Rolling bearing life prediction method based on classification regression diffusion model

Abstract

The embodiment of the invention belongs to the field of fault diagnosis and maintenance of mechanical equipment, and particularly relates to a rolling bearing service life prediction method based on a classification regression diffusion model, in the prediction method, initial degradation point division is not carried out, and remaining service life labels are directly created; optimal hyper-parameter configuration is obtained by using an early stop method and cross validation, and a Transform model is pre-trained on the basis of the optimal hyper-parameter; and splicing the extracted vibration signal features and the prediction result of the pre-training model to obtain condition input, embedding a sampling timestamp coded by an embedded layer into the input features through a Hadarcard product, modeling conditional error distribution during prediction of the pre-training model by the classification regression model on the basis of the original pre-training model, training the classification regression diffusion model, and obtaining a classification result of the pre-training model. And a mean square error loss function and an optimization algorithm are utilized to optimize and train weight parameters of the prediction model, and the residual life of the bearing is predicted according to the prediction model subjected to optimization training. The residual service life of the bearing can be predicted without dividing the initial degradation point.

Description

Technical Field

[0001] The embodiments of the present application belong to the technical field of mechanical equipment fault diagnosis and maintenance, and in particular to a rolling bearing life prediction method based on a classification regression diffusion model. Background Art

[0002] Rolling bearings, as the core supporting components of rotating machinery, achieve efficient and stable operation of rotating bodies by reducing the friction coefficient. Their health status is directly related to production safety and equipment reliability. They are commonly used in two application scenarios: one is the precision manufacturing field with strict precision requirements, such as optical component processing, micro-electromechanical systems and nano-level processing equipment; the other is high-load power transmission systems, including turbines, wind turbines, automobile engines and high-speed rail traction motors.

[0003] The wide range of applications makes the reliability of rolling bearings extremely important. In practical applications, rolling bearings face numerous potential failure factors, such as installation deviations, manufacturing process flaws, or substandard bearing material strength, all of which pose potential risks. Therefore, research on predicting the remaining useful life of bearings has become a hot topic in the fields of mechanical equipment fault diagnosis and predictive maintenance.

[0004] In related technologies, bearing performance degradation is a complex nonlinear process throughout its life cycle. Affected by the coupling of multiple physical fields, traditional linear models have difficulty constructing a mapping relationship between its degradation characteristics and remaining life. The traditional degradation starting point-remaining life mapping model relies on the physical degradation characteristics of the signal. However, due to the diversity of fault types and different characteristic patterns, the model is easily affected by the fault type. The prediction accuracy decreases in industrial scenarios with consistent working conditions but unknown fault types. At the same time, to solve the problem of uneven data distribution, traditional methods often discard data in the healthy and slowly degraded stages. Although this alleviates the uneven distribution, it loses the initial key information of the bearing fault and faces challenges such as non-stationary characteristics and complex feature relationships. Summary of the Invention

[0005] The purpose of the embodiments of the present application is to provide a rolling bearing life prediction method based on a classification regression diffusion model, so as to at least solve the technical problem that when the traditional prediction model does not divide the initial degradation point, the prediction results of the remaining service life of the rolling bearing are nonlinear and have low generalization ability, resulting in the stability and prediction accuracy of the prediction model being unable to meet the real-time and rapid response requirements of the engineering.

[0006] To achieve the above objectives, the present invention provides the following technical solutions.

[0007] An embodiment of the present invention provides a rolling bearing life prediction method based on a classification regression diffusion model, the prediction method comprising the following steps:

[0008] Use acceleration sensors to pick up rolling bearing accelerated life test data at each measuring point;

[0009] Extract the time-frequency domain features of the sampling points, normalize the extracted feature data, and use the processing results as the input of the Transformer prediction model; do not perform initial degradation point division, directly create the remaining service life label, and divide it into training data set and test data set;

[0010] Build a Transformer model and a classification regression diffusion model, use early stopping and cross-validation to obtain the optimal hyperparameter configuration, and pre-train the Transformer model based on the optimal hyperparameters;

[0011] The extracted vibration signal features and the prediction results of the pre-trained model are concatenated to obtain the conditional input. The sampling timestamps encoded by the embedding layer are embedded into the input features through the Hadarmard product. The classification regression model is used to model the conditional error distribution of the pre-trained model prediction based on the original pre-trained model. The classification regression diffusion model is trained, and the Transformer model is frozen during the training of the classification regression diffusion model.

