Fault detection and positioning method and system for train transmission system under variable working conditions

Abstract

The invention belongs to the technical field of intelligent operation and maintenance of train transmission systems, and discloses a train transmission system fault detection and positioning method under variable working conditions. The sample diversity is enhanced by constructing a pseudo fault sample library; eliminating noise interference in the data by using a residual shrinkage feature extraction module, and extracting representative features; a self-adaptive alignment mechanism is designed to map different working condition health samples to a unified feature space, and the influence of working condition fluctuation on data distribution is relieved; health and pseudo fault feature pairs are constructed, working condition attributes and essential attributes of health states are revealed, and relation scores of the feature pairs are calculated by using a Colmogorov-Arnodel network relation. According to the method, the train transmission system fault is accurately detected and positioned under variable working conditions, the dependence of a data driving method on fault data and the dependence on data distribution consistency are overcome, and important engineering application value is provided for high-performance, long-life and safe and stable operation of the train transmission system in China.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault detection, and more specifically, relates to a method and system for fault detection and location of a train drive system under variable working conditions. Background Art

[0002] As a core component of the railway transportation network, the stable operation of the train drive system is directly related to transportation safety and operation efficiency. However, during actual service, the system needs to cope with various complex working conditions such as speed fluctuations, load changes, extreme temperatures, and component wear. These factors are likely to induce component performance degradation and significantly increase the risk of faults. How to accurately detect and locate faults under variable working conditions has become a key problem in ensuring the safe operation of trains and reducing maintenance costs.

[0003] Currently, data-driven fault detection technologies have attracted much attention due to their powerful pattern recognition capabilities. However, these methods generally rely on two basic assumptions: one is that a large number of real fault samples are required for model training, and the other is that the distributions of training data and test data need to be consistent. For the train drive system, it is difficult to meet these two conditions. Firstly, due to extremely high safety requirements and the normalization of planned maintenance, real fault samples are extremely scarce. Secondly, the train operating conditions are frequently changed due to factors such as speed, load, environmental conditions, and vehicle configuration, resulting in data distribution drift. This distribution inconsistency makes traditional data-driven methods prone to false detection or missed detection when facing complex working conditions, and the detection performance is greatly reduced. Therefore, there is an urgent need to develop a fault detection and location technology that can work under zero-fault sample conditions and adapt to variable working conditions, so as to deeply explore the essential characteristics of the healthy state, effectively cope with data distribution differences, and thus significantly improve the robustness and adaptability of the model, providing a reliable guarantee for the safe and stable operation of the train drive system. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art and actual needs, the present invention proposes a method and system for fault detection and location of a train drive system under variable working conditions, aiming to achieve intelligent fault identification and location without fault samples and under variable working conditions.

[0005] To achieve the above object, the present invention provides a method for fault detection and location of a train drive system under variable working conditions, including:

[0006] S1. Construction of a pseudo-fault sample library: Based on the healthy samples collected from the high-speed train drive system under various operating conditions, a pseudo-fault sample generator is designed to generate pseudo-fault samples outside the distribution of healthy samples. At the same time, diverse fault samples are collected from equipment with similar functions to the drive system and are adaptively transformed to generate pseudo-fault samples applicable to this system. These two types of pseudo-fault samples jointly constitute a diverse pseudo-fault sample library, effectively alleviating the problem that the model is overly sensitive to minor perturbations in the healthy state when only relying on healthy samples, and laying a solid data foundation for subsequent fault detection. On this basis, the pseudo-fault sample library and most of the historical healthy samples are used as the training data set, and a part of the historical healthy samples are used as the validation data set. Time-frequency domain features are extracted from the training data set and the validation data set respectively to capture the time and frequency characteristics of the signals, providing high-quality inputs for model training.

