Method for monitoring cross-rotating-speed state of brake pad of high-speed train influenced by vibration noise under long ramp

Abstract

The invention belongs to the technical field of data monitoring, and discloses a method for monitoring the cross-rotating-speed state of a high-speed train brake pad influenced by vibration noise under a long ramp, and the method comprises the steps: firstly reconstructing source domain data to obtain an uncertain set, and then constructing a feature extraction network of group offset generalization in combination with a set network parameter objective function; and monitoring the state of the brake pad of the high-speed train according to the constructed group offset generalization feature extraction network. According to the method, the influence of group offset on the accuracy is reduced as much as possible through the constructed group offset generalization network; furthermore, by means of elastic network regularization, the complexity of the model is reduced, meanwhile, the influence of noise on the model is reduced, and the cross-rotating-speed state monitoring of the brake pad under the influence of wheel-rail vibration noise is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of data monitoring, and relates to high-speed train running status monitoring, and in particular to a method for monitoring the cross-speed status of a high-speed train brake pad affected by vibration and noise on a long and steep slope. Background Art

[0002] As a key component of friction braking of high-speed trains, brake pads play an important role in stable speed regulation, accurate parking and emergency braking of trains. The quality of their braking performance is the basis for ensuring the safe operation of trains. Under long and steep slope conditions, the train brakes for a long time under a large braking force, which leads to aggravated damage to the brake pad interface and different damage states of the friction block (such as eccentric wear, cracks and peeling). The vibration signals of the friction blocks in different states contain rich friction interface characteristics, so a reasonable analysis of the signals can be used to know the health state of the friction interface. However, under long and steep slope conditions, in order to stabilize the speed, the train adopts a periodic and continuous braking method, and the speed changes frequently. The difference in signal characteristics at different speeds causes obvious domain offset, making it difficult to extract invariant features to achieve cross-speed state monitoring. At the same time, wheel-rail vibration noise can easily excite higher-order modal vibrations of the braking system, affect the tribological behavior of the braking interface, make the vibration signal characteristics more complex and changeable, further lead to group offset, and bring challenges to cross-speed state monitoring. Therefore, studying the vibration signals affected by wheel-rail vibration and noise under long and steep slope conditions and realizing cross-speed status monitoring of high-speed train brake pads are of great significance for early warning of brake failures and reducing the maintenance costs of high-speed trains.

[0003] At present, many scholars at home and abroad have conducted simulation research on brake wear and brake temperature on long and steep slopes. Zhou et al. established an image recognition model through an adaptive grayscale mapping method to calculate the wear of brake pads. Finally, it was found that the wear of brake pads on long and steep slopes increased by nearly 50% compared with that on traditional slopes (see Zhou Hangbo, "Analysis of Wear Law of Brake Pads of High-speed Trains on Long and Slopes," Henan Science, vol. 41, no. 9, pp. 1367-1374, 2023). Zhou et al. simulated and calculated aluminum-embedded steel structure brake discs and all-steel brake discs using finite element software, and obtained the temperature field and thermal stress distribution of the brake discs. The results showed that aluminum-embedded steel brake discs can significantly reduce the temperature and temperature gradient of the brake disc surface while achieving lightweight (see Zhou Suxia, Shao Jing, and Tong Xin, "Research on New Brake Discs Based on Continuous Braking on Long and Slopes," Journal of Mechanical Engineering, vol. 58, no. 20, pp. 391-398, 2022). The above method provides theoretical guidance for braking on long and steep slopes. However, long-term braking on long and steep slopes can cause wear of the brake pads, causing the brake system to fail, posing a huge hidden danger to train driving safety. Therefore, real-time monitoring of the health status of the brake pads on long and steep slopes can effectively prevent brake failure and ensure train driving safety.

[0004] Data-driven condition monitoring methods have begun to be introduced for high-speed train brake pad state recognition. However, the brake pad signal data under long and steep slope conditions in actual braking is easily affected by braking time, braking pressure, and the elastic-plastic deformation, surface cracks, and surface morphology of the material. Even the vibration signal characteristics collected by the friction blocks at the same speed are different, resulting in category shift. In particular, the vibration law caused by friction at different speeds on long and steep slopes is more complex and changeable. At the same time, the wheel-rail vibration noise interferes with the original feature distribution, which leads to domain shift and attribute shift, greatly reducing the diagnostic accuracy. In practice, it is impossible to collect braking data at all speeds in advance through experiments, and the interference of wheel-rail vibration noise is obvious. Therefore, learning the damage state characteristics at different speeds, generalizing the limited data collected by the experiment to unknown scenarios, and reducing the impact of category, domain, and attribute shifts on the model are effective methods to solve the above problems.

