Self-adaptive network learning method suitable for monitoring state of brake pad of long ramp high-speed train

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a self-adaptive network learning method suitable for monitoring the state of a brake pad of a long ramp high-speed train, and the method comprises the steps: carrying out the learning through employing a source domain data set and more than two target domain data sets; the method comprises the following steps: firstly, pre-training a prediction model formed by a feature extractor and a classifier by using a source domain data set; configuring a corresponding prediction model for each target domain, performing initialization by using the pre-trained prediction model, and testing each target domain data set through the prediction model to obtain a corresponding classification result; and constructing a loss function according to the test result of the data set of each target domain, and optimizing the classifier of each target domain by using the constructed loss function. The method has obvious advantages in solving the multi-target problem, shows huge potential in actual train brake monitoring application, and provides support for monitoring the cross-rotating-speed state of the brake pad of the long and large ramp high-speed train.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and relates to a multi-target domain adaptive network learning method, in particular to an adaptive network learning method applicable to the state monitoring of the brake pads of high-speed trains on long and steep slopes. Background Art

[0002] During the operation of high-speed trains, compared with the braking process on normal slopes, the increased braking time of trains on long and steep slope routes poses higher requirements for the braking performance of trains. In recent years, many scholars have conducted relevant research on the thermodynamic characteristics, wear mechanism and performance evolution law of train brake pads during the braking process under the conditions of long and steep slopes. Li et al. studied the air braking of trains on the long and steep slopes of the Lanxin Line through simulation and bench tests, and analyzed the correlation between the temperature of the brake pads and wear [Zhou Suxia, Shao Jing, Tong Xin. Research on a new type of brake disc based on continuous braking on long and steep slopes [J]. Chinese Journal of Mechanical Engineering, 2022, 58(20): 391-398]. Kang et al. established a braking heat load calculation model to study the influence of trains going downhill at different speeds on the temperature rise of brake discs and brake pads on long and steep slope routes [Kang Jinghui, Lyu Baojia, Jiao Biaoqiang, et al. Research on the braking heat load of EMU brake discs under the conditions of long and steep slopes [J]. Railway Locomotive & Car, 2020, 40(06): 19-23]. Zhou et al. simulated the influence of the designed aluminum-embedded steel structure brake disc on the temperature and thermal stress distribution of the braking interface under the condition of continuous braking of trains on long and steep slopes through the finite element method [Li Wanxin. Thermal simulation analysis of the brake disc of the EMU on the long and steep slopes of the Lanxin Line [J]. Railway Locomotive & Car, 2017, 37(01): 1-3+9]. The above research mainly focuses on the influence of braking on long and steep slopes on the brake pads, and explains the temperature, force and wear changes of the brake pads during braking. However, these studies are difficult to effectively judge the damage state of the brake pads during the friction process and cannot meet the needs of the state monitoring of the brake pads on long and steep slopes.

[0003] The vibration caused by friction during train braking can reflect the tribological state of the braking interface, and there are significant differences in the vibration signals of brake blocks with different damage forms. Therefore, the state monitoring method based on vibration data has gradually attracted attention. Collecting and deeply analyzing the vibration signals of brake blocks during braking has become an effective way to solve the problem of brake pad state monitoring. Hu et al. [6] used wavelet transform and deep subdomain generalization network to analyze the vibration acceleration signals of brake blocks during braking, realizing the intelligent state monitoring of high-speed train brake pads [[6]R. Hu, M. Zhang, X. Meng, and Z. Kang, "Deep subdomain generalisation network for health monitoring of high-speed train brake pads," Engineering Applications of Artificial Intelligence, vol. 113, 2022, doi: 10.1016 / j.engappai.2022.104896]. Zhang et al. proposed a friction fault diagnosis method based on one-dimensional convolutional neural network and GraphSAGE network, considering the correlation between the vibration signals of different brake block faults, and realizing the effective identification of brake block faults under unbalanced data [M. Zhang, X. Li, Z. Xiang, J. Mo, and S. Xu, "Diagnosis of brake friction faults in high-speed trains based on 1D CNN and GraphSAGE under data imbalance," Measurement: Journal of the International Measurement Confederation, vol. 207, 2023, doi: 10.1016 / j.measurement.2022.112378]. However, due to the long braking time caused by the large slope of the long and steep road section, the distribution of vibration data collected at different rotational speeds is significantly different, resulting in the phenomenon of domain shift. Domain shift poses higher requirements for the cross-task learning ability and robustness of deep learning models, becoming the main challenge in the state monitoring of brake pads on long and steep slopes.