[0007] S2. Construct a unified domain health relationship learning framework: The unified domain health relationship learning framework includes a residual shrinkage feature extraction module, an adaptive alignment mechanism, and an attribute feature mining mechanism. The residual shrinkage feature extraction module extracts more representative healthy state and pseudo-fault state features by removing noise interference in the training data. The adaptive alignment mechanism takes the healthy features under a certain operating condition as the anchor point, selects healthy features from another operating condition as positive samples, and uses pseudo-fault features as negative samples. By designing a boundary reconstruction loss, the distance between the positive sample and the anchor point is shortened, and the distance between the negative sample and the anchor point is increased, thereby unifying and aligning the healthy features under different operating conditions to a shared space boundary, while isolating pseudo-fault samples and eliminating the influence of operating condition fluctuations on the distribution of healthy features. The attribute feature mining mechanism concatenates the healthy state features of multiple operating conditions to generate feature pairs to reveal the operating condition attributes of the healthy state, and forms new feature pairs by concatenating the healthy state features and pseudo-fault samples to reveal the essential attributes of the healthy state. The concatenated feature pairs are input into the Kolmogorov-Arnold network relationship evaluator to mine the internal correlations of the feature pairs and calculate the relationship scores, comprehensively characterizing the relationship between the healthy state and the pseudo-fault state.

[0008] S3. According to the core components of the train drive system (such as traction motors, transmission gearboxes, and axle boxes), adopt a divide-and-conquer strategy to repeat the processes of steps S1 and S2 for each component, respectively training exclusive unified domain health relationship learning frameworks to avoid interference from complex dependency relationships between components on the detection accuracy. The predicted relationship scores of each component are integrated into a relationship score matrix for subsequent optimization and status judgment.

[0009] S4. Using the boundary reconstruction loss and the error between the true relationship score matrix and the predicted relationship score matrix as the joint optimization objective, iteratively optimize the framework using the training dataset; input the validation dataset into the trained model, and with the goal of maximizing the model accuracy, determine the healthy boundary value of each component as the threshold for state discrimination;

[0010] S5. For the sample to be detected and its corresponding historical healthy samples, extract time-frequency domain features and then input them into the healthy relationship learning framework of the corresponding components respectively to calculate the similarity scores; compare the scores with the healthy boundary values. If the relationship score is greater than or equal to the boundary value, it is determined to be in a healthy state; otherwise, it is determined to be in a faulty state. Through this process, accurate judgment of the states of each component is achieved.

[0011] Furthermore, the pseudo-fault sample generator generates samples that are near the feature space of healthy samples but outside its distribution range by applying perturbations along random directions and introducing noise related to the operating conditions based on healthy state samples under different operating conditions. To adapt to the feature changes of healthy samples under different operating conditions, a working condition adaptive rejection probability mechanism is introduced to dynamically adjust the offset boundary, ensuring that the generated pseudo-fault samples neither deviate too much from the healthy distribution nor are too close to it, thus forming working condition specific pseudo-fault clusters;

[0012] Furthermore, the number of pseudo-fault categories generated by the pseudo-fault sample generator for each component is greater than 15, and the number of fault categories of equipment with functions similar to the transmission system is greater than 20.

[0013] Furthermore, the adaptive alignment mechanism constructs triplets from healthy samples and pseudo-fault samples under different operating conditions to achieve effective alignment and differentiation of features. In the design of each triplet, the adaptive alignment mechanism selects a healthy sample from a specific operating condition as the anchor point, simultaneously selects a healthy sample from another operating condition as the positive sample, and includes multiple pseudo-fault samples as negative samples, and designs an isometric constraint function. Taking the pseudo-fault samples as the adjustment reference points, dynamically optimize the spatial positions of the anchor point and the positive sample: when the distance between the anchor point and the positive sample is exactly equal to their respective distances from multiple negative samples, the system will move the anchor point and the positive sample to a more compact sub-region, similar to precisely calibrating them onto the same hypersphere; at the same time, by restricting the excessive closeness between negative samples, while enhancing the close connection between healthy samples, the clear boundary between similar samples and dissimilar samples is maintained.

[0014] Furthermore, an activation function is used in the Irmogorov-Arnold network relationship evaluator to transfer information in each layer, capture the complex dependencies between feature pairs, and reveal the common characteristics in the healthy state, thereby distinguishing the subtle feature differences between the healthy state and the pseudo-fault state; through the weighted aggregation of the first few layers of information, the relationship score of the feature pair is generated, providing a reliable quantitative basis for fault detection. Through the weighted aggregation of the first few layers of information, the relationship score of the feature pair is generated.

[0015] Furthermore, the healthy boundary is 0.9-0.99.

[0016] In general, compared with the prior art, the above technical solution conceived by the present invention can achieve the following beneficial effects.

