBO:

[0195] ;

[0196] in, As the lower bound of evidence, It is a standard normal prior. To provide a given original input Post-latent variables The posterior distribution, In distribution The expectations below Given latent variables Post-raw input The likelihood, Let KL divergence be denoted as KL divergence.

[0197] To embed physical constraints, an instantiated physical constraint loss function is added to ELBO. This forms the total loss function:

[0198] ;

[0199] in, It is the reconstruction loss. It is the KL divergence regularization term. and yes and The tradeoff coefficient.

[0200] During training, the three-dimensional dynamic feature tensor Inputting the network, forward propagation computes the reconstructed tensor. and potential distribution parameters and physical constraint loss By employing backpropagation and gradient descent algorithms, not only is the reconstruction error and the deviation of the latent distribution from the prior minimized, but the minimum value is also forced to be minimized. This process forces the encoder to learn latent feature representations. It is necessary to balance data reconstruction with the satisfaction of physical rules, so as to "embed" physical constraints into feature representation and make the feature encoding of latent space conform to physical laws.

[0201] After the variational encoder is trained, it is used for feature extraction. The 3D dynamic feature tensor of all training data is input into the encoder. Obtain the latent feature vector of each sample from the bottleneck layer. (Usually the average is taken) The set of feature vectors obtained at this time. It's preliminary. Let be the i-th potential feature vector in the feature vector set, which may contain some noise or pseudo-features that do not conform to physical laws.

[0202] For the selection process, a physical rule compliance check is required. For each potential feature vector... The input is then used to obtain the reconstructed features from the decoder, or through an auxiliary mapping network, the corresponding interpretable physical quantity, such as the dominant frequency, is derived. Energy distribution Then, these derived physical quantities are compared with the physical rule template corresponding to the current sample. The physical constraint loss value of the sample is calculated. Set a deviation threshold. The filtering criteria are: Only eigenvectors that satisfy this condition The eigenvectors that are retained constitute the primary physical constraint eigenvector set. , This step ensures that the features in the set have a high degree of consistency with the known physical laws of wheel-rail systems. Primary Physical Constraint Feature Vector Set There may be redundancy, meaning that multiple eigenvectors actually originate from the same physical vibration source, only differing slightly in amplitude or phase.

[0203] In order to remove redundancy and extract the essential features representing different physical vibration sources, Perform spectral clustering analysis. First, calculate the pairwise similarity between all feature vectors in the set and construct a similarity matrix. ,in , Similarity matrix The similarity between the i-th latent feature vector and the j-th latent feature vector. It is an exponential function. for and Euclidean distance, It is a scale parameter.

[0204] Next, the degree matrix is ​​calculated. It is a diagonal matrix, and the i-th diagonal element .

[0205] Then, calculate the normalized Laplacian matrix. , .right Perform eigenvalue decomposition and select its first eigenvalue. The eigenvectors corresponding to the smallest non-zero eigenvalues ​​form a new feature matrix. , ,in That's the number of samples. For this new feature matrix... The row vectors (i.e., the coordinates of each sample in the new feature space) are used to apply the K-means clustering algorithm to divide the samples into groups. Number of clusters. This can be determined by observing the inflection points of the eigenvalues ​​of the Laplacian matrix. After clustering, all eigenvectors belonging to the same cluster are considered to originate from the same physical mechanism. For each cluster, the centroids (mean vectors) of all its eigenvectors are calculated. These centroid vectors constitute a matrix such as... Figure 8 The set of physical constraint feature vectors after redundancy removal shown , , Let be the eigenvector of the i-th physical constraint. (Set) Each centroid vector in This represents the essential characteristics of a potential independent vibration source learned from data, conforming to physical laws, and providing a clear and non-redundant input for subsequent decoupling of vibration components.

[0206] Furthermore, step S3 includes the following steps:

[0207] Step S31: Using the physical constraint feature vector set as input, construct an interpretable adversarial separation network architecture containing a generator network and multiple discriminator networks, wherein each discriminator corresponds to a potential vibration component to be separated;

[0208] Step S32: The generator network reconstructs a synthetic feature tensor that is consistent with the original three-dimensional dynamic feature tensor in terms of physical constraint loss but whose dimensions are separable, based on the input physical constraint feature vector.

[0209] Step S33: Each discriminator network focuses on determining whether the input synthetic feature tensor contains its corresponding vibrational component, and forces the generator to separate vibrational features of different physical mechanisms into different dimensions or channels of the synthetic tensor through adversarial training;

[0210] Step S34: After the adversarial training is stabilized, freeze the generator network, input the physical constraint feature vector set again to generate the corresponding synthetic feature tensor, perform channel-level Softmax normalization on the synthetic feature tensor to obtain the probability that each potential vibration source feature vector belongs to each separated vibration component, and form a probability distribution matrix.

[0211] Step S35: Calculate the gradient of the probability value of each separated vibration component in the probability distribution matrix with respect to each local feature in the original three-dimensional dynamic feature tensor, and generate a decoupling contribution map based on the gradient value.

[0212] In this embodiment of the invention, the final generated physical constraint feature vector set is used... As the core input, a specially designed interpretable adversarial segregation network is constructed. This network architecture includes a generator network. and A discriminator network , Let i be the discriminator network, where This represents the number of potential independent vibrational components to be separated, either predefined or determined by prior knowledge; for example, it can be set... =3, corresponding to "periodic polygonal wear component of wheels", "periodic corrugation component of rails", and "random irregularities and impact components of rails", respectively. Generator It is a deep feedforward neural network whose input layer dimension is the same as the i-th physical constraint feature vector. The dimensions are consistent, while the output layer dimension is designed to be... ,in This is the dimension of the synthetic feature subspace. This means that the generator's output is structurally divided into... Each data block corresponds to a synthetic feature representation of a potential independent component. Each discriminator... It is also a neural network whose input dimension is the same as the dimension of a single data block output by the generator (i.e., The output is a scalar representing the discriminator's assessment that its input feature belongs to the "true dimensionality". The probability of "vibrational components". This "one generator to multiple discriminators" architecture is the foundation for achieving multi-channel adversarial separation. The generator's goal is to "deceive" all discriminators, causing their output features to be judged as true features of the corresponding category; while each discriminator's goal is to accurately distinguish between the features synthesized by the generator and the feature prototypes extracted from real data that belong to that category. During the training phase, the generator network... Receive a physical constraint feature vector As input. The forward propagation process of the generator is to take... Mapped to a dimensional synthetic eigenvectors , To meet the requirement of "consistency with the original 3D dynamic feature tensor in terms of physical constraint loss," a physical constraint consistency loss term must be added to the generator's training objective, in addition to the adversarial loss. The specific implementation involves: synthesizing the feature vector output by the generator. This is achieved through a pre-trained, lightweight "feature-physical quantity inverse mapping network". Mapping back to a space of computable physical quantities, such as estimating dominant frequencies. and energy distribution Then, these estimates are compared with... Physical rule template for corresponding samples Physical constraint loss between Its calculation method is different from the previously defined Similarly, this loss term forces the features output by the generator, even after separation transformation, to maintain consistency with the physical constraints of the original data in their implicit physical properties (frequency, energy). This prevents the generator from generating physically meaningless features simply to deceive the discriminator. Simultaneously, the generator must also minimize the reconstruction loss. It is expressed by the following formula:

[0213] ,in It is an aggregate function (such as summation or averaging) used to... The separated feature blocks are re-aggregated into one that is consistent with the input. A vector of the same dimension. This loss ensures that the generation process does not lose key information from the input features. Therefore, the generator's output... It is data that is identical to the original data in terms of physical properties, but is explicitly divided into different structural categories. The synthetic feature representation of each part (dimension or channel) provides the structural foundation for subsequent adversarial separation. First, it is necessary to perform feature analysis on each discriminator. Prepare "real samples". This is done by extracting physical constraint feature vector sets. In this process, based on prior knowledge or auxiliary clustering results, a small group of clusters is manually or semi-automatically selected that is believed to be mainly composed of the first... Samples dominated by vibrational components constitute a small set of real samples. In each training iteration, for the Each discriminator's training data consists of two parts: one part comes from... real samples The other part is the generator. Current input for random batch The resulting synthesized output The Data blocks (Right now The corresponding number in the middle Each component (Vacuum slices).

[0214] Discriminator The training objective is to maximize its ability to distinguish between real and generated samples, and its loss function adopts the standard binary cross-entropy form: , For the first The loss function of a discriminator.

[0215] generator Compared to the first The adversarial loss of the individual discriminator aims to ensure that the corresponding data blocks it generates are indistinguishable from the real ones.

[0216] , For generator Compared to the first The adversarial loss of the discriminator.