[0012] The mean square error loss function and optimization algorithm are used to optimize the weight parameters of the training prediction model, and the remaining life of the bearing is predicted based on the optimized trained prediction model.

[0013] Furthermore, in the step of extracting the time-frequency domain features of the sampling points, the time-frequency domain features include the current signal power density spectral entropy, spectral mean, kurtosis, root mean square, skewness, peak, permutation entropy and Pearson similarity of the sampling signal spectrum at time 0.

[0014] Furthermore, in the constructed Transformer model, the Transformer model uses a 2-layer encoder network to capture sequence dependencies;

[0015] In the constructed classification regression diffusion model, the denoising network of the classification regression diffusion model is composed of 4 embedding layers plus a linear layer stacked together. The first embedding layer is used to encode the sampling time step and embed the splicing conditional vector through the Hadarmard product; the subsequent embedding layer is used to encode the time step where the noise is located and embed it into the middle hidden layer of the regression diffusion model through the Hadarmard product.

[0016] Furthermore, in the step of using early stopping and cross-validation to obtain the optimal hyperparameter configuration, the Transformer model encoder uses early stopping combined with cross-validation to optimize the hyperparameters during training;

[0017] The patience value of the early stopping method is set to 20 iterations, the learning rate search range of the loss function is set between 1e-3 and 1e-4, the batch size is 32, and the iterations are set to 300. The Adma optimizer and the mean squared error loss function are used.

[0018] Furthermore, in the step of embedding the sampling timestamp encoded by the embedding layer into the input feature by the Hadarmard product, the sampling timestamp T is raised to a 9-dimensional vector by the embedding layer. , obtained by Hadarmard product , so that the timestamp information is embedded in the conditional input of the classification regression diffusion model, and the randomness of the results is suppressed to a certain extent through continuous input.

[0019] Furthermore, in the step of modeling the conditional error distribution of the classification regression model when predicting the pre-trained model based on the original pre-trained model, the conditional embedding mechanism of the classification regression diffusion model in the process of forward diffusion and noise addition of data is utilized to make the noise conditional distribution include the prior information obtained by the pre-trained model and the result of the previous denoising step.

[0020] Furthermore, the noise intensity during the diffusion process is expressed as:

[0021]

[0022] in, Indicates time The intensity of the noise added when ; for The value range is between 0 and 1, and it is a length of T represents the number of noise-added time steps in the forward diffusion model.

[0023] The distribution of the noised data at any time step including t=1 is expressed as:

[0024]

[0025]

[0026]

[0027] In the reverse process corresponding to the forward process, the distribution to be predicted is expressed as follows:

[0028]

[0029]

[0030]

[0031] Where, is the initial variable, is a normal distribution, is the identity matrix, is the noise intensity; Represents the prediction result of the pre-trained model, represents the denoising result at time step t, for Denoising results for the time step.

[0032] Furthermore, in the forward diffusion noise adding process, a noise list is generated by cosine noise adding, and noise is randomly selected from the list. , the random number seed is set to 22.

[0033] Furthermore, in the step of optimizing the weight parameters of the training prediction model using the mean square error loss function and the optimization algorithm, the training learning rate is 1e-4, the batch size is 64, the diffusion time step is 1000 steps, and gradient clipping is performed when the gradient is greater than 1.

[0034] Compared with the prior art, the rolling bearing life prediction method based on the classification regression diffusion model of the present invention has the following beneficial effects:

[0035] First, the present invention considers the degradation process of the bearing throughout its entire life cycle, and conducts joint modeling of time-frequency domain characteristics as they change with material failure and operating conditions. It uses a classification regression diffusion model combined with a pre-trained Transformer model for prediction, enabling prediction of the remaining useful life of the bearing without the need to identify an initial degradation point.

[0036] Second, compared with the classical physics-based prediction model and the data-driven prediction model, the present invention combines the advantages of the Transformer network and the classification regression diffusion model. With low computational complexity, it can effectively model the distribution relationship between features and remaining life. The training process only considers noise condition modeling. The network structure is simple and is not affected by the gradient vanishing and explosion problems. The algorithm runs fast. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention.

