Bearing remaining life prediction system and method based on large language model

Through the large language model of STFT transformation and freezing of partial layers, the time-frequency characteristics of the bearing are extracted, which solves the problem of low accuracy of bearing life prediction under different working conditions, realizes high accuracy prediction in noise environment, and reduces data demand. Improves the robustness of predictions.

CN118820756BActive Publication Date: 2024-11-22HANGZHOU INTERNATIONAL INNOVATION INSTITUTE OF BEIHANG UNIVERSITY +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411303275.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-19
Publication Date
2024-11-22
Estimated Expiration
2044-09-19

AI Technical Summary

Technical Problem

The prior art is difficult to accurately predict the remaining life of bearings under different operating conditions, and the data coverage of the training model is incomplete, and it is impossible to effectively deal with external interference and noise influence, resulting in low prediction accuracy.

Method used

The large language model of STFT transformation and frozen part of the layers is used to segment the vibration signals through the sliding window, time-frequency characteristics are extracted and normalized patching layer, input mapping layer, frozen large language model, anti-normalization layer, and STFT inverse transformation are used. Modules and smoothing filters predict the remaining life of the bearing and adapt to noise environments.

Benefits of technology

It realizes more accurate bearing residual life prediction under different operating conditions, reduces the need for a large amount of degraded data, and effectively utilizes pre-trained models to improve the accuracy and robustness of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118820756B_ABST
    Figure CN118820756B_ABST
Patent Text Reader

Abstract

A system and method for predicting the remaining life of a bearing based on a large language model, belonging to the technical field of life prediction. In the system, a data acquisition module is configured to obtain a vibration signal of a healthy period of a bearing to be predicted; a transformation module is configured to obtain L short signal sequences by segmenting the acquired vibration signal through a sliding window, and perform STFT transformation on each short signal to obtain a time-frequency feature matrix; a remaining life prediction model is configured to predict the remaining life of the bearing according to the time-frequency feature matrix, and the remaining life prediction model is trained by a large language model with some layers frozen. The present invention can predict the remaining life of a bearing and the vibration characteristics of the remaining period of the bearing by a large language model with some layers frozen.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a bearing remaining life prediction system and method based on a large language model, belonging to the technical field of life prediction. Background Art

[0002] Bearings are almost the most important type of rotating machinery in industrial systems. Due to long-term operation under harsh conditions such as high speed and heavy load, bearings are prone to failure. Damage to core mechanical components such as bearings often causes the paralysis of the entire transmission system, affecting a wider range of complex engineering systems that could have operated normally, resulting in serious safety accidents, human and economic losses. PHM technology aims to monitor the parameters of the object through an industrial sensor system, and effectively identify the operating status of the object through an intelligent algorithm, so as to obtain the health status information of the object and ultimately achieve condition-based maintenance. Once proposed, PHM technology has been adopted and used in many industrial fields, effectively improving the reliability and safety of various equipment while reducing maintenance costs. An important link in the PHM system developed around rotating machinery such as bearings is the remaining life prediction (RUL). Accurate RUL prediction can effectively estimate the state of bearing degradation, which is of great significance to various applications in actual industry.

[0003] The Chinese invention patent application with publication number CN118010350A discloses a method for predicting the remaining life of a bearing based on probabilistic sparse attention, including step 1: using an acceleration sensor to collect the vibration signals of the bearing from healthy to completely degraded throughout its life cycle, and normalizing them as the input of an end-to-end diagnostic model; step 2: using VMD decomposition to decompose the original vibration signal into multiple components, selecting multiple components for signal reconstruction, and determining the number of components superimposed in the final reconstructed signal based on the correlation coefficient between the reconstructed signal and the original signal to achieve denoising; step 3: dividing the training set and the test set, and using two groups of bearing full life cycle signals under the same working conditions as the training set and the test set respectively. set; Step 4: Calculate the RMS value of the denoised original signal in a sliding window manner, and construct the bearing health index based on the RMS value of the entire life cycle; Step 5: Input the original signal into the bearing remaining life prediction model based on probability sparse attention in a sliding window manner, and extract the enhanced time and frequency domain features through the main frequency perception module; Step 6: Input the obtained enhanced time-frequency features into the probability sparse time-frequency cross attention module for feature fusion, and output the deeply fused time-frequency fusion features; Step 7: Input the time-frequency fusion features into the single-channel attention network for further feature extraction; Step 8: Input it into the fully connected network for feature dimensionality reduction and fit the bearing health index.

