Incremental fault diagnosis method for heavy haul train bogie

Abstract

The invention discloses a similar incremental fault diagnosis method for a heavy-load train bogie, which relates to the technical field of heavy-load train bogie fault diagnosis and comprises the steps of data preparation, model construction, incremental training and online diagnosis. Incremental training updates a basic diagnosis model by combining empirical playback of forgetting risk weighting, a decoupling classification loss function including old class retention items and asymmetric spherical regularization. According to the method, catastrophic forgetting in incremental learning is effectively inhibited, the recognition capability of the model for learned old fault categories is maintained, meanwhile, the new fault categories can be accurately learned, and online, self-adaptive and long-term stable operation of the heavy-load bogie fault diagnosis model is realized.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for heavy-haul train bogies, and in particular to a quasi-incremental fault diagnosis method for heavy-haul train bogies. Background Technology

[0002] Heavy-haul railway vehicles operate under long-term high loads and complex track conditions. As a key load-bearing and guiding structure, the bogie is prone to fatigue, stiffness reduction, and leakage in its primary and secondary suspension springs and dampers. If these faults are not identified in time, they can lead to increased vibration, accelerated damage to structural components, and even reduced wheel-rail contact safety margin.

[0003] Existing intelligent fault diagnosis methods are mostly based on offline modeling, assuming that all fault categories have been mastered during the training phase. However, in actual heavy-haul railway operations, fault categories appear gradually over time, and the differences in operating conditions between different lines and vehicles lead to the continuous emergence of new faults. Traditional models must be retrained when encountering new categories, otherwise they are prone to forgetting knowledge of old categories; at the same time, the number of samples of new and old categories is extremely imbalanced, causing the classifier to be biased towards new categories; in addition, the feature space is prone to drift during multiple incremental learning processes, causing the representation of old categories to gradually become ineffective.

[0004] Based on the above problems, there is an urgent need for an incremental fault diagnosis technology that can continuously learn, continuously diagnose, and maintain long-term stability in order to adapt to the actual operating scenarios of heavy-haul railway bogies. Summary of the Invention

[0005] To address the issues of existing offline diagnostic models' inability to identify novel faults and their tendency to forget old knowledge during incremental learning, this invention proposes a quasi-incremental fault diagnosis method for heavy-load train bogies. By constructing an online incremental learning framework, the diagnostic model can update and retain knowledge without retraining the entire dataset when new fault categories appear, thus solving the aforementioned problems.

[0006] This application discloses a quasi-incremental fault diagnosis method for bogies of heavy-haul trains, comprising the following steps: S1. Collect the vibration acceleration and displacement signals of the bogie, and perform preprocessing to obtain time series samples; S2. Construct a basic diagnostic model that includes a feature extractor and a classifier; S3. Based on the experience playback of the forgetting risk weighted reservoir sampling, establish and maintain a fixed-capacity memory buffer to retain old class samples; S4. Construct a decoupled classification loss function that combines the old class preservation mechanism. The decoupled classification loss function includes a binary cross-entropy loss function and an old class output preservation term. S5. Add an asymmetric spherical regularization term to the loss function to constrain the magnitude of the feature vectors and enhance the inter-class angular separability. S6. Jointly optimize the decoupled classification loss function and the asymmetric spherical regularization term, update the network parameters, and obtain the incrementally updated diagnostic model. S7. Use the incrementally updated diagnostic model to classify the bogie signals acquired in real time.

[0007] Preferably, the basic diagnostic model includes a feature extractor and a classifier; The feature extractor uses a one-dimensional convolutional network, including multiple convolutional layers, ReLU activation layers, and max pooling layers, to map the input time series into a high-dimensional feature vector; The classifier includes a fully connected layer connected to the back end of the feature extractor, and the number of nodes in the classifier output layer dynamically expands as the number of fault categories increases.

[0008] Preferably, the experience replay based on the forgetting risk-weighted reservoir sampling in S3 includes: For the One arriving sample The forgetting risk score is:

[0009] in, Describing the L1 norm, For the current model to samples Predicted probability distribution For the previous stage model of samples The predicted probability distribution; sample The probability of actually replacing a random sample in the buffer is:

[0010] in, For the buffer capacity, This is a sample index variable used to iterate through all samples reached up to the current time. In order to achieve the first step in the incremental learning process One arriving data sample, For the first bogie sample The forgetting risk score.

