Urban rail train gearbox type incremental fault diagnosis method capable of resisting label noise

Abstract

The invention discloses an anti-label noise urban rail train gearbox incremental fault diagnosis method, which comprises the following steps of: segmenting and standardizing collected gearbox fault data of an urban rail train bogie according to a preset length to obtain a fault sample; based on the fault sample, constructing a data set which can be used for class incremental learning; constructing a fault diagnosis model, and performing initial training on the fault diagnosis model by using the data set to obtain a fault diagnosis model after initial training; performing incremental training on the fault diagnosis model after the initial training to obtain a trained fault diagnosis model; and performing fault identification on a to-be-detected vibration signal by using the trained fault diagnosis model to obtain a diagnosis result of the gearbox. According to the method, disastrous forgetting and noise interference can be effectively handled, and the accuracy of fault diagnosis of the gearbox of the urban rail train is improved.

Description

Technical Field

[0001] This invention relates to the field of urban rail train fault detection, and specifically to a method for diagnosing incremental faults in urban rail train gearboxes that is resistant to tag noise. Background Technology

[0002] The efficient operation of urban rail trains relies on the stable functioning of key components. As one of the critical components of the urban rail train's transmission system, the gearbox's performance significantly impacts the safety and reliability of train operation. With continuous operation and under long-term alternating stress, the gearbox is highly susceptible to fatigue wear, potentially leading to serious problems such as gear fracture and bearing damage. Therefore, to ensure the safety and stability of train operation, fault diagnosis technology is urgently needed to monitor the gearbox's condition, thereby reducing risks and preventing serious safety accidents.

[0003] In recent years, deep learning technology has become an effective tool for fault diagnosis in train transmission systems, demonstrating high accuracy in identifying various faults. However, existing deep learning methods still face two major challenges in the actual operation of train gearbox fault diagnosis: 1. Train fault data is often not generated all at once, but continuously. When new fault categories appear, the trained model will be unable to adapt to these new fault categories, leading to a rapid decline in model performance. 2. Existing research often assumes that fault sample labels are completely accurate, neglecting the label noise problem caused by labeling delays and errors in the train data acquisition and labeling process. Label noise can significantly reduce the accuracy of fault diagnosis models.

[0004] In traditional train gearbox fault diagnosis methods, while replay mechanisms and regularization methods can alleviate the model forgetting problem to some extent, their robustness is insufficient in scenarios where the number of new fault categories exceeds that of old fault categories. Furthermore, noise mitigation strategies such as sample selection and robust loss are not well-suited to the incremental learning mechanism for fusing new and old fault knowledge, making it easy for noise to accumulate during iterations.

[0005] Therefore, in order to solve the problem that catastrophic amnesia and tag noise are difficult to deal with in a coordinated manner during the fault diagnosis process, resulting in low fault diagnosis accuracy, there is a need for an incremental fault diagnosis method for urban rail train gearboxes that is resistant to tag noise. This method can effectively deal with catastrophic amnesia and noise interference, and improve the accuracy of fault diagnosis for urban rail train gearboxes. Summary of the Invention

[0006] In view of this, the purpose of this invention is to overcome the deficiencies in the prior art and provide an incremental fault diagnosis method for urban rail train gearboxes that is resistant to tag noise, which can effectively cope with catastrophic forgetting and noise interference and improve the accuracy of fault diagnosis for urban rail train gearboxes.

[0007] The present invention provides an incremental fault diagnosis method for urban rail train gearboxes that is resistant to tag noise, comprising:

[0008] The collected gearbox fault data of urban rail train bogies are divided and standardized according to a preset length to obtain fault samples.

[0009] Based on fault samples, a dataset that can be used for incremental learning is constructed;

[0010] Construct a fault diagnosis model, and use the dataset to perform initial training on the fault diagnosis model to obtain the initially trained fault diagnosis model.

[0011] Incremental training is performed on the initially trained fault diagnosis model to obtain a well-trained fault diagnosis model.

[0012] The trained fault diagnosis model is used to identify faults in the vibration signal under test, and the diagnostic results of the gearbox are obtained.

[0013] Furthermore, the collected gearbox fault data of the urban rail train bogies is segmented and standardized according to a preset length, specifically including:

[0014] Acquire triaxial acceleration signal data from measurement points near the input and output shafts of the gearbox. The sampling frequency of this data is uniformly set to [value missing]. kHz, including normal conditions and several types of gear faults, each operating condition combination corresponds to Continuous recording for seconds, total One sampling point;

[0015] For each type of gearbox fault, a preset [method / mechanism] is used. The continuous signal is segmented without overlap using sampling points as the sample length, resulting in... The segment signals were then standardized using Z-scores to ensure that each segment had a mean of 0 and a variance of 1, ultimately forming a 1×1000 dimensional signal. Normalized fault samples.

[0016] Furthermore, based on fault samples, a dataset suitable for incremental learning is constructed, specifically including:

[0017] Determine the global truth labels for faulty samples and construct the asymmetric label transition matrix corresponding to the global truth labels. , with matrix The row vectors are used as the class probability distribution to generate global noise labels corresponding to fault samples; where, the matrix for The prior transition matrix has diagonal elements representing the probability of maintaining the correct label, and off-diagonal non-zero elements only appear between faults of equal or adjacent severity. It is forbidden to relabel severe faults as mild or normal states.

[0018] The global noise labels are converted into local class numbers through the global and local mapping table of the current incremental stage, thus completing the asymmetric label noise injection of the training set, while keeping the test set labels as true values ​​for performance evaluation.