[0005] In recent years, scholars at home and abroad have proposed using domain generalization to solve the problem of cross-speed condition monitoring. They use multiple related known data sets to train the model, making the diagnostic knowledge universal and generalizing it to unknown new data sets. Zheng et al. combined the prior diagnostic information of the fault signal to find the common feature space of multiple source domains, and successfully used domain generalization to realize cross-domain fault diagnosis of rolling bearings (see H. Zheng, Y. Yang, J. Yin, Y. Li, R. Wang, and M. Xu, "Deep Domain Generalization Combining A Priori Diagnosis Knowledge Toward Cross-Domain Fault Diagnosis of Rolling Bearing," IEEE Trans. Instrum. Meas., vol. 70, pp. 1-11, 2021, doi: 10.1109 / TIM.2020.3016068). Zhao et al. proposed an adversarial mutual information-guided single domain generalization network for mechanical fault diagnosis, and designed a domain generation module to generate pseudo target domains with significant distribution differences from the source domain, which improved the generalization performance and finally performed well in bearing and gearbox fault diagnosis (see C. Zhao and W. Shen, “Adversarial Mutual Information-Guided Single Domain Generalization Network for Intelligent Fault Diagnosis,” IEEE Trans. Ind. Inf., vol. 19, no. 3, pp. 2909-2918, Mar. 2023, doi: 10.1109 / TII.2022.3175018). Hu et al. learned domain-invariant representations of different domains through sub-domain generalization and successfully applied domain generalization to brake pad condition monitoring and bearing fault diagnosis of EMUs (see 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, p. 104896, Aug. 2022, doi: 10.1016 / j.engappai.2022.104896).The above method finds the common feature space of multiple source domains and learns the domain-invariant features of different source domains, which has good generalization performance without noise. However, the brake pad data collected under long and steep slopes are easily disturbed by wheel-rail vibration noise, and the noise itself may be mistakenly identified as a fault feature, which brings great challenges to the cross-speed state monitoring of brake pads under long and steep slope conditions of high-speed trains.

[0006] To solve the problem of noise interference, many scholars have proposed to use methods such as signal reconstruction and filtering to eliminate noise. For example, Li et al. proposed to reconstruct the signal spectrum through spectrum fusion to effectively eliminate the random features in the noise. The results showed that it performed well in bearing fault classification under strong noise interference (see Y.Li, G.Cheng, and C.Liu, “Research on bearing fault diagnosis based on spectrum characteristics under strong noise interference,” Measurement, vol. 169, p. 108509, Feb. 2021, doi: 10.1016 / j.measurement.2020.108509). Zhao et al. proposed a deep residual shrinkage network, combining the attention mechanism with soft threshold filtering to achieve adaptive threshold filtering, and achieved good results under the influence of strong interference noise (see M. Zhao, S. Zhong, X. Fu, B. Tang, and M. Pecht, "Deep Residual Shrinkage Networks for Fault Diagnosis," IEEE Trans. Ind. Inf., vol. 16, no. 7, pp. 4681-4690, Jul. 2020, doi: 10.1109 / TII.2019.2943898). The above method has achieved good results in reducing noise interference at the signal processing level. However, in the brake pad status monitoring of high-speed trains under long and steep slope conditions, the dynamic changes of noise induced by train wheel-rail vibration are sensitive due to wheel wear, and the influence of noise cannot be eliminated through signal processing, which increases the difficulty of recognition.

[0007] In summary, in view of the large difference in vibration signals of high-speed trains at different speeds on long and steep slopes and the susceptibility to wheel-rail vibration noise, the differences in categories, attributes (noise) and fields (speed) will lead to group shift, which will cause the feature classifier to change, causing it to learn false related features, confusing important information for distinguishing fault categories, and making it difficult to effectively monitor the status of the brake pads. Summary of the invention

[0008] In view of the above technical problems existing in the prior art, the present invention provides a method for monitoring the cross-speed state of the brake pad of a high-speed train affected by vibration and noise on a long and steep slope. By constructing a group shift generalization network (GSGN), the influence of group shift on the accuracy is minimized as much as possible. Furthermore, elastic network regularization is used to reduce the complexity of the model and the influence of noise on the model, thereby realizing the cross-speed state monitoring of the brake pad under the influence of wheel-rail vibration and noise.

