Label correction intelligent diagnosis method for axle box bearing of rail train under label noise

By using collaborative training and label correction technology of two deep neural network models in the track train axle box bearing fault diagnosis, the overfitting problem caused by label noise is solved, and a higher accuracy fault diagnosis is achieved.

CN119064012APending Publication Date: 2024-12-03XI AN JIAOTONG UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411242147.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-05
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing intelligent diagnostic technology is affected by label noise in the fault diagnosis of axle box bearings of track trains, resulting in model overfitting and erroneous fault judgment, making it difficult to achieve accurate diagnosis.

Method used

Two deep neural network models with different structures are trained simultaneously, and noise label data is filtered through gradient exchange and reverse update, and label correction is performed through high-precision models. The loss function is optimized to balance correct and correct label data, and the generalization of the model is improved. ability.

Benefits of technology

It effectively reduces the impact of noise labels, improves the accuracy of fault diagnosis of rail train axle box bearings, and enhances the credibility of data and the generalization ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119064012A_ABST
    Figure CN119064012A_ABST
Patent Text Reader

Abstract

An intelligent diagnosis method for label correction of a rail train axle box bearing under label noise comprises the steps that two network models of different structures are constructed at the same time, and correct label data are extracted through the difference of unique visual angles and decision boundaries of the network models; through gradient exchange and reverse updating, interaction is realized, the influence of noise label data is avoided, and a high-precision network model is obtained; correcting the noise tag by using a high-precision network model, so that the tag of the original data set is more credible; training the high-precision network model again to obtain more implicit features so as to improve the generalization ability of the model; according to the method, the problem of label noise in the industrial data set is solved, and the correction of the noise label and the improvement of the fault diagnosis precision are finally realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a label correction intelligent diagnosis method for a rail train axle box bearing under label noise. Background Art

[0002] The rapid development of rail transit and the surge in operating mileage have put forward higher requirements for safety and stability. According to statistics, 30%-40% of major safety accidents in rail trains are caused by mechanical failures in the train, among which mechanical failures caused by bogies account for the highest proportion, and once a train bogie fails, it will almost inevitably cause a train failure. As a key component of the bogie, the working condition of the axle box bearing is directly related to the driving safety of the train, and has an extremely important impact on the stable and efficient operation of the train. Moreover, the bogie axle box bearing works in a high-speed and heavy-load environment, and is one of the components with the worst working conditions among all the equipment on the train. Due to the long-term and repeated action of contact stress, the working surface of the bearing is very prone to wear, cracks or indentations, and severe cases may even cause the bearing to break, which in turn causes serious accidents, endangers passenger safety, causes economic losses and adverse social impacts.

[0003] Existing intelligent diagnosis ([1] Lei Yaguo, Jia Feng, Zhou Xin, et al. Big data health monitoring method for mechanical equipment based on deep learning theory [J]. Journal of Mechanical Engineering, 2015, 51(21): 49-56. [2] Wang Jing, Xu Zhiwei, Liu Wenjing, et al. A review of research on intelligent health monitoring and fault diagnosis mechanism of rolling bearings [J]. Computer Science and Technology, 2024, 18(04): 878-898.) uses machine learning technology to automatically extract fault features from axle box bearing monitoring data, and then establishes a direct mapping relationship between features and bearing faults. It is the key to ensure the safety of the current big data era. Intelligent diagnosis technology is an important means for the safe operation of axle box bearings; however, intelligent diagnosis technology relies on high-quality labeled data. Due to the lag in labeling or human misjudgment caused by the complexity of the system, there is label noise in the obtained data, which reduces the application of intelligent diagnosis methods in the actual train axle box bearing fault diagnosis. Because once the data contains a large number of noise labels, the deep learning model will be affected by the noise labels. Since the model cannot distinguish between clean labeled data and noisy labeled data, during the training process, the model will overfit to the noisy labeled data, resulting in incorrect fault judgment during actual fault diagnosis, causing serious harm.

[0004] In summary, due to the low quality of data labels obtained from actual industrial sites, the model will overfit to the incorrectly labeled data, resulting in the existing fault diagnosis technology being difficult to achieve accurate fault diagnosis of rail train axle box bearings. Summary of the invention

[0005] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to propose an intelligent diagnosis method for label correction of rail train axle box bearings under label noise, reduce the adverse effects of label noise data, and realize intelligent identification of the health status of rail train axle box bearings and correction of erroneous labeling under the condition of erroneous labeling.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A label correction intelligent diagnosis method for a rail train axle box bearing under label noise, comprising the following steps:

[0008] Step 1: Get a labeled noisy dataset D, including a correct labeled dataset and mislabeled datasets In the formula, represents the mth data in the correct label data set in the label noise data set D, is the corresponding correct label, M C is the number of correctly labeled data; It represents the mth data in the wrong label data set in the label noise data set D. is its corresponding error label, M N is the number of incorrectly labeled data; each type of data sample should have correctly labeled data;