[0019] Furthermore, a fault diagnosis model is constructed, and the model is initially trained using a dataset, specifically including:

[0020] S31. Construct a convolutional transform network model based on multi-scale DCT spatial attention, embed a multi-scale DCT spatial attention module in the feature extraction stage, and use the Transformer encoder to model the long-range dependencies between features, and set an scalable linear classification head at the end.

[0021] S32. Input the initial category vibration signal into the convolutional transform network model, use multi-scale soft label and entropy joint loss for end-to-end training, measure the difference between the network model's predicted output and the soft label, and update the network model parameters through backpropagation until the network model converges.

[0022] S33. After the initial training phase is completed, the network model parameters are saved as the teacher model baseline to provide old class knowledge reference for the subsequent incremental learning phase, and to guide the student model to retain the ability to identify old class faults.

[0023] Furthermore, step S31 specifically includes:

[0024] S311. Use a CNN module to perform feature mapping on the input one-dimensional signal data, automatically learn and extract local features important for fault diagnosis from the original data, specifically including:

[0025] For one-dimensional original signal Using kernel-size set Parallel processing of one-dimensional convolutional kernels and feature fusion:

[0026] ;

[0027] ;

[0028] in, This represents a 1D convolution operation. This indicates that the kernel size is Feature maps extracted by 1D convolution, The feature fusion result is obtained from three parallel 1D convolutional branches;

[0029] S312. The local features extracted by the CNN module are concatenated through channels and fed into the multi-scale DCT spatial attention module to enhance key frequency responses. After compression and activation operations, the enhanced features are obtained, specifically including:

[0030] Calculate the DCT convolution kernel to obtain the independent convolution kernel corresponding to each frequency index:

[0031] ;

[0032] in, The temporal length of the 1D CNN output. For frequency component index, For time-series location index; These are the weight values ​​corresponding to the frequency and kernel position, i.e., the elements of the convolution kernel;

[0033] Given a temporal feature sequence of size B×C×L, a DCT convolution kernel is used to perform convolution operations on the features and channel dimension compression, reducing the number of channels from C dimensions to 1 dimension. The Sigmoid function is then applied to obtain the weight distribution between feature channels. Finally, the attention weights are multiplied by the original input features and then added back to the original input features to achieve feature enhancement at key temporal positions.

[0034] ;

[0035] ;

[0036] ;

[0037] in, It is a 1D time series feature. For the first The first of the scales Eigenvalues ​​of each frequency component It is a fully connected layer. For single-channel temporal attention weights, These are the attention-weighted temporal features;

[0038] S313. The enhanced features are input into the Transformer encoder module to adaptively enhance fault information in key fault frequency bands and suppress redundant and interference components, specifically including:

[0039] Given the features output by a multi-scale DCT spatial attention module The shape is B×C×L, which is transformed into B×L×C through a dimension permutation operation, and a learnable positional embedding is introduced:

[0040] ;

[0041] in, Output feature sequences for the multi-scale DCT spatial attention module; For learnable location embedding matrices, each location Corresponding to an embedding vector ; The input sequence after adding location information;

[0042] The Transformer encoder consists of multiple stacked encoder layers. Each encoder layer comprises a multi-head self-attention sublayer and a feedforward neural network sublayer. Training stability is ensured through residual connections and layer normalization.

[0043] ;

[0044] ;

[0045] in, It is a multi-head attention mechanism that uses multiple parallel attention heads to capture the dependencies between different subspaces in the feature sequence; To normalize, the mean and variance of the input data are normalized. The feedforward network performs an independent nonlinear transformation on the features at each location; This is the output of the encoder layer.

[0046] Furthermore, in step S32, end-to-end training is performed using a joint loss of multi-scale soft labels and entropy, specifically including:

[0047] S321. During training, the classifiers of each CNN branch of the convolutional transform network model predict the input samples respectively. The reciprocal of the prediction entropy output of each scale branch is used as the weight, and a weighted average is performed with the prediction probability value to obtain the prediction probability of the current sample. Specifically, this includes:

[0048] ;

[0049] ;

[0050] ;

[0051] in, For global average pooling, The Softmax activation function is used. For each scale at the 1st Round epoch for the first The predicted probability of the class of each sample. For each scale at the 1st Round epoch for the first Each sample belongs to Predicted probability of the category Number of categories; The entropy value represents the branch prediction. The higher the entropy value, the more uncertain the branch prediction is; the lower the entropy value, the higher the confidence level of the branch prediction result. For the first The sample at the th Multi-scale weighted prediction probability of each epoch;

[0052] S322. Combining a temporal ensemble strategy, the training process is divided into a warm-up phase and an optimization phase. Different soft-label ensemble strategies are adopted in different phases, and the joint loss is calculated based on the soft labels and the predicted probabilities:

[0053] ;

[0054] ;

[0055] in, This represents the dividing round between the warm-up and optimization phases in the training process; The original label for the sample; This is a momentum parameter used to control the degree to which historical tags are retained; For the first The sample at the th The fusion of soft tags for each epoch; For the first The sample at the th The corresponding epoch in the round Category fusion soft tags; For the first The sample at the th In the first epoch Predicted probability of the category; For the first The predicted entropy value of each sample; This represents the total number of samples used in training. This represents the joint loss function for soft labels.

