Health state monitoring system and construction method of high-speed train brake pad under variable working conditions and noise influence

Abstract

The application discloses a health state monitoring system and a construction method of a high-speed train brake pad under the influence of noise and variable working conditions, which comprises an input layer, which is used for preprocessing the received brake pad signal; a feature extraction module, which is used for extracting the brake pad signal obtained through the input layer; an IAM mixed module, which comprises a plurality of extraction units arranged in parallel, and is used for re-extracting the features extracted by the feature extraction module respectively, and splicing the features extracted by each extraction unit; and a classifier, which is used for classifying according to the splicing result output by the IAM mixed module, and determining the brake pad fault type. The IAM module can learn the multiple domain invariant representations of the extracted features and align them, the distance between the source domain and the target domain is reduced through the loss function, the transfer between the source domain and the target domain is realized, and thus the monitoring state of the brake pad can be better monitored under the influence of noise and variable working conditions.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of rail transit technology, and relates to high-speed train brake pad monitoring, in particular to health state monitoring technology of high-speed train brake pad under variable working conditions with noise influence. BACKGROUND

[0002] Disc brake is an important part of high-speed train braking system, which is mainly realized by friction between brake pad and brake disc. Its performance change is closely related to the service safety of the train. The health state of the friction block on the brake pad has an important influence on the braking performance of the high-speed train. Friction block damage can lead to reduced service life of the brake pad, accelerated wear of the brake disc, and even endanger the safety of train operation. During the braking process of the high-speed train, as the working condition of the friction block changes continuously, the collected braking friction signals will have different data distributions. It is of great significance to realize the state monitoring of the brake pad under unknown working conditions. At the same time, the noise signals generated by the interference factors such as vehicle body vibration and electromagnetic signals bring great challenges to the feature extraction and state recognition of the brake pad.

[0003] In recent years, deep learning algorithms driven by big data have been applied in brake state monitoring. Through a large amount of labeled data collected, the diagnostic model is trained to effectively identify the state. Liu et al. proposed a heavy-haul train braking system state monitoring model based on multidimensional feature fusion and ensemble learning. Through principal component analysis and feature fusion, the redundant information of the important feature set is eliminated, and the fault recognition rate and training efficiency are improved (Multidimensional Feature Fusion and Ensemble Learning-Based Fault Diagnosis for the Braking System of Heavy-Haul Train, LIU Z, ZHANG M, LIU F, et al. [J]. Ieee Transactions on Industrial Informatics, 2021, 17(1): 41-51). Hu et al. applied the subfield generalization method to realize state monitoring of different wear states of the friction block in the brake pad (Deep subdomain generalisation network for health monitoring of high-speed train brake pads, HU R, ZHANG M, MENG X, et al. [J]. Engineering Applications of Artificial Intelligence, 2022, 113). The above researches on train braking system and brake pad state monitoring have achieved good results, but they do not consider the brake state monitoring under variable working conditions. Changes in working conditions will change the distribution probability of the collected friction block data, and then cause domain shift phenomenon. In practice, the working conditions such as train speed and load are not constant, and it is impossible to collect friction data under all working conditions. Therefore, it has become a research hotspot to solve the brake pad state monitoring under unknown working conditions through known working condition data.

[0004] Transfer learning has achieved good results in solving the problem of variable working condition state monitoring, but in the case of noise interference, the brake data deviates from the true value, and then the distribution of the sample is disturbed, which cannot effectively distinguish samples with different labels. Especially in the process of high-speed train braking with variable working environment, due to the multi-working condition change state of the brake pad, the noise presents multiple intensities, which makes it difficult to effectively monitor the brake pad based on traditional monitoring technology. SUMMARY

[0005] The application aims to provide a high-speed train brake pad health state monitoring system under variable working conditions with noise influence, map the extracted features to different low-dimensional feature spaces, and make the system have better robustness and realize train brake pad state monitoring under variable working conditions with different intensity noise influence.

[0006] Another object of the application is to provide a construction method of the high-speed train brake pad health state monitoring system under variable working conditions with noise influence.