[0038] Figure 1 This is a system architecture diagram of a bearing remaining service life prediction method based on a classification regression diffusion model according to the present invention;

[0039] Figure 2 Schematic diagram of the implementation process of the method for predicting the remaining useful life of a bearing based on the classification regression diffusion model of the present invention;

[0040] Figure 3 This is a schematic diagram of the vibration acceleration signal of a bearing throughout its life cycle according to an embodiment of the present invention;

[0041] Figure 4 Schematic diagram of the overall structure of the Transformer model and classification regression diffusion model according to an embodiment of the present invention;

[0042] Figure 5 Schematic diagram of similarity characteristics between the first sampling point spectrum and the current sampling point spectrum of the bearing signal according to an embodiment of the present invention;

[0043] Figure 6 This is a characteristic diagram of the power density spectrum entropy of a bearing signal according to an embodiment of the present invention;

[0044] Figure 7 This is a mean characteristic diagram of the bearing signal spectrum according to an embodiment of the present invention;

[0045] Figure 8 This is a peak characteristic diagram of a bearing signal according to an embodiment of the present invention;

[0046] Figure 9 This is a root mean square characteristic diagram of a bearing signal according to an embodiment of the present invention;

[0047] Figure 10 Arrange entropy characteristic diagram of bearing signal according to an embodiment of the present invention;

[0048] Figure 11 This is a characteristic diagram of bearing signal skewness according to an embodiment of the present invention;

[0049] Figure 12 This is a characteristic diagram of the kurtosis of a bearing signal according to an embodiment of the present invention;

[0050] Figure 13 This is a schematic diagram of the bearing remaining service life prediction process of the prediction model proposed in the present invention;

[0051] Figure 14 This is a schematic diagram of the prediction results of the remaining service life of a bearing based on the prediction model proposed in this invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0054] In one embodiment of the present invention, a rolling bearing life prediction method based on a classification regression diffusion model is provided. The present invention does not require the division of initial degradation points and can directly predict the remaining life of the bearing, providing theoretical support for fault diagnosis and predictive maintenance of mechanical equipment systems.

[0055] Specifically, such as Figure 1 and Figure 2 As shown, the prediction method provided in the embodiment of the present application includes the following steps:

[0056] S1. Using an acceleration sensor to pick up the rolling bearing accelerated life test data at each measuring point;

[0057] In step S1 of the present invention, an acceleration sensor is used to pick up the rolling bearing accelerated life test data of each vibration measurement point of the rotating mechanical device; specifically, the present invention uses an existing public data set of bearing accelerated life test (Wang B, Lei Y, Li N, et al.: A hybrid prognostics approach for estimating remaining useful life of rolling element bearings, IEEE Transactions on Reliability, IEEE, 2020, 69 (1): 401–412), the present invention takes the rotating bearing 1-1 as the research object, the sampling frequency of the sensor in the experiment is 25.6 kHz, the bearing experimental working conditions are a rotation speed of 2100 rpm, a radial force of 12000 N, and a sample is collected every 1 minute. The sample length in one collection is 32768 sample points, that is, 1.28s of data;

[0058] Figure 3 Figure 1 is the vibration acceleration signal of bearing 1-1 over its entire life cycle. It can be seen that the vibration signal waveform of bearing 1-1 exhibits irregular fluctuations from its initial operation to complete destruction. The amplitude fluctuates more dramatically at the beginning and end of the period, with distinct peaks and valleys, while the fluctuation is relatively gentle in the middle period.

[0059] Furthermore, the rolling bearing life prediction method provided in the embodiment of the present application further includes the steps of:

[0060] S2. Extract the time-frequency domain features of the sampling points, normalize the extracted feature data, and use the processing results as the input of the prediction model (the constructed Transformer model). Without performing the initial degradation point division, the remaining service life labels are directly created and divided into training and test data sets.

[0061] In one implementation, Figure 5-Figure 12As shown, in step S2 of the present invention, the time-frequency domain features include the current signal power density spectral entropy, spectrum mean, kurtosis, root mean square, skewness, peak value, permutation entropy and Pearson similarity of the spectrum of the sampling signal at time 0;

[0062] In the examples of this application, by observing Figure 5-Figure 12 Therefore, the present invention uses the Min-Max normalization method (i.e., deviation standardization) to normalize the accelerated life test data used and combine them into The eight-dimensional vector of the lifetime label is not divided into initial degradation points. Instead, the total lifetime minus the current sampling time is directly used to create the remaining lifetime label, which is then normalized between 0 and 1 to reduce the impact of the time scale.

[0063] At the same time, the present invention divides the training data set and the test data set. The training set uses bearings 3-1 and bearings 3-2, and the test set uses bearing 1-1.