[0004] With the increasing demand for cross - domain data analysis, many experts have proposed domain adaptation methods for the domain shift problem caused by distribution differences. Wang et al. proposed a triplet loss - guided adversarial domain adaptation method, which uses an adversarial method based on the Wasserstein distance and minimizes the distances between the same class centers and different domains for data - level and class - level alignment to achieve bearing fault diagnosis under variable rotational speeds and variable loads [X. Wang and F. Liu, "Triplet loss guided adversarial domain adaptation for bearing fault diagnosis," Sensors (Switzerland), vol. 20, no. 1, 2020, doi: 10.3390 / s20010320]. An et al. proposed a Gaussian mixture variational - based domain adaptation fault diagnosis method, which projects the features learned between different domains into a common space by considering the global and local features of fault signals to achieve bearing fault diagnosis under variable working conditions [Y. An, K. Zhang, Y. Chai, Z. Zhu, and Q. Liu, "Gaussian Mixture Variational - Based Transformer Domain Adaptation Fault Diagnosis Method and Its Application in Bearing Fault Diagnosis," IEEE Transactions on Industrial Informatics, vol. 20, no. 1, pp. 615 - 625, 2024, doi: 10.1109 / TII.2023.3268750]. The above - mentioned domain adaptation methods aim to narrow the differences between the source domain and a single target domain. However, due to the fact that the data of different rotational speed target domains show multiple distributions under long - gradient braking, how to achieve domain alignment for multiple target domains has become the key to the cross - rotational - speed status monitoring of brake pads. Summary of the Invention

[0005] The purpose of the present invention is to provide an adaptive network learning method suitable for the status monitoring of brake pads of high - speed trains on long gradients, aiming at the above - mentioned problems existing in the prior art. It can strengthen its domain adaptation ability by learning the data distributions at multiple rotational speeds and effectively achieve the cross - rotational - speed status monitoring of brake pads of high - speed trains on long gradients.

[0006] To achieve the above - mentioned purpose, the present invention provides an adaptive network learning method suitable for the status monitoring of brake pads of high - speed trains on long gradients, which uses a source - domain data set and two or more target - domain data sets for learning, and includes the following steps:

[0007] S1. Pre-train the prediction model composed of a feature extractor and a classifier using the source domain dataset.

[0008] S2. Configure a corresponding prediction model for each target domain, initialize it using the prediction model pre-trained in step S1, and test each target domain dataset using the prediction model to obtain the corresponding classification results. For the target domain dataset t, perform operations in the following sub-steps:

[0009] S21. Test the target domain dataset t using the prediction model, sort it in descending order according to the confidence level output by the prediction model, and extract the samples with a confidence level greater than the set confidence level threshold range from the target domain dataset t to form the dataset F. Classify according to the pseudo-labels of each sample tested in the dataset F, and perform weighted aggregation on the corresponding sample features extracted by the feature extractor to construct several class centroids.

[0010] S22. Calculate the cosine similarity between the remaining samples of the target domain dataset t in step S21 and each class centroid, sort it in descending order according to the similarity, and extract the dataset F′ with a similarity greater than the set similarity threshold range from the remaining samples of the target domain dataset t, and assign the same class pseudo-labels as the class centroids.

[0011] S23. Construct a KNN classifier based on the dataset F′, and assign class pseudo-labels to the remaining samples of the target domain dataset t in step S22.

[0012] S3. Construct a loss function based on the test results of each target domain dataset, and use the constructed loss function to optimize the classifier of each target domain. Return to step S2, and repeat steps S2 - S3 until the classifiers of each target domain converge.

[0013] In the present invention, first perform short-time Fourier transform on the source domain dataset and the target domain dataset, and use the source domain dataset and the target domain dataset after short-time Fourier transform as the input of the prediction model.

[0014] The above step S1 includes the following sub-steps:

[0015] S11. Perform random mixup processing on the samples in the source domain dataset, and add the processed samples to the source domain dataset.

[0016] S12. Pre-train the prediction model using the source domain dataset obtained in step S11.

[0017] In the above step S2, for the target domain dataset t, first test the target domain dataset t using the pre-trained prediction model, extract the samples with a confidence level greater than the confidence level threshold according to the set confidence level threshold, perform random mixup processing on them, and add the processed samples to the target domain dataset t.

[0018] In the present invention, the process of random mixing enhancement is as follows: the sample to be processed and the noise n~P that follows the normalized distribution are N(0,1) input into the encoder together to achieve feature encoding, and the AdaIN is used to perform different conversions on the encoded features. Finally, the decoder maps the features back to the original space, which is expressed as follows:

[0019]

[0020] where e ξ is the encoder, d ζ is the decoder, and l IN is the element batch standard layer; are the two linear conversion layers of AdaIN respectively; R i (x) is the result of the i-th conversion, i = 1,..., Nang, and Nang is the number of conversions;

[0021] The converted sample and the original sample are randomly weighted and mixed under the normalized distribution Finally, the enhanced sample is obtained after scaling through the sigmoid function, and its generation process is as follows:

[0022]

[0023] where w0 is the weight of the original sample, and w i is the weight of the result of the i-th conversion.