[0007] To achieve the above object, the application adopts the following technical solutions.

[0008] The high-speed train brake pad health state monitoring system under variable working conditions with noise influence provided by the application comprises:

[0009] An input layer is used to preprocess the received brake pad signal.

[0010] A feature extraction module is used to extract the features of the brake pad signal obtained through the input layer.

[0011] An IAM mixing module comprises a plurality of parallel extraction units, is used to re-extract the features extracted by the feature extraction module, and splices the features extracted by the extraction units.

[0012] A classifier is used to classify the spliced results output by the IAM mixing module and determine the brake pad fault type.

[0013] The input layer is used to preprocess the received brake pad signal and convert the brake pad signal into feature data consistent with the input format of the feature extraction module. In the application, the input layer mainly comprises a convolution layer, a BN (Batch Normalization) layer, a ReLu function and a maximum pooling layer.

[0014] The feature extraction module is used to extract the features of the signal sample and protect the integrity of the information. Therefore, the ResNet residual network is used in the application. The ResNet residual network comprises four kinds of residual blocks connected in sequence, each residual block comprises three convolution layers (CN1, CN2 and CN3), and the number of each residual block is 2-4. The first convolution layer of each residual block is provided with a parallel downsampling layer, and the downsampling layer comprises a convolution layer and a batch normalization layer, and the convolution kernel size and the output channel number of the convolution layer are consistent with the third convolution layer of the parallel residual block.

[0015] The IAM hybrid module is used to extract more features and further improve the recognition performance of the system.

[0016] The classifier includes a full connection layer for reorganizing the plurality of features and a softmax layer for outputting a classification result.

[0017] The high-speed train brake pad health state monitoring system further comprises an optimization module configured to obtain a system loss function during system training and optimize system network parameters according to the loss value.

[0018] The system loss function includes a classification loss and a domain adaptation loss, and the loss function formula is as follows:

[0019]

[0020] In the formula, represents a cross-entropy loss function, represents a domain adaptation loss, represents a system prediction result, represents an i-th signal sample in a source domain, represents an i-th signal sample label in the source domain, n s represents the number of signal samples in the source domain, h j represents a j-th extraction unit, n r represents the number of extraction units, g represents an input layer and a feature extraction module, X s represents a source domain sample signal, X t represents a target domain sample signal, represents a trade-off parameter.

[0021]

[0022] In the formula, represents a representation and a representation , C represents the number of label categories, and c represents a label category. represents the number of samples with the label c in the source domain, represents a k-th sample with the label c in the source domain, represents a sample set with the label c in the source domain, represents the number of samples with the label c in the target domain, represents an l-th sample with the label c in the target domain, denote a set of samples in the target domain with label c, denote a set of samples in the target domain with label c, denote a set of samples in the target domain with label c,

[0023] The optimization module adopts a stochastic gradient descent (SGD) algorithm to optimize the system network parameters.

[0024] The application further provides a construction method of the high-speed train brake pad health state monitoring system under the noise influence of the variable working condition, which comprises the following steps:

[0025] S1 divide the training data into a plurality of batch training subsets, each batch training subset containing the same number of source domain data sets and target domain data sets;

[0026] S2 train the high-speed train brake pad health state monitoring system using each batch training subset in turn;

[0027] S3 obtain a system loss function after completing each batch training, the system loss function comprising a classification loss and a domain adaptation loss, and the loss function formula being as follows:

[0028]

[0029] In the formula, denote a cross-entropy loss function, denote a domain adaptation loss, denote a system prediction result, denote an i-th signal sample in the source domain, denote an i-th signal sample label in the source domain, n s denote a source domain signal sample number, h j denote a j-th extraction unit, n r denote an extraction unit number, g denote an input layer, X s denote a source domain sample signal, X t denote a target domain sample signal, denote a trade-off parameter;

[0030] S4 optimize the system network parameters based on the obtained loss value.

[0031] In the step S1, the source domain data set is composed of brake pad sample signals with a plurality of labels; and the target domain data set is composed of brake pad sample signals without labels.