[0009] The present invention is implemented by adopting the following technical solutions.

[0010] The present invention provides a method for monitoring the cross-speed state of a high-speed train brake pad under the influence of vibration and noise on a long and steep slope, which comprises the following steps:

[0011] S1 builds a feature extraction network for group shift generalization; this step includes the following sub-steps:

[0012] S11 defines a number of groups based on the source domain data of the brake pad vibration signal for training, and constructs a group set;

[0013] S12 arbitrarily combines the groups in the group set to construct an uncertain set;

[0014] S13 uses the uncertain set to train the feature extraction network; and optimizes the feature extraction network parameters according to the following objective function:

[0015]

[0016] Among them, θ represents the feature extraction network parameters, Ψ represents the group set; ψ represents any group; x represents the feature extraction network input feature, y represents the prediction result; l represents the total loss based on the uncertain set; represents the empirical distribution of all source domain data of the ψth group; Show obedience The expected loss under the distribution; C represents the capacity constant; n ψ represents the number of source domain data samples of the ψth group;

[0017] S2 inputs the brake pad vibration signal to be detected into the trained feature extraction network to obtain the brake pad signal category recognition result.

[0018] In the above-mentioned method for monitoring the cross-speed status of the brake pads of a high-speed train affected by vibration and noise on a long and steep slope, the feature extraction network is ResNet, VGG or CNN, etc.

[0019] In the above-mentioned method for monitoring the cross-speed state of the brake pad of a high-speed train affected by vibration noise on a long and steep slope, in step S11, the collected brake pad vibration signal is reconstructed according to the signal category, signal attribute and speed to obtain the source domain data of the training brake pad vibration signal; the signal attribute refers to the type of noise in the signal (including Gaussian white noise or wheel vibration noise of different-order wheel polygon wear); then a number of vibration signals are randomly collected from the source domain data of the training brake pad vibration signal to form a group; the operation is repeated to obtain a number of groups; all groups constitute a group set, expressed as Ψ={P1,P2,...,P ψ ,...,P m}, P ψ represents the true probability distribution of the ψth group, and m represents the number of groups.

[0020] The source domain data of the training brake pad vibration signal is constructed as follows: the brake pad vibration signals under different speed conditions are collected, and the noises with different signal attributes are fused into the collected brake pad vibration signals to complete the reconstruction of the brake pad vibration signals; the reconstructed brake pad vibration signals are used as the source domain data of the training brake pad vibration signals.

[0021] The signal category refers to the brake pad health status category; the brake pad health status category includes normal, eccentric wear, crack, peeling, etc.

[0022] In the above method for monitoring the cross-speed state of the brake pad of a high-speed train under the influence of vibration and noise on a long and steep slope, in step S12, the uncertainty set Q is expressed as:

[0023]

[0024] Among them, q ψ represents the weight corresponding to the ψth group; △ m is a (m-1)-dimensional probability simplex; P ψ represents the true probability distribution of the ψth group.

[0025] In the above method for monitoring the cross-speed state of the brake pad of a high-speed train under the influence of vibration and noise on a long and steep slope, in step S13, the total loss l is expressed as:

[0026]

[0027] Among them, l′ represents the original loss value, which is calculated by weighting the loss values ​​corresponding to each group in the uncertain set (the worst group in the uncertain set is given the largest weight); λ1 and λ2 represent the regularization parameters that control the strength of L1 and L2 regularization, respectively, and w represents the parameter vector.

[0028] The optimization operation of the feature extraction network parameters based on the total loss is as follows:

[0029] (1) Update the weight of the ψth group in the tth iteration according to the following formula:

[0030]

[0031] Among them, “∧” represents the exponential operator; η q Indicates setting step length; represents the weight of the ψth group obtained in the t-1th iteration process; l(θ (t-1) ; (x, y) represents the total loss obtained in the tth iteration process;

[0032] (2) Update the weight distribution q in the tth iteration according to the following formula (t) :

[0033]

[0034] (3) Update the feature extraction network parameter θ in the tth iteration according to the following formula (t) :

[0035]

[0036] Among them, θ (t-1) represents the feature extraction network parameters obtained in the t-1th iteration; η θ Indicates the setting step size.