[0056] Furthermore, the fault diagnosis model after initial training is incrementally trained, including:

[0057] A joint strategy of replaying old samples and knowledge distillation is adopted to dynamically expand and fine-tune the classification head of the initially trained fault diagnosis model under the constraint of gradient absolute value-based augmentation loss. Specifically, this includes:

[0058] S41. Based on the confidence scores of the soft labels generated during the current training phase, select replays with priority given to those with a maximum confidence score of ≥ 100%. If the sample size is insufficient, supplement with samples having a confidence level of [insert confidence level here]. ~ To avoid exacerbating forgetting during the incremental stage, samples from the previous stage are used to avoid replaying noisy samples, which would exacerbate forgetting during the incremental stage. In subsequent stages, new class data and replay data are combined for training.

[0059] S42. In the incremental learning phase, the fault diagnosis model after initial training is frozen as the teacher model, and the fault data of the old and new classes are input into the student model. The logits output of the student model for the old class faults is temperature scaled and then KL divergence is calculated with the soft labels output by the teacher model. The student model is forced to retain the decision boundary of the old class fault categories through distillation loss.

[0060] S43. Input the old-class fault data into the teacher model, use the logit prediction value output by the teacher model as the training target, calculate the difference between the logit prediction value of the old-class faults and the student model, and calculate the absolute value of the gradient of the old-class fault samples, use it as a dynamic weight, amplify the loss only for old-class fault samples with negative difference, and the higher the absolute value of the gradient, the greater the amplification factor of the fault sample loss, and strengthen the constraint in backpropagation.

[0061] in, ;

[0062] ;

[0063] ;

[0064] in, For the current student model, the categories The predicted probability; The absolute value of the gradient is the i-th The absolute value of the difference between the predicted probability of an old-class fault sample and its true label; This represents the number of old-type fault samples. For the teacher model to the first The fault samples in Gejiu were predicted as category [missing information] in the previous stage. The output score; It is a collection of old-type fault samples; For the current student model, the first The predicted fault samples in Gejiu are categorized as follows: The output score; The logits output value of the student model is used to constrain it to be no less than that of the teacher model. These are the weighting coefficients. For gradient-aware enhancement loss function, This is the total loss function.

[0065] Furthermore, the logits output of the student model for old-class faults, after temperature scaling, are compared with the soft labels output by the teacher model using KL divergence calculation, specifically including:

[0066] ;

[0067] in, This is a temperature parameter used to control the smoothness of the Softmax output; The output logits of the teacher model; Output the old class `logits` for the student model; The Softmax function maps numerical values ​​to the interval [0,1]. KL divergence measures the difference between two probability distributions. and differences express , express ; This represents the distillation loss function.

[0068] The beneficial effects of this invention are as follows: This invention discloses an incremental fault diagnosis method for urban rail train gearboxes that is resistant to label noise. It proposes a multi-scale spatial attention mechanism to enhance the feature extraction of train gearbox data on DCT convolution kernels at different scales, thereby enhancing the model's ability to capture key fault information. In the incremental learning stage, an enhancement loss strategy based on the absolute value of gradient is adopted for old-class fault samples. The weights are dynamically adjusted according to the predicted values ​​of the samples, and old-class knowledge is appropriately retained to reduce the degree of forgetting. The normalized weights are obtained by calculating the reciprocal of the branch prediction entropy to dynamically weight and fuse the prediction probabilities of each scale, generating soft labels. The labels are then corrected by combining a temporal integration strategy. Finally, in both incremental and label noise scenarios for urban rail train gearboxes, this invention effectively alleviates the problems of catastrophic forgetting and noise interference, and improves the accuracy of fault diagnosis for urban rail train gearboxes. Attached Figure Description

[0069] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0070] Figure 1 This is a schematic diagram of the fault diagnosis method of the present invention;

[0071] Figure 2 This is a diagram of the convolutional transform network model architecture of the present invention;

[0072] Figure 3 This is a diagram of the multi-scale DCT spatial attention module architecture of the present invention;

[0073] Figure 4 This is a diagram illustrating the gearbox fault diagnosis process in the incremental scenario of the present invention. Detailed Implementation

[0074] The present invention will be further described below with reference to the accompanying drawings, as shown in the figures:

[0075] This embodiment discloses a method for diagnosing incremental faults in gearboxes of urban rail trains that is resistant to tag noise, including the following steps:

[0076] S1. The collected gearbox fault data of the urban rail train bogie is divided into segments according to a preset length and standardized to obtain fault samples;

[0077] S2. Based on fault samples, construct a dataset that can be used for incremental learning.

[0078] S3. Construct a fault diagnosis model, use the dataset to perform initial training on the fault diagnosis model, and obtain the initially trained fault diagnosis model.

[0079] S4. Perform incremental training on the initially trained fault diagnosis model to obtain a well-trained fault diagnosis model.

[0080] S5. Use the trained fault diagnosis model to identify faults in the vibration signal under test and obtain the diagnostic results of the gearbox.

[0081] In this embodiment, step S1 involves dividing and standardizing the collected gearbox fault data of the urban rail train bogie according to a preset length, specifically including:

[0082] Acquire triaxial acceleration signal data from measurement points near the input and output shafts of the gearbox. The sampling frequency of this data is uniformly set to [value missing]. kHz, including normal conditions and 8 types of gear faults, each operating condition combination corresponds to Continuous recording for seconds, total One sampling point;

[0083] To expand the sample size and meet the incremental learning requirements, each type of gearbox fault is assigned a preset value. The continuous signal is segmented without overlap using sampling points as the sample length, resulting in... The segment signals were then standardized using Z-scores to ensure that each segment had a mean of 0 and a variance of 1, ultimately forming a 1×1000 dimensional signal. Normalized fault samples.