[0032] In the step S3,

[0033]

[0034] In the formula, denote a representation with the condition distribution difference between the representation C represents the number of label categories, c represents a label category, represents the number of samples with label c in the source domain, represents the k-th sample with label c in the source domain, represents the set of samples with label c in the source domain, represents the number of samples with label c in the target domain, represents the l-th sample with label c in the target domain, represents the set of samples with label c in the target domain, represents the feature space to which the sample is mapped by the mapping function .

[0035] In the above step S4, the system network parameters are optimized by using a stochastic gradient descent (SGD) algorithm.

[0036] Taking the completion of training of all batch training subsets as an iteration cycle, the above steps S2-S4 are repeated until the number of iterations is reached or the loss value tends to be stable.

[0037] The sample signal of the brake pad to be tested is input into the trained high-speed train brake pad health state monitoring system, and the corresponding brake pad fault type can be output.

[0038] The high-speed train brake pad health state monitoring system under the influence of noise and under varying conditions provided by the application has the following beneficial effects:

[0039] The IAM module can learn multiple domain-invariant representations of the extracted features and align them, and the distance between the source domain and the target domain is reduced by the proposed loss function, realizing the transfer between the source domain and the target domain, so that the brake pad monitoring state can be better monitored under the influence of noise and under varying conditions. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The data processing flow of the high-speed train brake pad health state monitoring system for the source domain and the target domain.

[0041] Figure 2 The construction process schematic diagram of the high-speed train brake pad health state monitoring system under the influence of noise and under varying conditions in the application example.

[0042] Figure 3 The four health state schematic diagram of the friction block.

[0043] Figure 4 The spalling state friction block sample vibration signal spectrum of different signal-to-noise ratios.

[0044] Figure 5Test results of classification accuracy of different models in different migration tasks.

[0045] Figure 6 Test results of state recognition accuracy of different models in different migration tasks.

[0046] Figure 7 Confusion matrix of recognition results obtained by different models in the same migration task.

[0047] Figure 8 T-sne dimensionality reduction visualization results of target domain features of the last hidden layer of different models in different migration tasks. DETAILED DESCRIPTION

[0048] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings and embodiments.

[0049] Embodiment 1

[0050] The noise-affected variable-condition high-speed train brake pad health state monitoring system provided in the embodiment includes an input layer, a feature extraction layer, an IAM hybrid module, and a classifier.

[0051] The input layer is used for preprocessing the received brake pad signal and converting the brake pad signal into feature data (in a format consistent with the input format of the feature extraction module). Here, the input layer mainly consists of a convolution layer (kernel size of 3, channel number of 128), a BN (Batch Normalization) layer, a ReLu function, and a maximum pooling layer.

[0052] The feature extraction module is used for feature extraction of the brake pad signal obtained through the input layer and protection of the integrity of the information. Here, the feature extraction module uses a ResNet residual network. The ResNet residual network includes four kinds of residual blocks (block1-block4) connected in sequence, and the number of each kind of residual block is 3, 4, 3, and 2, respectively. Each residual block includes three convolution layers (CN1, CN2, and CN3), and a batch normalization layer is provided after each convolution layer; the parameter settings of the four kinds of residual blocks are shown in Table 1.

[0053] Table 1: Convolution layer parameters of four kinds of residual blocks

[0054]

[0055] The first convolution layer CN1 of each residual block is provided with a down-sampling layer in parallel, and the down-sampling layer consists of a convolution layer and a batch normalization layer, and the convolution kernel size and the output channel number of the convolution layer are consistent with the third convolution layer CN3 of the residual block in parallel.

[0056] The IAM mixing module includes a plurality of extraction units (the number of extraction units in the embodiment is 4) arranged in parallel, which are used to re-extract the features extracted by the feature extraction module respectively, and splice the features extracted by each extraction unit. The IAM mixing module can extract more features, and further improve the recognition performance of the system. Each extraction unit uses a different convolutional neural network structure, and a global average pool is arranged after each extraction unit, which is used to convert the obtained feature data into one-dimensional data to obtain a feature vector. The convolutional neural network structure of the four extraction units is shown in Table 2, and the convolutional neural network includes 1-3 convolutional layers, and an average pooling layer can also be arranged after the convolutional layer.