[0024] In the above step S21, the construction process of the data set F is expressed as:

[0025]

[0026] where I k is the target domain sample selected from the k-th category; K is the number of all categories; is the number of target domain samples; g ω ,k(·) is the k-th element output by the classifier; f θ (·) is the feature extractor; is the set confidence threshold range;

[0027] According to the sample confidence level, the sample features in the data set F are weighted and aggregated to construct several category centroids, which are expressed as follows:

[0028]

[0029] In the above step S22, the construction process of the data set F' is as follows:

[0030]

[0031] Among them, I k ′ is the target domain sample selected from the k-th category; K is the number of all categories; is the remaining sample number of the target domain dataset t in step S21; g ω,k (·) is the k-th element output by the classifier; f θ (·) is the feature extractor; is the set similarity threshold range.

[0032] In the above step S23, the process of assigning categories to the remaining samples of the target domain dataset t in step S22 by the KNN classifier is expressed as: The process of assigning categories is expressed as:

[0033]

[0034] Among them, is the assigned category pseudo-label; r″ top is the variable controlling the number of neighbors; Euclidean represents using the Euclidean distance as the distance metric.

[0035] In the above step S3, the loss function includes the cross-entropy loss L CE , the prototype contrast learning loss L PCL and the knowledge distillation loss L DIS , which is expressed as:

[0036] L = L CE + L PCL + L DIS ;

[0037] Among them, the cross-entropy loss L CE is obtained by comparing the classification result of the current target domain with the pseudo-label; the prototype contrast learning loss L PCL is obtained by comparing the similarity between the classifiers of the current target domain and the adjacent target domains; the knowledge distillation loss L DIS is obtained by comparing the classification result of the previous target domain dataset in the corresponding classifier with the classification result of the current target domain classifier.

[0038] The cross-entropy loss L CE is expressed as follows:

[0039]

[0040] Among them, is the distribution of the t-th target domain, x i is the data sample of the t-th target domain, K is the number of categories, is the weight parameter of the k-th class in the t-th target domain, is the pseudo-label of the sample, is the indicator function that determines whether the sample belongs to the k-th class. When the predicted class k is the same as the pseudo-label, the value of

[0041] the prototype contrastive learning loss L PCL is expressed as follows:

[0042]

[0043] where x j is another sample that belongs to the same t-th target domain as x i ; is the indicator function that determines whether two samples belong to different classes. When the pseudo-label classes of the two samples are different, the value of

[0044] the knowledge distillation loss L DIS is expressed as follows:

[0045]

[0046] where D KL (·) is the KL divergence; t is the target domain, and s is the source domain.

[0047] Compared with the prior art, the adaptive network learning method provided by the present invention for monitoring the state of the brake pads of high-speed trains on long and steep slopes has the following beneficial effects:

[0048] 1. By introducing the centroid similarity pseudo-label and the prototype contrastive learning method, etc., the prediction model of the present invention can better adapt to the change of the data distribution in the target domain, maintain a high recognition accuracy, has obvious advantages in solving multi-target problems, shows great potential in the actual application of train braking monitoring, and provides support for realizing the cross-rotation state monitoring of the brake pads of high-speed trains on long and steep slopes;

[0049] 2. By proposing the method for constructing the centroid and similarity pseudo-label, the present invention can improve the reliability of the pseudo-label in the target domain, and further improve the discrimination ability of unlabeled samples;

[0050] 3. By performing random mixing enhancement processing on the samples with high confidence in the source domain dataset and the target domain dataset, the present invention can effectively improve the ability of the model to extract class features under the change of data distribution;

[0051] 4. The present invention adopts the prototype contrastive learning method to align different domains and realizes the smooth migration from the source domain to multiple changing target domains. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1Schematic diagram of the adaptive network learning method for monitoring the status of brake pads of high-speed trains on long and steep slopes;

[0053] Figure 2 Schematic diagram of the training process for the target domain dataset;

[0054] Figure 3 Schematic diagram of the principle of the adaptive network learning method for monitoring the status of brake pads of high-speed trains on long and steep slopes;

[0055] Figure 4 Prediction accuracy curves for different domains;

[0056] Figure 5 Confusion matrix diagrams for different domains;

[0057] Figure 6 t-SNE visualization analysis results for different domains; where different colors in (a) represent samples of datasets A - D, (b) corresponds to the total classification results of the source domain C and target domains D, B, A, (c) corresponds to the classification results of target domain D, (d) corresponds to the classification results of target domain B, and (e) corresponds to the classification results of target domain A. Detailed implementation manners

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0059] Embodiment 1

[0060] This embodiment is directed to the scenario of monitoring the status of brake pads of high-speed trains on long and steep slopes. It analyzes the vibration acceleration signals collected during the braking process of high-speed trains. The status of the brake pads mainly includes normal, uneven wear, spalling, and cracks.