[0037] Compared with the prior art, the method for monitoring the cross-speed state of the brake pad of a high-speed train affected by vibration and noise on a long and steep slope provided by the present invention has the following technical effects:

[0038] 1) The present invention proposes a group offset generalization network; by minimizing the loss of the worst group, the model is prevented from learning pre-specified false correlations, thereby improving the network's effective monitoring of the brake pad status;

[0039] 2) The present invention takes into account the group offset caused by the influence of different rotation speeds, wheel-rail vibration noise, etc. on the characteristic distribution of the original vibration signal, and enables the model to effectively align the signal characteristic distribution of the friction block at different rotation speeds affected by wheel-rail noise;

[0040] 3) The present invention reduces the model's reliance on non-correlated elements through elastic network regularization, reduces the impact of noise on the model, and further improves the generalization accuracy of the model;

[0041] 4) The present invention has conducted extensive experiments on a brake friction dataset affected by wheel-rail vibration noise on long and steep slopes; the experimental results show that GSGN can identify the health status of the target friction block under wheel-rail vibration noise with an average accuracy of about 97%, which is better than other generalization methods in the field. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the brake disc and friction block structure;

[0043] Figure 2 A flow chart of a method for monitoring the cross-speed state of brake pads of a high-speed train under the influence of vibration and noise on a long and steep slope;

[0044] Figure 3 Schematic diagram of the process of building a feature extraction network (GSGN) for group shift generalization;

[0045] Figure 4 Construct a block diagram for the Group Shift Generalized Feature Extraction Network (GSGN).

[0046] Figure 5 This is a schematic diagram of the group offset principle;

[0047] Figure 6 Schematic diagram of the high-speed train brake pad status diagnosis results under different noise environments. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] Example 1

[0050] In this embodiment, the vibration acceleration signal of the friction block is obtained according to the following conditions:

[0051] (1) To simulate a steep slope, the braking force is 500N (equivalent to the 35kN braking force of a high-speed train); to simulate long-distance braking, the braking time is 20 minutes; to simulate periodic continuous braking, the driving speed is controlled within a certain range through braking. That is, under a braking force of 500N, the speed is kept within a certain range for 20 minutes through intermittent braking, thereby simulating the braking of a high-speed train on a long slope.

[0052] (2) According to the braking strategy, under the condition of periodic continuous braking on a long slope, the vehicle speed will fluctuate periodically in a certain range of low speed. Under low speed and strong interface friction, the braking system will produce high-frequency and high-intensity friction-induced vibrations, resulting in eccentric wear, cracks, and peeling damage. The braking system of the CRH380A EMU is taken as the research object. The embodiment uses 100rpm as the interval and 200rpm as the speed fluctuation range. A total of five periodic braking speeds including 300rpm→100rpm and 700rpm→500rpm are selected to collect vibration acceleration signals of the friction block in three damage states and normal state. The tangential acceleration of the three-dimensional vibration acceleration is selected as the research object. During the test, the ambient temperature was 25°C, and sampling was performed after 20 minutes of continuous braking. The sampling time was 10s in total, the sampling frequency was 50kHz, and there were 500,000 time points in total. After each sampling, the brake disc was cooled to the ambient temperature to eliminate the influence of temperature.

[0053] (3) Use the friction blocks and brake discs of CRH380A EMU that have been scaled down to a certain extent, with the shapes and sizes as follows: Figure 1 As shown in the figure. The brake pad is composed of multiple friction blocks. However, changes in factors such as the arrangement of friction blocks on the brake pad, the installation angle, the shape of friction blocks, and the workload can cause large-scale shedding or abnormal wear of the friction block surface material, making its stress distribution uneven. The damage forms of high-speed train friction blocks include eccentric wear, cracks, and spalling, which are the result of the combined effects of internal factors (material problems) and external factors (stress problems).

[0054] In this embodiment, the test takes normal, eccentric wear, peeling and cracked friction blocks as the research objects, the material is powder metallurgy, and 4 friction blocks with different health conditions are designed. The specific design parameters of the friction block are as follows:

[0055] (1) Normal: No damage occurs.

[0056] (2) Eccentric wear: To collect more comprehensive eccentric wear damage information, the uneven wear height was set to 1 mm, and the material and other dimensions were consistent with those of the normal friction block.

[0057] (3) Crack: Select the crack depth as 0.5mm. Starting from the center circle, the machine tool processes a crack with a width of 0.5mm and a depth of 0.5mm on the surface of the friction block. The material and other dimensions are consistent with the normal friction block.