[0084] In this embodiment, step S2 involves constructing a dataset for incremental learning based on fault samples, specifically including:

[0085] Determine the global truth labels for faulty samples and construct the asymmetric label transition matrix corresponding to the global truth labels. , with matrix The row vectors are used as the class probability distribution to generate global noise labels corresponding to fault samples; where, the matrix for The prior transition matrix has diagonal elements representing the probability of maintaining the correct label, and off-diagonal non-zero elements only appear between faults of equal or adjacent severity. It is forbidden to relabel severe faults as mild or normal states.

[0086] The global noise labels are converted into local class numbers through the global and local mapping table of the current incremental stage, thus completing the asymmetric label noise injection of the training set, while keeping the test set labels as true values ​​for performance evaluation.

[0087] In this embodiment, step S3 involves constructing a fault diagnosis model and performing initial training on the fault diagnosis model using a dataset, specifically including:

[0088] S31. For example Figure 2 As shown, a Convolution Transformer with Multiscale DCT-based Spatial Attention (CT-MDSA) network model is constructed. A Multiscale DCT-based Spatial Attention (MDSA) module is embedded in the feature extraction stage. The Transformer encoder is used to model long-range dependencies between features, and an scalable linear classification head is set at the end to adapt to the dynamic expansion requirements of fault categories in incremental scenarios. The architecture diagram of the Multiscale DCT Spatial Attention module is shown below. Figure 3 As shown.

[0089] S32. Input the initial category vibration signal into the convolutional transform network model, use multi-scale soft label and entropy joint loss for end-to-end training, measure the difference between the network model's predicted output and the soft label, and update the network model parameters through backpropagation until the network model converges.

[0090] S33. After the initial training phase is completed, the network model parameters are saved as the teacher model baseline to provide old class knowledge reference for the subsequent incremental learning phase, and to guide the student model to retain the ability to identify old class faults.

[0091] In this embodiment, step S31 specifically includes:

[0092] S311. Use a CNN module to perform feature mapping on the input one-dimensional signal data, automatically learn and extract local features important for fault diagnosis from the original data, specifically including:

[0093] For one-dimensional original signal Using kernel-size set Parallel processing of one-dimensional convolutional kernels and feature fusion:

[0094] ;

[0095] ;

[0096] in, This represents a 1D convolution operation. This indicates that the kernel size is Feature maps extracted by 1D convolution, The feature fusion result is obtained from three parallel 1D convolutional branches. It is 1D temporal feature data, that is, the input temporal feature tensor; It is a single temporal feature element; 1D represents 1-dimensional, which will not be elaborated here.

[0097] S312. The local features extracted by the CNN module are concatenated through channels and fed into the multi-scale DCT spatial attention module to enhance key frequency responses. After compression and activation operations, the enhanced features are obtained, specifically including:

[0098] Calculate the DCT (Discrete Cosine Transform) convolution kernel to obtain the independent convolution kernel corresponding to each frequency index:

[0099] ;

[0100] in, The temporal length of the 1D CNN output. For frequency component index, For time-series location index; These are the weight values ​​corresponding to the frequency and kernel position, i.e., the elements of the convolution kernel;

[0101] To capture key information at different temporal scales, given a temporal feature sequence of size B×C×L, a DCT convolution kernel is used to perform convolution operations on the features and channel dimension compression, reducing the number of channels from C dimensions to 1 dimension. The Sigmoid function is then applied to obtain the weight distribution between feature channels. Finally, the attention weights are multiplied by the original input features and then added back to the original input features to achieve feature enhancement at key temporal locations. The specific operation is as follows:

[0102] ;

[0103] ;

[0104] ;

[0105] in, It is a 1D time series feature. For the first The first of the scales Eigenvalues ​​of each frequency component It is a fully connected layer. For single-channel temporal attention weights, These are the attention-weighted temporal features;

[0106] S313. The enhanced features are input into the Transformer encoder module to adaptively enhance fault information in key fault frequency bands and suppress redundant and interference components, specifically including:

[0107] Given the features output by a multi-scale DCT spatial attention module The shape is B×C×L, which is transformed into B×L×C through a dimension permutation operation, and a learnable positional embedding is introduced:

[0108] ;

[0109] in, Output feature sequences for the multi-scale DCT spatial attention module; For learnable location embedding matrices, each location Corresponding to an embedding vector ; The input sequence after adding location information;

[0110] The Transformer encoder consists of multiple stacked encoder layers. Each encoder layer comprises a multi-head self-attention sublayer and a feedforward neural network sublayer. Training stability is ensured through residual connections and layer normalization.

[0111] ;

[0112] ;

[0113] in, It is a multi-head attention mechanism that uses multiple parallel attention heads to capture the dependencies between different subspaces in the feature sequence; To normalize, the mean and variance of the input data are normalized to avoid gradient vanishing or exploding. The feedforward network performs an independent nonlinear transformation on the features at each location; Output for encoder layer; It is the intermediate output in the encoder layer. It is the result of fusing the features after multi-head attention processing with the residuals of the original input features and then normalizing them. It will be used as the input of the feedforward neural network sublayer.