[0061] This embodiment provides an adaptive network learning method applicable to monitoring the status of brake pads of high-speed trains on long and steep slopes, as Figures 1 - 3As shown in the figure, learning is carried out using the source domain dataset and more than two target domain datasets. The source domain dataset includes several vibration acceleration signal samples and corresponding brake pad status labels. The target domain dataset only includes several vibration acceleration signal samples. First, perform short-time Fourier transform (STFT) on the source domain dataset and the target domain datasets, and use the source domain dataset and the target domain datasets after short-time Fourier transform as the input of the prediction model. The collected vibration acceleration is generally an unsteady signal. As an extended form of Fourier transform, short-time Fourier transform is suitable for analyzing the time-frequency characteristics of unsteady signals. STFT uses a window to divide a non-linear steady signal into multiple time periods, and then performs Fourier transform within each time period to obtain the spectral information of the signal within each short time domain. The obtained spectral information is spliced before and after to form the time-frequency domain information distribution of the entire signal. The size of the window function is selected according to the frequency of the signal to balance the time and frequency resolutions. The mathematical expression of STFT is as follows;

[0062]

[0063] Among them, ω is the signal frequency, w(τ) is the window function, and X(ω,τ) is the contribution of the signal with frequency ω at time τ in the window. Here, the window size is set to 512, and the vibration data of each sample is converted into a two-dimensional time-frequency signal using STFT to provide rich time-frequency information, improve the accuracy of feature extraction, and improve the performance of the monitoring model.

[0064] The adaptive network learning method for monitoring the brake pad status of high-speed trains on long uphill slopes provided in this embodiment, as Figures 1 - 3 shown, includes the following steps:

[0065] S1. Pre-train the prediction model composed of a feature extractor and a classifier using the source domain dataset.

[0066] In the present invention, there are no specific limitations on the feature extractor and the classifier. The feature extractor can be Resnet (such as ResNet18 / 34 / 50), VGG (such as VGG16 / VGG19), DenseNet (such as DenseNet121 / 169), etc. The classifier can be a Softmax classifier; an SVM classifier, etc.

[0067] The above step S1 includes the following sub-steps:

[0068] S11. Perform random mixing and enhancement processing on the samples in the source domain dataset, and add the processed samples to the source domain dataset.

[0069] Through random mixup augmentation, complex and diverse new samples can be generated, continuously improving the model's ability to extract class features under data distribution changes, thereby reducing the network's adaptation process to the target domain. The random mixup augmentation process is as follows: The sample to be processed and the noise n~P that follows a normalized distribution are N(0,1) input into the encoder (a convolutional layer is used here) together to achieve feature encoding. The AdaIN is used to transform the encoded features in different ways. Finally, the decoder (a transposed convolutional layer is used here) maps the features back to the original space, which is expressed as follows:

[0070]

[0071] where, e ξ is the encoder, d ζ is the decoder, l IN is the element-wise batch normalization layer; are the two linear transformation layers of AdaIN respectively; R i (x) is the result of the i-th transformation, i = 1, …, Nang, and Nang is the number of transformations;

[0072] The transformed sample and the original sample are randomly weighted and mixed under the normalized distribution , and finally, the enhanced sample is obtained after scaling through the sigmoid function. Its generation process is as follows:

[0073]

[0074] where, w0 is the weight of the original sample, and w i is the weight of the result of the i-th transformation.

[0075] S12. Use the source domain dataset obtained in step S11 to pre-train the prediction model.

[0076] In this embodiment, the cross-entropy loss function is used, SGD is used as the optimizer, and the number of iterations is set to 10, and the batch size is 16; then use the source domain dataset obtained in step S11 to train the prediction model, and the pre-trained prediction model is obtained after training. The classifier weight parameters obtained at this time are used as the source domain prototypes.

[0077] S2. Configure the corresponding prediction model for each target domain, and initialize it with the prediction model pre-trained in step S1. Test each target domain dataset through the prediction model to obtain the corresponding classification results.