[0057] Table 2: Convolutional neural network structure of four extraction units

[0058]

[0059] The classifier is used to classify the spliced results output by the IAM mixing module to determine the brake pad fault type. The classifier includes a fully connected layer for recombining a plurality of features and a softmax layer for outputting a classification result.

[0060] The high-speed train brake pad health state monitoring system described above further includes an optimization module for obtaining a system loss function during system training, and optimizing system network parameters according to the loss value.

[0061] The system loss function includes a classification loss and a domain adaptation loss, and the loss function formula is as follows:

[0062]

[0063] In the formula, represents a cross-entropy loss function, represents a domain adaptation loss, represents a system prediction result, represents the i-th signal sample of the source domain, represents the i-th signal sample label of the source domain, n s represents the number of signal samples of the source domain, h j represents the j-th extraction unit, n r represents the number of extraction units, g represents the input layer and the feature extraction module, X s represents the source domain sample signal, X t represents the target domain sample signal, represents a trade-off parameter.

[0064]

[0065] In the formula, represents a feature and a feature The conditional distribution differences between them, where C represents the number of label types and c represents the label types. This represents the number of samples with label c in the source domain. Let c represent the k-th sample in the source domain with label c. Let c represent the set of samples labeled c in the source domain. This represents the number of samples with label c in the target domain. Let c represent the l-th sample in the target domain with label c. Let c represent the set of samples labeled c in the target domain. Indicates through mapping function The feature space to which the samples are mapped. Here, the mapping function... It refers to a neural network consisting of an input layer, a feature extraction module, and an IAM hybrid module.

[0066] The optimization module described above uses the stochastic gradient descent (SGD) algorithm to optimize the system network parameters.

[0067] The following is combined Figure 1 The construction method of the high-speed train brake pad health status monitoring system under varying operating conditions and noise influence is explained in detail, specifically including the following steps:

[0068] S1 divides the training data into several batches of training subsets, each batch containing the same number of source domain datasets and target domain datasets.

[0069] In this step, the source domain dataset consists of several labeled brake pad sample signals; the target domain dataset consists of several unlabeled brake pad sample signals.

[0070] S2 sequentially trains the high-speed train brake pad health status monitoring system using each batch of training subsets.

[0071] In this step, the monitoring system is trained using a mini-batch training method. The training data is divided into several batches of training subsets, and one training subset is used for each training session. After each training session is completed, steps S3 and S4 are executed.

[0072] During training, both the source domain dataset and the target domain dataset are simultaneously input into the monitoring system. The two datasets share the system's network parameters.

[0073] S3 obtains the system loss function after each batch of training is completed. The system loss function includes classification loss and domain adaptation loss, and the formula for the loss function is as follows:

[0074]

[0075] In the formula, denotes a cross-entropy loss function, denotes a domain adaptation loss, denotes a system prediction result, denotes an i-th signal sample of a source domain, denotes an i-th signal sample label of a source domain, n s denotes a source domain signal source domain number, h j denotes a j-th extraction unit, n r denotes an extraction unit number, g denotes an input layer, X s denotes a source domain sample signal, X t denotes a target domain sample signal, denotes a trade-off parameter.

[0076]

[0077] S4, based on the obtained loss value, optimizing the system network parameters.

[0078] This step, based on the loss value obtained in step S3, uses the stochastic gradient descent (SGD) algorithm to optimize the system network parameters (including the network parameters of the input layer, the feature extraction layer, and the IAM hybrid module).

[0079] Taking the completion of training of all batch training subsets as an iteration cycle, repeating the above steps S2-S4 until the number of iterations is reached or the loss value tends to be stable.