[0009] Step 2: Train two deep neural networks f and g at the same time, input the noisy label dataset D obtained in step 1, and cooperate with each other in the learning process; specifically, in each small sample data, the two deep neural networks f and g will evaluate their respective losses and select the data with smaller loss as valid and reliable knowledge; the sample data extraction process is as follows:

[0010]

[0011] In the above formula, and is the screened data, D' is its specific sample, is the input data, l is the cross entropy loss function, and R(T) is the function that controls the sample size, which is dynamically set to:

[0012]

[0013] In the above formula, T represents the current iteration round, τ and T k are the hyperparameters set according to the memory effect;

[0014] Step 3: When training on a dataset D containing labeled noise, the classifiers can show different label noise filtering capabilities; if the selected instance and The data labels are not completely clean. The two deep neural networks f and g will adaptively correct the training errors through the peer-to-peer network, and will not pass them to themselves to affect the next model training. The sample parameter update method is as follows:

[0015]

[0016] In the above formula, w f and w g are the network parameters of the deep neural network f and g respectively, η is the learning rate, is the gradient operator;

[0017] Step 4: Select the CNN model in Co-teaching, the BiLSTM model in the field of deep learning, and the Resnet models with different layers for training. The training target is minl(f,D'). First, train the single model according to steps 2 and 3. After reaching the set training rounds, select the two network models with the highest test accuracy; the two high-precision network models are interactively trained;

[0018] Step 5: After reaching the set training round, there are two high-precision network model choices when labeling the wrongly labeled data; when correcting, select the category with the highest predicted probability for each unlabeled sample as its true label:

[0019]

[0020] In the above formula, y N is the corrected label of the Nth sample, i is its label category f i (x) is the probability of belonging to the i-th class calculated based on the input data;

[0021] Step 6: Correct the incorrectly labeled data and input the correct labeled data into the high-precision network model again for training. When designing the loss function, ensure that the data of the two high-precision network models can benefit equally; the loss function combines the loss of the correctly labeled data and the loss of the corrected label data, and ensures a balance between the two through weight adjustment, thereby optimizing the training effect of the high-precision network model. Therefore, the loss function is:

[0022]

[0023] In the above formula, and are the label prediction values ​​of the mth data in the correct label data set and the wrong label data set, respectively, and α(t) is the weight for balancing the correct label and the corrected label;

[0024] The optimization goal is minL, and then the Adam optimization algorithm is used to update the network parameter w to minimize the objective function;

[0025] Step 7: Repeat steps 2 to 6 to iteratively optimize the network training-label correction-network retraining process under the noisy labels to achieve the correction of the noisy labels and the improvement of the accuracy of the high-precision network model.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention proposes a label correction intelligent diagnosis method for rail train axle box bearings under label noise. The method simultaneously constructs two network models with different structures, and uses their unique perspectives and decision boundaries to extract correct label data; through gradient exchange and reverse update, interaction is achieved to avoid the influence of noise label data and obtain a high-precision model; the high-precision model is then used to correct the noise label to make the original data set label more credible; and then the network model is trained again to obtain more implicit features to improve the model generalization ability. The method of the present invention overcomes the label noise problem in industrial data sets, and ultimately achieves the correction of noise labels and the improvement of fault diagnosis accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of an embodiment of the present invention.

[0029] Figure 2 This is a bar chart comparing the diagnosis results of the embodiment of the present invention with other methods.

[0030] Figure 3 : This is a feature distribution diagram of data dimensionality reduction according to an embodiment of the present invention, where (a) is the feature distribution of original label noise data, and (b) is the feature distribution of data after label correction. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below with reference to the accompanying drawings.

[0032] Reference Figure 1 , a label correction intelligent diagnosis method for a rail train axle box bearing under label noise, comprising the following steps:

[0033] Step 1: Get a labeled noisy dataset D, including a correct labeled dataset and mislabeled datasets In the formula, represents the mth data in the correct label data set in the label noise data set D, is the corresponding correct label, M C is the number of correctly labeled data; It represents the mth data in the wrong label data set in the label noise data set D. is its corresponding error label, M N is the number of incorrectly labeled data. It should be noted that each type of data sample should have correctly labeled data. During the test, common noise types in actual industrial sites were simulated and different noise rates were set to fully simulate the actual situation.

