Wheel degradation sample generation method based on diffusion process

Through the wheel degradation sample generation method based on the diffusion process, the problems of insufficient wheel wear degradation samples and unexplainability of the GAN network are solved, the sample quality and generation efficiency are improved, and the operation and maintenance efficiency and safety of rail transit are enhanced.

CN118916694APending Publication Date: 2024-11-08XIAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410952991.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

In complex operating environments, train wheel wear and deterioration fails and degradation samples are insufficient, resulting in low intelligent operation and maintenance efficiency of rail transit, and the uninterpretation of the GAN network leads to the black box problem of sample generation.

Method used

Using the wheel degradation sample generation method based on the diffusion process, samples are collected through sensors, forward diffusion process and reverse denoising process are constructed, and new wheel degradation samples are generated using the root mean square error loss function and recurrent neural network, which solves the sample generation. problems of unexplainability and inadequate quality.

Benefits of technology

It improves the quality and generation efficiency of wheel degradation samples, reduces the operation and maintenance costs of intelligent high-speed rail, and enhances the basis for train operation safety decisions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916694A_ABST
    Figure CN118916694A_ABST
Patent Text Reader

Abstract

The invention discloses a wheel degradation sample generation method based on a diffusion process. Firstly, a wheel wear degradation sample is collected through a sensor or a recorder; then determining noise distribution, and constructing a degraded sample forward diffusion process; establishing a reverse denoising process in combination with a root-mean-square error loss function to obtain a wheel degradation sample after reverse denoising, and realizing generation of a wheel sample; and finally, combining the obtained de-noised wheel degradation sample with an original wheel sample, and predicting the wheel sample through GRU so as to realize a generated sample test. The method is helpful for improving the quality of the data set at the early stage of intelligent operation and maintenance of rail transit, and provides a basis for rail transit safety decision making.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of degradation sample generation, and in particular relates to a wheel degradation sample generation method based on a diffusion process. Background Art

[0002] In order to improve the efficiency of intelligent operation and maintenance of rail transit systems, it is necessary to timely predict and maintain the degradation of key components of the vehicle body. However, due to the complex operating environment, train wheels are easily affected by different stresses and wear, resulting in wheel wear degradation failure. On the other hand, the detection process of a single wheel degradation amount will also be affected by the quality of the sensor, resulting in insufficient degradation samples and other problems. Therefore, in order to accurately model and verify the degradation process of train wheels and further ensure that the train does not derail during high-speed operation, when predicting the wheel degradation amount, it is necessary to first generate samples of wear degradation samples. GAN networks are widely used in sample generation, but due to the poor interpretability of network models, black box problems are formed. The diffusion model is a sample generation method that adds noise forward and gradually denoises in reverse. It mainly uses a mathematical model to describe and sample the sample distribution, adds noise to the sample multiple times, and trains the sample recovery model to obtain a generated sample that is close to the original sample. Therefore, the present invention proposes a wheel degradation sample generation method based on a diffusion process to further improve the intelligent operation and maintenance process of rail transit. Summary of the invention

[0003] The purpose of the present invention is to provide a wheel degradation sample generation method based on a diffusion process, which is helpful to improve the quality of the early data set of rail transit intelligent operation and maintenance and provide a basis for rail transit safety decision-making.

[0004] The technical solution adopted by the present invention is a method for generating wheel degradation samples based on a diffusion process, which is specifically implemented according to the following steps:

[0005] Step 1: Collect wheel wear degradation samples through sensors or recorders;

[0006] Step 2: Determine the noise distribution and construct the forward diffusion process of the degraded samples;

[0007] Step 3: Establish a reverse denoising process in combination with the root mean square error loss function to obtain the wheel degradation sample after reverse denoising, thereby realizing wheel sample generation;

[0008] Step 4: Combine the denoised wheel degradation samples obtained in step 3 with the original wheel samples, and use GRU to predict the wheel samples, thereby generating sample tests.