[0080] Application Example

[0081] Friction between the friction block and the brake disc will cause vibration. The change rule of vibration shows different characteristics with the change of the health state of the friction block. According to this principle, the vibration signal collected by the acceleration sensor can be analyzed to realize the health state monitoring of the friction block. The vibration acceleration of the friction block is collected on the high-speed train brake scaling test bench. A three-dimensional acceleration sensor (sensitivity 10 mV / g, range ±500g) installed on the brake disc above the back of the friction block can collect vibration acceleration signals in three directions.

[0082] The braking process of the high-speed train brake pad is simulated by using the high-speed train brake scaling test bench. The experimental sampling frequency is 50KHz, and the vibration data of the friction block under three brake disc rotation speeds (350rpm, 400rpm, 450rpm) are collected. The noise-affected variable working condition high-speed train brake pad health state monitoring system (MRAN) provided in embodiment 1 is trained and tested. As shown in Figure 2 The overall scheme of this application example is composed of four main steps of signal collection, data processing, system training, verification and application.

[0083] (1) Collect the vibration data (tangential acceleration) of the brake disc under different rotating speeds during the braking process of the friction block under different fault conditions,

[0084] (2) Standardize the vibration data using the mean and variance, then segment it into samples of equal length. Divide the data set (training set and test set) according to the rotating speed and noise intensity.

[0085] (3) Train the high-speed train brake pad health state monitoring system (MRAN) under variable working conditions provided in Example 1 with the training set to obtain a transfer model with the smallest distance between the source domain and the target domain and save it. It can apply the knowledge of the friction block state learned in the source domain to the target domain for state monitoring.

[0086] (4) Input the target samples in the test set into the saved monitoring system to obtain the label discrimination result of the target samples and realize brake pad state monitoring.

[0087] In the experiment, to simulate different braking fault states, according to the previous literature research and data collection, four kinds of actual fault types were selected for processing, including normal, crack, eccentric wear and peeling, as shown in Figure 3 Based on the characteristics of the leading edge of different friction blocks, the corresponding specific setting parameters are set as follows:

[0088] (1) Normal: the friction block is in normal working condition, as shown in Figure 3 (a).

[0089] (2) Crack: high-frequency vibration can cause changes in local contact pressure on the surface of the friction block. Repeated normal and tangential stress oscillation under high-frequency vibration can cause cracks on the surface of the friction block; here, a 1mm wide and 1mm deep crack is machined on the friction surface to simulate the crack state of the friction block, as shown in Figure 3 (b).

[0090] (3) Eccentric wear: during braking, a moment is formed between the friction block and the brake disc, causing uneven distribution of contact surface shear stress and different degrees of wear, i.e. eccentric wear phenomenon; here, a 1mm eccentric wear friction block is designed, as shown in Figure 3 (c).

[0091] (4) Peeling: during the friction process of the friction block and the brake disc, stress is concentrated near the contact leading edge, forming uneven areas on the surface of the friction block and accumulating wear particles. Repeated wear can cause peeling and scratches on the friction block. A 5mm 2 area of 1mm deep peeling area is formed on the surface of the friction block, as shown in Figure 3 (d).

[0092] The tangential vibration acceleration of the friction blocks was used as the signal data for friction block condition monitoring. 5 × 10⁻⁶ samples were collected for each type of friction block. 5 The data was collected at 10 sampling points and standardized. Based on the standardized data, 500 random, non-repeating samples were generated, each containing 512 sampling points. In practice, there are many types of noise; in this application example, representative Gaussian white noise was selected as the contamination noise for the vibration signal. Gaussian white noise with different signal-to-noise ratios (10dB, 5dB, 0dB, -5dB, and -10dB) was added to the vibration data to generate samples containing noise. The signal-to-noise ratio is defined as:

[0093]

[0094] Repeat the above steps to generate three datasets with different signal-to-noise ratios under three rotational speed conditions, each dataset containing 2000 samples. Dataset information is shown in Table 3.

[0095] Figure 4 The time-domain vibration signals of friction block samples in the peeling state with different signal-to-noise ratios when the brake disc rotation speed is equal to 450 rpm; Figure 4 (a) in the image shows vibration data without added noise, clearly showing periodic pulses. As the signal-to-noise ratio decreases, the added noise increases its interference with the effective information in the signal, making it more difficult to monitor the condition of the friction block.