[0034] Step 2: Train two deep neural networks f and g at the same time, input the noisy label dataset D obtained in step 1, and cooperate with each other in the learning process; specifically, in each small sample data, the two deep neural networks f and g will evaluate their respective losses and select the data with smaller loss as valid and reliable knowledge; the sample data extraction process is as follows:

[0035]

[0036] In the above formula, and is the screened data, D' is its specific sample, is the input data, l is the cross entropy loss function, and R(T) is the function that controls the sample size, which is dynamically set to:

[0037]

[0038] In the above formula, T represents the current iteration round, τ and T k are the hyperparameters set according to the memory effect;

[0039] Step 3: Since different deep neural networks f and g tend to produce different decision boundaries based on their unique structures and parameter settings, these decision boundaries reflect their unique learning capabilities and feature recognition methods; therefore, when training on a dataset D containing labeled noise, the classifiers can exhibit their own different label noise filtering capabilities; each deep neural network f or g may have stronger robustness to certain types of noisy labels due to its unique learning characteristics, thereby more effectively removing these noises during the training process and improving the generalization performance of the model; if the selected instance and The data labels are not completely clean. The two deep neural networks f and g will adaptively correct the training errors through the peer-to-peer network, and will not pass them to themselves to affect the next model training. The sample parameter update method is as follows:

[0040]

[0041] In the above formula, w f and w g are the network parameters of the deep neural network f and g respectively, η is the learning rate, is the gradient operator;

[0042] Step 4: Select the CNN model in Co-teaching, the BiLSTM model popular in the field of deep learning, and the Resnet model with different layers for training. The training target is min l(f,D'). First, train the single model according to steps 2 and 3. After reaching the set training rounds, select the two network models with the highest test accuracy as high-precision network model 1 and high-precision network model 2; then, in order to maximize the use of structural differences to screen data, select the two high-precision network models with different structures with the highest accuracy after training for interactive training;

[0043] Step 5: After reaching the set training round, since two high-precision network models with different decision boundaries are trained, there are two high-precision network models 1 and 2 to choose from when labeling the wrongly labeled data; when correcting, select the category with the largest prediction probability for each unlabeled sample as its true label:

[0044]

[0045] In the above formula, y N is the corrected label of the Nth sample, i is its label category f i (x) is the probability of belonging to the i-th class calculated based on the input data;

[0046] Step 6: Before updating the label, directly remove the wrongly labeled data and do not participate in the training, which will result in incomplete information. Therefore, the wrongly labeled label data is corrected and input into the high-precision network model again with the correct labeled label (because the network parameters of the previous and subsequent models have changed, they are recorded as model A and model B) for training. In order to ensure that the high-precision network model can benefit equally from the two types of data, this balance factor is specially considered when designing the overall loss function; through fine adjustment of the loss function, it is ensured that the high-precision network model fully absorbs and utilizes various types of data information during the optimization process, thereby improving its final performance; the overall loss function will combine the loss of correctly labeled data and the loss of corrected label data, and ensure the balance between the two through appropriate weight adjustment, thereby optimizing the training effect of the high-precision network model, so the loss function is:

[0047]

[0048] In the above formula, and are the label prediction values ​​of the mth data in the correct label data set and the wrong label data set respectively, and θ(t) is the weight to balance the correct label and the corrected label;

[0049] The optimization goal is minL, and then the Adam optimization algorithm is used to interactively update the network parameters w to minimize the objective function;

[0050] Step 7: Repeat steps 2 to 6 to iteratively optimize the network training-label correction-network retraining process under the noisy labels to achieve the correction of the noisy labels and the improvement of the accuracy of the high-precision network model.

[0051] This embodiment uses the PHM Beijing high-speed rail bogie dataset and the XJTU high-speed rail bogie dataset to verify the method of the present invention, as shown in Table 1 for details.

[0052] The experimental test data of the PHM Beijing high-speed rail bogie data set is selected from the right end cover of the axle box far away from the bogie reduction gearbox. The acceleration sensor speed is set to 1200rmp, the acquisition frequency is 64kHZ, and a total of 2 types of gear faults, 2 types of bearing faults, and 1 type of normal are selected, with a total of 5 types of data types, and the number of samples for each state is 200.

[0053] The XJTU high-speed rail bogie data set is driven by dual motors, with a single motor power of 2.2kw and a motor speed of 3600rmp, simulating an actual operating speed of about 223km / h. The data is collected by the end cover acceleration sensor, with a collection frequency of 12.8kHZ. A total of 3 types of bearing faults and 1 type of normal are selected, with a total of 4 types of data types, and the number of samples for each state is 200.

[0054] Table 1 Bearing vibration signal sample set

[0055]

[0056] Randomly select samples from each category of samples in the data set according to the noise rate for label conversion. The conversion process satisfies the following formula: Given noise Flip from clean y, where T ij Represents the element in the i-th row and j-th column of the noise matrix T. The specific settings are shown in Table 2.