[0009] The present invention is also characterized in that:

[0010] Step 2 is implemented as follows:

[0011] Step 2.1: Assume there are n wheel samples Where k is the detection time of each sample, and the nth wheel sample It is expressed as: represents the sample of the nth wheel at the kth moment;

[0012] Step 2.2: Determine that the added noise is Gaussian noise and obeys the standard normal distribution ε~N(0,1), record the number of noise diffusion as T, and the speed of noise diffusion as 1-α T When the diffusion times are equal to 0, that is, T = 0, the diffusion speed is equal to 0, α T =1, when the diffusion times tend to infinity, T→+∞, the diffusion speed tends to 1, α T →1; By for each original wheel sample Gradually add noise to get the noisy wheel sample Where t = 0, .., T, the sample forward diffusion process formula is shown in (1):

[0013]

[0014] Through the additive and multiplicative properties of the standard normal distribution, we can get the original sample before adding noise. and The relationship is shown in formula (2):

[0015]

[0016] Among them, α t '=α t ×α t-1 ×α t-2 ×...×α1;

[0017] Formula (2) indicates that the relationship model between the original wheel sample and the noisy wheel sample is established by using the diffusion value of all the noise added forward.

[0018] The forward diffusion process is represented by a random process: Using the properties of normal distribution, we know that the forward diffusion process also obeys normal distribution. Combined with the sampling method, the wheel samples after noise can be obtained

[0019] Step 3 is implemented as follows:

[0020] Step 3.1: Obtain the wheel sample after noise through step 2 After that, the original wheel sample needs to be deduced The reverse denoising process is shown in formula (3):

[0021]

[0022] in, Given the wheel sample after t-1 steps of noise and the original wheel sample, the probability value of the wheel sample after noise is obtained; Given the original wheel sample, the probability value of the wheel sample after t-1 steps of noise is obtained; Given the original wheel sample, the probability value of the wheel sample after noise is obtained;

[0023] Combining the probability density distribution with the properties of the standard normal distribution, the probability density function result of the reverse denoising process is obtained, as shown in formula (4):

[0024]

[0025] The reverse denoising process is further obtained in the probability density function of formula (4): The mean μ and variance σ of are shown in formula (5):

[0026]

[0027] Step 3.2: New samples are obtained by randomly sampling the mean and variance in the reverse denoising process. Since the diffusion speed 1-α in formula (5) t and α t 'All are known, just need to determine The wheel samples after reverse denoising can be gradually obtained by adding the noise ε The specific calculation process is shown in formula (6):

[0028]

[0029] Therefore, only the wheel samples after input noise need to be constructed through the neural network with the output training noise ε t Since the wheel samples are time series samples, in order to be closer to the actual degradation distribution characteristics, a recurrent neural network is used to construct With training noise ε t Relationship function Where rnn represents a recurrent neural network, and ω represents a network training parameter;

[0030] Using the root mean square error as the loss function is more suitable for time series samples and has a negative impact on the training noise ε t Make a prediction, and then obtain the reverse denoised samples through random sampling, that is, generate samples to achieve wheel sample generation.

[0031] The generation process of the generated samples in step 3.2 is as follows:

[0032] Step 3.2.1: Fit the noised wheel samples through the neural network RNN With original wheel sample Relationship function, get the noise result ε1;

[0033] Step 3.2.2: Obtain the trained wheel samples through the random sampling process of formula (4)

[0034] Step 3.2.3: Noise-modified wheel sample With the wheel sample after training As the input of the neural network rnn, the noise result ε2 is obtained, and steps 3.2.1 to 3.2.3 are repeated;

[0035] Step 3.2.4: When the trained wheel samples are gradually obtained The generation process ends when the noise-modified wheel samples are combined With training noise ε t , substituted into formula (5), the mean and variance of the reverse denoising process are sampled to obtain the wheel degradation samples after reverse denoising That is, the new sample Implement wheel sample generation.

[0036] Step 4 is implemented according to the following steps:

[0037] Use step 3 to get all wheel samples after training in The wheel degradation sample after removing all noise is called the generated sample In order to test the sample generation effect, the original wheel sample in the nth sample And generate samples The first j samples in the training model are used as the input of the training model, that is, The gated recurrent unit GRU is used to train the wheel sample prediction model, and the samples from the jth moment to the kth moment in the nth generated sample are used as the test input, that is, The root mean square error is used as the prediction loss function, and the calculation process is shown in formula (7):

[0038]

[0039] When the LOSS value is closer to 0, it means that the generated sample is closer to the real sample, indicating that the generation effect is better.