[0017] Existing intelligent fault detection technology usually relies on a certain amount of fault data, and requires that the distribution of training data and test data be consistent in order to achieve effective fault detection. However, in actual engineering, fault samples of train transmission systems are extremely scarce. At the same time, the operating conditions are dynamically affected by various factors such as speed, load, environment and train configuration, which makes it difficult to maintain consistent data distribution. This seriously limits the application of intelligent detection technology in train transmission systems. The present invention proposes an innovative solution that can achieve accurate state identification without relying on a large number of fault samples. By constructing a pseudo-fault sample library, the diversity of samples is significantly enhanced, the model's over-sensitivity to health status is effectively reduced, and the ability to identify fault status is greatly improved; in response to the problem of inconsistent distribution of healthy samples caused by operating condition fluctuations, an adaptive alignment mechanism is designed to map healthy samples under different operating conditions to a unified feature space, thereby reducing the impact of operating condition changes on sample distribution; by constructing feature pairs between healthy features and pseudo-fault samples, the operating condition attribute characteristics and the essential attribute characteristics of the health status are deeply explored to improve the model's adaptability to operating condition changes; for the core components of the train transmission system (such as traction motors, transmission gearboxes and axle boxes), a divide-and-conquer strategy is adopted to train exclusive unified domain health relationship learning frameworks separately to avoid interference with detection accuracy caused by complex dependencies between components; the feature pairs of online monitoring samples of each component and their healthy samples and their similarity measurements are used to achieve accurate detection and positioning of the train transmission system status type. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to make the technical details of this technical solution clearer and easier to understand, the following will briefly describe the basic situation of the drawings related to this technical solution. It should be noted that the drawings shown here are only schematic diagrams of some typical embodiments of this technology, and technicians in the relevant field can derive the supporting drawings required for other implementation methods based on the illustrated content without creative work.

[0019] Figure 1 It is a schematic diagram of constructing a pseudo-fault sample library provided by the implementation of the present invention;

[0020] Figure 2 It is a schematic diagram of the process of the offline training stage of the fault detection and location method for the train drive system based on the unified domain health relationship learning framework provided by the implementation of the present invention.

[0021] Figure 3 is Figure 2 a schematic diagram of the residual shrinkage network involved in the unified health domain relationship learning framework in

[0022] Figure 4 is Figure 2 a schematic diagram of the adaptive alignment mechanism involved in the unified health domain relationship learning framework in

[0023] Figure 5 is Figure 2 a schematic diagram of the attribute feature mining mechanism involved in the unified health domain relationship learning framework in

[0024] Figure 6 is Figure 2 a schematic diagram of the Kolmogorov-Arnold network relationship evaluator involved in the unified health domain relationship learning framework in

[0025] Figure 7 It is a schematic diagram of the process of the online detection stage of the fault identification and location method for the train drive system of the unified health domain relationship learning framework provided by the implementation of the present invention. Detailed implementation manners

[0026] To make the objectives, technical solutions and advantages of the present invention clearer and more thorough, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be emphasized that the specific embodiments described herein are only for explaining the present invention, rather than limiting its scope. In addition, the technical features involved in each implementation manner can be flexibly combined without conflict to give full play to their synergistic effects.

[0027] Please refer to Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 、 Figure 5 、 Figure 6 and Figure 7 , the fault detection and location method for the train drive system based on the unified domain health relationship learning framework provided by the present invention. The present invention mainly includes an offline training stage and an online fault detection stage:

[0028] 1) Offline training stage

[0029] Step 1: Systematically collect historical health samples of multiple key components (such as motors, gearboxes, axle boxes, etc.) of the train drive system under different working conditions, denoted as where C K represents the Kth operating condition and is divided into a training set and a validation set according to a scientific ratio; based on the health samples, a pseudo-fault sample generator is used to generate pseudo-fault samples distributed outside the boundaries of the health samples. At the same time, diverse fault samples are collected from equipment with similar functions to the drive system and adapted and transformed to generate pseudo-fault samples applicable to this system;

[0030] Specifically, as Figure 1 shown, the pseudo-fault sample library D F is composed of external equipment fault samples D E and out-of-distribution samples D O . The external equipment fault samples D E are constructed by migrating historical fault data of similar equipment (such as train drive systems of different models), covering typical fault modes such as gear fractures and bearing wear, ensuring the extensiveness of the samples. The out-of-distribution samples D O are generated based on the health state samples under different operating conditions through an improved soft Brownian shift method, where the generation mechanism process:

[0031]

[0032] In the formula, α j is the perturbation radius, is the normalized random direction vector, and σn j is the noise related to the operating condition. Using the perturbation radius, the normalized random direction vector, and the noise related to the operating condition, the offset boundary is dynamically adjusted to generate pseudo-fault samples D O near the healthy feature space. In order to adapt to the characteristic distribution of health samples under variable operating conditions, a condition-adaptive rejection probability is introduced:

[0033]

[0034] In the formula, d * is the current offset distance, d - is the historical minimum offset distance, and the recommended parameters are k = 7 and σ ∈ [0, 1]. This design ensures that the pseudo-fault samples form a condition-adaptive clustering along the outer edge of the healthy distribution, being both close to the potential fault state and avoiding confusion with the healthy samples. Integrate the external fault samples D E with the out-of-distribution samples D O to form the pseudo-fault sample library D F .

[0035] Step 2: Use continuous wavelet transform to convert the health samples and pseudo-fault sample library of the train drive system into wavelet diagrams;

[0036] Specifically, from the healthy sample library D of the train drive system H and the pseudo-fault sample library D F mine the representative time-frequency domain bands of their data and convert them into wavelet diagrams. The conversion process of the wavelet diagrams is as follows:

[0037]

[0038] In the formula, the variables and τ are the scale and offset respectively. f(t) represents the vibration signal, represents the wavelet basis function.

[0039] Step 3, as Figure 2 shown, the unified domain health relationship learning framework includes a residual shrinkage feature extraction module, an adaptive alignment mechanism, and an attribute feature mining mechanism. Use the residual shrinkage network to remove the noise interference in the healthy sample library and pseudo-fault sample library of the train drive system and extract highly representative high-dimensional features;

[0040] Specifically, use the soft threshold function of the residual shrinkage network to denoise the input signal, decompose the input signal into wavelet feature maps, and through the designed threshold filtering mechanism, gradually approach the noise components hidden in the feature maps to zero. Dependent operations:

[0041]

[0042] In the formula, x is the input, y is the output, and τ is the softening threshold. Through the decomposition of the input signal, as well as the in-threshold filtering and reconstruction of the decomposed signal, irrelevant interference information is effectively removed. In addition, the soft threshold function will take the derivative of the above formula, so that the gradient after derivation is only 0 and 1, avoiding the situations of gradient explosion and gradient disappearance. Derivation operation:

[0043]

[0044] As Figure 3 shown, the residual shrinkage unit uses a unique module to calculate the softening threshold. This unique module uses GAP to calculate the absolute value of the wavelet diagram x to generate a one-dimensional vector. Then, the one-dimensional vector is propagated by two fully connected layers to generate a scaling parameter. The sigmoid function can expand the scaling parameter to the range of 0 to 1.

[0045]

[0046] In the formula, a is the scaling parameter of z corresponding to the input of the two fully connected layers. Then this softening threshold τ can be calculated through the following calculation formula

[0047]

[0048] In the formula, i, j, and c are the width index, height index, and channel index of the wavelet feature map x respectively, average is the average calculation, so that the softening threshold τ is within the range of 0 to 1. For this purpose, the residual shrinkage network can extract representative high-dimensional features of the input samples. For the healthy sample library of the train drive system and the pseudo-fault sample library D F as the input, the corresponding output of the residual shrinkage network is the health state features of the train drive system and the fault features of the pseudo-fault sample library

[0049] Step three, as Figure 4 shown, the adaptive alignment mechanism uses the health features under a certain working condition as the anchor point, selects the health features from another working condition as the positive samples, and at the same time uses the pseudo-fault features as the negative samples. By designing the boundary reconstruction loss, it shortens the distance between the positive samples and the anchor point and pushes the distance between the negative samples and the anchor point, unifies and aligns the health features under different working conditions to a shared space boundary, and isolates the pseudo-fault samples at the same time, effectively eliminating the influence of working condition fluctuations on the distribution of health features;