[0078] For the target domain dataset t, in order to improve the ability of the model to extract class features under the change of the target domain data distribution, random mixing augmentation is also performed on the samples in the target domain data. In this embodiment, first, the pre-trained prediction model is used to test the target domain dataset t, and according to the set confidence threshold (the confidence threshold in this embodiment is 0.8), the samples with a confidence greater than the confidence threshold are extracted for random mixing augmentation (the process is as described above), and the processed samples are added to the target domain dataset t. The classifier for this target domain is trained based on the target domain dataset with the samples added by random mixing augmentation.

[0079] For the target domain dataset t, as Figure 2 shown, the following steps are performed:

[0080] S21, test the target domain dataset t through the prediction model, arrange the samples in descending order of the confidence output by the prediction model, and extract the samples with a confidence greater than the set confidence threshold range from the target domain dataset t to form the dataset F; classify according to the pseudo-labels of the corresponding samples obtained by testing the samples in the dataset F, and perform weighted aggregation on the corresponding sample features extracted by the feature extractor to construct several class centroids.

[0081] In this step, the reliability of sample classification is measured by the softmax confidence. According to the descending order of confidence, the first 50% of the samples are extracted from the target domain dataset t to construct the dataset F, that is, r top = 2. The construction process of the dataset F is expressed as:

[0082]

[0083] where, I k is the target domain sample selected in the k-th class; K is the number of all classes; N Tt is the number of target domain samples; g ω ,k(·) is the k-th element output by the classifier; f θ (·) is the feature extractor; is the set confidence threshold range.

[0084] According to the sample confidence, the sample features in the dataset F are weighted aggregated to construct several class centroids, which is expressed as follows:

[0085]

[0086] S22, by calculating the cosine similarity between the remaining samples of the target domain dataset t in step S21 and the class centroids of each class, arrange them in descending order of similarity, and extract the dataset F' with a similarity greater than the set similarity threshold range from the remaining samples of the target domain dataset t, and assign the same class pseudo-labels as the class centroids.

[0087] In this step, by calculating the cosine similarity between the centroid and each unlabeled target domain sample, arranging them in descending order of similarity, 50% of the samples are extracted from the remaining samples of the target domain dataset t obtained in step S21 to construct the dataset F′, that is, r t ′ op =2. The construction process of the dataset F′ is as follows:

[0088]

[0089] where I′ k is the target domain sample selected in the k-th category; K is the number of all categories; is the number of remaining samples in the target domain dataset t in step S21; g ω,k (·) is the k-th element output by the classifier; f θ (·) is the feature extractor; is the set similarity threshold range.

[0090] S23. Based on the dataset F′, construct a KNN classifier and assign class pseudo-labels to the remaining samples of the target domain dataset t in step S22.

[0091] Fit a K-Nearest Neighbors (KNN) classifier through the sample set F′ [see A. Fazli and J. Poshtan, "Wind turbine fault prognosis using SCADA measurements, pre-fault labeling, and KNN classifiers robust against data imbalance," Measurement, vol. 243, Feb 2025, Art no. 116202, doi: 10.1016 / j.measurement.2024.116202]. The KNN classifier assigns pseudo-labels to the target domain samples according to the k-nearest neighbor class distribution of the unlabeled samples, effectively reducing the misclassification risk when the samples have highly overlapping clusters.

[0092] Assign classes to the remaining samples of the target domain dataset t in step S22 through the KNN classifier The process is expressed as:

[0093]

[0094] where is the assigned class pseudo-label; r″ top is a variable controlling the number of neighbors. For example, when r″top When k = 5, calculate which category has the largest number among the categories to which the 5 assigned label samples closest to the to-be-assigned label sample belong, and then assign the same category pseudo-label to the to-be-assigned label sample; Euclidean means using Euclidean distance as the distance metric.

[0095] S3. Construct a loss function based on the test results of each target domain dataset, and use the constructed loss function to optimize the classifiers of each target domain. The weight parameters of each classifier serve as the prototypes of the corresponding target domains.

[0096] The loss function includes the cross-entropy loss L CE , the prototype contrastive learning loss L PCL and the knowledge distillation loss L DIS , which is expressed as:

[0097] L = L CE + L PCL + L DIS (8).

[0098] The cross-entropy loss L CE is obtained by comparing the classification result of the current target domain with the pseudo-label; it is expressed as follows:

[0099]

[0100] Among them, is the distribution of the t-th target domain, x i is the data sample of the t-th target domain, K is the number of categories, is the weight parameter of the k-th category in the t-th target domain, is the transpose, is the pseudo-label of the sample, is the indicator function to judge whether the sample belongs to the k-th category. When the predicted category k is the same as the pseudo-label, has a value of 1, and 0 otherwise.