[0096] Table 3 Friction block dataset information

[0097]

[0098] Choose any two datasets (with the same signal-to-noise ratio) from Table 3 and use them as the source domain dataset and the target domain dataset, respectively. Use the labeled source domain dataset and the unlabeled target domain dataset as the training set, and use the labeled target domain dataset as the test set.

[0099] The monitoring system was trained according to the construction method provided in Example 1. Based on the construction method given above, each training set was further divided into 64 batch training subsets, and the maximum number of iterations was set to 100, with a learning rate of 5×10⁻⁶. -2 The SGD optimization algorithm is used to optimize the network parameters of the system.

[0100] Then, the monitoring effect of the constructed monitoring system was tested using a test set. To reduce randomness, each experiment was performed 10 times and the average value was taken. The experimental results are shown in Table 4, where the signal-to-noise ratios of the source and target domains are the same.

[0101] Table 4. Migration tasks under varying noise levels and accuracy (%)

[0102]

[0103] As can be seen from Table 4, in the 6 groups of experiments without adding noise, the classification accuracy of C to A is the highest, which is 99.31%, and the average classification accuracy is 98.75%. It shows that under the condition without noise interference, the monitoring system provided by the application has good migration ability and label discrimination ability for the friction block data set under different brake disc rotating speed conditions. With the decrease of signal-to-noise ratio, the classification accuracy of the monitoring system in the variable rotating speed migration task presents a decreasing trend, and the average accuracy of the 6 migration tasks decreases from 98.75% to 88.65%. The results show that the monitoring system provided by the application has good robustness, and when the source domain and target domain data are interfered by the same size of noise, the monitoring system can still achieve good results in the cross-domain identification task of variable working condition friction block data.

[0104] To fully demonstrate the superiority of the monitoring system provided by the present application, the state recognition accuracy of different domain adaptive networks in the six transfer tasks in Table 3 is compared. The feature extraction part of the comparison model uses ResNet consistent with the network parameters in the application example. The first comparison model uses DAN (SUN H, CHEN X, WANG L, et al. C(2)DAN: An Improved Deep Adaptation Network with Domain Confusion and Classifier Adaptation [J]. Sensors (Basel), 2020, 20(12)), which uses a three-layer adaptive network and MK-MMD for domain adaptation. The second comparison model uses DDC (ZHANG W W, WANG F, JIANG Y, et al. Cross-Subject EEG-Based Emotion Recognition with Deep Domain Confusion [Z]. INTELLIGENT ROBOTICS AND APPLICATIONS, ICIRA 2019, PT I. 2019: 558-70.10.1007 / 978-3-030-27526-6_49), which uses a one-layer adaptive network and MMD to align the representations between the domains. The third comparison model uses DANN (LI F, TANG T, TANG B, et al. Deep convolution domain-adversarial transfer learning for fault diagnosis of rolling bearings [J]. Measurement, 2021, 169), which uses adversarial learning to reduce domain differences.

[0105] The comparison network is tested for 10 times per group of experiments, and the experimental results are shown in Figure 5 The horizontal axis in the figure is the signal-to-noise ratio of the source domain and the target domain data, and the vertical axis is the classification accuracy. It can be seen that the accuracy of the monitoring system (proposed) provided by the present application is generally better than that of the three comparison models in each transfer task. In addition, as the signal-to-noise ratio decreases, the accuracy of the four networks in the transfer task decreases, but it can be seen that the accuracy of the monitoring system (proposed) provided by the present application is less affected. The results show that when the source domain and the target domain are affected by the same intensity of noise, MRAN can extract more effective state features from the friction block vibration data, and has better robustness and label discriminability.

[0106] To fully verify the domain adaptation capability of the monitoring system provided by the application under different intensity noise interference under different working conditions, the application further sets six migration tasks as shown in Table 5 and conducts experiments. As can be seen from the table, the accuracy of MRAN is the highest 99.38% in migration experiment A ∞ - B 10 , and the accuracy is the lowest 78.48% in migration experiment A 10 -C -10 .