[0058] (4) Spalling: The normal friction block works under a given load until the block spalling phenomenon occurs naturally. The material and other dimensions are the same as those of the normal friction block.

[0059] Then, the vibration acceleration signal of the friction block (referred to as vibration signal) is collected on a brake friction test bench disclosed in the art.

[0060] The vibration signals collected under different speed conditions are randomly repeated and sampled in steps of 1024, and divided into 1500 samples for each condition, for a total of 7500 samples. Each sample contains 1024 time points, and the data set is divided into AE according to different working conditions, as shown in Table 1. Each data set contains four health states: normal, eccentric wear, spalling, and crack.

[0061] Table 1 Dataset names and health status abbreviations

[0062]

[0063] Based on the brake pad vibration signal collected above, the present embodiment provides a method for monitoring the cross-speed state of the brake pad of a high-speed train affected by vibration noise on a long slope, such as Figure 2-4 As shown, it includes the following steps:

[0064] S1 builds a feature extraction network for group shift generalization;

[0065] S2 inputs the brake pad vibration signal to be detected into the trained feature extraction network to obtain the brake pad signal category recognition result.

[0066] The above steps S1-S2 are explained in detail below.

[0067] In the above step S1, the feature extraction network uses ResNet18. Figure 4 As shown in FIG. 1 , the feature extraction network includes a first feature extraction module, a second feature extraction module, a third feature extraction module, and a fourth feature extraction module, which are arranged in sequence; the four feature extraction module structures all include a convolution unit and a maximum pooling layer, and the output of the convolution unit is input into the maximum pooling layer after layer normalization (BN) and activation function ReLU. The first feature extraction module, the second feature extraction module, the third feature extraction module, and the fourth feature extraction module are specifically arranged as follows Figure 4 As shown; the convolution unit size of the first feature extraction module is 64×1, and the number of output channels is 16; the convolution unit size of the second feature extraction module is 16×1, and the number of output channels is 32; the convolution unit size of the third feature extraction module is 5×1, and the number of output channels is 64; the convolution unit size of the fourth feature extraction module is 5×1, and the number of output channels is 64; the maximum pooling layer window size of the first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module is 2×1.

[0068] In order to construct a group shift generalization network, the source domain data of the brake pad vibration signal for training is first grouped, and an uncertain set is constructed, and the objective function of the network parameters is defined; then the feature extraction network is optimized based on the uncertain set combined with the objective function.

[0069] In a specific implementation, step S1 includes the following sub-steps:

[0070] S11 defines a number of groups based on the source domain data of the brake pad vibration signal for training, and constructs a group set.

[0071] In the classic classification problem setting, there are training data from multiple categories, and the number of samples and data distribution in each category are different, which is called category shift. Figure 5 As shown in the figure, in addition to the categories, there are also offsets in attributes, such as different noise types in the braking signal (in this embodiment, it refers to Gaussian white noise or wheel vibration noise caused by different orders of wheel polygon wear). In the problem of domain generalization, there are also domain offsets, such as the different speeds of high-speed trains when braking on long slopes. The above offsets are collectively referred to as group offsets.

[0072] Deep learning models often perform poorly when there is a group offset in the training data. In the task of identifying the brake pad status on a long ramp of a high-speed train, if noise is introduced, it will induce higher-order modal vibrations of the brake system, resulting in false correlations between the brake signal and the noise signal. To this end, the present invention defines discrete groups based on a combination of categories, attributes, and fields, and tries to ensure that the worst performance in the group is good enough, or ensures that the performance in all groups is the same and good, so as to improve the robustness of the model when the group is offset.

[0073] In a specific implementation, Gaussian white noise or wheel-rail vibration noise is introduced and fused into the collected brake pad vibration signal to reconstruct the brake pad vibration signal.

[0074] The amplitude of Gaussian white noise obeys the normal distribution, and its probability density function is expressed as:

[0075]

[0076] x is a random variable, representing the amplitude of white noise, μ is the mean; σ 2 is the variance, which indicates the strength of the noise.