[0101] The weight parameter ω of the classifier is expressed as:

[0102]

[0103] Each row ω k = [w k,1 w k,2 ...w k,d is the weight parameter p k of the k-th category.

[0104] The prototype contrastive learning loss L PCLIt is obtained by comparing the similarity between classifiers of the current target domain and the adjacent target domain (in this embodiment, the previous target domain is taken) among various categories. This comparison mechanism helps the model maintain consistent class discrimination ability when facing data from different domains. In this way, the model can strengthen the learning of data distribution characteristics among multiple target domains, thereby improving the ability of cross-domain adaptation. The prototype contrast learning loss L PCL is expressed as follows:

[0105]

[0106] where x j is another sample that belongs to the same t-th target domain as x i ; is an indicator function that determines whether two samples belong to different categories. When the pseudo-label categories of the two samples are different, is 1, and when they are the same, it is 0.

[0107] The knowledge distillation loss L DIS is obtained by comparing the classification results of the previous target domain dataset in the corresponding classifier with the classification results of the current target domain classifier, and is used to measure the model difference to ensure that the model does not forget the knowledge learned previously when learning the target domain. The knowledge distillation loss L DIS is expressed as follows:

[0108]

[0109] where D KL (·) is the KL divergence; t is the target domain, and s is the source domain; Equation (13) represents calculating the KL divergence based on the classification results of the source domain dataset in the source domain classifier and the classification results in the target domain classifier when t = 1; Equation (14) represents calculating the KL divergence based on the classification results of the target domain dataset t - 1 in the corresponding target domain classifier and the classification results in the target domain classifier when t.

[0110] After optimizing the weight parameters of each target domain classifier using the above loss function L, return to step S2, and repeat steps S2 - S3 until the classifiers of each target domain converge. In the present invention, when the number of training iterations reaches the upper limit, it indicates that the classifier converges.

[0111] The method provided by the present invention will be illustrated by way of example using the test data of the long-gradient cross-rotational speed braking test of the friction block in different states (normal, eccentric wear, spalling, and crack) of the brake pad. The friction system used in this test includes a brake disc, a friction block, and a brake caliper; the friction block is fixed to the caliper of the friction system by a fixture; the brake disc is driven to rotate by a flywheel.

[0112] The test adopts the drag braking method, and the specific parameters are as follows: the braking pressure P = 500 N (corresponding to the actual braking force of 36.25 KN), the friction radius R = 130 mm, and the braking disc rotation speed range is set to 400 - 700 rpm. In the braking test simulating the long and steep slope condition, the braking disc rotation speeds are set to 700, 600, 500, and 400 rpm in sequence, representing different speed stages during braking. First, the rotation speed is set to 700 rpm, and the clamp is started to press the brake pad to make it fit well with the braking disc. Then, the clamp is closed and the motor is started until the rotation speed reaches the set value, and then the clamp is started again to start the formal test. Considering the continuous braking state in the long and steep slope, after braking for 15 minutes according to the set conditions, the vibration acceleration data is collected. After this experiment, different damaged friction blocks are replaced to repeat the test. When the tests of four states are completed, the rotation speed value is changed to repeat the above test steps. To ensure the reliability of the collected data, each drag braking test is repeated 4 times, and the collection time for each time is 1 minute. During the braking process, the vibration acceleration signal generated by the test is obtained through the three-axis vibration acceleration sensor installed on the fixture, and the sampling frequency is 50 KHz. The vibration acceleration data collected under each working condition is shown in Table 1.

[0113] Table 1 Multi-target domain migration dataset

[0114]

[0115] The sample data in each dataset is first processed according to the short-time Fourier transform given above.

[0116] To verify the effectiveness of the proposed model for multi-target domain migration, multiple groups of multi-target domain migration experiments are conducted here to evaluate the monitoring performance of the model at multiple different rotation speeds under the long and steep slope working condition. The specific experimental settings are shown in Table 2. Experiment D→C, B, A means taking the vibration data at 700 rpm as the source domain, and the vibration data at the other three different rotation speeds as consecutive multiple target domains. The feature extractor uses Resnet18, and the classifier uses the Softmax classifier; the optimizer selected is SGD, and the initial learning rate is 5×10 -3 , the network iterates 10 times for each target domain, and the batch size is 16. 80% of the data in the source domain dataset is randomly selected as the training set, and 20% as the test set; all the data in the target domain is used for training and testing. First, the feature extractor and classifier are pre-trained using the training set in the source domain according to the steps S1 given in the adaptive network learning method described above; then, the classifiers for each target domain are trained and tested according to steps S2 and S3.