[0114] In this embodiment, step S32 employs end-to-end training using a combination of multi-scale soft labels and entropy loss, specifically including:

[0115] S321. During training, the classifiers of each CNN branch of the convolutional transform network model predict the input samples respectively. The reciprocal of the prediction entropy output of each scale branch is used as the weight, and a weighted average is performed with the prediction probability value to obtain the prediction probability of the current sample. Specifically, this includes:

[0116] ;

[0117] ;

[0118] ;

[0119] in, Global average pooling is a method that performs average pooling on the features output by the convolutional layer across the entire dimension. The Softmax activation function transforms the input values ​​into a probability distribution where the total value is 1. For each scale at the 1st Round epoch for the first The predicted class probability of each sample; For each scale at the 1st Round epoch for the first Each sample belongs to The predicted probability of a category is The specific probability value corresponding to class c; Number of categories; The entropy value represents the branch prediction. The higher the entropy value, the more uncertain the branch prediction is; the lower the entropy value, the higher the confidence level of the branch prediction result. For the first The sample at the th Multi-scale weighted prediction probability of each epoch;

[0120] S322. Combining a temporal ensemble strategy, the training process is divided into a warm-up phase and an optimization phase. Different soft-label ensemble strategies are adopted in different phases, and the joint loss is calculated based on the soft labels and the predicted probabilities:

[0121] ;

[0122] ;

[0123] in, This represents the dividing round between the warm-up and optimization phases in the training process; The original label for the sample; This is a momentum parameter used to control the degree to which historical tags are retained; For the first The sample at the th The fusion of soft tags for each epoch; For the first The sample at the th The corresponding epoch in the round Category fusion soft tags; For the first The sample at the th In the first epoch The predicted probability of the category corresponds to what was mentioned earlier. The Middle The specific value of the class; For the first The predicted entropy value of each sample; This represents the total number of samples used in training. This represents the soft-label joint loss function, used to guide the optimization objective during model training.

[0124] In this embodiment, step S4 involves incremental training of the initially trained fault diagnosis model, including:

[0125] A joint strategy of replaying old samples and knowledge distillation is adopted. Under the constraint of gradient absolute value-based augmentation loss, the classification head of the initially trained fault diagnosis model is dynamically expanded and its parameters are fine-tuned to achieve the injection of new class knowledge. Specifically, this includes:

[0126] S41. Based on the confidence of the soft labels generated during the current training process, the replay is selected. Samples with a maximum confidence of ≥0.8 are given priority. If the number of samples is insufficient at this time, samples with a confidence between 0.7 and 0.8 are selected to avoid replaying noisy samples from exacerbating the forgetting in the incremental stage. In subsequent stages, the new class data and the replay data are integrated for training.

[0127] S42. In the incremental learning phase, the fault diagnosis model after initial training is frozen as the teacher model, and the fault data of the old and new classes are input into the student model. The logits output of the student model for the old class faults is temperature scaled and then KL divergence is calculated with the soft labels output by the teacher model. The student model is forced to retain the decision boundary of the old class fault categories through distillation loss.

[0128] Specifically, the logits output of the student model for old-class faults, after temperature scaling, are compared with the soft labels output by the teacher model to calculate the KL divergence, which includes:

[0129] ;

[0130] in, This is a temperature parameter used to control the smoothness of the Softmax output; it can be set to 2. The output logits of the teacher model; Output the old class `logits` for the student model; The Softmax function maps numerical values ​​to the interval [0,1]. KL divergence measures the difference between two probability distributions. and differences express , express ; This represents the distillation loss function.

[0131] S43. Input the old-class fault data into the teacher model, use the logit prediction value output by the teacher model as the training target, calculate the difference between the logit prediction value of the old-class faults and the student model, and calculate the absolute value of the gradient of the old-class fault samples, use it as a dynamic weight, amplify the loss only for old-class fault samples with negative difference, and the higher the absolute value of the gradient, the greater the amplification factor of the fault sample loss, and strengthen the constraint in backpropagation.

[0132] in, ;

[0133] ;

[0134] ;

[0135] in, For the current student model, the categories The predicted probability; The absolute value of the gradient is the i-th The absolute value of the difference between the predicted probability of an old-class fault sample and its true label; This represents the number of old-type fault samples. For the teacher model to the first The fault samples in Gejiu were predicted as category [missing information] in the previous stage. The output score; It is a collection of old-type fault samples; For the current student model, the first The predicted fault samples in Gejiu are categorized as follows: The output score; The logits output value of the student model is used to constrain it to be no lower than that of the teacher model. These are weighting coefficients, which can take values ​​of 0.4, 0.3, and 0.3 respectively. This is a gradient-aware enhancement loss function used to mitigate the catastrophic forgetting problem caused by the number of samples of the new class being far greater than that of the old class; Let be the total loss function, which is the objective function to be minimized during model training. (Formula) The value 1 in the table represents the ideal probability corresponding to the true label, that is, the probability corresponding to the true category should be 1.

[0136] Additionally, in step S4, the predicted probabilities at each scale can be fused using the normalized value of the inverse of the prediction entropy of each branch to generate soft labels. These labels are then corrected using a temporal ensemble strategy to avoid interference from label noise. This mechanism is similar in principle to the label correction mechanism used in step S32, aiming to ensure that the data labels of the input model are always as clean as possible, thereby suppressing the interference of noisy labels on parameter updates. Further details will not be elaborated here.

[0137] In this embodiment, in step S5, the test vibration signal is standardized and fed into the trained fault diagnosis model, and the diagnosis result is obtained through multi-scale feature extraction and Transformer sequence modeling.

[0138] Specifically, the test vibration signal is standardized by using the statistical parameters of the training set to perform Z-score standardization, forming a single-channel sequence sample with a dimension of 1×2048;

[0139] The process of feeding the standardized test vibration signal into the trained fault diagnosis model to obtain the fault diagnosis result includes: inputting the standardized test set into the main network saved during the training phase, performing forward inference to obtain the posterior probability of various faults, and using the category corresponding to the maximum posterior probability as the gearbox fault diagnosis result.