[0064] For further information, please refer to Figure 1 、 Figure 2 and Figure 4 , the prediction method provided by the embodiment of the present invention further includes:

[0065] S3. Build a Transformer model and a classification regression diffusion model, use early stopping and cross-validation to obtain the optimal hyperparameter configuration, and pre-train the Transformer model based on the optimal hyperparameters.

[0066] S4. Concatenate the extracted vibration signal features (the results of the normalization process in step S2) and the prediction results of the pre-trained model to obtain conditional input, embed the sampling timestamps encoded by the embedding layer into the input features through Hadarmard product, model the conditional error distribution of the pre-trained model prediction based on the original pre-trained model, train the classification regression diffusion model, and freeze the Transformer model during the training of the classification regression diffusion model;

[0067] Wherein, in the step of embedding input features, the embedded input features are the conditional inputs obtained above;

[0068] S5. Optimize the weight parameters of the training prediction model using the mean square error loss function and the optimization algorithm, and predict the remaining life of the bearing based on the optimized trained prediction model.

[0069] In step S3 of the present invention, the optimal hyperparameter configuration includes learning rate, number of iterations, etc.

[0070] The embodiment of the present application pre-trains the Transformer based on the optimal parameters to obtain robust weight parameters;

[0071] Please refer to Figure 4 In the embodiment of the present invention, in the constructed Transformer model structure, the Transformer model uses a two-layer encoder network to capture sequence dependencies. During the training of the Transformer model encoder, the early stopping method is combined with cross-validation to optimize the hyperparameters. The patience value of the early stopping method is set to 20 iterations. The loss function uses a learning rate search range set between 1e-3 and 1e-4. The batch size is 32, the iterations are set to 300, and the Adma optimizer and mean square error loss function are used. Cross-validation uses 90% of the training set data. The optimal learning rate obtained by optimization is 2.43e-4, and the early stopping rounds during training are limited to 129.

[0072] Furthermore, in the constructed classification regression diffusion model, the denoising network of the classification regression diffusion model is composed of 4 embedding layers plus a linear layer stacked together. The first embedding layer is used to encode the sampling time step and embed the splicing conditional vector through the Hadarmard product; the subsequent embedding layer is used to encode the time step where the noise is located and embed it into the middle hidden layer of the regression diffusion model through the Hadarmard product; among them, the size of the middle hidden layer of the classification regression diffusion model is set to 128.

[0073] The experiments of the present invention were carried out in the Windows 11 operating system. Specifically, Python 3.9 and CUDA 11.3 were used as the experimental environment. The computer configuration was as follows: the CPU model was AMD Ryzen 5 5600, and the video memory was NVIDIA GeForce RTX 3060 GPU with 12GB.

[0074] In step S4, the extracted vibration signal features and the prediction results of the pre-trained model are combined to obtain the conditional input The first embedding layer embeds the sampling timestamp T encoded by the embedding layer into the input feature through Hadarmard product. The classification regression model models the conditional error distribution of the pre-trained model prediction based on the original pre-trained model, thereby improving the accuracy of conditional distribution modeling during model prediction.

[0075] Therefore, the present invention uses the embedding layer to upgrade the sampling timestamp T to a 9-dimensional vector , obtained by Hadarmard product , X represents the signal feature, which embeds the timestamp information into the conditional input of the classification regression diffusion model, and suppresses the randomness of the result to a certain extent through the continuous input.

[0076] Furthermore, in the step of modeling the conditional error distribution of the classification regression model when predicting the pre-trained model based on the original pre-trained model, the embodiment of the present application utilizes the conditional embedding mechanism of the classification regression diffusion model in the process of forward diffusion and noise addition of data, so that the noise conditional distribution includes the prior information obtained by the pre-trained model and the result of the previous denoising step;

[0077] Among them, the noise intensity in the diffusion process is expressed as:

[0078]

[0079] in, Indicates time The intensity of the noise added when ; for The value range is between 0 and 1, and it is a length of T represents the number of noise-added time steps in the forward diffusion model.