[0011] Preferably, during incremental training, old class samples are drawn from the buffer to construct training batches, and the drawing process satisfies the following: Control the balance of the number of samples in each old category; Prioritize selecting samples that are larger in distance from the corresponding class center in the feature space; Increase the probability of samples with high forgetting risk scores being selected.

[0012] Preferably, the binary cross-entropy loss function is as follows:

[0013] in, For category index variables, This represents the total number of categories in the model at the current stage. For the first The indicator value of the real label, The model represents the first The original output of the class, Use the Sigmoid activation function; The old class output retention items are as follows:

[0014] in, For the first to participate in the current training computation Index of each sample, This is a summary of the samples that participated in the loss calculation in the current training iteration. This is the set of old classes that have already been learned in the previous incremental phase. For the old class collection Category index in For the current model to samples The logit output, For the previous stage of freezing the model on samples The logit output.

[0015] Preferably, the asymmetric spherical regularization term constrains the magnitude of the eigenvector, and the calculation formula is as follows:

[0016] in, For the first Each sample feature vector For sample feature vectors The length of the mold, Let be the dynamic threshold radius, representing the th . The feature modulus threshold radius of the sub-incremental stage; The formula for calculating the dynamic threshold radius is as follows:

[0017] in, The initial threshold radius, This is the attenuation coefficient.

[0018] Preferably, the asymmetric spherical regularization term is subject to an asymmetric penalty, specifically: For modulus greater than the current dynamic radius threshold The gradient penalty applied to the eigenvector is greater than the threshold for modulus length less than the current dynamic radius. The intensity of the gradient penalty applied to the eigenvectors.

[0019] Preferably, the loss function after joint optimization by S6 is:

[0020] in, and This is the balance coefficient.

[0021] Preferably, the incremental update process employs a two-stage training mechanism, including: Rapid adaptation phase: Freeze the parameters of the feature extractor and update only the parameters of the classifier to quickly obtain the initial ability to identify new fault categories; Stabilization and consolidation phase: Unfreeze all network parameters, and perform joint optimization training on the basic diagnostic model by combining the old class samples replayed from the memory buffer and the asymmetric spherical regularization term.

[0022] Preferably, when constructing training batches in incremental training, a class balancing strategy is used to dynamically adjust the ratio of new class samples to old class samples in the batch, so that the gradient contributions of new and old classes in training are relatively balanced.

[0023] The beneficial effects of this invention are: (1) This invention can effectively suppress catastrophic forgetting and maintain the ability to identify old faults. Experience replay combined with BCE decoupled update enables the model to strengthen the old class representation while learning new faults, solving the problem that traditional offline models cannot continuously diagnose.

[0024] (2) This invention enhances the ability to balance learning between new and old categories. BCE avoids the problem of excessive gradients for new categories caused by Softmax, so that the model is no longer biased towards fault categories that appear in the incremental stage.

[0025] (3) The feature distribution of the present invention is more stable and the separability is stronger. ASR restricts the features to the vicinity of the sphere, so that each category is mainly distinguished by the angle, which improves the interpretability and robustness of the model in long-term operation.

[0026] (4) This invention is applicable to streaming data scenarios in actual heavy-haul railways. The reservoir sampling provides unbiased, low-storage historical data replay, which is suitable for the limited storage conditions in railway sites.

[0027] (5) The overall diagnostic method of the present invention has the advantages of online learning, self-adaptation and low maintenance. Attached Figure Description

[0028] Figure 1 This is a flowchart of an incremental fault diagnosis method for heavy-load train bogies according to an embodiment of the present invention. Figure 2 This is the feature distribution of the present invention that has not undergone ASR constraints. Figure 3 This is the ASR-constrained feature distribution of an embodiment of the present invention; Figure 4 This is a structural diagram of the overall incremental diagnosis in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments.

[0030] This application discloses a quasi-incremental fault diagnosis method for bogies of heavy-haul trains, the process of which is as follows: Figure 1 As shown, it includes the following steps: S1. Collect the vibration acceleration and displacement signals of the bogie, and perform preprocessing to obtain time series samples.