[0217] The total adversarial loss of the generator is The sum of the losses corresponding to each discriminator: , This represents the total adversarial loss of the generator.

[0218] The generator's complete training loss is a weighted sum of the reconstruction loss, the physical constraint consistency loss, and the total adversarial loss: .

[0219] This is the complete training loss for the generator.

[0220] By alternately optimizing the generator (minimizing) ) and each discriminator (minimizing their respective The generator is forced to input physical features The mixed information contained within is decoupled and encoded separately into the output. Different data blocks In this process, each data block focuses on mimicking the characteristic patterns of a specific type of vibrational component to meet the "realism" requirement of the corresponding discriminator. When the adversarial training process reaches a stable state—that is, when the loss function values ​​of the generator and discriminator no longer exhibit trend changes and oscillate over multiple training cycles—the generator network is frozen. The parameters. Then, the entire physical constraint feature vector set... All vectors The corresponding synthetic feature vector sets are obtained by sequentially inputting the frozen generator. For each synthesized feature vector It is structurally viewed as indivual Concatenation of 3D vectors: To quantize each physical constraint feature vector In the mixed vibrational state represented, the "proportion" or "attribution probability" of each independent component, for each Perform channel-level Softmax normalization. Specifically, this involves: ... of Each data block is considered as The unnormalized scores (logits) of each "channel" or "category". Then, along this... The dimension consisting of data blocks is applied using the Softmax function.

[0221] For the Each data block has a corresponding probability value. The calculation is as follows: .here, yes No. A summary scalar of data blocks, typically taken as... The norm (such as L2 norm) or mean of a dimensional vector. For iterating over variables.

[0222] This calculation process ensures that for each input... They all ,and ∈[0,1]. The probability vector of all samples. Arranged in rows, forming a probability distribution matrix ,in yes The number of eigenvectors in the matrix. The Line 1 Column elements , indicating the first The physical constraint feature vector (corresponding to the vibration state of a certain time window) belongs to the th The probability of each independent vibration component. To understand the probability distribution matrix... Interpretability, that is, understanding which specific features in the original input data drive the final probability result, requires calculating the gradient of the probability value with respect to the original input. The original input here refers to the generated three-dimensional dynamic feature tensor. In, with physical constraint feature vectors The corresponding feature slice of that local time window The probability distribution matrix is ​​calculated using automatic differentiation techniques. Each element (No. The sample belongs to the first The probability of each component is relative to the original input features. The gradient. The resulting gradient tensor is denoted as... Its dimensions and same, This is for tensor shrinking operations. The gradient tensor specifies the original features. Even a tiny change in each tiny feature unit can lead to a change in the assignment probability. To what extent does the change occur? The location with the largest absolute gradient value corresponds to the original feature unit that judges the sample. Belongs to ingredients "This decision made a significant contribution. For the first..." For each vibrational component, the mean or weighted sum of the absolute values ​​of the gradients of all samples (or a subset of samples belonging to that component with a high probability) can be calculated to obtain an aggregated contribution map. , ,in This is the number of samples participating in the aggregation. This graph... With a structure similar to the original feature tensor slices, it intuitively and quantitatively reveals which specific regions, frequency bands, or combinations of operating conditions in the time domain, frequency domain, and space-condition correlation features are important for "separating the first" feature. The "vibrational component" played a key role, thereby enabling the visualization and interpretation of the separation process.

[0223] Furthermore, step S34 includes the following steps:

[0224] During the stable phase of adversarial training, the internal neuron patterns activated by the generator network to reconstruct each physical constraint feature vector are recorded, generating a pattern activation map.

[0225] The physical constraint feature vector set is input into the frozen generator, and a probe layer is inserted into the middle layer of the network to capture the intermediate representation sequence during the feature transformation process. Attention weights are assigned to the intermediate representation sequence, and the importance score of each position in the sequence to each channel of the final output synthetic feature tensor is calculated to generate an attention-guided channel association weight matrix.

[0226] Based on the channel association weight matrix, the synthesized feature tensor output by the generator is weighted, fused, and recombined into channels to enhance the independence between different channels.

[0227] The recombined synthetic feature tensor is subjected to a Softmax operation along the channel dimension to convert the response value of each channel into a probability form, thereby obtaining the probability distribution of each vector in the physical constraint feature vector set corresponding to each separate vibration component, and finally integrating them into a probability distribution matrix.

[0228] In this embodiment of the invention, during the stable phase of adversarial training, in addition to monitoring the loss, the generator network is simultaneously recorded. The internal neuron activation states when processing a batch of representative physical constraint feature vectors. Specifically, the outputs of several convolutional or fully connected layers in the generator are selected as "probe layers". When a feature vector is input... During forward propagation, the activation values ​​of all neurons in these probe layers are recorded. For each probe layer, the activation values ​​of all its neurons are arranged according to their spatial location (for convolutional layers) or channel order to form a two-dimensional activation map. Then, for the final output of the generator... For each data block, calculate its gradient with respect to the activation map of that layer during backpropagation. Using the basic idea of ​​class activation mapping, for each output data block... Calculate its gradient relative to the activation map of the probe layer, average the gradient along the channel dimension, and then perform a weighted combination with the original activation map to generate a result for the probe layer. The "pattern activation map" of each output component at this layer. This graph visually shows which neuron patterns within the generator network contribute to generating the representative... The feature blocks of each vibrational component play a crucial role. Averaging the mode activation maps of all representative samples yields a set of stable mode maps reflecting the internal working mechanism of the generator. In the generator network... After training is complete and the network is frozen, non-parametric probe layers are formally inserted at several key locations in the network structure (such as after each residual block and before the pooling layer) to analyze the feature transformation path in greater detail. These probe layers do not participate in training; they only serve as data recording points. When the physically constrained feature vector set is... When inputting into the network, each input During forward propagation, an intermediate feature representation is generated at each probe layer. Assume that a total of [number] features are set in the generator. With one probe layer, then for the input... This will capture an intermediate representation sequence. : ,in Indicates the first The output of each probe layer. This sequence fully characterizes the input features. The "physically constrained hybrid representation" is gradually decoupled and transformed into the final, structured synthetic features. The complete path. For the captured intermediate representation sequence. An attention mechanism is applied to quantify the importance of intermediate representations at different locations (different network depths) in the sequence to the final output channels. Specifically, the final synthesized features... of Each data block is used as The first "query". For the first One query Calculate its relationship with the sequence All intermediate representations Attention score. First, through a learnable linear transformation matrix. and , will query and all keys Projected onto the query vector respectively and key vector : , Then calculate the scaled dot product attention score. : ,in This is the dimension of the key vector. Next, for all probe layer positions... The scores are then Softmax normalized to obtain the attention weights. , This weight Quantified the first The intermediate representations are used to form the final... The importance of each output component is determined. After calculating for all samples, an attention-guided channel association weight tensor can be generated, whose elements roughly reflect the association strength patterns from different network depths to different output components. This is then used in conjunction with the attention weight matrix generated in the previous step. (or some kind of statistical summary thereof, such as a matrix obtained by averaging the attention weights over all samples) The original features of the generator are obtained by directly linearly transforming the last hidden layer. The process can be "intervened" in and "enhanced." One approach is to no longer simply use the output of the last hidden layer, but instead use the intermediate representation sequence. Perform a weighted summation to obtain a context-aware aggregated representation. : ,here Indicates to Each output channel is in the layer The attention weights are treated as a weight vector. Then, this aggregated representation... Input a lightweight reassembly network (e.g., a small multilayer perceptron), which is explicitly trained to maximize its output. The independence between channels can be achieved, for example, by introducing a decorrelation loss based on covariance. , ,in It is the output of the reorganized network. Characteristics of each channel It is the covariance matrix. It is the identity matrix. It uses the Frobenius norm. The output of this recombined network undergoes independent feature enhancement. This process utilizes attention mechanisms to identify key intermediate features and further "cleans" and strengthens the separability between different output channels through forced decorrelation. The resulting synthetic feature tensor, after weighted fusion and channel recombination, exhibits enhanced independence. Channel-level Softmax normalization is then performed to generate the final probability distribution. The processing is similar to the previous steps, but the input is... .calculate Each channel Summary scalar (e.g., L2 norm), and then calculate the k-th attribution probability vector. : This yields the set of attribution probability vectors. The probability calculated based on enhanced features theoretically offers higher interpretability and separation clarity than the probability calculated directly using the original generator output. The final probability vectors of all samples are aggregated to form the improved probability distribution matrix for subsequent anomaly pattern recognition. .