[0040] The beneficial effect of the present invention is that the wheel degradation sample generation method based on the diffusion process utilizes the forward diffusion process and the reverse denoising process of the diffusion model to generate new wheel samples, wherein a recurrent neural network is used to train and learn the noise parameters, and the root mean square error is used as the loss function for training the noise model and the sample prediction model, so that the newly generated samples can be closer to the real distribution, and the unexplainable shortcomings of the neural network in sample generation are solved, the sample generation efficiency is improved, and the operation and maintenance costs of the intelligent high-speed railway are further reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is an overall flow chart of a method for generating wheel degradation samples based on a diffusion process of the present invention;

[0042] Figure 2 The invention discloses a generation process of a wheel degradation sample generation method based on a diffusion process. DETAILED DESCRIPTION

[0043] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] The present invention is based on the diffusion process of the wheel degradation sample generation method, the flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0045] Step 1: Collect wheel wear degradation samples through sensors or recorders;

[0046] Step 2: Determine the noise distribution and construct the forward diffusion process of the degraded samples;

[0047] Step 2 is implemented as follows:

[0048] Step 2.1: Assume there are n wheel samples Where k is the detection time of each sample, and the nth wheel sample It is expressed as: represents the sample of the nth wheel at the kth moment;

[0049] Step 2.2: Determine that the added noise is Gaussian noise and obeys the standard normal distribution ε~N(0,1), record the number of noise diffusion as T, and the speed of noise diffusion as 1-α T When the diffusion times are equal to 0, that is, T = 0, the diffusion speed is equal to 0, α T =1, when the diffusion times tend to infinity, T→+∞, the diffusion speed tends to 1, α T →1; By for each original wheel sample Gradually add noise to get the noisy wheel sample Where t = 0, .., T, the sample forward diffusion process formula is shown in (1):

[0050]

[0051] Through the additive and multiplicative properties of the standard normal distribution, we can get the original sample before adding noise. and The relationship is shown in formula (2):

[0052]

[0053] Among them, α t '=α t ×α t-1 ×α t-2 ×...×α1;

[0054] Formula (2) indicates that the relationship model between the original wheel sample and the noisy wheel sample is established by using the diffusion value of all the noise added forward.

[0055] The forward diffusion process is represented by a random process: Using the properties of normal distribution, we know that the forward diffusion process also obeys normal distribution. Combined with the sampling method, the wheel samples after noise can be obtained

[0056] Step 3: Establish a reverse denoising process in combination with the root mean square error loss function to obtain the wheel degradation sample after reverse denoising, thereby realizing wheel sample generation;

[0057] Step 3 is implemented as follows:

[0058] Step 3.1: Obtain the wheel sample after noise through step 2 After that, the original wheel sample needs to be deduced The reverse denoising process is shown in formula (3):

[0059]

[0060] in, Given the wheel sample after t-1 steps of noise and the original wheel sample, the probability value of the wheel sample after noise is obtained; Given the original wheel sample, the probability value of the wheel sample after t-1 steps of noise is obtained; Given the original wheel sample, the probability value of the wheel sample after noise is obtained;

[0061] Combining the probability density distribution with the properties of the standard normal distribution, the probability density function result of the reverse denoising process is obtained, as shown in formula (4):

[0062]

[0063] The reverse denoising process is further obtained in the probability density function of formula (4): The mean μ and variance σ of are shown in formula (5):

[0064]

[0065] Step 3.2: New samples are obtained by randomly sampling the mean and variance in the reverse denoising process. Since the diffusion rate 1-α in formula (5) t and α t 'All are known, just need to determine The wheel samples after reverse denoising can be gradually obtained by adding the noise ε The specific calculation process is shown in formula (6):

[0066]

[0067] Therefore, only the wheel samples after input noise need to be constructed through the neural network with the output training noise ε t Since the wheel samples are time series samples, in order to be closer to the actual degradation distribution characteristics, a recurrent neural network is used to construct With training noise ε t Relationship function Where rnn represents a recurrent neural network, and ω represents a network training parameter;

[0068] In order to make the sample distribution after reverse denoising closer to the real sample distribution, the root mean square error is used as the loss function, which is more suitable for time series samples. t Make a prediction, and then obtain the reverse denoised samples through random sampling, that is, generate samples to achieve wheel sample generation.