[0031] Vibration acceleration and displacement signals (such as vertical / lateral acceleration of the car body and frame) of heavy-haul railway bogies are collected using sensors. The raw signals are sliced ​​using a sliding window to obtain time-series samples of fixed length. The sliced ​​samples are then standardized (Z-score normalization) to eliminate dimensional differences between different sensors.

[0032] S2. Construct a basic diagnostic model that includes a feature extractor and a classifier.

[0033] A one-dimensional convolutional neural network (1D-CNN) is constructed as the feature extractor, with a fully connected layer at the back end serving as the classifier. The feature extractor uses a one-dimensional convolutional network, including multiple convolutional layers, ReLU activation layers, and max-pooling layers, to map the input time series into a high-dimensional feature vector. The number of nodes in the classifier's output layer dynamically expands as the number of fault categories increases.

[0034] S3. Based on the experience playback of the forgetting risk weighted reservoir sampling, establish and maintain a fixed-capacity memory buffer to retain old class samples.

[0035] To preserve old knowledge, a fixed-capacity memory buffer is established. Unlike traditional experience replay, this embodiment, considering the long-term operation, large data volume, and limited storage of heavy-haul railways, adds a class-balanced resampling mechanism, a representative sample retention mechanism, and a forgetting risk measurement mechanism based on model output changes to the existing reservoir sampling strategy. When certain fault categories occur with extremely low frequency, a sample selection strategy based on feature diversity is adopted, and samples are weighted and filtered using forgetting risk scores. This ensures that the old class samples in the buffer maintain distribution diversity and their importance to model stability within a limited space, further reducing the risk of forgetting old classes during incremental training.

[0036] The buffer is maintained using a reservoir sampling algorithm: for the first... If a sample arrives and the buffer is not full, it is stored directly; otherwise, the buffer is full (its capacity is...). ), then The probability of replacing a random sample in the buffer is [a certain value].

[0037] Building upon the traditional "equal probability replacement" principle of water storage tanks, this embodiment further introduces a forgetting risk score to characterize the likelihood of the sample being forgotten in the current incremental phase, and adjusts its actual write / replacement probability accordingly. Specifically, for the first... One arriving sample The probability distributions are calculated using the previous stage frozen model and the current model, respectively, and denoted as follows: and Then the forgetting risk score of this sample is defined as:

[0038] in, Describing the L1 norm, For the current model to samples Predicted probability distribution For the previous stage model of samples The predicted probability distribution.

[0039] A higher forgetting risk score indicates a greater change in the current model's prediction for that sample compared to the previous stage, meaning the sample is more likely to be "forgotten" during incremental learning or its classification boundary will be significantly perturbed. Based on this, this embodiment sets the actual replacement probability of a sample as a weighted sum of the base probabilities of a traditional reservoir, as shown in the following formula:

[0040] in, For the buffer capacity, This is a sample index variable used to iterate through all samples reached up to the current time. In order to achieve the first step in the incremental learning process One arriving data sample, For the first bogie sample The forgetting risk score.

[0041] The maximum forgetting risk among the observed samples in the current stage is used for normalization. This ensures that, while maintaining the overall unbiased sampling property, samples with higher forgetting risk are more likely to be written into or retained in the buffer, while samples with lower risk and less contribution to model stability are more likely to be replaced.

[0042] The above strategy not only ensures that the samples in the buffer maintain an approximately unbiased distribution over a long period of time, avoiding overfitting to a specific category, but also prioritizes the retention of historical samples that are more critical to combating catastrophic forgetting through a forgetting risk weighting mechanism.

[0043] When training a new task, old class samples are drawn from the buffer and mixed with new class samples for the current task to form a mini-batch. Unlike the traditional method of "randomly drawing old class samples from the buffer," this embodiment comprehensively considers class balance, feature diversity, and forgetting risk scores during the drawing process. At the category level, the approximate balance of samples from each old category in the mini-batch is controlled to avoid some old categories from lacking updates for a long time.

[0044] At the feature level, samples that are far from the class center in the feature space are preferred to improve the diversity and representativeness of the samples in the buffer.