[0077] The expression of wheel-rail vibration noise is:

[0078]

[0079] Where x represents the circumferential position of the wheel; Represents the phase, generated by random numbers; A n represents the nth order non-circularization amplitude, L n is the roughness corresponding to the nth-order polygonal wear, which can be obtained through harmonic analysis based on the roughness level data of the wheel; r is the wheel radius; N is the wheel polygonal wear order, N = 17, 18, 19, 20; f n represents the nth order wheel self-excited vibration frequency caused by wheel polygon wear; λ n Represents the nth order wavelength of wheel polygon wear.

[0080] In this embodiment, the brake pad vibration signals under different speed conditions are first mixed, and then the Gaussian white noise or wheel-rail vibration noise of different orders is added to the collected brake pad vibration signal by weighted fusion to obtain a reconstructed brake pad vibration signal. The reconstructed brake pad vibration signal is used as the source domain data of the brake pad vibration signal for training.

[0081] Then, we use the prior knowledge of false correlation to define groups on the source domain data, that is, we randomly collect several vibration signals from the source domain data of the training brake pad vibration signal to form a group; repeat this operation to obtain several groups, and all groups constitute a group set. Then the probability distribution P of the source domain is composed of m groups P ψ The set of mixed groups is represented by Ψ={P1,P2,...,P ψ ,...,P m}, P ψ represents the true probability distribution of the ψth group.

[0082] S12 arbitrarily combines the groups in the group set to construct an uncertain set.

[0083] In this step, an uncertain set Q is determined based on the group set Ψ defined in step S11. The uncertain set Q is defined as any combination of these groups, that is: Where △m is a (m-1)-dimensional probability simplex. The Ψ selected by this method can make the learned model more robust to group shift.

[0084] S13 uses the uncertain set to train the feature extraction network; and optimizes the feature extraction network parameters according to the following objective function.

[0085] Since the optimal solution of linear programming is always at the vertex of the feasible region, the risk in the worst case of equation (5) is equal to the maximum expected loss of each group,

[0086]

[0087] Assume that the source domain distribution source group is known, that is, the source domain data contains triples of (x, y, ψ), and ψ is unknown in the target domain, so the model cannot use ψ directly. Instead, it minimizes the empirical worst group risk Parameters of training the group-shift generalized feature extraction network (GSGN):

[0088]

[0089] Among them, each group is the empirical distribution of all source domains (x, y, ψ′), and ψ′ = ψ (or equivalently a subset of training samples drawn from P). The above cross-group learning model has a good worst-group source domain training loss. This does not necessarily mean the best worst-group target domain test loss, because the worst-group generalization gap Here, elastic network regularization is used to reduce the generalization gap. This regularization combines L1 regularization and L2 regularization, absorbs their advantages, and can simultaneously obtain sparsity and parameter normalization. The general formula of elastic network regularization is expressed as:

[0090]

[0091] Where l′ represents the original loss value, which is calculated by weighting the loss values ​​corresponding to each group in the uncertain set (the worst group in the uncertain set is given the largest weight), that is, l′=α1l1+α2l2+...+α ψ l ψ +...+α m l m , where l ψ represents the loss function corresponding to the ψth group. This embodiment uses the cross-entropy loss function. ψ represents the loss weight corresponding to the ψth group; λ1 and λ2 represent the regularization parameters that control the strength of L1 and L2 regularization, respectively; ||·||1 and ||·||2 represent the L1 norm and L2 norm, respectively; and w is the model weight parameter.

[0092] However, due to the complexity of group shift, even under regularization, the generalization gap between different groups is still relatively large. To this end, the present invention narrows the generalization gap between different groups by directly minimizing the estimated upper bound of the worst group objective loss. Specifically, each group ψ has its own generalization gap:

[0093]

[0094] In order to approximately optimize the worst group test loss R(θ), the present invention proposes to use parameter-independent heuristic Among them, n ψ represents the group size of ψ, and C represents the model capacity constant, which is considered as a hyperparameter here. Therefore, the parameter estimator of the adjusted GSGN is:

[0095]

[0096] Among them, θ represents the feature extraction network parameters, Ψ represents the group set; ψ represents any group; x represents the feature extraction network input feature, y represents the prediction result; l represents the total loss based on the uncertain set; represents the empirical distribution of all source domain data of the ψth group; Show obedience The expected loss under the distribution; C represents the capacity constant; n ψ Represents the number of source domain data samples in the ψth group.