[0107] Table 5 Accuracy (%) of variable noise intensity and variable working condition migration tasks

[0108]

[0109] As described above, to fully prove the superiority of the monitoring system provided by the application, the state recognition accuracy of different domain adaptation networks in the six migration tasks in Table 4 is compared. Each set of comparison experiments is conducted 10 times to take the average value, and the experimental results are shown in Figure 6 . It can be seen that the monitoring system (proposed) provided by the application has the highest accuracy in all migration tasks. Taking (c) in Figure 6 as an example, as the signal-to-noise ratio of the target domain data decreases, the difference between the accuracy of the three comparison models and the monitoring system provided by the application in the migration task becomes larger and larger. When the SNR of the target domain is-10dB, the accuracy of the monitoring system (proposed) can still reach 81.68%, which is 8.57%, 14.34% and 15.8% higher than the comparison models DAN, DDC and DANN respectively, proving that the monitoring system provided by the application has better robustness and domain adaptation capability in this migration task.

[0110] By comparing the experimental results of migration tasks Y1, Y3 and Y5, it can be seen that as the signal-to-noise ratio of the source domain decreases, the accuracy of the four networks in the migration task also decreases, but the decrease of the signal-to-noise ratio of the source domain has less effect on the monitoring system (proposed) in the friction block state recognition. The results show that MRAN can still learn effective friction block state features from the source domain with noise interference and apply them to the migration task.

[0111] In order to more intuitively reflect the classification effect of the proposed method, three migration tasks are set as shown in Table 6. The target domain sample recognition results of migration task U3 are shown in the confusion matrix as shown in Figure 7 . In the figure, labels 1-4 represent four friction block states of wear, peeling, normal and cracking, respectively. The abscissa represents the predicted label, and the ordinate represents the true label. From the (a) confusion matrix in Figure 7 , it can be seen that most of the misjudgments are normal and peeling recognition errors.Figure 7 The large amount of peeling in (b), (c), (d) in (a) and the normal sample are misjudged, which shows that in the migration task U3, the label discrimination ability of the three comparative models is not as good as the monitoring system (proposed) provided by the application. Four networks can correctly identify all 500 samples in the eccentric wear state, which shows that under the influence of noise, MRAN can still effectively identify samples in different working conditions in the eccentric wear state.

[0112] Table 6 Variable noise intensity variable working condition friction block data migration task

[0113]

[0114] The target domain features of the last hidden layer of each model in the migration task of Table 6 are subjected to t-sne dimension reduction visualization processing, and the results are shown in Figure 8 . Figure 8 (a)-(c) in (a)-(c) are the visualization results of the monitoring system provided by the application in the migration tasks U1, U2 and U3. Figure 8 (d)-(f) in (d)-(f) are the visualization results of the three networks DAN, DDC and DANN in the migration task U3. From Figure 8 (a)-(c) in (a)-(c) can be seen, as the signal-to-noise ratio of the target domain data decreases, different kinds of samples still maintain good clustering, but a small amount of samples will be confused into other kind clusters. From Figure 8 (c)-(f) in (c)-(f) can be seen, the three comparative models fail to effectively distinguish the peeling, normal and crack three states, the monitoring system (proposed) provided by the application has better clustering, and maintains the label discrimination while reducing the domain difference.

[0115] The application proposes to use MRAN to monitor the health state of high-speed train brake pad under the influence of noise and variable working conditions, realize cross-domain state recognition of friction block under unknown working conditions by using known working condition friction block data. MRAN uses labeled source data and unlabeled target data for model training, the ResNet contained in the network can extract deep features of the data, IAM can align and extract multiple domain invariant representations of the learned features, and better realize the health state monitoring of the friction block under unknown working conditions and noise influence. The network measures and reduces the distance between the source domain and the target domain through CMMD, and realizes the transfer between the source domain and the target domain. At the same time, the network learns the friction block features of different states in the source domain, and realizes the identification of the health state of the friction block. The application verifies the effectiveness of the proposed method on the brake friction data set. The experimental results show that in the health state monitoring of the brake pad under the influence of noise and variable working conditions, the monitoring system provided by the application can better realize unsupervised state monitoring, and show better clustering and label discriminability for experimental data. The application has certain practicability in real-time state monitoring of train brake pad and actual engineering scene, and the state monitoring of high-speed train friction block in more working conditions is the research goal in the future.