[0140] To verify the effectiveness and advancement of the fault diagnosis method of this invention, a comparative experiment was conducted using gearbox data from a publicly available dataset simulating faults in a train bogie transmission system in a certain literature. The experiment selected data from "WC6 condition" (motor speed 60Hz, lateral load +10kN) out of the nine operating conditions in the dataset for research. The gearbox vibration signal data is described in Table 1 below. This vibration signal data includes normal conditions and eight types of gear faults (including tooth root cracks, broken teeth, bearing inner and outer ring faults, etc.). Each fault type consists of 360 training samples and 240 test samples, with each sample containing 2048 data points.

[0141] Table 1

[0142]

[0143] Based on the above data, in order to verify the effectiveness of the gradient-aware loss enhancement method proposed in this invention in mitigating catastrophic forgetting, it is compared with a variety of incremental learning methods; at the same time, the proposed dynamic integration method of convolutional entropy weight soft labels is compared with the noise reduction method to verify the effectiveness of the noise reduction module.

[0144] All methods use the same network skeleton, and the implementation details are as follows:

[0145] EWC: It estimates the importance of parameters to the old task using the Fisher information matrix and applies a quadratic regularization constraint to high-importance weights to suppress update amplitude, thereby mitigating catastrophic forgetting. When the number of new categories is significantly greater than that of old categories, important weights are easily overwritten, leading to memory failure.

[0146] LWF: This method utilizes the soft labels output by the old model to construct a distillation loss, which is then jointly optimized with the cross-entropy of the new task to achieve knowledge transfer. This approach does not require storing the original data, but relies on the prediction quality of the old model.

[0147] iCaRL: A small number of representative samples are retained for each class and retrained periodically, while the mean classifier is used instead of the fully connected layer to reduce class bias.

[0148] LUCIR: Reduces amplitude bias by using a cosine classifier and introduces a joint constraint of margin ranking loss and feature distillation to preserve the topology of the old class feature space.

[0149] IC: Calculate the absolute value of the gradient of the output score corresponding to the input fault sample, and incorporate it as the sample weight into the cross-entropy loss. Adaptively assign smaller weights to the class with more samples and larger weights to the class with fewer samples.

[0150] GCE combines cross-entropy loss and mean squared error, achieving a balance between the two by controlling the hyperparameters.

[0151] ASL: A robust loss function for class imbalance and label noise. Its core is to optimize loss calculation through class asymmetric weights and probability threshold constraints.

[0152] ELR introduces a regularization term that uses the model's historical ensemble predictions as the learning objective. By utilizing early model outputs as reliable targets, it reduces the tendency to memorize noisy labels.

[0153] CCR: Dynamically adjusts the loss weight based on the noise rate of each class to reduce the contribution of high noise classes.

[0154] ROBOT: First, use a small loss threshold to screen credible samples to estimate the noise transition matrix, and then perform reverse correction on the loss. However, the dimension of the transition matrix changes as the class space expands during the incremental process, and it needs to be re-estimated continuously.

[0155] This invention compares the performance of different incremental learning methods in a scenario with 20% label noise. The results are shown in Table 2, comparing the performance of CT-MDSA with other incremental learning methods. CT-MDSA maintains high accuracy at all stages, achieving an initial accuracy of 97.28%, and 94.67%, 92.40%, and 88.31% in incremental stages 1, 2, and 3, respectively, demonstrating its effectiveness in mitigating catastrophic forgetting. Traditional incremental learning methods (such as EWC and LWF) achieve certain accuracy in the initial stage, but their performance significantly declines as the incremental stages progress. For example, EWC, as a weight regularization method, lacks a sample replay mechanism, causing important weights of old classes to be continuously overwritten during updates, resulting in an accuracy drop to 20.89% in stage 3, a decrease of nearly 70 percentage points from the initial stage. The iCaRL, LUCIR, and IC methods mitigate catastrophic forgetting to some extent by introducing strategies such as knowledge distillation and sample replay. Therefore, they can maintain relatively good performance in the incremental stage. For example, LUCIR achieves an accuracy of 84.72% in stage 3, but its performance still lags behind CT-MDSA.

[0156] Table 3 shows the performance comparison of CT-MDSA with other noise reduction methods, illustrating the performance of various noise reduction methods at each stage. Compared with CT-MDSA of this invention, other methods without the introduction of the convolutional entropy weighted soft label dynamic integration module exhibit poor performance. With increasing stages, the accuracy of both GCE and CCR shows a continuous downward trend, dropping to 69.73% and 72.60% respectively in the third stage, reflecting their difficulty in continuously and effectively handling label noise during incremental dynamic learning. The interference of noise on the model accumulates continuously with the incremental process. ASL and ROBOT, through asymmetric modulation or bidirectional correction, suppressed overfitting to some extent, ultimately maintaining accuracy at 81.60% and 84.53% respectively. CT-MDSA significantly outperforms other comparative methods at each incremental stage. This is because the proposed module reduces the cumulative effect of mislabeled samples in incremental training through sample-level historical momentum updates and prediction entropy weighting, thereby improving the robustness and stability of the model in noisy scenarios.

[0157] Table 2

[0158]

[0159] Table 3

[0160]

[0161] Meanwhile, the present invention conducted ablation experiments, comparing the results of incremental fault diagnosis and gradient-aware loss enhancement without the addition of the MDSA module, as well as the experimental results of the dynamic integration strategy without the addition of convolutional entropy weight soft labels, as shown in Table 4.