[0050] Specifically, the adaptive alignment mechanism weaves the health features from different working conditions with the pseudo-fault features into a dynamic triple structure In each triple, a healthy sample is selected from a specific working condition as the anchor point A healthy sample is selected from another working condition as the positive sample F and several pseudo-fault samples N are introduced as the negative samples. In order to further ensure that the similarity structure between healthy samples does not distort during cross-working condition adjustment, a boundary reconstruction mechanism is introduced. Taking the pseudo-fault samples as the reference points, by precisely regulating the sample distribution in the feature space, it promotes the healthy samples to gather closely and maintain a coordinated and consistent structure. When the distances from the anchor point to the same negative sample are denoted as d1 and d2 respectively, if d1 = d2, the anchor point and the positive sample will be guided to a more compact sub-region, pulling the two to the hypersphere centered on the negative sample, which not only enhances the internal connection between healthy samples, but also effectively maintains the clarity and compactness of the relationship between similar samples by restricting the excessive closeness between negative samples. The core of the boundary reconstruction mechanism lies in the isometric constraint function, and its mathematical expression:

[0051]

[0052] Wherein, K is the total number of operating conditions, and d(·, ·) represents the distance metric in the feature space. By quantifying the difference between the distances from the anchor points and positive samples to negative samples, the healthy samples tend to be evenly distributed in the feature space.

[0053] Step Three, Figure 5 As shown, the attribute feature mining mechanism splices the health state features of multiple operating conditions to generate feature pairs to reveal the condition attributes of the health state, and forms new feature pairs by splicing the health state features with pseudo-fault samples to reveal the essential attributes of the health state;

[0054] Specifically, splicing the health features of different conditions with each other reveals the condition attributes of the health state under various operating conditions. The splicing operation is as follows:

[0055]

[0056] Wherein is to perform feature splicing in the dimension. This splicing operation generates feature pairs to reveal the condition attributes of the health state under different operating conditions. Subsequently, by splicing the health state features and pseudo-fault features the essential attribute features of the health state are revealed. The splicing operation is as follows:

[0057]

[0058] generates feature pairs to reveal the essential attributes of the health state. These feature pairs not only reflect the multi-condition attributes of the health state but also enhance the model's ability to distinguish between the health state and the pseudo-fault state.

[0059] Step Five, as Figure 6 shown, construct a Kolmogorov-Arnold network relationship evaluator to mine the internal associations between features and calculate the relationship scores to comprehensively characterize the relationship between the health state and the pseudo-fault state;

[0060] Specifically, the Kolmogorov-Arnold network relationship evaluator gradually mines the internal relationships of feature pairs through multi-layer non-linear transformations, and passes information through the designed activation function at each layer of the network. Its expression is:

[0061]

[0062] Wherein g (l，i) represents the activation value of the i-th neuron in the l-th layer, It is the corresponding activation function. This iterative process enables the Kolmogorov - Arnold network relationship evaluator to deeply explore the potential relationships between feature pairs, not only accurately capturing the common features within the healthy state but also effectively distinguishing the significant differences between the healthy state and the pseudo - fault state, providing a reliable basis for fault detection. The activation function of the Kolmogorov - Arnold network relationship evaluator is a composite form of the basis function and the spline function:

[0063] φ(g) = ω b (g)+ω s ·spline(g)

[0064] In the formula, ω b and ω s are the weight coefficients of the basis function and the spline function respectively. The basis function b(g) selects the SiLU activation function, and its expression is:

[0065]

[0066] While the spline function SiLU(g) is realized through the linear combination of B - spline basis functions:

[0067] spline(g) = ∑ i c i B i (g)

[0068] In the formula, c i is the trainable coefficient, and B i (g) represents the i - th B - spline basis function. The Kolmogorov - Arnold network relationship evaluator calculates the relationship score by aggregating the information of the previous layers, and its calculation process is

[0069]

[0070] In the formula, GELU is used as the activation function, BN represents batch normalization, and Concat is the feature concatenation operation, which depicts the interaction between features. The value range of the relationship score r is limited between [0, 1], and it is used to quantify the possibility that the feature pair belongs to the same category. For different operating conditions, the Kolmogorov - Arnold network relationship evaluator generates the corresponding relationship score matrix:

[0071]

[0072] In the formula represents the predicted relationship score matrix under the n - th operating condition, reflecting the relationship distribution between the healthy state and the pseudo - fault state. Further construct the true similarity relationship label:

[0073]

[0074] Next, the mean square error is used to calculate the error of the shrinkage attention relationship network, and the calculation process is as follows

[0075]

[0076] Combine and to form the total loss Then, the optimizer Adme is used to optimize the unified domain health relationship learning framework to minimize the loss function to the minimum.