[0069] The generation process of the generated samples in step 3.2 is as follows:

[0070] Step 3.2.1: Fit the noised wheel samples through the neural network RNN With original wheel sample Relationship function, get the noise result ε1;

[0071] Step 3.2.2: Obtain the trained wheel samples through the random sampling process of formula (4)

[0072] Step 3.2.3: Noise-modified wheel sample With the wheel sample after training As the input of the neural network rnn, the noise result ε2 is obtained, and steps 3.2.1 to 3.2.3 are repeated;

[0073] Step 3.2.4: When the trained wheel samples are gradually obtained When , the generation process ends. The above generation process is as follows Figure 2 shown.

[0074] By combining the noisy wheel samples With training noise ε t , substituted into formula (5), the mean and variance of the reverse denoising process are sampled to obtain the wheel degradation samples after reverse denoising That is, the new sample Implement wheel sample generation.

[0075] Step 4: Combine the denoised wheel degradation samples obtained in step 3 with the original wheel samples, and use GRU to predict the wheel samples, thereby generating sample tests.

[0076] Step 4 is implemented according to the following steps:

[0077] Use step 3 to get all wheel samples after training in The wheel degradation sample after removing all noise is called the generated sample In order to test the sample generation effect, the original wheel sample in the nth sample And generate samples The first j samples in the training model are used as the input of the training model, that is, The gated recurrent unit GRU is used to train the wheel sample prediction model, and the samples from the jth moment to the kth moment in the nth generated sample are used as the test input, that is, The root mean square error is used as the prediction loss function, and the calculation process is shown in formula (7):

[0078]

[0079] When the LOSS value is closer to 0, it means that the generated sample is closer to the real sample, indicating that the generation effect is better.

[0080] In the face of the problem of modeling the train wheel degradation process, it is necessary to ensure the quality of the wheel degradation samples. However, due to the cumbersome collection process and insufficient sensor monitoring, the wheel degradation samples are insufficient. Therefore, in order to improve the accuracy of the wheel degradation process modeling, new wheel samples are generated to improve the sample quality. The main process of generating samples is: first, collect wheel wear degradation samples through sensors or recorders; second, determine the noise distribution and construct the forward diffusion process of degradation samples; then, establish a reverse denoising process in combination with the root mean square error loss function to obtain the wheel degradation samples after reverse denoising; finally, combine the denoised wheel degradation samples obtained in step 3 with the original wheel samples, and predict the wheel samples through GRU, so as to achieve the generation of sample tests. The present invention helps to improve the quantity and overall quality of wheel samples, use mathematical models to solve the shortcomings of the unexplainable nature of neural networks, ensure the efficiency of intelligent operation and maintenance of trains, and thus improve the safety level of train operation.

[0081] Example 1

[0082] The present invention is based on the diffusion process of the wheel degradation sample generation method, the flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0083] Step 1: Collect wheel wear degradation samples through sensors or recorders;

[0084] Step 2: Determine the noise distribution and construct the forward diffusion process of the degraded samples;

[0085] Step 3: Establish a reverse denoising process in combination with the root mean square error loss function to obtain the wheel degradation sample after reverse denoising, thereby realizing wheel sample generation;

[0086] Step 4: Combine the denoised wheel degradation samples obtained in step 3 with the original wheel samples, and use GRU to predict the wheel samples, thereby generating sample tests.

[0087] Example 2

[0088] The present invention is based on the diffusion process of the wheel degradation sample generation method, the flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0089] Step 1: Collect wheel wear degradation samples through sensors or recorders;

[0090] Step 2: Determine the noise distribution and construct the forward diffusion process of the degraded samples;

[0091] Step 2 is implemented as follows:

[0092] Step 2.1: Assume there are n wheel samples Where k is the detection time of each sample, and the nth wheel sample It is expressed as: represents the sample of the nth wheel at the kth moment;

[0093] Step 2.2: Determine that the added noise is Gaussian noise and obeys the standard normal distribution ε~N(0,1), record the number of noise diffusion as T, and the speed of noise diffusion as 1-α T When the diffusion times are equal to 0, that is, T = 0, the diffusion speed is equal to 0, α T =1, when the diffusion times tend to infinity, T→+∞, the diffusion speed tends to 1, α T →1; By for each original wheel sample Gradually add noise to get the noisy wheel sample Where t = 0, .., T, the sample forward diffusion process formula is shown in (1):

[0094]

[0095] Through the additive and multiplicative properties of the standard normal distribution, we can get the original sample before adding noise. and The relationship is shown in formula (2):

[0096]

[0097] Among them, α t '=α t ×α t-1 ×α t-2 ×...×α1;

[0098] Formula (2) indicates that the relationship model between the original wheel sample and the noisy wheel sample is established by using the diffusion value of all the noise added forward.