[0080] The distribution of the noised data at any time step including t=1 is expressed as:

[0081]

[0082]

[0083]

[0084] Where, is the initial variable, is a normal distribution, is the identity matrix, is the noise intensity; Represents the prediction result of the pre-trained model, express Time step denoising results; represents the original data that is not corrupted by noise at time step t; Represents the original data from the initial time step to time step t;

[0085] Furthermore, in the embodiment of the present application, in the reverse process corresponding to the forward process, the distribution form to be predicted is expressed as follows:

[0086]

[0087]

[0088]

[0089] In the formula, in the formula, is the initial variable, is a normal distribution, is the identity matrix, is the noise intensity; Represents the prediction result of the pre-trained model, represents the denoising result at time step t, for Denoising results for the time step; Represents the mean of the denoising results from time t to t-1; Indicates the residual degree of noise when denoising from time step t to t-1.

[0090] Furthermore, in the forward diffusion noise adding process, a noise list is generated by cosine noise adding, and noise is randomly selected from the list. , the random number seed is set to 22.

[0091] In addition, the Transformer pre-training model is frozen during the classification regression diffusion model training in this embodiment to prevent overfitting.

[0092] Furthermore, in step S5 of the embodiment of the present invention, the root mean square error loss function and the Adma optimization algorithm are used to optimize the weight parameters of the training prediction model to complete the prediction of the remaining service life of the bearing; wherein, the training learning rate is 1e-4, the batch size is 64, the diffusion time step is 1000 steps, and gradient clipping is performed when the gradient is greater than 1.

[0093] Furthermore, the remaining service life of the bearing is predicted based on the trained prediction model. The prediction diffusion regression process of the present invention is as follows: Figure 13 As shown in Figure 1, the model denoising theory is based on the prediction value of the pre-trained model, by introducing conditional random noise, combining the error between the predicted value and the true value and the randomness of the input change to model the model.

[0094] Therefore, the denoising process can be regarded as interpolating between the true value and the value predicted by the pre-trained model. It can be seen that during the denoising process, the prediction results of the pre-trained model have large errors. The classification regression diffusion model gradually reduces the root mean square error by continuously predicting the distribution of errors. Finally, when the denoising process is completed, the root mean square error has dropped to 0.08. This shows that through the training of the denoising model, the prediction deviation caused by noise and random fluctuations can be effectively reduced, thereby improving the model's ability to describe uncertainty. This process shows that by introducing the denoising function of the diffusion model, the model's fitting ability in fluctuating data is improved, effectively improving the prediction performance;

[0095] Finally, the life prediction results of the embodiment of the present application after Min-Max denormalization are as follows: Figure 14 As shown in the figure, the root mean square error between the predicted result and the true value is 9.67, the curve is close to the true value, the degradation trend is similar, and the predicted points fluctuate above and below the actual degradation line.

[0096] Therefore, the present invention considers the degradation process of the bearing throughout its entire life cycle, and conducts joint modeling of the time-frequency domain characteristics as they change with material failure and operating conditions. The classification regression diffusion model is combined with a pre-trained Transformer model for prediction, which can predict the remaining service life of the bearing without the need to divide the initial degradation point.

[0097] Compared with the classical physics-based prediction model and the data-driven prediction model, the present invention combines the advantages of the Transformer network and the classification regression diffusion model. With low computational complexity, it can effectively model the distribution relationship between features and remaining life. The training process only considers noise condition modeling. The network structure is simple and is not affected by the gradient vanishing and explosion problems. The algorithm runs fast.

[0098] In summary, the prediction method of the present invention is used to predict the remaining service life of a bearing, and based on the obtained prediction results, predictive maintenance can be performed on mechanical system equipment.

[0099] Accordingly, another embodiment of the present invention provides a computer device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor;

[0100] When the processor executes the computer program, a method for predicting the remaining useful life of a bearing taking into account the service degradation diffusion phenomenon provided by the above embodiment is implemented.

[0101] Furthermore, an embodiment of the present invention provides a storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute a bearing remaining service life prediction method considering the service degradation diffusion phenomenon described in the above-mentioned embodiment of the invention.

[0102] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0103] Those skilled in the art will clearly understand that for the sake of convenience and brevity, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0104] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and exemplary embodiments. They can be applied to a variety of fields suitable for the present invention. Further modifications will be readily apparent to those skilled in the art. Therefore, the present invention is not limited to the specific details and illustrations shown and described herein without departing from the general concept defined by the claims and their equivalents.