[0045] At the stability level, by combining the forgetting risk score, the probability of high-risk (more easily forgotten) old class samples being drawn into the mini-batch is increased. Specifically, based on a relative comparison of the forgetting risk scores of the samples observed in the current stage, samples with relatively larger forgetting risk scores are considered high (forgotten) risk samples, and samples with relatively smaller forgetting risk scores are considered low (forgotten) risk samples.

[0046] For old samples in the buffer Forgetting risk score has been calculated Convert it to sampling weights:

[0047] In this embodiment Set it to 10-6 to prevent the weight from being 0; This represents the set of all old samples within the buffer. For any old sample in the buffer, This is the forgetting risk score for this sample. Sampling without replacement is performed based on probability; select... The probabilities of entering the mini-batch are as follows:

[0048] Due to sampling weight Follow As the number of samples increases monotonically, the older sample with the higher forgetting risk score is more likely to be included in the mini-batch.

[0049] S4. Construct a decoupled classification loss function that combines the old class preservation mechanism. The decoupled classification loss function includes a binary cross-entropy loss function and an old class output preservation term.

[0050] Instead of the traditional Softmax cross-entropy loss, we adopt the Binary Cross-Entropy (BCE) loss function for classification. Furthermore, we introduce an "old class preservation term" to ensure the stability of the old class output, thus constructing an improved loss function suitable for incremental class learning scenarios.

[0051] The binary cross-entropy loss function is as follows:

[0052] in, For category index variables, This represents the total number of categories in the model at the current stage. For the first The indicator value of the real label, The model represents the first The original output of the class, The Sigmoid activation function is expressed as: , soon real number Map to the interval (0, 1).

[0053] BCE transforms a multi-class classification problem into multiple independent binary classification problems, allowing the output channel of each class to be optimized independently. In this way, the competition caused by the normalization of probabilities of each class in Softmax is eliminated, and the problem of gradient being dominated by the new class due to the imbalance of the number of new and old samples in incremental learning is solved.

[0054] Building upon the aforementioned BCE decoupling loss, to further suppress the drift of the old class decision boundary during incremental learning, this invention introduces the old class output preservation term, Cold. Let Cold be the set of old classes already learned in the previous incremental stage. For each of these classes, the value of the frozen model in the previous stage is recorded in the sample... logit output And the logit output corresponding to the current model. The old class retainable item is defined as:

[0055] in, For the first to participate in the current training computation Index of each sample, This is a summary of the samples that participated in the loss calculation in the current training iteration. This is the set of old classes that have already been learned in the previous incremental phase. For the old class collection Category index in For the current model to samples The logit output, For the previous stage of freezing the model on samples The logit output.

[0056] This measure prevents excessive disturbance to the old class discrimination boundary when learning new fault categories by explicitly constraining the offset of the current model from the previous model in the old class output logit, thereby maintaining the stability of the old class recognition ability during parameter update.

[0057] S5. Add an Asymmetric Spherical Regularization (ASR) term to the loss function to constrain the magnitude of the feature vectors and enhance the angular separability between classes.

[0058] To prevent feature space drift, an ASR regularization term is added to the total loss function. : Define a regularization term: Set a feature magnitude threshold radius. (For example Only when the sample feature vector Length of the module Exceed Punishment will only be imposed at that time.

[0059] in, For the first Each sample feature vector For sample feature vectors The length of the module.

[0060] Asymmetric mechanisms: For smaller modulus (less than) For features of ), no penalty is imposed, preserving their plasticity to adapt to new fault categories; for modulus lengths larger than 100 mm, the following applies: The characteristics of the property are penalized by "pulling back" through gradients to limit its magnitude expansion.

[0061] Geometric effect: It forces the model to distinguish different fault categories by optimizing the direction (angle) of the feature vector, rather than simply relying on increasing the feature magnitude, thereby enhancing the angular separability of the features.

[0062] To improve the adaptability of ASR, this embodiment further proposes a dynamic radius threshold mechanism: a larger threshold is used in the initial incremental stage. To ensure the model's adaptability to new fault features, the threshold is gradually reduced as the learning epochs increase, causing the feature vectors to gradually approximate a sphere, thereby improving overall inter-class angular stability. The specific formula is as follows: Let the... The threshold radius for the next incremental stage is:

[0063] in, The initial threshold radius is set to 0.5 in this embodiment. This is the attenuation coefficient, used to gradually shrink the radius of the sphere.