[0117] Table 2 Multi-target domain migration tasks

[0118]

[0119] Taking experiment C→D, B, A as an example, the change of the accuracy curve of the classifier learned by the method of the present invention is as follows Figure 4 shown. Each curve represents the performance of the model on its corresponding domain at different training stages. From Figure 4 it can be seen that at the 11th and 21st iteration times, the accuracy drops suddenly in each domain and rises after maintaining for a period of time. This phenomenon is because as time goes by and more target domain data are introduced, the weight parameters of the classifier are constantly changing to adapt to the feature distributions of multiple target domains, resulting in the classifier forgetting the feature knowledge learned from the source domain data. Prototype contrast learning (PCL) optimizes the classifier to increase the accuracy by aligning the domains and increasing the discrimination between different classes. It not only strengthens the adaptability of the classifier on the target domain but also improves the ability of the classifier to retain the source domain knowledge during the long-term learning process. As the number of iterations increases, the performance of the classifier on multiple target domains gradually improves and remains stable.

[0120] To test the classification effect of the classifier in the source domain and each target domain, the test set is input into the trained model, and a confusion matrix is generated, as Figure 5 shown. The numbers 0 to 3 in the figure respectively represent the normal, eccentric wear, spalling, and crack states of the friction block. The average accuracies of the source domain and each target domain are 100%, 97.5%, 98.25%, and 98.5% respectively. The results show that the trained classifier can accurately identify the state types in different domains when facing the data distributions of multiple target domains and maintain a high accuracy in the source domain.

[0121] To demonstrate the effect of the method of the present invention on the cross-rotational speed state monitoring of the friction block and illustrate the change of data in the model, a visual analysis is performed on the data characteristics of experiment C→D, B, A, and the results are as Figure 6 shown. Figure 6 (a) Different colors represent the samples of the four datasets A - D, Figure 6 (b)-(e) Different colors represent the four states of the friction block: normal, eccentric wear, spalling, and crack. From Figure 6 (a), it can be observed that the sample data of each domain are represented by different colors, and the sample features of different domains form obvious aggregation states in the feature space, verifying the effectiveness of the proposed model in multi-target domain tasks. Figure 6 (c)-(e) further show the distribution of the state features of different friction blocks in each target domain. It can be seen from the figure that the distribution of the state features in each target domain is the same as Figure 6(b) Similarity means that, based on the alignment of domain features, the same friction block damage states in different domains are successfully mapped to similar positions in the feature space. This indicates that the model continuously learns from the data and strengthens effective feature alignment strategies, enabling the data from different domains to form a unified distribution in the feature space while maintaining the discrimination ability of features in different damage states.

[0122] Those of ordinary skill in the art will realize that the embodiments here are to help the reader understand the principles of the present invention and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations without departing from the essence of the present invention according to the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. An adaptive network learning method applicable to the state monitoring of brake pads of high-speed trains on long and steep slopes, which uses a source domain dataset and more than two target domain datasets for learning, and is characterized in that, It includes the following steps: S1. Pre-train the prediction model composed of a feature extractor and a classifier using the source domain dataset; S2. Configure a corresponding prediction model for each target domain, and initialize it using the prediction model pre-trained in step S1. Test each target domain dataset using the prediction model to obtain the corresponding classification results. For the target domain dataset t, perform operations in the following sub-steps: S21. Test the target domain dataset t using the prediction model, sort the samples in descending order according to the confidence level output by the prediction model, and extract the samples with a confidence level greater than the set confidence level threshold range from the target domain dataset t to form a dataset F. Classify according to the pseudo-labels obtained by testing each sample in the dataset F, and perform weighted aggregation on the corresponding sample features extracted by the feature extractor to construct several class centroids; S22. Calculate the cosine similarity between the remaining samples of the target domain dataset t in step S21 and each class centroid, sort in descending order according to the similarity, and extract the dataset F' with a similarity greater than the set similarity threshold range from the remaining samples of the target domain dataset t, and assign the same class pseudo-label as the class centroid; S23. Construct a KNN classifier based on the dataset F', and assign class pseudo-labels to the remaining samples of the target domain dataset t in step S22; S3. Construct a loss function based on the test results of each target domain dataset, and use the constructed loss function to optimize the classifier of each target domain. Return to step S2, and repeat steps S2 - S3 until the classifiers of each target converge.

2. The adaptive network learning method for monitoring the state of the brake pads of high-speed trains applicable to long and steep gradients according to claim 1, characterized in that, Perform short-time Fourier transform on the source domain dataset and the target domain dataset, and use the source domain dataset and the target domain dataset after short-time Fourier transform as the input of the prediction model.

3. The adaptive network learning method for monitoring the state of the brake pads of high-speed trains applicable to long and steep gradients according to claim 1, characterized in that, The step S1 includes the following sub-steps: S11. Perform random mixup processing on the samples in the source domain dataset, and add the processed samples to the source domain dataset; S12. Pre-train the prediction model using the source domain dataset obtained in step S11.