Claims

1. A high-speed train brake pad health state monitoring system under the influence of noise, characterized in that, The system comprises: an input layer for preprocessing the received brake pad signal; a feature extraction module for extracting features from the brake pad signal obtained through the input layer; an IAM mixing module comprising four extraction units arranged in parallel, for re-extracting the features extracted by the feature extraction module and splicing the features extracted by each extraction unit; each extraction unit uses a different convolutional neural network, and a global average pool is arranged after each extraction unit; the first extraction unit is provided with a convolutional layer with a kernel size of 1; the second extraction unit is provided with two convolutional layers, the first convolutional layer has a kernel size of 1, and the second convolutional layer has a kernel size of 5; the third extraction unit is provided with three convolutional layers, the first convolutional layer has a kernel size of 1, the second convolutional layer has a kernel size of 3, and the third convolutional layer has a kernel size of 3; the fourth extraction unit comprises a convolutional layer and an average pooling layer, and the convolutional layer has a kernel size of 1; a classifier for classifying the spliced results output by the IAM mixing module to determine the brake pad fault type; an optimization module for obtaining a system loss function during system training and optimizing system network parameters according to the loss value; Steps: S1 divide the training data into several batches of training subsets, each batch of training subsets containing the same number of source domain data sets and target domain data sets; S2 use each batch of training subsets to train the high-speed train brake pad health state monitoring system in turn; S3 obtain the system loss function after completing each batch of training; S4 optimize the system network parameters based on the obtained loss value; The system loss function comprises a classification loss and a domain adaptation loss, and the loss function formula is as follows: ； In the formula, denotes a cross-entropy loss function, denotes a domain adaptation loss, denotes a system prediction result, denotes an i-th signal sample of a source domain, denotes an i-th signal sample label of a source domain, n s denotes a number of signal samples of a source domain, h j denotes a j-th extraction unit, n r denotes a number of extraction units, g denotes an input layer and a feature extraction module, X s denotes a source domain sample signal, X t denotes a target domain sample signal, denotes a trade-off parameter; ； wherein, denotes the conditional distribution difference between the representations and the representations denotes the number of classes of labels, c denotes a class of labels, denotes the number of samples in the source domain with label c, denotes the k-th sample in the source domain with label c, denotes the set of samples in the source domain with label c, denotes the number of samples in the target domain with label c, denotes the l-th sample in the target domain with label c, denotes the set of samples in the target domain with label c, denotes the feature space into which the samples are mapped by the mapping function .

2. The noise-affected down-regulation working condition monitoring system for the high-speed train brake pad health state according to claim 1, characterized in that, The feature extraction module uses a ResNet residual network.

3. The noise-affected down-regulation condition monitoring system for the high-speed train brake pad of claim 2, wherein, The ResNet residual network comprises four residual blocks connected in sequence, and each residual block comprises three convolutional layers; the number of each residual block is 2-4; the first convolutional layer of each residual block is arranged in parallel with a downsampling layer, and the downsampling layer comprises a convolutional layer and a batch normalization layer, and the convolutional kernel size and output channel number of the convolutional layer are consistent with those of the third convolutional layer of the residual block connected in parallel.

4. The noise-affected down-regulation condition monitoring system for the high-speed train brake pad of claim 1, wherein, The classifier comprises a fully connected layer for recombining multiple features and a softmax layer for outputting classification results.

5. The noise-affected down-regulation condition monitoring system for high-speed train brake pad of claim 1, wherein, In step S1, the source domain data set is composed of brake pad sample signals with labels, and the target domain data set is composed of brake pad sample signals without labels.

Images