[0162] Table 4

[0163]

[0164] Overall, by sequentially introducing attention, dynamic ensemble, and loss enhancement modules, the model exhibits progressively improving robustness in incremental scenarios with 20% label noise. BM, lacking any regularization or correction mechanism, suffers a sharp drop in accuracy to 14.29% with each incremental stage. M1, by introducing a gradient-aware loss enhancement module, mitigates catastrophic forgetting, ultimately achieving an accuracy of 88.27%, but still lower than CT-MDSA. Both M2 and M3 show significant fluctuations, indicating that a single module cannot simultaneously handle both new and old class discrimination and mislabeling suppression. CT-MDSA alleviates the challenges faced by traditional incremental learning, reducing bias towards new classes and maintaining higher classification accuracy even with label noise.

[0165] In summary, the method proposed in this invention has good performance in incremental fault diagnosis scenarios with tag noise.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for diagnosing incremental faults in urban rail train gearboxes that resists tag noise, characterized in that: include: The collected gearbox fault data of urban rail train bogies are divided and standardized according to a preset length to obtain fault samples. Based on fault samples, a dataset that can be used for incremental learning is constructed; Construct a fault diagnosis model, and use the dataset to perform initial training on the fault diagnosis model to obtain the initially trained fault diagnosis model. Incremental training is performed on the initially trained fault diagnosis model to obtain a well-trained fault diagnosis model. The trained fault diagnosis model is used to identify faults in the vibration signal under test, and the diagnostic results of the gearbox are obtained.

2. The incremental fault diagnosis method for urban rail train gearboxes with anti-tag noise as described in claim 1, characterized in that: The collected gearbox fault data of urban rail train bogies is segmented and standardized according to a preset length, specifically including: Acquire triaxial acceleration signal data from measurement points near the input and output shafts of the gearbox. The sampling frequency of this data is uniformly set to [value missing]. kHz, including normal conditions and several types of gear faults, each operating condition combination corresponds to Continuous recording for seconds, total One sampling point; For each type of gearbox fault, a preset [method / mechanism] is used. The continuous signal is segmented without overlap using sampling points as the sample length, resulting in... The segment signals were then standardized using Z-scores to ensure that each segment had a mean of 0 and a variance of 1, ultimately forming a 1×1000 dimensional signal. Normalized fault samples.

3. The incremental fault diagnosis method for urban rail train gearboxes with anti-tag noise as described in claim 1, characterized in that: Based on fault samples, a dataset suitable for incremental learning is constructed, specifically including: Determine the global truth labels for faulty samples and construct the asymmetric label transition matrix corresponding to the global truth labels. , with matrix The row vectors are used as the class probability distribution to generate global noise labels corresponding to fault samples; where, the matrix for The prior transition matrix has diagonal elements representing the probability of maintaining the correct label, and off-diagonal non-zero elements only appear between faults of equal or adjacent severity. It is forbidden to relabel severe faults as mild or normal states. The global noise labels are converted into local class numbers through the global and local mapping table of the current incremental stage, thus completing the asymmetric label noise injection of the training set, while keeping the test set labels as true values ​​for performance evaluation.

4. The incremental fault diagnosis method for urban rail train gearboxes with anti-tag noise as described in claim 1, characterized in that: Constructing a fault diagnosis model and initially training it using a dataset includes: S31. Construct a convolutional transform network model based on multi-scale DCT spatial attention, embed a multi-scale DCT spatial attention module in the feature extraction stage, and use the Transformer encoder to model the long-range dependencies between features, and set an scalable linear classification head at the end. S32. Input the initial category vibration signal into the convolutional transform network model, use multi-scale soft label and entropy joint loss for end-to-end training, measure the difference between the network model's predicted output and the soft label, and update the network model parameters through backpropagation until the network model converges. S33. After the initial training phase is completed, the network model parameters are saved as the teacher model baseline to provide old class knowledge reference for the subsequent incremental learning phase, and to guide the student model to retain the ability to identify old class faults.

5. The incremental fault diagnosis method for urban rail train gearboxes with anti-tag noise as described in claim 4, characterized in that: Step S31 specifically includes: S311. Use a CNN module to perform feature mapping on the input one-dimensional signal data, automatically learn and extract local features important for fault diagnosis from the original data, specifically including: For one-dimensional original signal Using kernel-size set Parallel processing of one-dimensional convolutional kernels and feature fusion: ； ； in, This represents a 1D convolution operation. This indicates that the kernel size is Feature maps extracted by 1D convolution, The feature fusion result is obtained from three parallel 1D convolutional branches; S312. The local features extracted by the CNN module are concatenated through channels and fed into the multi-scale DCT spatial attention module to enhance key frequency responses. After compression and activation operations, the enhanced features are obtained, specifically including: Calculate the DCT convolution kernel to obtain the independent convolution kernel corresponding to each frequency index: ； in, The temporal length of the 1D CNN output. For frequency component index, For time-series location index; These are the weight values ​​corresponding to the frequency and kernel position, i.e., the elements of the convolution kernel; Given a temporal feature sequence of size B×C×L, a DCT convolution kernel is used to perform convolution operations on the features and channel dimension compression, reducing the number of channels from C dimensions to 1 dimension. The Sigmoid function is then applied to obtain the weight distribution between feature channels. Finally, the attention weights are multiplied by the original input features and then added back to the original input features to achieve feature enhancement at key temporal positions. ； ； ； in, It is a 1D time series feature. For the first The first of the scales Eigenvalues ​​of each frequency component It is a fully connected layer. For single-channel temporal attention weights, These are the attention-weighted temporal features; S313. The enhanced features are input into the Transformer encoder module to adaptively enhance fault information in key fault frequency bands and suppress redundant and interference components, specifically including: Given the features output by a multi-scale DCT spatial attention module The shape is B×C×L, which is transformed into B×L×C through a dimension permutation operation, and a learnable positional embedding is introduced: ； in, Output feature sequences for the multi-scale DCT spatial attention module; For learnable location embedding matrices, each location Corresponding to an embedding vector ; The input sequence after adding location information; The Transformer encoder consists of multiple stacked encoder layers. Each encoder layer comprises a multi-head self-attention sublayer and a feedforward neural network sublayer. Training stability is ensured through residual connections and layer normalization. ； ； in, It is a multi-head attention mechanism that uses multiple parallel attention heads to capture the dependencies between different subspaces in the feature sequence; To normalize, the mean and variance of the input data are normalized. The feedforward network performs an independent nonlinear transformation on the features at each location; This is the output of the encoder layer.