[0099] The forward diffusion process is represented by a random process: Using the properties of normal distribution, we know that the forward diffusion process also obeys normal distribution. Combined with the sampling method, the wheel samples after noise can be obtained

[0100] Step 3: Establish a reverse denoising process in combination with the root mean square error loss function to obtain the wheel degradation sample after reverse denoising, thereby realizing wheel sample generation;

[0101] Step 3 is implemented as follows:

[0102] Step 3.1: Obtain the wheel sample after noise through step 2 After that, the original wheel sample needs to be deduced The reverse denoising process is shown in formula (3):

[0103]

[0104] in, Given the wheel sample after t-1 steps of noise and the original wheel sample, the probability value of the wheel sample after noise is obtained; Given the original wheel sample, the probability value of the wheel sample after t-1 steps of noise is obtained; Given the original wheel sample, the probability value of the wheel sample after noise is obtained;

[0105] Combining the probability density distribution with the properties of the standard normal distribution, the probability density function result of the reverse denoising process is obtained, as shown in formula (4):

[0106]

[0107] The reverse denoising process is further obtained in the probability density function of formula (4): The mean μ and variance σ of are shown in formula (5):

[0108]

[0109] Step 3.2: New samples are obtained by randomly sampling the mean and variance in the reverse denoising process. Since the diffusion rate 1-α in formula (5) t and α t 'All are known, just need to determine The wheel samples after reverse denoising can be gradually obtained by adding the noise ε The specific calculation process is shown in formula (6):

[0110]

[0111] Therefore, only the wheel samples after input noise need to be constructed through the neural network with the output training noise ε t Since the wheel samples are time series samples, in order to be closer to the actual degradation distribution characteristics, a recurrent neural network is used to construct With training noise ε t Relationship function Where rnn represents a recurrent neural network, and ω represents a network training parameter;

[0112] In order to make the sample distribution after reverse denoising closer to the real sample distribution, the root mean square error is used as the loss function, which is more suitable for time series samples. tMake a prediction, and then obtain the reverse denoised samples through random sampling, that is, generate samples to achieve wheel sample generation.

[0113] The generation process of the generated samples in step 3.2 is as follows:

[0114] Step 3.2.1: Fit the noised wheel samples through the neural network RNN With original wheel sample Relationship function, get the noise result ε1;

[0115] Step 3.2.2: Obtain the trained wheel samples through the random sampling process of formula (4)

[0116] Step 3.2.3: Noise-modified wheel sample With the wheel sample after training As the input of the neural network rnn, the noise result ε2 is obtained, and steps 3.2.1 to 3.2.3 are repeated;

[0117] Step 3.2.4: When the trained wheel samples are gradually obtained When , the generation process ends. The above generation process is as follows Figure 2 shown.

[0118] By combining the noisy wheel samples With training noise ε t , substituted into formula (5), the mean and variance of the reverse denoising process are sampled to obtain the wheel degradation samples after reverse denoising That is, the new sample Implement wheel sample generation.

[0119] Step 4: Combine the denoised wheel degradation samples obtained in step 3 with the original wheel samples, and use GRU to predict the wheel samples, thereby generating sample tests.

[0120] Example 3

[0121] The present invention is based on the diffusion process of the wheel degradation sample generation method, the flow chart is as follows Figure 1 As shown, the specific implementation steps are as follows:

[0122] Step 1: Collect wheel wear degradation samples through sensors or recorders;

[0123] Step 2: Determine the noise distribution and construct the forward diffusion process of the degraded samples;

[0124] Step 2 is implemented as follows:

[0125] Step 2.1: Assume there are n wheel samples Where k is the detection time of each sample, and the nth wheel sample It is expressed as: represents the sample of the nth wheel at the kth moment;