[0097] Based on the above analysis, in the above step S13, before training, q is initialized (0) and θ (0) In the t-th iteration training process, the m groups in the uncertainty set Q are input into the feature extraction network, and the total loss is obtained according to the formula (7) given above. Then, the parameters of the feature extraction network are optimized according to the total loss. The operation is as follows:

[0098] (1) Update the weight of the ψth group in the tth iteration according to the following formula:

[0099]

[0100] Among them, “∧” represents the exponential operator; η q Indicates setting step length; represents the weight of the ψth group obtained in the t-1th iteration process; l(θ (t-1) ; (x, y) represents the total loss obtained in the tth iteration process;

[0101] (2) Update the weight distribution q in the tth iteration according to the following formula (t) :

[0102]

[0103] Among them, q (t-1) represents the weight distribution obtained in the t-1th iteration process;

[0104] (3) Update the feature extraction network parameter θ in the tth iteration according to the following formula (t) :

[0105]

[0106] Among them, θ (t-1) represents the feature extraction network parameters obtained in the t-1th iteration; η θ Indicates the setting step size.

[0107] Based on the update weight The uncertain set Q is updated, and then the above step S13 is repeated to perform iterative training until the upper limit of the number of iterations is reached, thereby completing the training of the feature extraction network.

[0108] In the above step S2, the brake pad vibration signal to be detected is input into the trained feature extraction network to obtain the brake pad signal category recognition result.

[0109] In order to evaluate the effectiveness of the proposed method in the generalization problem, several related intelligent diagnosis methods were selected for comparison. The hyperparameters of all compared methods were determined from relevant studies and experiments to achieve satisfactory performance. For all methods, the number of trainings and batch size were set to 100 and 64, respectively. The PyTorch platform was used for programming, and the Nvidia3060Ti GPU was used for accelerated computing. The average accuracy and standard deviation were evaluated after 10 trials to reduce the impact of randomness. To ensure fairness, the backbone network of all methods is the same, using ResNet18. Specifically, the following methods were selected.

[0110] (1) MLDG: By synthesizing a virtual test domain in each mini-batch to simulate the domain migration from the source domain to the target domain during training, it solves the domain shift problem while reducing the impact of noise and improving the model's adaptability to new tasks. See D. Li, Y. Yang, Y.-Z. Song, and TM Hospedales, "Learning to Generalize: Meta-Learning for Domain Generalization," Oct. 10, 2017, arXiv: arXiv: 1710.03463. Accessed: Nov. 16, 2023.

[0111] (2) Mixup: Mix input features and labels by linear interpolation, so that the model can smooth the output distribution, reduce overconfident predictions, and improve the generalization ability of the model. See H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, “mixup: Beyond Empirical Risk Minimization,” April 27, 2018, arXiv: arXiv:1710.09412. Accessed: Dec. 14, 2022.

[0112] (3) DANN: Introducing domain adversarial training effectively eliminates noise features, enabling the model to learn feature representations that are insensitive to domain differences. See Y. Ganin et al., “Domain-Adversarial Training of Neural Networks,” in Domain Adaptation in Computer Vision Applications, G. Csurka, Ed., in Advances in Computer Vision and Pattern Recognition., Cham: Springer International Publishing, 2017, pp. 189-209. doi: 10.1007 / 978-3-319-58347-1_10.

[0113] (4) Deep CORAL: Its core is to introduce CORAL loss. This loss calculates the difference between the feature covariance matrices of the source domain and the target domain, and minimizes this difference to achieve feature distribution alignment, see B. Sun and K. Saenko, “Deep CORAL: Correlation Alignment for Deep Domain Adaptation,” Jul.06, 2016, arXiv:arXiv:1607.01719. Accessed: Jan.17, 2024.

[0114] The experimental results of the generalization task under different noise environments of high-speed train long ramps are shown in Table 2 and Figure 6 As shown. It can be seen that the method provided by the present invention achieves an average accuracy of 97.64%, which is the best among the compared methods. A closer look reveals that the accuracy of the method of the present invention is not greatly affected by different noises, and the fluctuation is only about 2%, while other methods are greatly affected by noise, and the fluctuation is much higher than that of the method of the present invention. It is proved that the group optimization of the method of the present invention can well eliminate the sensitivity of the network to outliers after considering the grouping structure of the data set. In particular, by observing the Mixup model, it can be found that the generalized network generated based on data performs poorly in dealing with noise problems. This may be due to the fact that the model overfits the noise in the training data and lacks robustness to new noise.