6. The incremental fault diagnosis method for urban rail train gearboxes with anti-tag noise as described in claim 4, characterized in that: In step S32, end-to-end training is performed using a combination of multi-scale soft labels and entropy loss, specifically including: S321. During training, the classifiers of each CNN branch of the convolutional transform network model predict the input samples respectively. The reciprocal of the prediction entropy output of each scale branch is used as the weight, and a weighted average is performed with the prediction probability value to obtain the prediction probability of the current sample. Specifically, this includes: ； ； ； in, For global average pooling, The Softmax activation function is used. For each scale at the 1st Round epoch for the first The predicted probability of the class of each sample. For each scale at the 1st Round epoch for the first Each sample belongs to Predicted probability of the category Number of categories; The entropy value represents the branch prediction. The higher the entropy value, the more uncertain the branch prediction is; the lower the entropy value, the higher the confidence level of the branch prediction result. For the first The sample at the th Multi-scale weighted prediction probability of each epoch; S322. Combining a temporal ensemble strategy, the training process is divided into a warm-up phase and an optimization phase. Different soft-label ensemble strategies are adopted in different phases, and the joint loss is calculated based on the soft labels and the predicted probabilities: ； ； in, This represents the dividing round between the warm-up and optimization phases in the training process; The original label for the sample; This is a momentum parameter used to control the degree to which historical tags are retained; For the first The sample at the th The fusion of soft tags for each epoch; For the first The sample at the th The corresponding epoch in the round Category fusion soft tags; For the first The sample at the th In the first epoch Predicted probability of the category; For the first The predicted entropy value of each sample; This represents the total number of samples used in training. This represents the joint loss function for soft labels.

7. The incremental fault diagnosis method for urban rail train gearboxes with anti-tag noise as described in claim 1, characterized in that: Incremental training is performed on the initially trained fault diagnosis model, including: A joint strategy of replaying old samples and knowledge distillation is adopted to dynamically expand and fine-tune the classification head of the initially trained fault diagnosis model under the constraint of gradient absolute value-based augmentation loss. Specifically, this includes: S41. Based on the confidence scores of the soft labels generated during the current training phase, select replays with priority given to those with a maximum confidence score of ≥ 100%. If the sample size is insufficient, supplement with samples having a confidence level of [insert confidence level here]. ~ To avoid exacerbating forgetting during the incremental stage, samples from the previous stage are used to avoid replaying noisy samples, which would exacerbate forgetting during the incremental stage. In subsequent stages, new class data and replay data are combined for training. S42. In the incremental learning phase, the fault diagnosis model after initial training is frozen as the teacher model, and the fault data of the old and new classes are input into the student model. The logits output of the student model for the old class faults is temperature scaled and then KL divergence is calculated with the soft labels output by the teacher model. The student model is forced to retain the decision boundary of the old class fault categories through distillation loss. S43. Input the old-class fault data into the teacher model, use the logit prediction value output by the teacher model as the training target, calculate the difference between the logit prediction value of the old-class faults and the student model, and calculate the absolute value of the gradient of the old-class fault samples, use it as a dynamic weight, amplify the loss only for old-class fault samples with negative difference, and the higher the absolute value of the gradient, the greater the amplification factor of the fault sample loss, and strengthen the constraint in backpropagation. in, ; ； ； in, For the current student model, the categories The predicted probability; The absolute value of the gradient is the i-th The absolute value of the difference between the predicted probability of an old-class fault sample and its true label; This represents the number of old-type fault samples. For the teacher model to the first The fault samples in Gejiu were predicted as category [missing information] in the previous stage. The output score; It is a collection of old-type fault samples; For the current student model, the first The predicted fault samples in Gejiu are categorized as follows: The output score; The logits output value of the student model is used to constrain it to be no less than that of the teacher model. These are the weighting coefficients. For gradient-aware enhancement loss function, This is the total loss function.

8. The incremental fault diagnosis method for urban rail train gearboxes with anti-tag noise as described in claim 7, characterized in that: The logits output of the student model for old-class faults, after temperature scaling, are compared with the soft labels output by the teacher model to calculate KL divergence, specifically including: ； in, This is a temperature parameter used to control the smoothness of the Softmax output; The output logits of the teacher model; Output the old class `logits` for the student model; The Softmax function maps numerical values ​​to the interval [0,1]. KL divergence measures the difference between two probability distributions. and differences express , express ; This represents the distillation loss function.

Images