[0126] Step 2.2: Determine that the added noise is Gaussian noise and obeys the standard normal distribution ε~N(0,1), record the number of noise diffusion as T, and the speed of noise diffusion as 1-α T When the diffusion times are equal to 0, that is, T = 0, the diffusion speed is equal to 0, α T =1, when the diffusion times tend to infinity, T→+∞, the diffusion speed tends to 1, α T →1; By for each original wheel sample Gradually add noise to get the noisy wheel sample Where t = 0, .., T, the sample forward diffusion process formula is shown in (1):

[0127]

[0128] Through the additive and multiplicative properties of the standard normal distribution, we can get the original sample before adding noise. and The relationship is shown in formula (2):

[0129]

[0130] Among them, α t '=α t ×α t-1 ×α t-2 ×...×α1;

[0131] Formula (2) indicates that the relationship model between the original wheel sample and the noisy wheel sample is established by using the diffusion value of all the noise added forward.

[0132] The forward diffusion process is represented by a random process: Using the properties of normal distribution, we know that the forward diffusion process also obeys normal distribution. Combined with the sampling method, the wheel samples after noise can be obtained

[0133] Step 3: Establish a reverse denoising process in combination with the root mean square error loss function to obtain the wheel degradation sample after reverse denoising, thereby realizing wheel sample generation;

[0134] Step 3 is implemented as follows:

[0135] Step 3.1: Obtain the wheel sample after noise through step 2 After that, the original wheel sample needs to be deduced The reverse denoising process is shown in formula (3):

[0136]

[0137] in, Given the wheel sample after t-1 steps of noise and the original wheel sample, the probability value of the wheel sample after noise is obtained; Given the original wheel sample, the probability value of the wheel sample after t-1 steps of noise is obtained; Given the original wheel sample, the probability value of the wheel sample after noise is obtained;

[0138] Combining the probability density distribution with the properties of the standard normal distribution, the probability density function result of the reverse denoising process is obtained, as shown in formula (4):

[0139]

[0140] The reverse denoising process is further obtained in the probability density function of formula (4): The mean μ and variance σ of are shown in formula (5):

[0141]

[0142] Step 3.2: New samples are obtained by randomly sampling the mean and variance in the reverse denoising process. Since the diffusion rate 1-α in formula (5) t and α t 'All are known, just need to determine The wheel samples after reverse denoising can be gradually obtained by adding the noise ε The specific calculation process is shown in formula (6):

[0143]

[0144] Therefore, only the wheel samples after input noise need to be constructed through the neural network with the output training noise ε t Since the wheel samples are time series samples, in order to be closer to the actual degradation distribution characteristics, a recurrent neural network is used to construct With training noise ε t Relationship function Where rnn represents a recurrent neural network, and ω represents a network training parameter;

[0145] In order to make the sample distribution after reverse denoising closer to the real sample distribution, the root mean square error is used as the loss function, which is more suitable for time series samples. tMake a prediction, and then obtain the reverse denoised samples through random sampling, that is, generate samples to achieve wheel sample generation.

[0146] Step 4: Combine the denoised wheel degradation samples obtained in step 3 with the original wheel samples, and use GRU to predict the wheel samples, thereby generating sample tests.

[0147] Step 4 is implemented according to the following steps:

[0148] Use step 3 to get all wheel samples after training in The wheel degradation sample after removing all noise is called the generated sample In order to test the sample generation effect, the original wheel sample in the nth sample And generate samples The first j samples in the training model are used as the input of the training model, that is, The gated recurrent unit GRU is used to train the wheel sample prediction model, and the samples from the jth moment to the kth moment in the nth generated sample are used as the test input, that is, The root mean square error is used as the prediction loss function, and the calculation process is shown in formula (7):

[0149]

[0150] When the LOSS value is closer to 0, it means that the generated sample is closer to the real sample, indicating that the generation effect is better.

Claims

1. A wheel degradation sample generation method based on a diffusion process, characterized in that: Follow the steps below to implement it: Step 1: Collect wheel wear degradation samples through sensors or recorders; Step 2: Determine the noise distribution and construct the forward diffusion process of the degraded samples; Step 3: Establish a reverse denoising process in combination with the root mean square error loss function to obtain the wheel degradation sample after reverse denoising, thereby realizing wheel sample generation; Step 4: Combine the denoised wheel degradation samples obtained in step 3 with the original wheel samples, and use GRU to predict the wheel samples, thereby generating sample tests.