[0064] but:

[0065] Furthermore, this embodiment introduces an asymmetric penalty coefficient, allowing the feature modulus expansion and contraction processes to have different gradient magnitudes. For example, a stronger penalty is applied to features with excessively large modulus lengths, while a weaker penalty is applied to features with excessively small modulus lengths, promoting the rapid formation of new class representations without destroying old class features.

[0066] This asymmetric spherical constraint significantly enhances inter-class angular separability and effectively suppresses feature drift during the incremental process.

[0067] Figure 2 This diagram illustrates the feature distribution without ASR regularization as described in this embodiment. It shows that without spherical constraints, feature points are mainly dispersed along the vector magnitude direction, with differences primarily in feature length, while inter-class angular discrimination is weak. Due to the lack of unified geometric constraints, the distribution ranges of different categories in space tend to overlap, and inter-class boundaries are unclear. This illustrates the potentially disordered distribution pattern of the feature space under traditional incremental learning mechanisms.

[0068] Figure 3This diagram illustrates the feature distribution after ASR regularization in this embodiment. It shows that by applying a spherical constraint to the feature modulus, various features are confined to a uniform sphere and distributed along different angular directions, forming a cluster structure with significant directional differences. This diagram aims to illustrate that ASR can enhance the inter-class angular separation effect in the feature space, making different categories more stable and clearly distinguishable in geometric structure.

[0069] S6. Jointly optimize the decoupled classification loss function and asymmetric spherical regularization term, update the network parameters, and obtain the incrementally updated diagnostic model.

[0070] Overall optimization objective (total loss): ,in and The values ​​are the balance coefficients (0.001 and 0.1 in this embodiment).

[0071] Incremental update: Minimize the total loss using mixed data at each incremental stage. Update network parameters.

[0072] Online diagnostics: After training, the updated model is used to classify and output faults from the real-time collected bogie vibration data.

[0073] To further enhance the stability and adaptability of this invention in incremental fault diagnosis, this embodiment proposes an incremental training process optimization method that differs from existing technologies, including: A two-stage incremental training mechanism of "adaptation first, then stabilization" is introduced.

[0074] In each incremental phase, the model first freezes the feature extraction network in the "rapid adaptation phase" and only updates the classifier, enabling new fault categories to quickly acquire preliminary recognition capabilities. Then, in the "stabilization and consolidation phase", all parameters are unfrozen, and joint training is performed using empirical replay samples and ASR regularization, so that the learning of new classes does not destroy the feature structure of old classes, thereby achieving a better balance between plasticity and stability.

[0075] A balanced mini-batch build strategy is adopted.

[0076] To address the issue of a dominant number of new class samples and a very small number of old class samples in incremental learning, this invention dynamically adjusts the ratio of new and old class samples in the training batch to ensure a relatively balanced gradient contribution. This strategy effectively prevents the model from excessively shifting towards the new class, improving the persistence of old class recognition and overall diagnostic stability.

[0077] The optimization of the above training process is significantly different from the direct mixed training or simple playback methods in the prior art, and further improves the adaptability and anti-forgetting ability of the method proposed in this embodiment in incremental scenarios.

[0078] S7. Use the incrementally updated diagnostic model to classify the bogie signals acquired in real time.

[0079] like Figure 4 As shown, the method proposed in this application consists of multiple incremental stages, each stage including three parts: fault sample input, incremental training, and diagnostic output. The left side shows the fault vibration samples collected in the current stage, which are combined with the replay samples in the buffer as training data.

[0080] Figure 4 The "Memory Bank" module shown in the middle and lower left of the diagram is a fixed-capacity sample storage buffer used to store old-class fault samples retained from historical incremental phases. The Memory Bank is dynamically maintained using a reservoir-based sampling and forgetting risk weighting strategy to retain representative historical samples within limited storage space. The "Sample Replay" process, indicated by the arc arrow pointing from the "Memory Bank" to the training module, represents the process where, in the current incremental training phase, some old-class samples are extracted from the Memory Bank and combined with the newly collected fault samples to form training data, mitigating the problem of forgetting old-class knowledge during incremental learning. The "+" symbol between the "Memory Bank" and the fault samples in the current phase indicates that newly collected samples and old-class samples replayed from the Memory Bank are merged during the training phase to form a joint training sample set.