[0229] Furthermore, step S4 includes the following steps:

[0230] Step S41: Receive the probability distribution matrix and treat it as a set of points in a high-dimensional feature space, where each point represents the coordinates of a vibration sample in the separated component space;

[0231] Step S42: Initialize the abnormal pattern adaptive clustering engine, and use a cluster center self-discovery algorithm based on local density peaks and relative distance to automatically identify potential cluster center points from the probability distribution matrix without pre-setting the number of clusters;

[0232] Step S43: Based on the discovered cluster centers, construct a dynamic similarity measurement network based on graph neural networks. This network learns the similarity between point pairs and updates dynamically to reflect the manifold structure of abnormal patterns under complex working conditions.

[0233] Step S44: Using the similarity matrix calculated by the dynamic similarity measurement network, execute the spectral clustering algorithm to divide the points in the probability distribution matrix into different clusters, each cluster representing a specific vibration mode;

[0234] Step S45: Based on the probability distribution of vibration components under historical normal operating conditions, establish a normal mode baseline distribution model, perform KL divergence calculation on each newly divided vibration mode cluster and the normal mode baseline distribution model, mark the mode cluster with divergence value exceeding the dynamic threshold as an abnormal mode, assign a unique abnormal mode label, and finally generate an abnormal mode label set.

[0235] In this embodiment of the invention, the final generated probability distribution matrix is ​​received. The matrix has dimensions of ,in It is the total number of vibration samples. It represents the number of separate, independent vibrational components. Each row of the matrix, i.e., one... 3D probability vector , ,satisfy This vector represents the first... Each sample (corresponding to the vibration state of a specific time window) is in the... The coordinates are in a "separated component space" spanned by individual vibrational components. This space is... A standard simplex of dimension 1. The position of each sample point in this space is determined by the probability of it belonging to each component. For example, a point highly concentrated in the "wheel polygon" component (e.g., probability vector [0.9, 0.05, 0.05]) is located in the region of the simplex near the vertex of the "wheel polygon"; a point with significant probabilities in both the "ripple" and "random irregularity" components is located somewhere along the line connecting these two vertices. The entire probability distribution matrix... Consider it as part of this high-dimensional probability space The data points form a point set, and subsequent clustering and anomaly identification are performed within this feature space, which has a clear physical semantics. The anomaly pattern adaptive clustering engine is initialized using an algorithm based on local density peaks and relative distances to automatically identify potential cluster centers from the point set, without needing to pre-specify the number of clusters. The process consists of three steps. The first step is to calculate the value for each data point. Local density A common method for calculating local density is to use a truncated kernel or a Gaussian kernel: ,in It is a point and The Euclidean distance between them It is a cutoff distance, usually taken as the distance value of the top 2% of all point-to-point distances after sorting them from smallest to largest. Reflects the point The density of neighboring points. The second step is to calculate the density of each point. Minimum distance to all points with higher density than it This is called the relative distance: For the point with the highest density, its Defined as the maximum distance to all other points. A larger point indicates that it is a relatively isolated point within its local high-density region, possibly a boundary point between different clusters or the center of a cluster. The third step is to identify cluster centers. Typically, cluster centers also have high local density. and a large relative distance Therefore, a decision value is calculated for each point. , Click all of them. Values ​​are sorted in descending order. Decision values A clear "inflection point" will emerge, and the points preceding this inflection point are selected as candidate cluster centers. Another visualization method is to plot... x-axis The decision graph uses the vertical axis as the ordinate, located in the upper right corner of the graph (i.e., the height). ,high The points are determined manually or through heuristic rules (such as setting...). and Cluster centers are selected using a dual threshold (using a double threshold). Assume that C cluster centers are identified through this process. This number C represents the number of potential vibrational pattern categories revealed by the data itself. Based on the discovered C cluster centers, a dynamic similarity metric network based on a graph neural network is constructed to learn similarities that better reflect the inherent manifold structure of the data. First, using all... Using 10 data points as nodes, construct an initial... Nearest neighbor graph. For each node Find the closest Euclidean distance. Given 10 neighboring nodes, establish undirected edges between them, with initial weights for each edge. It can be set as a distance-based Gaussian similarity: (if yes of Neighbor or yes of (nearest neighbor, otherwise 0), where This is the scale parameter. This graph reflects the initial local geometry of the data. Then, a graph neural network is constructed. The input to this network is the features of each node (i.e., its probability vector). The network consists of multiple stacked graph convolutional layers. The operation of a single graph convolutional layer can be represented as: ,in It is the first The node feature matrix of the layer It is a normalized adjacency matrix with self-loops. , It is an adjacency matrix. It is the identity matrix. (It is a degree matrix) It is a learnable weight matrix. It is a non-linear activation function. Through multi-layer graph convolution, the features of each node are aggregated with information from its multi-hop neighbors, ultimately outputting a context-aware embedding representation for each node. The goal of a dynamic similarity metric network is to learn a function. , making This approach more accurately reflects the similarity of two points on the vibration pattern manifold under complex operating conditions. During training, discovered cluster centers are used as weak supervision signals. One approach is to design a loss function that encourages higher similarity between node pairs belonging to the same cluster (an initial cluster assigned by the center point through nearest neighbors) and lower similarity between node pairs not belonging to the same cluster. Contrastive loss, such as a variant of the InfoNCE loss, can be used. After training, the network can dynamically update the similarity metric based on the overall structure of the data, capturing complex patterns better than fixed Euclidean distance or Gaussian kernels. The final similarity matrix between all point pairs is calculated using the trained dynamic similarity metric network. ,in , for and similarity, Let be the similarity function. Then, based on this learned similarity matrix... The spectral clustering algorithm is executed to divide the point set into C clusters. The steps of spectral clustering are as follows.

[0236] The first step is to construct a similarity matrix. (Received).

[0237] The second step is to calculate the degree matrix. It is a diagonal matrix, with diagonal elements .

[0238] The third step is to calculate the normalized Laplacian matrix. , .

[0239] Step 4, Calculation The first C smallest eigenvalues ​​and their corresponding eigenvectors Arrange these feature vectors column by column to form a Feature matrix , .

[0240] Fifth step, convert the matrix Each row is considered a point in a C-dimensional space. These M points are then divided into C clusters using the K-means clustering algorithm. The number of cluster centers in the K-means clustering is set to C. After K-means clustering, the original probability space point set is divided into C distinct clusters. The sample points within each cluster have similar probability distribution patterns measured by a dynamic similarity network; therefore, each cluster represents a specific, potential vibration pattern (which could be a normal pattern or a specific abnormal pattern). After obtaining C vibration pattern clusters through spectral clustering, it is necessary to determine which clusters belong to abnormal patterns. This is achieved by comparing each newly discovered cluster with a baseline distribution model built from historical normal data. First, a large number of vibration samples accumulated during long-term historical operation under recognized normal operating conditions (such as new wheel and track status, smooth straight-line operation, and no abnormal alarm records) are collected to obtain their corresponding probability distribution vectors, forming a large database of normal pattern probability distributions. , This database is used to establish a baseline probability distribution model of normal vibration modes in the separated component space. Since multiple steady states exist even under normal operating conditions, and their probability distribution is multimodal, a Gaussian mixture model is used to analyze them. Modeling is performed. The probability density function of the GMM is: ,in The number of mixed components is determined by minimizing the Bayesian information criterion. It is the first The mixing weights of the Gaussian components satisfy... ; and They are the first The mean vector and covariance matrix of each Gaussian component. Model parameters. From the expectation maximization algorithm The value is estimated in the middle. This GMM model Q characterizes the probability density distribution of all possible vibration modes of the system in a healthy state. Secondly, for each newly divided mode cluster... (c=1,...,C), calculate the empirical probability distribution of this cluster. One approach is to discretize the probability space, or more directly, to divide the clusters... The probability vector of all sample points in the distribution is considered as derived from the distribution. Independent and identically distributed samples were drawn from the sample. Then, the following calculations were performed. Compared with the baseline distribution Kullback-Leibler divergence between As a quantitative indicator of the cluster's deviation from the normal pattern: In actual calculations, because and These can all be quite complex, and Monte Carlo integration is usually used as an approximation: ,in It uses kernel density estimation based on intra-cluster sample pairs Point probability density estimation. The value is non-negative; the larger the value, the larger the pattern cluster. With historical normal distribution The greater the difference, the higher the likelihood of it being an abnormal pattern. A threshold is needed to determine if it's an anomaly. This threshold is not fixed, but dynamically determined based on the "complex working condition type and intensity" corresponding to the current data segment being analyzed. A working condition-threshold mapping table is established, defining the recommended KL divergence thresholds for different working condition combinations (e.g., "high-speed straight line," "low-speed curve," "passing through a turnout") and different environmental intensities (e.g., "high altitude cold," "humidity"). For example, under the "high-speed straight-line" operating condition, it is abnormally sensitive to corrugation, and the threshold τ is set relatively low; under the "low-speed passing through turnout" operating condition, some impact modes are normal, and the threshold τ is set relatively low. Adjust accordingly. Finally, increase the KL divergence value for each cluster. With dynamically retrieved thresholds Compare. If Then the cluster The clusters are labeled as anomalous patterns. For each labeled anomalous cluster, a structured composite label is generated. This label contains the following fields: 1) Anomaly type encoding: determined based on the index of the dominant component in the probability vector for this cluster of samples, such as "ABN_TYPE_CORR" (wave-dominant anomaly). 2) Anomaly deviation: i.e., the calculated deviation. 3) Key Separation Component Index: Indicates the physical excitation component that the anomaly primarily manifests, such as "MAIN_COMP_2" (corresponding to the "rail corrugation" component). 4) Temporal range and spatial location of samples within a cluster. The labels of all anomaly clusters constitute the final anomaly pattern label set. , ,in This represents the total number of identified anomalous clusters. This set of tags provides a structured and quantitative basis for subsequent fault location, report generation, and maintenance decisions.