4. The adaptive network learning method applicable to the state monitoring of the brake pads of high-speed trains on long and steep slopes according to claim 1, wherein For the target domain dataset t, first test the target domain dataset t using the pre-trained prediction model, extract the samples with a confidence level greater than the confidence level threshold for random mixup processing, and add the processed samples to the target domain dataset t.

5. The adaptive network learning method for monitoring the state of the brake pads of high-speed trains applicable to long and steep gradients according to claim 3 or 4, characterized in that The random mixing enhancement processing procedure is as follows: the sample to be processed and the noise n~P that follows the normalized distribution are input into the encoder together to achieve feature encoding, the encoded features are transformed in different ways using AdaIN, and finally the decoder maps the features back to the original space, which is expressed as follows: N(0,1) They are input into the encoder together to achieve feature encoding, the encoded features are transformed in different ways using AdaIN, and finally the decoder maps the features back to the original space, which is expressed as follows: Among them, e ξ is the encoder, d ζ is the decoder, and l IN is the element-wise batch normalization layer; are the two linear transformation layers of AdaIN respectively; R i (x) is the result of the i-th transformation, where i = 1, …, Nang, and Nang is the number of transformations; The converted sample and the original sample are subjected to random weight mixing under the normalized distribution and finally scaled by the sigmoid function to obtain the enhanced sample. Its generation process is as follows: Among them, w0 is the weight of the original sample, and w i is the weight of the i-th conversion result.

6. The adaptive network learning method applicable to the state monitoring of the brake pads of high-speed trains on long and steep gradients according to claim 1, characterized in that, In step S21, the construction process of the dataset F is expressed as: where, I k is the target domain sample selected from the k-th category; K is the number of all categories; is the number of target domain samples; g ω,k (·) is the k-th element output by the classifier; f θ (·) is the feature extractor; is the set confidence threshold range; Weightedly aggregate the sample features in the dataset F according to the sample confidence level to construct several class centroids, which is expressed as follows:

7. The adaptive network learning method applicable to the state monitoring of the brake pads of high-speed trains on long and steep slopes according to claim 1 or 6, characterized in that, In step S22, the construction process of the dataset F' is as follows: where, I′ k is the target domain sample selected from the k-th category; K is the number of all categories; is the remaining sample number of the target domain dataset t in step S21; g ω,k (·) is the k-th element output by the classifier; f θ (·) is the feature extractor; is the set similarity threshold range.

8. The adaptive network learning method for monitoring the state of the brake pads of high-speed trains applicable to long and steep gradients according to claim 7, wherein In step S23, the process of assigning categories to the remaining samples of the target domain dataset t in step S22 by the KNN classifier is expressed as: Among them, is the assigned class pseudo-label; r t ′ o ′ p is the variable for controlling the number of nearest neighbors; Euclidean means using Euclidean distance as the distance metric.

9. The adaptive network learning method applicable to the state monitoring of the brake pads of high-speed trains on long and steep slopes according to claim 1, characterized in that, In step S3, the loss function includes the cross-entropy loss L CE , the prototype contrastive learning loss L PCL and the knowledge distillation loss L DIS , which is expressed as: L = L CE + L PCL + L DIS ; Among them, the cross-entropy loss L CE is obtained by comparing the classification result of the current target domain with the pseudo-label; the prototype contrastive learning loss L PCL is obtained by comparing the similarity between the classifiers of the current target domain and the adjacent target domains in each category; the knowledge distillation loss L DIS is obtained by comparing the classification result of the previous target domain dataset in the corresponding classifier with the classification result in the current target domain classifier.

10. The adaptive network learning method applicable to the status monitoring of the brake pads of high-speed trains on long and steep gradients according to claim 9, wherein, Cross-entropy loss L CE is expressed as follows: wherein, is the distribution of the t-th target domain, x i is the data sample of the t-th target domain, K is the number of classes, is the weight parameter of the k-th class in the t-th target domain, is the pseudo-label of the sample, is the indicator function to judge whether the sample belongs to the k-th class. When the predicted class k is the same as the pseudo-label, has a value of 1, otherwise 0; Prototype contrastive learning loss L PCL is expressed as follows: where x j is another sample that belongs to the same t-th target domain as x i ; is an indicator function that determines whether two samples belong to different classes. When the pseudo-label classes of the two samples are different, it is 1, and when they are the same, it is 0; Knowledge distillation loss L DIS is expressed as follows: Among them, D KL (·) is the KL divergence; t is the target domain, and s is the source domain.

Images