[0077] Step 6: To determine the health boundary value of the unified domain health relationship learning framework, the validation dataset is used to set the health boundary value, which is convenient for judging the fault state and occurrence location of the train drive system in the online detection stage.

[0078] Specifically, first define the health boundary value range as 0.9 - 0.99, then input the validation dataset into the unified domain health relationship learning framework, and compare the output relationship score of the unified domain health relationship learning framework with the health boundary value. If it is greater than the health boundary value, it is considered a healthy state; if it is less than, it is considered an unhealthy state. Then, the accuracy evaluation index is used to find the health boundary value that maximizes the accuracy of the unified domain health relationship learning.

[0079] 2) Online fault detection stage:

[0080] Step 1: As Figure 7 shown, the monitoring signals of each key part of the train drive system are collected in real time as the dataset to be detected where represents the monitoring data of the Kth key component. At the same time, the system introduces historical health state samples as a reference, denoted as where contains the known health state data of the Kth key component under various operating conditions;

[0081] Step 2: The system uses the residual shrinkage feature extraction module to process the health state signal and the real-time monitoring signal respectively, to generate the corresponding health state features and

[0082] Step 3: The health state feature and the monitoring feature are concatenated into a feature pair and fed into the Kolmogorov - Arnold network relationship evaluator to deeply mine the internal relationship of the features;

[0083] Step 4: The Kolmogorov - Arnold network relationship evaluator calculates the relationship score between the health state features and the monitoring features.

[0084] Step 5: Compare the calculated relationship score with the health boundary value. If it is greater than or equal to the health boundary value, the sample to be detected is considered to be in a healthy state; otherwise, it is in a faulty state. Perform the above operations on each key component of the transmission system to achieve fault detection and location of the transmission system.

[0085]

[0086] To highlight the advantages of the method of the present invention in fault detection and location, the present invention conducts comparative experiments with deep convolutional neural networks, domain adversarial neural networks, deep sub - domain adaptation networks, and one - dimensional convolutional autoencoders. Table 1 shows the comparison results of the detection accuracy and recall rate among them. It can be seen from the table that the fault detection accuracy and recall rate of the present invention are significantly higher than those of the other five methods. Method Detection accuracy Recall rate Deep convolutional neural network 42.65 42.98 Domain adversarial neural network 62.35 63.12 Deep sub-domain adaptation network 63.45 64.05 One-dimensional convolutional autoencoder 73.56 74.12 The method of the present invention 92.26 93.24

Claims

1. A method and system for detecting and locating faults in a train transmission system under variable operating conditions, characterized in that: include: S1. Construction of pseudo-fault sample library: Based on the healthy samples collected from the high-speed train transmission system under various operating conditions, a pseudo-fault sample generator is designed to generate pseudo-fault samples outside the distribution of healthy samples; at the same time, a variety of fault samples are collected from devices with similar functions to the transmission system, and pseudo-fault samples suitable for this system are generated through adaptive conversion; these two types of pseudo-fault samples together constitute a diverse pseudo-fault sample library, which effectively alleviates the problem that the model is too sensitive to small disturbances in the health state when relying only on healthy samples, and lays a solid data foundation for subsequent fault detection; On this basis, the pseudo-fault sample library and most of the historical healthy samples are used as training data sets, and a part of the historical healthy samples are used as verification data sets; Extract time-frequency domain features from the training data set and validation data set respectively to capture the time and frequency characteristics of the signal and provide high-quality input for model training; S2. Construct a unified domain health relationship learning framework; The unified domain health relationship learning framework includes a residual shrinkage feature extraction module, an adaptive alignment mechanism, and an attribute feature mining mechanism; the residual shrinkage feature extraction module extracts more representative health status and pseudo-fault status features by eliminating noise interference in training data; The adaptive alignment mechanism uses the healthy features under a certain working condition as anchor points, selects healthy features from another working condition as positive samples, and uses pseudo-fault features as negative samples. By designing boundary reconstruction loss, the distance between positive samples and anchor points is shortened, and the distance between negative samples and anchor points is pushed away. Thus, the healthy features under different working conditions are uniformly aligned to a shared spatial boundary, while pseudo-fault samples are isolated, eliminating the impact of working condition fluctuations on the distribution of healthy features. The attribute feature mining mechanism splices the health status features of multiple operating conditions to generate feature pairs to reveal the condition attributes of the health status, and splices the health status features with pseudo-fault samples to form new feature pairs to reveal the essential attributes of the health status. The spliced ​​feature pairs are input into the Kolmogorov-Arnold network relationship evaluator to mine the intrinsic associations between the features and calculate the relationship scores, comprehensively characterizing the relationship between the health status and the pseudo-fault status. S3. Based on the core components of the train transmission system (such as traction motors, transmission gearboxes, and axle boxes), a divide-and-conquer strategy is adopted to repeat steps S1 and S2 for each component, and a dedicated unified domain health relationship learning framework is trained separately to avoid the interference of complex dependencies between components on detection accuracy; the predicted relationship scores of each component are integrated into a relationship score matrix for subsequent optimization and state judgment; S4. Take the boundary reconstruction loss and the error between the true relationship score matrix and the predicted relationship score matrix as the joint optimization objectives, and use the training data set to iteratively optimize the framework; input the validation data set into the trained model, and determine the healthy boundary value of each component as the threshold for state discrimination with the goal of maximizing the model accuracy; S5. For the sample to be tested and its corresponding historical healthy sample, extract the time-frequency domain features and input them into the unified domain health relationship learning framework of the corresponding component to calculate the similarity score; compare the score with the healthy boundary value. If the relationship score is greater than or equal to the boundary value, it is judged as a healthy state; otherwise, it is judged as a fault state. This process can achieve accurate judgment of the status of each component.