[0241] Furthermore, step S45 includes the following steps:

[0242] Collect and store the probability distribution data of all vibration components obtained through steps S1 to S3 under historical normal operating conditions to form a normal mode probability distribution database;

[0243] Gaussian mixture model is used to model the normal mode probability distribution database to characterize the multi-peak probability density distribution of normal vibration modes in the separated component space and generate a normal mode baseline distribution model; the probability distribution of all sample points in each vibration mode cluster to be identified is extracted and the empirical probability distribution of the cluster is calculated.

[0244] Calculate the KL divergence between the empirical probability distribution of each vibration mode cluster and the normal mode baseline distribution model, and use it as an index of the abnormal deviation of the mode cluster.

[0245] Based on the type and intensity of the current complex working conditions, the anomaly judgment threshold of KL divergence is dynamically retrieved and determined from the working condition-threshold mapping table. Pattern clusters whose anomaly deviation index exceeds the dynamic threshold are judged as anomalies. A composite label containing anomaly type code, deviation degree and main separation component index is generated for each anomaly cluster, and the labels are summarized to form an anomaly pattern label set.

[0246] In this embodiment of the invention, segments representing the health status of the system are rigorously selected from long-term accumulated operational data. The selection criteria include: no known fault alarms recorded in the wheel-rail system during the specified time period; wear and tear on key vehicle and track components within maintenance standard limits; stable train operation on straight or large-radius curve sections; and no interference from extreme environmental factors. For these selected normal operating condition data segments, the complete processing flow from steps S1 to S3 is executed sequentially. Step S1 performs time-frequency analysis on the original vibration signal and constructs a three-dimensional feature tensor. Step S2 extracts physical constraint feature vectors from the latent space of the tensor using a physical constraint embedding model. Step S3 uses an interpretable adversarial separation network to decouple these feature vectors, generating a probability distribution vector corresponding to each time window. This vector describes the probability of vibration energy at that moment belonging to different physical excitation sources. All these probability distribution vectors calculated from the normal data, along with their corresponding metadata such as operating speed, track location, and acquisition time, are stored in a structured database. This database is the "Normal Mode Probability Distribution Database," which defines the range of vibration modes in the "Separated Component Space" under healthy conditions. A Gaussian mixture model is used to model all probability distribution vectors in this database. Since normal operating conditions themselves contain multiple states, their probability distributions will form multiple clustered regions in space. The Gaussian mixture model can effectively describe this multimodal distribution characteristic. The modeling process automatically determines the number of mixture components, which is determined by the model selection criteria. One approach is to try fitting the data with different numbers of mixture components and calculate the Bayesian information criterion value for each model, selecting the model order that minimizes this value. After determining the order, the model parameters are estimated using the expectation-maximization algorithm, including the mean vector, covariance matrix, and mixing weights for each Gaussian component. The mean vector defines the central position of each typical normal mode in the feature space, the covariance matrix describes the distribution range and shape of the mode, and the mixing weights characterize the relative frequency of the mode in historical data. The resulting parameterized Gaussian mixture model is the "Normal Mode Baseline Distribution Model," which fully characterizes the probability density distribution of historical normal vibration modes. For each vibration mode cluster newly partitioned by the clustering algorithm, its empirical probability distribution needs to be calculated for comparison with the baseline model. Assume there exists a mode cluster C to be evaluated, which contains... There are 10 sample points, and each sample point is a... A probability vector of dimension . Calculate the empirical distribution of this cluster. One direct approach is to use kernel density estimation. At each point in the probability space... At a given point, the probability density is estimated as the average of the kernel function values ​​of all sample points within that cluster. The kernel function is typically a Gaussian kernel, and its bandwidth parameter needs to be selected based on the data characteristics. Another more convenient method in subsequent calculations is to use a base model... In comparison, the calculation of KL divergence can be based on a Monte Carlo approximation of the samples, in which case it is not necessary to explicitly write out the formula. The analytical form. All sample points within cluster C can be considered as derived from the distribution. Independent and identically distributed samples drawn from the sample, any information about The expected values ​​are approximated using the sample mean. Both methods provide a statistical description of the current data-driven pattern cluster. To quantify the difference between the new pattern cluster C and historical normal patterns, its empirical distribution needs to be calculated. Compared with the baseline distribution model The KL divergence between two probability distributions. The KL divergence is an asymmetric measure of the difference between two probability distributions. Its calculation involves an integral, which is calculated by considering the distribution... On the support set, for The probability density multiplied by it and Integrating the logarithmic density difference. Because and These processes can be quite complex, and the calculations are typically approximated using a sample-based Monte Carlo method. Specifically, all sample points are extracted from the current cluster C. For each sample point... Perform the following calculations: First, calculate... In the benchmark Gaussian mixture model The logarithmic probability density is given. This requires calculation. exist The logarithmic density is calculated for each Gaussian component, and then the exponential density of each component is weighted, mixed, and the logarithm is taken again. Next, it is necessary to estimate... The empirical distribution of its own cluster The logarithmic probability density under the given conditions. If Kernel density estimation can be used for direct calculation; if a sample representation is used, it can be calculated through... The nearest neighbor method estimates the local probability density at this point. Then, the log density difference for each sample point is calculated. Finally, the current cluster C is compared with the baseline model. The KL divergence between the samples is approximately the arithmetic mean of the logarithmic density differences of all sample points. This calculated KL divergence value... The value is non-negative; the larger the value, the greater the overall difference between the vibration mode represented by the current cluster and the historical normal mode. This difference may originate from the shift of the distribution center, changes in the distribution shape, or the emergence of new probability concentration areas. This value is formally defined as the "abnormal deviation index" for this mode cluster. The calculated abnormal deviation index is used to determine anomalies for each mode cluster. The determination does not use a single global threshold, but rather a dynamic threshold mechanism associated with the "condition type and intensity" of the currently analyzed data. To this end, the system maintains a predefined "condition-threshold mapping table." This table is developed based on extensive historical data analysis and domain expert experience, with each row defining a specific condition combination and its corresponding KL divergence threshold. The dimensions of the condition combination include track geometry, operating speed range, traction and braking status, environmental conditions, and their intensity levels. For example, the "high-speed straight section dry environment" condition is associated with a lower threshold because, under such ideal conditions, any significant deviation from the normal mode is more likely to indicate anomalies. The "low-speed passage through a small-radius curve turnout in a wet and slippery environment" condition is associated with a higher threshold because, under complex conditions, the system's intrinsic vibrations become more amplified and diversified, and some significant mode deviations may still remain within acceptable normal ranges. In application, the system dynamically retrieves matching KL divergence anomaly judgment thresholds from the mapping table based on the main operating condition characteristics of the current data batch. The retrieval process supports fuzzy matching or interpolation. Subsequently, the deviation index of each mode cluster is... The deviation is compared with a retrieved dynamic threshold. If the deviation exceeds the threshold, the pattern cluster is determined to be an anomalous pattern cluster. For each determined anomalous cluster, the system automatically generates a structured composite label. The label contains the following core fields: First, the anomalous type code, which is determined based on the analysis of the centroid vector of the anomalous cluster. The independent component index corresponding to the highest probability value in the centroid vector defines the main manifestation of the anomalous pattern, such as "CORR dominant" or "MIX polygon_wavy". Second, the specific anomalous deviation value. Third, the index of the main separating components. In addition, it may also include the sample size, time span, and spatial location information of the cluster. The composite labels of all anomalous clusters are summarized to form the final "annomous pattern label set", providing direct and quantitative basis for subsequent fault location, report generation, and maintenance decisions.