2. The method for generating wheel degradation samples based on diffusion process according to claim 1, characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1: Assume there are n wheel samples Where k is the detection time of each sample, and the nth wheel sample It is expressed as: represents the sample of the nth wheel at the kth moment; Step 2.2: Determine that the added noise is Gaussian noise and obeys the standard normal distribution ε~N(0,1), record the number of noise diffusion as T, and the speed of noise diffusion as 1-α T When the diffusion times are equal to 0, that is, T = 0, the diffusion speed is equal to 0, α T =1, when the diffusion times tend to infinity, T→+∞, the diffusion speed tends to 1, α T →1; By for each original wheel sample Gradually add noise to get the noisy wheel sample Where t = 0, .., T, the sample forward diffusion process formula is shown in (1): Through the additive and multiplicative properties of the standard normal distribution, we can get the original sample before adding noise. and The relationship is shown in formula (2): Among them, a t '=a t ×a t-1 ×a t-2 ×...×α1; Formula (2) indicates that the relationship model between the original wheel sample and the noisy wheel sample is established by using the diffusion value of all the noise added forward. The forward diffusion process is represented by a random process: Using the properties of normal distribution, we know that the forward diffusion process also obeys normal distribution. Combined with the sampling method, the wheel samples after noise can be obtained 3. The method for generating wheel degradation samples based on diffusion process according to claim 2, characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1: Obtain the wheel sample after noise through step 2 After that, the original wheel sample needs to be deduced The reverse denoising process is shown in formula (3): in, Given the wheel sample after t-1 steps of noise and the original wheel sample, the probability value of the wheel sample after noise is obtained; Given the original wheel sample, the probability value of the wheel sample after t-1 steps of noise is obtained; Given the original wheel sample, the probability value of the wheel sample after noise is obtained; Combining the probability density distribution with the properties of the standard normal distribution, the probability density function result of the reverse denoising process is obtained, as shown in formula (4): The reverse denoising process is further obtained in the probability density function of formula (4): The mean μ and variance σ of are shown in formula (5): Step 3.2: New samples are obtained by randomly sampling the mean and variance in the reverse denoising process. Since the diffusion speed 1-α in formula (5) t and α t 'All are known, just need to determine The wheel samples after reverse denoising can be gradually obtained by adding the noise ε The specific calculation process is shown in formula (6): Therefore, only the wheel samples after input noise need to be constructed through the neural network with the output training noise ε t Since the wheel samples are time series samples, in order to be closer to the actual degradation distribution characteristics, a recurrent neural network is used to construct With training noise ε t Relationship function Where rnn represents a recurrent neural network, and ω represents a network training parameter; Using the root mean square error as the loss function is more suitable for time series samples and has a negative impact on the training noise ε t Make a prediction, and then obtain the reverse denoised samples through random sampling, that is, generate samples to achieve wheel sample generation.

4. The method for generating wheel degradation samples based on diffusion process according to claim 3, characterized in that: The generation process of the sample generated in step 3.2 is as follows: Step 3.2.1: Fit the noised wheel samples through the neural network RNN With original wheel sample Relationship function, get the noise result ε1; Step 3.2.2: Obtain the trained wheel samples through the random sampling process of formula (4) Step 3.2.3: Noise-modified wheel sample With the wheel sample after training As the input of the neural network rnn, the noise result ε2 is obtained, and steps 3.2.1 to 3.2.3 are repeated; Step 3.2.4: When the trained wheel samples are gradually obtained The generation process ends when the noise-modified wheel samples are combined With training noise ε t , substituted into formula (5), the mean and variance of the reverse denoising process are sampled to obtain the wheel degradation samples after reverse denoising That is, the new sample Implement wheel sample generation.

5. The method for generating wheel degradation samples based on diffusion process according to claim 4, characterized in that: The step 4 is specifically implemented according to the following steps: Use step 3 to get all wheel samples after training in The wheel degradation sample after removing all noise is called the generated sample In order to test the sample generation effect, the original wheel sample in the nth sample And generate samples The first j samples in the training model are used as the input of the training model, that is, The gated recurrent unit GRU is used to train the wheel sample prediction model, and the samples from the jth moment to the kth moment in the nth generated sample are used as the test input, that is, The root mean square error is used as the prediction loss function, and the calculation process is shown in formula (7): When the LOSS value is closer to 0, it means that the generated sample is closer to the real sample, indicating that the generation effect is better.