[0081] Figure 4 The middle section shows the incremental training process. The model is initialized with the pre-trained weights from the previous stage and its parameters are updated through joint training with both new and old samples. ASR regularization is introduced during training to constrain the feature space structure and maintain the stability of the inter-class angular distribution. The right side shows the diagnostic model after training. Its identifiable fault categories expand continuously with each incremental stage, while the discrimination boundaries of the original categories remain stable.

[0082] Figure 4 The "Classifier" module on the right represents the classification network located at the back end of the feature extractor. It is used to classify the extracted feature vectors. The number of its output nodes expands dynamically as the incremental stage progresses to adapt to the recognition needs of new fault categories.

[0083] In a specific embodiment, to verify the effectiveness of the method proposed in this application, a multibody dynamics model of a heavy-haul railway bogie was constructed based on the SIMPACK multibody dynamics platform. Five typical operating conditions were set: normal state, primary spring failure, secondary spring failure, primary longitudinal damper failure, and secondary lateral damper failure. Vibration and acceleration signals from 49 sensor channels were collected at a sampling frequency of 200 Hz, and 600 samples were generated for each condition. The obtained samples were divided into training, validation, and test sets in an 8:1:1 ratio. To simulate the gradual emergence of fault categories in actual operation, the five operating conditions were divided into three incremental learning stages, gradually introducing new fault categories, so that the model could retain its ability to identify known faults while learning new faults. All comparison methods used the same convolutional neural network structure and training parameter settings to ensure fairness in the comparison; the comparison algorithms included regularization methods such as GR, LWF, and EWC, as well as representative experience replay and hybrid incremental learning methods such as ER, iCaRL, FastICARL, and DER.

[0084] Experimental results show that the method proposed in this application significantly outperforms existing technologies in terms of overall recognition accuracy and anti-forgetting ability. After completing all incremental stages, the method proposed in this application achieves a final average accuracy of 95.18%, significantly higher than iCaRL's 90.56% and FastICARL's 83.33%; the average forgetting rate is only 2.50%, significantly better than ER (8.61%), FastICARL (13.61%), and iCaRL (6.67%), fully demonstrating that this application can effectively suppress the catastrophic forgetting that is prevalent in incremental learning. Furthermore, in the final test results of the three incremental tasks, the method proposed in this application achieves recognition accuracies of 93.89%, 98.33%, and 93.33%, respectively, all the highest among all methods, demonstrating a comprehensive advantage in both new class adaptation and old class retention.

[0085] Ablation experiments further validated the effectiveness of the method proposed in the embodiments of this application. The average accuracy of the model using only empirical replay was 90.00%, which increased to 91.67% after adding BCE decoupling loss, and to 91.30% after adding ASR. The average accuracy of the model proposed in this application, which combines all three methods, was significantly improved to 95.18%, and the average forgetting rate was reduced to 2.50%. This indicates that empirical replay, BCE loss, and ASR regularization have a significant synergistic effect in suppressing forgetting and enhancing class separability.

[0086] In summary, this embodiment fully demonstrates the effectiveness and significant technical advantages of this application in incremental fault diagnosis tasks for heavy-haul railway bogies. It not only surpasses existing methods in identification accuracy but also maintains stable diagnostic performance during long-term incremental operation, showing promising prospects for engineering applications.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A kind of for heavy haul train bogie incremental type fault diagnosis method, it is characterized in that, The method comprises the following steps: S1, collecting vibration acceleration signals and displacement signals of a bogie and pre-processing to obtain time series samples; S2, constructing a basic diagnostic model comprising a feature extractor and a classifier; S3, based on the experience replay of the forgetting risk weighted reservoir sampling, establishing and maintaining a fixed capacity memory buffer to retain old class samples; S4, constructing a decoupled classification loss function combined with an old class maintenance mechanism, the decoupled classification loss function comprising a binary cross-entropy loss function and an old class output maintenance term; S5, adding an asymmetric spherical regularization term to the loss function to constrain the length of the feature vector and enhance the inter-class angle separability; S6, jointly optimizing the decoupled classification loss function and the asymmetric spherical regularization term to update the network parameters to obtain an incrementally updated diagnostic model; S7, using the incrementally updated diagnostic model to classify the real-time collected bogie signals.