[0247] Furthermore, step S5 includes the following steps:

[0248] Step S51: Receive the set of abnormal mode labels and the decoupling contribution map, and establish a mapping relationship between the labels and the map, that is, each abnormal mode label is associated with the spatial contribution area of ​​its corresponding mode cluster on the decoupling contribution map.

[0249] Step S52: On the decoupled contribution map, highlight the spatial contribution region associated with each anomaly mode label, and overlay the time-frequency information restored from the three-dimensional dynamic feature tensor to generate a visualization map with time-space-frequency anomaly localization.

[0250] Step S53: Combine the anomaly deviation index and the main separation component index in the anomaly pattern label set to analyze the main physical source of the anomaly and match the corresponding fault or degradation pattern description text from the knowledge base;

[0251] Step S54: Integrate the visualized spectrum, the parsed physical source, the matched descriptive text, and the key statistics of the original signal. Based on a preset report template, automatically generate a structured vibration anomaly separation and identification report, such as... Figure 10 As shown;

[0252] Step S55: The generated vibration anomaly separation and identification report, especially the newly discovered anomaly patterns and their representation in the physical constraint feature space, is used as an incremental learning sample and fed back to the vibration source physical constraint embedding model in step S2. This is used to fine-tune the physical constraint loss function or expand the physical rule template, thereby forming an adaptive and enhanced identification closed loop for unknown anomaly patterns.

[0253] In this embodiment of the invention, the generated abnormal pattern label set is received. Contribution map of decoupling (k=1,...,K). Establish the mapping relationship between labels and the graph. For anomalous pattern labels... It is associated with an anomalous cluster. All samples within this cluster are in the original three-dimensional dynamic feature tensor. There is a corresponding time window index. This is found in the decoupling contribution map. In this process, local contribution slices corresponding to these specific time window indices can be extracted. Because the contribution map quantifies the impact of different original features on the separation of the first... The importance of each component, through analysis of anomalous clusters In which components of the sample The contribution map shows high-contribution regions, which can identify which separate component channels the anomalous pattern is mainly "active" on, and the specific locations of these high-contribution regions in the time domain, frequency domain, and operating condition dimensions. This establishes a mapping chain of "anomaly label → anomalous sample time window set → hotspot regions of contribution for each component". Based on the above mapping, a comprehensive visualization map is generated.

[0254] First, on the decoupling contribution map, for each anomaly label The contribution hotspots of the associated sample set in the main anomaly component channels are highlighted with specific colors. These hotspots may appear as stripes on the time-frequency plane, clusters in the operating condition dimension, etc. Next, from the original three-dimensional dynamic feature tensor... In this process, the time-frequency information corresponding to these hotspot regions is reconstructed. This is typically achieved by calculating the magnitude of the short-time Fourier transform, i.e., calculating the spectral amplitude of the vibration signal for each time window. The contribution hotspot area (indicated by "reason" or "key feature") is merged and displayed with the time-frequency energy cloud map (displayed as "performance" or "original signal morphology") overlaid on it.

[0255] Simultaneously, train speed curves and track markings (curve start points, turnout locations, etc.) are annotated below the timeline. The final result is a multi-layered overlay map, displayed simultaneously on the same time-frequency coordinate plane using different colors and transparency levels.

[0256] 1) Abnormal contribution areas (such as red semi-transparent coverage);

[0257] 2) Time-frequency energy distribution of the original vibration signal (represented by grayscale or color contour plots);

[0258] 3) Operating timeline at the bottom.

[0259] This atlas achieves integrated spatiotemporal and frequency-based three-dimensional localization and visualization of anomalous patterns. It also incorporates a set of anomalous pattern labels. The information provided allows for in-depth analysis of the anomaly's primary physical source. The "Primary Separating Component Index" in the label indicates which physical stimulus source the anomaly is primarily associated with (e.g., the "wheel polygon" component). The "Anomaly Deviation" value in the label... The degree of anomaly was quantified. By analyzing the centroid position of the anomaly cluster in the probability distribution space, it can be determined whether it is dominated by a single component or coupled with multiple components.

[0260] Furthermore, by combining the time-frequency characteristics of anomalous contribution hotspots in the visualized map—for example, whether the hotspots are concentrated on frequency ridges linearly related to vehicle speed, or whether they are fixed at a specific location on the track—more specific physical inferences can be made. For instance, if the main component index of an anomaly label is "wheel polygon," and the frequency of its hotspot region is strictly multiple of the vehicle speed and distributed across all bogies, then it can be inferred to be "high-order wheel polygon wear."

[0261] Another key component of the tag is "rail corrugation," where the frequency of its hotspot area is fixed and its spatial location precisely corresponds to a section of curved rail, thus indicating "curved section rail corrugation." These physical inferences are matched against a pre-built fault knowledge base, which stores textual descriptions, possible causes, effects, and maintenance recommendations for various typical fault modes. Upon successful matching, the corresponding descriptive text is extracted to form a physical mechanism explanation and qualitative description of the currently identified anomaly. All intermediate results and the final analysis are integrated, and a structured vibration anomaly separation and identification report is automatically generated according to a preset XML or JSON template. The report typically includes the following sections:

[0262] 1) Execution Summary: List the total number of anomalous patterns identified in this analysis, the type and severity level of each anomalous pattern (based on...). Deviation classification).

[0263] 2) Anomaly Details List: Lists detailed information for each anomaly tag in tabular form, including: anomaly ID, type, start time, end time, corresponding line mileage range, main associated vibration components, KL divergence deviation value, and estimated equivalent fault parameters (such as polygon order and wavelength range).

[0264] 3) Key visualization map: Embedded spatiotemporal frequency anomaly localization map.

[0265] 4) Physical analysis and diagnostic recommendations: Write the generated physical source resolution text and the maintenance recommendation text matched from the knowledge base.

[0266] 5) Raw data snapshot: Attach key statistics of the raw vibration signal during the abnormal period, such as peak value, RMS value, and kurtosis index.

[0267] 6) Method Confidence Level Description: Separation effectiveness evaluation indicators such as Γ can be attached. The report is output in document form for maintenance personnel to review and for use by the decision support system. The generated vibration anomaly separation and identification report will include, in particular, anomaly patterns confirmed as "newly discovered" (i.e., those patterns appearing in the historical normal database). (and not adequately recorded in the existing fault knowledge base), serving as valuable incremental learning samples, which are fed back into the vibration source physical constraint embedding model.

[0268] The feedback process includes:

[0269] 1) Feature backtracking: Based on the time window index of the new abnormal sample, backtrack to its corresponding physical constraint feature vector. This constitutes a new subset of abnormal features. .

[0270] 2) Constraint Derivation: Analysis Based on the statistical properties of the data and their performance on physical rules, we attempt to deduce in reverse the new constraint forms or parameter ranges that the data may satisfy, which are not yet covered by existing physical rule templates.

[0271] 3) Model fine-tuning: As a "new abnormal positive sample", from the normal database The corresponding features are sampled as "negative samples" to form incremental learning data pairs. The variational encoder is fine-tuned using a contrastive learning loss function. as follows:

[0272] ;

[0273] in, To rebuild the losses, For KL divergence loss, These are the weighting coefficients for the KL divergence loss. To compare the losses, To compare the weighting coefficients of the loss, New anomalous samples are encouraged to move closer to each other in the latent space while moving away from normal samples.

[0274] 4) Knowledge Base Update: The new physical constraints derived through reverse derivation are formalized and added as new entries to the physical rule template library. Simultaneously, the description, features, and labels of this new anomaly mode are added to the fault knowledge base. Through this closed loop, the vibration source physical constraint embedding model and the entire identification system gain the ability to learn and adapt to new types of anomaly modes, achieving adaptive enhancement of the "discovery-learning-identification" process for unknown fault modes. This allows the system to continuously evolve with the accumulation of operational data, improving its long-term robustness and generalization ability.

[0275] Furthermore, step S55 includes the following steps:

[0276] From the newly generated vibration anomaly separation and identification report, extract the new anomaly pattern labels and their physical source descriptions that are not covered by the historical knowledge base; trace back the physical constraint feature vectors corresponding to the new anomaly pattern, and locate the source feature vector subset that caused the anomaly in the physical constraint feature vector set;

[0277] Based on the aforementioned subset of source feature vectors, a new form or parameter range of the physical constraint conditions it satisfies in the vibration source physical constraint embedding model is derived in reverse.

[0278] Using the subset of source feature vectors as positive samples and the normal mode feature vectors as negative samples, an incremental learning dataset is constructed to fine-tune the variational encoder embedded in the vibration source physical constraint model, enabling the model to encode such new abnormal features. The new physical constraint conditions derived in reverse are formalized and added as new entries to the physical rule template library, and the physical rule template generation logic corresponding to the working condition is updated, thereby optimizing the physical constraint conditions in step S2 of the subsequent identification process and realizing the continuous evolution of the identification capability.