Claims

1. A rolling bearing life prediction method based on classification regression diffusion model, characterized in that: The following steps are involved: Use acceleration sensors to pick up rolling bearing accelerated life test data at each measuring point; Extract the time-frequency domain features of the sampling points, normalize the extracted feature data, and use the processing results as the input of the Transformer prediction model; do not perform initial degradation point division, directly create the remaining service life label, and divide it into training data set and test data set; Build a Transformer model and a classification regression diffusion model, use early stopping and cross-validation to obtain the optimal hyperparameter configuration, and pre-train the Transformer model based on the optimal hyperparameters; The extracted vibration signal features and the prediction results of the pre-trained model are concatenated to obtain the conditional input. The sampling timestamps encoded by the embedding layer are embedded into the input features through the Hadarmard product. The classification regression model is used to model the conditional error distribution of the pre-trained model prediction based on the original pre-trained model. The classification regression diffusion model is trained, and the Transformer model is frozen during the training of the classification regression diffusion model. The mean square error loss function and optimization algorithm are used to optimize the weight parameters of the training prediction model, and the remaining life of the bearing is predicted based on the optimized trained prediction model.

2. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 1 is characterized in that: In the step of extracting the time-frequency domain features of the sampling points, the time-frequency domain features include the current signal power density spectral entropy, spectrum mean, kurtosis, root mean square, skewness, peak, permutation entropy and Pearson similarity of the spectrum of the sampling signal at time 0.

3. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 2 is characterized in that: In the constructed Transformer model, the Transformer model uses a 2-layer encoder network to capture sequence dependencies; In the constructed classification regression diffusion model, the denoising network of the classification regression diffusion model is composed of 4 embedding layers plus a linear layer stacked together. The first embedding layer is used to encode the sampling time step and embed the splicing conditional vector through the Hadarmard product; the subsequent embedding layer is used to encode the time step where the noise is located and embed it into the middle hidden layer of the regression diffusion model through the Hadarmard product.

4. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 3 is characterized in that: In the step of using early stopping and cross-validation to obtain the optimal hyperparameter configuration, the Transformer model encoder uses early stopping combined with cross-validation to optimize the hyperparameters during training. The patience value of the early stopping method is set to 20 iterations, the learning rate search range of the loss function is set between 1e-3 and 1e-4, the batch size is 32, and the iteration number is set to 300. The Adma optimizer and the mean squared error loss function are used.

5. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 4 is characterized in that: In the step of embedding the sampling timestamp encoded by the embedding layer into the input feature by the Hadarmard product, the sampling timestamp T is upgraded to a 9-dimensional vector by the embedding layer. , obtained by Hadarmard product , so that the timestamp information is embedded in the conditional input of the classification regression diffusion model, and the randomness of the result is suppressed through continuous input.

6. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 5 is characterized in that: In the step of modeling the conditional error distribution of the classification regression model when predicting the pre-trained model based on the original pre-trained model, the conditional embedding mechanism of the classification regression diffusion model in the process of forward diffusion and noise addition of data is utilized to make the noise conditional distribution include the prior information obtained by the pre-trained model and the result of the previous denoising step.

7. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 6 is characterized in that: The noise intensity during the diffusion process is expressed as: ； in, Indicates time The intensity of the noise added when ; for The value range is between 0 and 1, and it is a length of T represents the number of noise-added time steps in the forward diffusion model. For the distribution of noised data at any time step including t=1 Obey, expressed as: ; ； ； Where, is the initial variable, is a normal distribution, is the identity matrix, is the noise intensity; Represents the prediction result of the pre-trained model, express Time step denoising results; represents the original data that is not corrupted by noise at time step t; Represents the original data from the initial time step to time step t; In the reverse process corresponding to the forward process, the distribution to be predicted is It is expressed as follows: ； ； ； Where, is the initial variable, is a normal distribution, is the identity matrix, is the noise intensity; Represents the prediction result of the pre-trained model, represents the denoising result at time step t, for Denoising results for the time step; Represents the mean of the denoising results from time t to t-1; Indicates the residual degree of noise when denoising from time step t to t-1.

8. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 7 is characterized in that: In the forward diffusion noise adding process, a noise list is generated by cosine noise adding method, and noise is randomly selected from the list. , the random number seed is set to 22.

9. The rolling bearing life prediction method based on the classification regression diffusion model according to claim 8, characterized in that: In the step of optimizing the weight parameters of the training prediction model using the mean square error loss function and the optimization algorithm, the training learning rate is 1e-4, the batch size is 64, the diffusion time step is 1000 steps, and the gradient clipping is performed when the gradient is greater than 1.

Images