[0115] Table 2 Experimental results under different noise environments (accuracy / %)

[0116]

[0117] Note: Task BCD→A means mixing the B, C, and D datasets as source domain data and the A dataset as target domain data. Other tasks are similar.

[0118] Adding Gaussian white noise means adding Gaussian white noise to the vibration acceleration signal of each data set by weighted fusion to obtain a reconstructed brake pad vibration signal;

[0119] Adding wheel-rail noise means adding the 17th, 18th, 19th and 20th order wheel-rail vibration noise to the vibration acceleration signal collected by each data set by weighted fusion to obtain the reconstructed brake pad vibration signal;

[0120] Here the weight w signal =0.5,w noise =0.5 simulates a strong noise environment.

[0121] Those skilled in the art will appreciate that the embodiments herein are intended to help readers 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 skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes, characterized in that: The following steps are involved: S1 builds a feature extraction network for group shift generalization; This step includes the following sub-steps: S11 defines a number of groups based on the source domain data of the brake pad vibration signal for training, and constructs a group set; S12 arbitrarily combines the groups in the group set to construct an uncertain set; S13 uses the uncertain set to train the feature extraction network; and optimizes the feature extraction network parameters according to the objective function: Among them, θ represents the feature extraction network parameters, Ψ represents the group set; ψ represents any group; x represents the feature extraction network input feature, and y represents the prediction result; Represents the total loss based on the uncertain set; represents the empirical distribution of all source domain data of the ψth group; Show obedience The expected loss under the distribution; C represents the capacity constant; n ψ represents the number of source domain data samples of the ψth group; S2 inputs the brake pad vibration signal to be detected into the trained feature extraction network to obtain the brake pad signal category recognition result.

2. The method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes according to claim 1 is characterized in that: The feature extraction network is ResNet, VGG or CNN.

3. The method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes according to claim 1 is characterized in that: In step S11, the collected brake pad vibration signal is reconstructed according to the signal type, signal attribute and rotation speed to obtain the source domain data of the brake pad vibration signal for training; the signal attribute refers to the type of noise in the signal; Then, a number of vibration signals are randomly collected from the training brake pad vibration signal source domain data to form a group; this operation is repeated to obtain a number of groups; all groups constitute a group set, which is expressed as Ψ={P1,P2,...,P ψ ,...,P m }, P ψ represents the true probability distribution of the ψth group, and m represents the number of groups.

4. The method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes according to claim 3 is characterized in that: The brake pad vibration signals under different speed conditions are collected, and the noises with different signal attributes are fused into the collected brake pad vibration signals to complete the reconstruction of the brake pad vibration signals; the reconstructed brake pad vibration signals are used as the source domain data of the brake pad vibration signals for training.

5. According to the method for monitoring the cross-speed status of high-speed train brake pads affected by vibration noise on a long and steep slope as described in claim 4, the noise type includes Gaussian white noise or wheel vibration noise caused by polygonal wear of wheels of different orders.

6. The method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes according to claim 3 is characterized in that: The signal category refers to the brake pad health status category; the brake pad health status category includes normal, eccentric wear, crack, and peeling.

7. The method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes according to any one of claims 1 to 6, characterized in that: In step S12, the uncertain set Q is expressed as: Among them, q ψ represents the weight corresponding to the ψth group; △m is the (m-1)-dimensional probability simplex; P ψ represents the true probability distribution of the ψth group.

8. The method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes according to claim 7, characterized in that: In step S13, the total loss It is expressed as: in, Represents the original loss value, which is calculated by weighting the loss values ​​corresponding to each group in the uncertain set; λ1 and λ2 represent the regularization parameters that control the strength of L1 and L2 regularization, respectively, and w represents the network weight parameter vector.

9. The method for monitoring the cross-speed state of brake pads of high-speed trains under the influence of vibration and noise on long and steep slopes according to claim 8, characterized in that: In step S13, the feature extraction network parameter optimization operation is performed according to the total loss as follows: (1) Update the weight of the ψth group in the tth iteration according to the following formula: Among them, "∧" represents the exponential operator; η q Indicates setting step length; represents the weight of the ψth group obtained in the t-1th iteration process; (x,y) represents the total loss obtained in the tth iteration process; (2) Update the weight distribution q in the tth iteration according to the following formula (t) : (3) Update the feature extraction network parameter θ in the tth iteration according to the following formula (t) : Among them, θ (t-1) represents the feature extraction network parameters obtained in the t-1th iteration; η θ Indicates the setting step size.

Images