[0279] In this embodiment of the invention, upon receiving the newly generated vibration anomaly separation and identification report, which details the discovered anomaly patterns in a structured format, the system maintains an internal historical knowledge base that records metadata for all previously identified and analyzed anomaly patterns, including the typical feature vector range, probability distribution centroid location, and associated physical source description keywords for each pattern. The system parses the new report, calculates the probability distribution centroid vector for each anomaly record in the report, and performs similarity matching with the features of all historical patterns in the knowledge base. The matching algorithm calculates cosine similarity. If the similarity between the features of a record and any historical pattern feature in the knowledge base is below a set acceptance threshold, and the anomaly deviation index of its report is sufficiently high, then the record is determined to be a "new anomaly pattern not covered by the historical knowledge base." The system then fully extracts the composite tag and its physical description text of the new anomaly pattern. For example, the label might display "Anomaly Type: High-Frequency Fixed-Position Resonance, Dominant Component Index: 3, Deviation: 2.5," and the descriptive text might be "A narrow-band vibration lasting approximately 850 Hz was detected in the straight section from K50+100 to K50+200, independent of vehicle speed, with a significant vertical acceleration response in the axle box." After confirming the new anomaly, the system initiates a data backtracking process. Based on the unique cluster identifier associated with the new anomaly label, the system retrieves a list of time indices for all samples constituting the anomaly cluster from the clustering result storage in step S4. Using this time index list, the system locates the physical constraint feature vector set finally generated in step S2. The physical constraint feature vector set is a large storage structure, where each index corresponds to a specific time window. The system precisely extracts the physical constraint feature vectors corresponding to all time indices through querying. These extracted vectors constitute the "subset of source feature vectors that triggered the new anomaly." This subset represents the lowest-level, physically filtered feature expression that drives the downstream network to classify it as a new anomaly. The system creates a new data archive for this subset, records its mapping relationship with the new anomaly pattern, and labels it as the "Newly Discovered Anomaly Positive Sample Set". Based on the acquired source feature vector subset, the system attempts to infer the physical mechanism from the features, aiming to discover constraints not covered in the existing physical rule template. This process first performs statistical analysis on the subset, calculating its distribution center and covariance in the latent space. Next, using the decoder of the trained vibration source physical constraint embedding model, these anomaly latent variables are reconstructed into three-dimensional feature tensors. The reconstructed feature tensors are analyzed to observe their common patterns. Combined with the decoupling contribution map generated in step S3, the system analyzes which original signal features contribute the most to the formation of this anomaly pattern. Integrating the feature performance at the signal level with domain knowledge, the system generates hypotheses about the physical root cause of the new anomaly.For example, if analysis shows that the anomalous features are mainly vibrations at fixed spatial locations and at fixed high frequencies, with energy transferred between the track and the vehicle, it might be hypothesized that a new physical phenomenon exists, such as "structural resonance at a fixed location caused by local stiffness anomalies in the track foundation (such as track bed slab compaction or support failure)." The next step is to formalize this hypothesis into new physical constraints. Formalization may take several forms: defining a new loss term that penalizes features in the feature space that conform to the pattern of "energy in a specific frequency band concentrated at a fixed location"; or modifying the allowable range of a parameter (such as local support stiffness) in the existing physical model to extend it to include such anomalous states. This derivation process aims to generate new rule expressions or parameterization suggestions that are computable and can be integrated into the existing framework. Injecting new anomalous knowledge into the feature extraction model is the core of the self-evolutionary closed loop. Construct an incremental learning dataset. Use a subset of feature vectors from new anomalous sources as positive samples. Randomly sample feature vectors from the "normal feature library," which stores massive amounts of historical normal sample physical constraint feature vectors, with the exact same number of positive samples, as negative samples. Positive and negative samples are paired to form a fine-tuning dataset. The core of the vibration source physical constraint embedding model—the variational encoder—is then fine-tuned. The goal of fine-tuning is to enable it to generate highly discriminative latent space representations for new anomalous features. In each training iteration, mini-batch data is sampled from both the positive and negative sample sets and then merged into the variational encoder. The weighted sum of the three loss terms is calculated as the total loss. The reconstruction loss ensures that the model's basic encoder-decoder capability does not degrade. The physical constraint loss ensures that model updates do not violate known fundamental physical laws. The newly introduced contrastive learning loss is crucial for fine-tuning; its design goal is to optimize the structure of the latent space, making the latent representations of all new anomalous positive samples close together to form a compact cluster, while simultaneously pushing this cluster away from the latent representations of all normal negative samples in the feature space, forming a clear separation boundary. The network parameters of the variational encoder are iteratively updated using backpropagation and gradient descent algorithms. The fine-tuning process continues until two convergence criteria are met: the reconstruction loss on the validation set stabilizes and no longer decreases; and the separation index between the new anomalous feature cluster and the normal feature cluster in the latent space consistently exceeds a preset target value. Once the criteria are met, fine-tuning stops, and the discriminative knowledge of the new anomaly pattern is solidified into the updated variational encoder parameters. While the model completes knowledge solidification, the formalized new knowledge needs to be integrated into the system's rule system. The new physical constraints, derived in reverse and formalized, are encoded into executable logical modules or configuration files. For example, the rule "penalizing high-frequency resonance at a fixed location" is written as a new loss calculation function. The code module of this new rule and its metadata (including triggering conditions and associated anomaly types) are registered as a new record in the system's "physical rule template library." Next, the generation engine logic for the "operating condition-specific physical rule template" is updated. This engine originally selected and combined constraints from the rule library based on the current operating condition.After the update, when the engine encounters operating conditions similar to the new anomaly pattern again, it will automatically select this new rule from the library and incorporate it into the physical constraint template instantiated for the current operating condition. Therefore, in subsequent vibration analysis processes, the physical constraint embedding model in step S2 will be guided by both existing physical laws and this new knowledge when extracting features. Through a complete closed loop of "discovering new anomalies -> backtracking analysis -> deriving rules -> fine-tuning the model -> updating the rule base," the entire anomaly separation and identification system achieves the goal of continuously learning from data and autonomously evolving its core knowledge base and processing capabilities, enabling it to adapt to potentially more complex and varied wheel-rail system failure modes in the future.

[0280] Furthermore, the fine-tuning of the variational encoder in the vibration source physical constraint embedding model includes the following steps:

[0281] A batch of vectors is randomly sampled from the source feature vector subset, and an equal number of negative sample vectors are sampled from the historical normal pattern feature vector library to form the incremental learning data pairs for the current batch.

[0282] The incremental learning data is embedded into the variational encoder of the model to input the physical constraints of the vibration source, and its reconstruction loss and physical constraint loss are calculated. A contrastive learning loss term is introduced, which makes the feature representations of new abnormal positive samples in the latent space closer to each other, while moving away from the feature representations of normal negative samples, thereby enhancing the model's ability to distinguish new abnormal patterns.

[0283] The reconstruction loss, physical constraint loss and contrastive learning loss are weighted and summed to form the total loss function in the fine-tuning stage. The gradient descent algorithm is then used to update the network parameters of the variational encoder.

[0284] The model is iterated and updated until the reconstruction error of the model on the new abnormal samples is stable and the distance between the cluster of new abnormal feature vectors in the latent space and the cluster of normal feature vectors reaches the preset separation degree requirement. The model is then fine-tuned, thereby solidifying the knowledge of the new abnormal pattern into the vibration source physical constraint embedded model.