2. A method and system for detecting and locating faults in a train transmission system under variable operating conditions according to claim 1, characterized in that: The pseudo-fault sample generator is based on health status samples of different operating conditions. It applies disturbances in random directions and introduces noise related to the operating conditions to generate samples that are near the feature space of healthy samples but beyond their distribution range. In order to adapt to the characteristic changes of healthy samples under different operating conditions, a condition-adaptive rejection probability mechanism is introduced to dynamically adjust the offset boundary to ensure that the generated pseudo-fault samples neither deviate too much from the healthy distribution nor are too close to it, thereby forming condition-specific pseudo-fault clustering.

3. A method and system for detecting and locating faults in a train transmission system under variable operating conditions according to claim 1, characterized in that: The number of pseudo fault categories generated by the pseudo fault sample generator for each component is greater than 15, and the number of fault categories of equipment with similar functions to the transmission system is greater than 20.

4. A method and system for detecting and locating faults in a train transmission system under variable operating conditions according to claim 1, characterized in that: The adaptive alignment mechanism constructs triplets from healthy samples and pseudo-fault samples under different working conditions to achieve effective alignment and differentiation of features. In the design of each triplet, the adaptive alignment mechanism selects healthy samples from a specific working condition as anchor points, selects healthy samples from another working condition as positive samples, and includes multiple pseudo-fault samples as negative samples. An isometric constraint function is designed, and the pseudo-fault samples are used as the reference point for adjustment to dynamically optimize the spatial position of the anchor points and positive samples: when the distance between the anchor point and the positive sample is exactly equal to the distance between each of them and multiple negative samples, the system will migrate the anchor point and the positive sample to a more compact sub-region, similar to calibrating them precisely on the same hypersphere; at the same time, by limiting the excessive proximity between negative samples, the close connection between healthy samples is enhanced while maintaining a clear boundary between similar samples and heterogeneous samples.

5. The method and system for detecting and locating faults in a train transmission system under variable operating conditions according to claim 1, characterized in that: An activation function is used in the Ilmogorov-Arnold network relationship evaluator to transfer information in each layer, capture the complex dependencies between feature pairs, and reveal the common characteristics in the healthy state, thereby distinguishing the subtle feature differences between the healthy state and the pseudo-fault state; By weighted aggregation of the first few layers of information, the relationship scores of feature pairs are generated, providing a reliable quantitative basis for fault detection. By weighted aggregation of the first few layers of information, the relationship scores of feature pairs are generated.

6. A method and system for detecting and locating faults in a train transmission system under variable operating conditions according to claim 1, characterized in that: The healthy boundary is 0.9-0.

99.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, a method for detecting and locating faults in a train transmission system under variable operating conditions as described in any one of claims 1 to 6 is implemented.

Images