[0057] Table 2 Dataset noise settings

[0058]

[0059] The symmetric noise is defined as the probability that a correct label is mislabeled as another class label. The noise transfer matrix T is defined as:

[0060]

[0061] The method is experimentally verified using the PHMBeijing high-speed rail bogie dataset and the XJTU high-speed rail bogie dataset. Method 1 is to perform model interaction training only without data correction; Method 2 is to perform model interaction training and data self-correction, that is, the corrected training model and the corrected model are the same model; Method 3 is to perform model interaction training and data other correction, that is, the corrected training model and the corrected model are different models. The model accuracy is shown in Table 3. By comparing Method 2, Method 3 with Method 1, it can be seen that Method 2 and Method 3 both improve the model accuracy, proving the effectiveness of the method of the present invention.

[0062] Table 3 Model test accuracy before and after label correction

[0063]

[0064] Figure 2 A bar chart comparing the diagnostic results of the embodiment with other methods is given. Figure 2 It can be seen from the results in that the method of the present invention significantly improves the accuracy of the model in the label noise scenario; in addition, when compared with the popular Co-teaching method and the widely used Resnet and BiLSTM models, the method of the present invention has a significant improvement in model accuracy, especially showing an accuracy of 88.3% at a label noise rate of 0.45, which confirms the superiority of the method of the present invention.

[0065] Figure 3 The data distribution of the labeled noisy data before and after correction is given. It can be clearly seen from Figure (a) that the original noise data features present a highly mixed distribution, making it difficult to distinguish the class boundaries between labels 0 and 1, 0 and 3, and 1 and 2. The overlap of data features makes it difficult for the model to establish a clear decision boundary that can reliably assign label types. Figure (b) shows the model's ability to self-correct labels. Although there are still some overlaps in the data after label correction, significant improvements have been made in the classification decision boundary compared to the original labeled noise data. This enhancement significantly improves the availability of the data.

Claims

1. A label correction intelligent diagnosis method for rail train axle box bearings under label noise, characterized in that: The following steps are involved: Step 1: Get a labeled noisy dataset D, including a correct labeled dataset X C = and mislabeled datasets In the formula, represents the mth data in the correct label data set in the label noise data set D, is the corresponding correct label, M C is the number of correctly labeled data; It represents the mth data in the wrong label data set in the label noise data set D. is its corresponding error label, M N is the number of incorrectly labeled data; each type of data sample should have correctly labeled data; Step 2: Train two deep neural networks f and g at the same time, input the noisy label dataset D obtained in step 1, and cooperate with each other in the learning process; specifically, in each small sample data, the two deep neural networks f and g will evaluate their respective losses and select the data with smaller loss as valid and reliable knowledge; the sample data extraction process is as follows: In the above formula, and is the screened data, D′ is its specific sample, is the input data, l is the cross entropy loss function, and R(T) is the function that controls the sample size, which is dynamically set to: In the above formula, T represents the current iteration round, τ and T k are the hyperparameters set according to the memory effect; Step 3: When training on a dataset D containing label noise, the classifiers can show different label noise filtering capabilities; if the selected instance and The data labels are not completely clean. The two deep neural networks f and g will adaptively correct the training errors through the peer-to-peer network, and will not pass them to themselves to affect the next model training. The sample parameter update method is as follows: In the above formula, w f and w g are the network parameters of the deep neural network f and g respectively, η is the learning rate, is the gradient operator; Step 4: Select the CNN model in Co-teaching, the BiLSTM model in the field of deep learning, and the Resnet models with different layers for training. The training target is min l(f, D′). First, train the single model according to steps 2 and 3. After reaching the set training rounds, select the two network models with the highest test accuracy. The two high-precision network models are interactively trained. Step 5: After reaching the set training round, there are two high-precision network model choices when labeling the wrongly labeled data; when correcting, select the category with the highest predicted probability for each unlabeled sample as its true label: In the above formula, y N is the corrected label of the Nth sample, i is its label category f i (x) is the probability of belonging to the i-th class calculated based on the input data; Step 6: Correct the incorrectly labeled data and input the correct labeled data into the high-precision network model again for training. When designing the loss function, ensure that the data of the two high-precision network models can benefit equally; the loss function combines the loss of the correctly labeled data and the loss of the corrected label data, and ensures a balance between the two through weight adjustment, thereby optimizing the training effect of the high-precision network model. Therefore, the loss function is: In the above formula, and are the label prediction values ​​of the mth data in the correct label data set and the wrong label data set, respectively, and α(t) is the weight for balancing the correct label and the corrected label; The optimization goal is minL, and then the Adam optimization algorithm is used to update the network parameter w to minimize the objective function; Step 7: Repeat steps 2 to 6 to iteratively optimize the network training-label correction-network retraining process under the noisy labels to achieve the correction of the noisy labels and the improvement of the accuracy of the high-precision network model.

Citation Information

Cited By

  • Unmanned aerial vehicle intelligent fault diagnosis method for sample-label noise coupling scene

    CN120216888A