[0285] In this embodiment of the invention, a small batch of new anomalous positive sample vectors is obtained by using a pseudo-random number generator to perform non-repeating random sampling from the "source feature vector subset" corresponding to the newly discovered anomalous patterns. Each vector in this subset is a feature representation in the latent space of the vibration segment previously identified as a new anomalous pattern. Simultaneously, feature vectors of exactly equal size to the current positive sample batch are randomly sampled from the "historical normal pattern feature vector library" storing all feature vectors under historical normal operating conditions, forming a negative sample batch. The positive and negative samples are kept equal in quantity to ensure that the model will not be biased due to data imbalance during the learning process. The two small batches of samples are merged to form an incremental learning data pair batch containing an equal number of positive and negative samples. This batch is the basic data unit for the current iteration training of the model. The constructed incremental learning data pair batch is input into the variational encoder network of the vibration source physical constraint embedding model to be fine-tuned for forward propagation. The variational encoder network receives the input feature vectors, performs multiple nonlinear transformations, and outputs the probability distribution parameters of each input vector in the latent space, namely the mean vector and the log-variance vector. Using these parameters, a specific latent space representation vector is obtained through reparameterization techniques. Based on this latent space representation, three core loss terms need to be calculated. The first is the reconstruction loss. The sampled latent space representation is input into a pre-trained decoder network paired with the variational encoder. The decoder attempts to reconstruct an output that is as close as possible to the original input feature vector. The reconstruction loss calculates the mean squared error between the reconstructed output and the original input, and its value reflects the degree to which the model maintains its basic data reconstruction ability. The second is the physical constraint loss. This loss term aims to ensure that the learning process does not violate known physical laws. Through an auxiliary physical mapping network, incalculable estimates of physical quantities, such as dominant frequencies and energy distribution proportions, are derived from the latent space representation. These estimates are then compared with theoretical values ​​calculated based on current operating parameters and a template of basic physical rules, and the differences between the two are calculated, for example, using the mean squared error. This loss forces the model to maintain consistency with the fundamental dynamics of the system when adapting to new anomalies. The third is the contrastive learning loss. This is the core optimization objective introduced in this fine-tuning. Its design intent is to reshape the structure of the latent space. For all new anomalous positive samples within a batch, the loss term encourages them to move closer together, forming a compact, cohesive "cluster" in the feature space, by calculating a function of the cosine similarity or Euclidean distance between their feature representations in the latent space. Simultaneously, for each new anomalous positive sample, the loss term calculates its distance to the feature representations of all normal negative samples within the batch and penalizes smaller distances, thus "pushing" a clear gap between the anomalous cluster and the normal sample region, enhancing the distinguishability of the features. A typical implementation uses triplet loss or supervised contrastive loss.After calculating the three scalar values—reconstruction loss, physical constraint loss, and contrastive learning loss—they are linearly added together using preset weighting coefficients to form the total loss function for this round of fine-tuning. The weighting coefficients determine the relative importance of each loss in the overall optimization objective and need to be set based on prior knowledge to balance the relationship between "preserving old capabilities," "following physical laws," and "learning new features." Then, the backpropagation process is initiated. The gradient of the total loss function with respect to all weight parameters of the variational encoder network is calculated. The gradient indicates the direction and magnitude of adjustment for each parameter to reduce the total loss. A gradient-based optimization algorithm, such as stochastic gradient descent with momentum, is used to update the variational encoder network parameters using the calculated gradients. The parameter update formula is the learning rate multiplied by the negative gradient direction. In incremental learning, the learning rate is usually set smaller than that used in the initial training and may be layered; for example, a smaller learning rate is used for the first few layers of the encoder to protect learned general features, while a slightly larger learning rate is used for later layers to adapt to new knowledge more quickly. The above processes of data batch preparation, forward propagation, loss calculation, backpropagation, and parameter updating constitute a complete training iteration. This iterative process is repeated cyclically, with each iteration using a new batch of data randomly resampled from a subset of the source feature vectors and the normal dataset. Fine-tuning continues until preset convergence criteria are met. There are two main criteria. The first criterion focuses on performance stability: monitoring the reconstruction loss on an independent validation set (composed of reserved new anomalous samples and normal samples), when the loss value no longer shows a decreasing trend over multiple consecutive training epochs and its fluctuation range is less than a threshold, the model is considered to have sufficiently adapted to and stabilized its ability to reconstruct new data. The second criterion focuses on the feature space structure: periodically calculating the Euclidean distance between the centroids of the clusters formed by all new anomalous positive samples in the latent space and the centroids of the clusters formed by normal negative samples. Simultaneously, the average distance between samples within a cluster is calculated. A target separation is set, for example, requiring the inter-cluster distance to be greater than a certain multiple of the average intra-cluster distance. When the calculated separation metric consistently exceeds the preset target and remains stable in subsequent evaluations, the new anomalous features are considered to have achieved clear separation from normal features in the latent space. The fine-tuning process terminates when both criteria are met simultaneously. At this point, the network weights of the variational encoder have been updated, and the discriminative feature knowledge of the new anomaly pattern has been "solidified" in its parameters. The fine-tuned model can map future anomalous vibration signals with similar physical origins to feature regions in the latent space adjacent to the current new anomaly cluster, thereby providing more discriminative feature inputs for downstream anomaly separation and recognition modules, ultimately achieving continuous and accurate recognition of new anomaly patterns.

[0286] Example 2:

[0287] This embodiment provides a vibration anomaly separation and identification system for wheel-rail systems under complex operating conditions, including:

[0288] The first module is used to acquire multi-source heterogeneous temporal vibration signals of the wheel-rail system under complex working conditions, perform time-space-frequency three-dimensional feature mapping on the multi-source heterogeneous temporal vibration sequence numbers, and generate a three-dimensional dynamic feature tensor.

[0289] The second module is used to input the three-dimensional dynamic feature tensor into the pre-constructed vibration source physical constraint embedding model to obtain the physical constraint feature vector set output by the model.

[0290] The third module is used to decouple the vibration components of the physical constraint feature vector set to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum.

[0291] The fourth module is used to input the probability distribution matrix into the anomaly pattern adaptive clustering engine to obtain the anomaly pattern clusters in the independent vibration components output by the engine, and to generate an anomaly pattern label set based on the anomaly pattern clusters.

[0292] The fifth module is used to fuse the abnormal pattern label set with the decoupling contribution map to generate a wheel-rail system vibration anomaly separation and identification report;

[0293] Specifically, after generating the wheel-rail system vibration anomaly separation and identification report, the parameters of the vibration source physical constraint embedding model are adjusted based on the wheel-rail system vibration anomaly separation and identification report.

[0294] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0295] Example 3:

[0296] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0297] Example 4:

[0298] This embodiment provides a computer device, including:

[0299] Memory, used to store computer programs / instructions;

[0300] A processor for executing the computer program / instructions to implement the steps of the method described in Embodiment 1.

[0301] Example 5:

[0302] This embodiment provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in Embodiment 1.

[0303] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0304] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0305] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0306] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0307] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0308] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A method for separating and identifying vibration anomalies in wheel-rail systems under complex operating conditions, characterized in that, include: Acquire multi-source heterogeneous temporal vibration signals of the wheel-rail system under complex working conditions, perform time-space-frequency three-dimensional feature mapping on the multi-source heterogeneous temporal vibration sequence numbers, and generate a three-dimensional dynamic feature tensor. The three-dimensional dynamic feature tensor is input into the pre-constructed vibration source physical constraint embedding model to obtain the physical constraint feature vector set output by the model; The vibration components of the physical constraint feature vector set are decoupled to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum. The probability distribution matrix is ​​input into the abnormal pattern adaptive clustering engine to obtain the abnormal pattern clusters in the independent vibration components output by the engine, and an abnormal pattern label set is generated based on the abnormal pattern clusters. By fusing the abnormal pattern label set with the decoupling contribution map, a vibration anomaly separation and identification report for the wheel-rail system is generated. Specifically, after generating the wheel-rail system vibration anomaly separation and identification report, the parameters of the vibration source physical constraint embedding model are adjusted based on the wheel-rail system vibration anomaly separation and identification report.

2. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 1, characterized in that, The method for generating the three-dimensional dynamic feature tensor includes: Simultaneously acquire multi-channel vibration acceleration signals, wheel-rail force signals, and operating status signals of the wheel-rail system under complex working conditions including curve sections, turnout areas, and different wheel-rail interface states, forming a multi-source heterogeneous time-series vibration signal stream; The multi-source heterogeneous temporal vibration signal stream is subjected to spatiotemporal alignment and scale normalization to generate a synchronous standardized signal matrix. For each signal channel in the synchronous normalized signal matrix, the amplitude change of the time-series signal is encoded into a symbol sequence, and the symbol transition probability is statistically analyzed to generate a time-domain recursive dynamic feature vector for each channel. Adaptive wavelet ridge extraction is performed on the synchronous normalized signal matrix to generate a time-frequency joint distribution feature matrix. Based on the time-domain recursive dynamic feature vector and the time-frequency joint distribution feature matrix, a three-dimensional dynamic feature tensor is constructed by fusing the spatial location and operating condition labels in the operating status signal.

3. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 1, characterized in that, The method for generating the physical constraint feature vector set includes: From the three-dimensional dynamic feature vector, the wheel-rail contact geometric constraints and system dynamic boundary conditions under the current working condition are parsed according to the operating state signal, and the physical rule template corresponding to the working condition is generated. A physical constraint loss function is constructed that includes vibration energy conservation, momentum transfer and structural modal coupling. The parameters in the physical constraint loss function are instantiated using the physical rule template corresponding to the working condition. The three-dimensional dynamic feature tensor is input into a pre-trained variational encoder. The training objective of the variational encoder is to minimize the physical constraint loss function while reconstructing the input features, so that the encoder learns feature representations that satisfy physical laws in the latent space. From the bottleneck layer of the variational encoder, feature vectors of the latent space are extracted, and feature vectors with a deviation from the physical constraint loss function calculation below a preset threshold are selected through physical rule compliance test to form a primary physical constraint feature vector set. Spectral clustering analysis is performed on the primary physical constraint feature vector set to merge feature clusters generated by the same physical mechanism, and the centroid vector of each feature cluster is extracted to generate a redundancy-free physical constraint feature vector set. Each vector in this set represents a potential vibration source feature that conforms to physical laws.

4. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 1, characterized in that, The vibration component decoupling is performed on the physical constraint feature vector set to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum, including: The physical constraint feature vector set is input into a pre-constructed and trained generative adversarial network to obtain the synthetic feature vector output by the generative adversarial network. The synthesized features are subjected to channel-level Softmax normalization to obtain the probability that each potential vibration source feature vector belongs to each separated vibration component, forming a probability distribution matrix; Calculate the gradient of the probability value of each separated vibration component in the probability distribution matrix with respect to each local feature in the original three-dimensional dynamic feature tensor, and generate a decoupling contribution map based on the gradient value; The generative adversarial network (GAN) includes a generator network and multiple discriminator networks. Each discriminator network corresponds to a potential vibration component to be separated. The generator network reconstructs a synthetic feature tensor that is consistent with the original three-dimensional dynamic feature tensor in terms of physical constraint loss but is separable in each dimension, based on the input physical constraint feature vector. Each discriminator network is used to determine whether the synthetic feature tensor contains its corresponding vibration component, and through adversarial training, the generator network separates vibration features of different physical mechanisms into different dimensions or channels of the synthetic tensor. During the stable phase of adversarial training of the generative adversarial network, the internal neuron patterns activated by the generator network to reconstruct each physical constraint feature vector are recorded, generating a pattern activation map. The physical constraint feature vector set is input into the frozen generator, and a probe layer is inserted into the middle layer of the network to capture the intermediate representation sequence during the feature transformation process. Attention weights are assigned to the intermediate representation sequence, and the importance score of each position in the sequence to each channel of the final output synthetic feature tensor is calculated to generate an attention-guided channel association weight matrix. Based on the channel association weight matrix, the synthesized feature tensor output by the generator is weighted, fused, and recombined with channels to obtain the recombined synthesized feature tensor. The recombined synthetic feature tensor is subjected to a Softmax operation along the channel dimension to convert the response value of each channel into a probability form, thereby obtaining the probability distribution of each vector in the physical constraint feature vector set corresponding to each separated vibration component, and generating a probability distribution matrix.

5. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 1, characterized in that, The data processing procedure of the anomaly pattern adaptive clustering engine includes: The probability distribution matrix is ​​regarded as a set of points in a high-dimensional feature space. Each point in the set represents the coordinates of a vibration sample in the separated component space. A cluster center self-discovery algorithm based on local density peaks and relative distance is used to identify the cluster center points in the probability distribution matrix. The cluster center points are input into a pre-constructed dynamic similarity measurement network to obtain a similarity matrix; The similarity is subjected to a spectral clustering algorithm to divide the points in the probability distribution matrix into different clusters, resulting in several vibration mode clusters, each cluster representing a specific vibration mode; Several vibration mode clusters are compared with a pre-constructed normal mode baseline distribution model by KL divergence calculation. Vibration mode clusters with divergence values ​​exceeding a preset dynamic threshold are marked as abnormal modes to obtain abnormal mode clusters. The normal mode baseline distribution model is constructed based on the probability distribution of vibration components under historical normal operating conditions. An abnormal pattern label set is generated based on the abnormal pattern cluster, including: Each abnormal pattern cluster is assigned a unique abnormal pattern label, resulting in an abnormal pattern label set.

6. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 5, characterized in that, The data processing procedure for the normal pattern baseline distribution model includes: Extract the probability distribution of all sample points in each vibration mode cluster to be identified, and calculate the empirical probability distribution of the cluster. Calculate the KL divergence between the empirical probability distribution of each vibration mode cluster and the normal mode baseline distribution model, and use the KL divergence as an index of the abnormal deviation of the mode cluster. Based on the type and intensity of the current complex working conditions, the dynamic threshold for anomaly determination of KL divergence is determined based on a pre-constructed working condition-threshold mapping table, and pattern clusters whose anomaly deviation index exceeds the dynamic threshold are determined to be abnormal.

7. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 6, characterized in that, By fusing the abnormal pattern label set with the decoupling contribution map, a vibration anomaly separation and identification report for the wheel-rail system is generated, including: Establish a mapping relationship between the set of anomalous pattern labels and the decoupling contribution map to obtain the spatial contribution region of each anomalous pattern cluster on the decoupling contribution map; Based on the spatial contribution regions of each anomaly pattern cluster on the decoupling contribution map, the time-frequency information in the three-dimensional dynamic feature tensor is superimposed to obtain a visualization map with time-space-frequency anomaly localization. By combining the anomaly deviation index and the main separation component index in the anomaly pattern label set, the main physical source of the anomaly is analyzed, and the corresponding fault or degradation pattern description text is matched from the pre-built knowledge base. Based on the visualized spectrum, main physical sources, matched descriptive text, and key statistics of the original signal, an anomaly separation and identification report is generated according to a preset report template.

8. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 1, characterized in that, Based on the vibration anomaly separation and identification report of the wheel-rail system, the parameters of the vibration source physical constraint embedding model are adjusted, including: Extract new anomaly pattern labels and their main physical sources from the vibration anomaly separation and identification report that are not covered by the historical knowledge base, and determine the source feature vector subset of the new anomaly pattern labels based on the physical constraint feature vector set; Based on the aforementioned subset of source feature vectors, a new form or parameter range of the physical constraint conditions satisfied by the subset in the vibration source physical constraint embedding model is derived in reverse. Using the subset of source feature vectors as positive samples and the normal mode feature vectors as negative samples, an incremental learning dataset is constructed. The variational encoder in the vibration source physical constraint embedding model is fine-tuned using the incremental learning dataset. The new physical constraint conditions derived in reverse are formalized and added as new entries to the physical rule template library. The physical rule template generation logic corresponding to the working condition is also updated.

9. The method for separating and identifying vibration anomalies in wheel-rail systems under complex working conditions according to claim 8, characterized in that, Fine-tuning the variational encoder in the vibration source physical constraint embedding model using incremental learning datasets includes: Perform iterative steps until the model's reconstruction error on the new anomalous samples stabilizes and the distance between the clusters of new anomalous feature vectors in the latent space and the clusters of normal feature vectors reaches the preset separation requirement, thus completing the fine-tuning. The iterative steps include: A batch of vectors is randomly sampled from the source feature vector subset as positive samples, and an equal number of negative sample vectors are sampled from the historical normal pattern feature vector library to form the incremental learning data pairs for the current batch. The incremental learning data is used to embed the physical constraints of the input vibration source into the variational encoder of the model, and the reconstruction loss and physical constraint loss are calculated. A contrastive learning loss term is introduced, which makes the feature representations of positive samples in the latent space closer to each other and farther away from the feature representations of negative samples. The reconstruction loss, physical constraint loss, and contrastive learning loss are weighted and summed to form the total loss function in the fine-tuning stage. The gradient descent algorithm is then used to update the network parameters of the variational encoder based on the total loss function.

10. A vibration anomaly separation and identification system for wheel-rail systems under complex working conditions, characterized in that, include: The first module is used to acquire multi-source heterogeneous temporal vibration signals of the wheel-rail system under complex working conditions, perform time-space-frequency three-dimensional feature mapping on the multi-source heterogeneous temporal vibration sequence numbers, and generate a three-dimensional dynamic feature tensor. The second module is used to input the three-dimensional dynamic feature tensor into the pre-constructed vibration source physical constraint embedding model to obtain the physical constraint feature vector set output by the model. The third module is used to decouple the vibration components of the physical constraint feature vector set to obtain a probability distribution matrix containing multiple independent vibration components and a corresponding decoupling contribution spectrum. The fourth module is used to input the probability distribution matrix into the anomaly pattern adaptive clustering engine to obtain the anomaly pattern clusters in the independent vibration components output by the engine, and to generate an anomaly pattern label set based on the anomaly pattern clusters. The fifth module is used to fuse the abnormal pattern label set with the decoupling contribution map to generate a wheel-rail system vibration anomaly separation and identification report; Specifically, after generating the wheel-rail system vibration anomaly separation and identification report, the parameters of the vibration source physical constraint embedding model are adjusted based on the wheel-rail system vibration anomaly separation and identification report.

Images