2. The incipient fault diagnosis method for heavy haul train bogie according to claim 1, characterized in that, The basic diagnostic model comprises a feature extractor and a classifier; The feature extractor adopts a one-dimensional convolutional network comprising multiple convolutional layers, ReLU activation layers and maximum pooling layers to map the input time series to a high-dimensional feature vector; The classifier comprises a fully connected layer connected at the back end of the feature extractor, and the number of nodes of the classifier output layer is dynamically expanded with the increase of fault classes.

3. The incipient fault diagnosis method for heavy haul train bogie according to claim 2, characterized in that, The experience replay of the forgetting risk weighted reservoir sampling in S3 comprises: For the first sample to arrive , the forgetting risk score is: wherein, denotes the L1 norm, is the probability distribution predicted by the current model for the sample is the probability distribution predicted by the previous model for the sample is the probability distribution predicted by the previous model for the sample is the probability distribution predicted by the previous model for the sample Sample The probability of actually replacing a random sample in the buffer is: wherein, is the buffer capacity, is a sample index variable used to iterate through all samples reached up to the current time instant, is the forgetting risk score of the data sample reached in the incremental learning process, is the wagon sample .

4. The incipient fault diagnosis method for heavy haul train bogie according to claim 3, characterized in that, Extracting old class samples from the buffer to construct a training batch during incremental training, and the extraction process satisfies: Balancing the number of old class samples; Prioritize samples with a larger distance from the corresponding class center in the feature space; Increase the probability of extracting samples with a high forgetting risk score.

5. The incipient fault diagnosis method for heavy haul train bogie according to claim 4, characterized in that, The binary cross-entropy loss function is as follows: in, For category index variables, This represents the total number of categories in the model at the current stage. For the first The indicator value of the real label, The model represents the first The original output of the class, Use the Sigmoid activation function; The old class output maintenance term is as follows: in, For the first to participate in the current training computation Index of each sample, This is a summary of the samples that participated in the loss calculation in the current training iteration. This is the set of old classes that have already been learned in the previous incremental phase. For the old class collection Category index in For the current model to samples The logit output, For the previous stage of freezing the model on samples The logit output.

6. The incipient fault diagnosis method for heavy haul train bogie according to claim 5, characterized in that, The asymmetric spherical regularization term constrains the length of the feature vector, and the calculation formula is as follows: in, For the first Each sample feature vector For sample feature vectors The length of the mold, Let be the dynamic threshold radius, representing the th . The feature modulus threshold radius of the sub-incremental stage; The calculation formula of the dynamic threshold radius is as follows: wherein, is an initial threshold radius, is a decay coefficient.

7. The incipient fault diagnosis method for heavy haul train bogie according to claim 6, characterized in that, The asymmetric spherical regularization term applies an asymmetric penalty, specifically: the gradient penalty strength applied to the feature vector having a mode length greater than the current dynamic radius threshold is greater than the gradient penalty strength applied to the feature vector having a mode length less than the current dynamic radius threshold .

8. The incipient fault diagnosis method for heavy haul train bogie according to claim 7, characterized in that, The loss function after joint optimization in S6 is: wherein and is a balancing factor.

9. The incipient fault diagnosis method for heavy haul train bogie according to claim 8, characterized in that, The incremental update process adopts a two-stage training mechanism, comprising: Fast adaptation stage: freeze the parameters of the feature extractor and only update the parameters of the classifier to quickly obtain preliminary recognition ability for new fault classes; Stable consolidation stage: unfreeze all network parameters, combine the old class samples replayed from the memory buffer and the asymmetric spherical regularization term to jointly optimize and train the basic diagnostic model.

10. The incipient fault diagnosis method for heavy haul train bogie according to claim 9, characterized in that, During the construction of the training batch in incremental training, a class balancing strategy is used to dynamically adjust the proportion of new class samples and old class samples in the batch, so that the gradient contributions of new and old classes in training are relatively balanced.

Images