[0004] However, there is no method in the prior art for predicting the remaining life of a bearing and the characteristics of the remaining period of the bearing using STFT and a large language model with some layers frozen, nor is there a method for overcoming the following technical problems:

[0005] 1. The state of the bearing itself changes during the degradation process and is also subject to external interference. The distribution of degradation information under time-varying characteristics drifts. Therefore, even if the same model of bearings is under the same operating condition, the distribution of degradation characteristics is still different, and this distribution difference is further amplified under different operating conditions; 2. The data collected for training the model is almost impossible to cover all operating conditions, because the cost and price required to collect all operating conditions and full life cycle data in an experimental environment are almost unacceptable. What is more in line with the actual industrial scenario is that we have a certain amount of operating failure data from other operating conditions similar to the target domain, and a small amount of initial degradation data under the target condition; it can be seen that even under the same operating condition and the same failure behavior, the distribution of degradation characteristics is still different, and the difference is greater under different operating conditions. Few existing mainstream studies have considered the impact of such distribution changes. Summary of the invention

[0006] The purpose of the present invention is to provide a bearing remaining life prediction system and method based on a large language model. The present invention can predict the remaining life of the bearing and the characteristics of the remaining time period of the bearing through STFT and a large language model with some layers frozen; the remaining life of the bearing and the characteristics of the remaining time period of the bearing can be predicted more accurately under the influence of noise.

[0007] To achieve the above-mentioned object, the present invention provides a bearing remaining life prediction system based on a large language model, which comprises:

[0008] A data acquisition module, configured to acquire a vibration signal of a bearing health period to be predicted;

[0009] The transformation module is configured to obtain L short signal sequences by segmenting the vibration signal of the bearing health period obtained by sliding window, and perform STFT transformation on each short signal to obtain a time-frequency feature matrix:

[0010] ;

[0011] In the matrix, The row represents the bearing health period to be predicted. D complex frequency signals of the vibration signal of a sliding window; Indicates The vibration signal of the i-th complex frequency signal of the sliding window; , ;

[0012] A remaining life prediction model is configured to predict the remaining life of a bearing according to a time-frequency feature matrix. The remaining life prediction model includes a normalized patch layer, an input mapping layer, a large language model with a portion of the layers frozen, a denormalized layer, an STFT inverse transformation module, a connection layer, and a smoothing filter, wherein the normalized patch layer is used to normalize the time-frequency feature matrix to obtain a normalized time-frequency feature matrix; the input mapping layer is used to perform a convolution operation on the normalized time-frequency feature matrix to align the input embedding layer dimension of the large language model with a portion of the layers frozen; the large language model with a portion of the layers frozen predicts the normalized remaining life matrix of the bearing according to the matrix input by the input mapping layer, and the denormalized layer denormalizes the normalized remaining life matrix of the bearing to obtain the remaining life matrix; the STFT inverse transformation module performs an STFT inverse transformation on each row of the remaining life matrix to obtain a time domain function of the t-th time period of the remaining life of the bearing, and the connection layer concatenates the time domain functions of the remaining T time periods of the bearing and filters them through a smoothing filter to obtain a time function of the remaining life of the bearing; the product of T and the time length of the window function is the remaining life of the bearing.

[0013] The present invention also provides a method for predicting the remaining life of a bearing based on the large language model, the method comprising the following steps:

[0014] S1-01: Obtain the vibration signal of the bearing health period to be predicted;

[0015] S1-02: The vibration signal of the bearing health period obtained by sliding window segmentation is obtained to obtain L short signal sequences, and each short signal is subjected to STFT transformation to obtain the time-frequency feature matrix:

[0016] ;

[0017] In the matrix, The row represents the bearing health period to be predicted. D complex frequency signals of the vibration signal of a sliding window; Indicates The vibration signal of the i-th complex frequency signal of the sliding window; , ;

[0018] S1-03: The time-frequency feature matrix is ​​input into the remaining life prediction model, and the remaining life prediction model predicts the remaining life of the bearing. The remaining life prediction model includes a normalized patch layer, an input mapping layer, a large language model with some layers frozen, a denormalized layer, an STFT inverse transform module, a connection layer and a smoothing filter, wherein the normalized patch layer is used to normalize the time-frequency feature matrix to obtain a normalized time-frequency feature matrix; the input mapping layer is used to perform a convolution operation on the normalized time-frequency feature matrix to align the input embedding layer dimension of the large language model with some layers frozen. degree; the large language model with some layers frozen predicts the normalized remaining life matrix of the bearing according to the matrix input by the input mapping layer, and the denormalization layer denormalizes the normalized remaining life matrix of the bearing to obtain the remaining life matrix; the STFT inverse transformation module performs STFT inverse transformation on each row of the remaining life matrix to obtain the time domain function of the remaining life of the bearing in the tth time period, and the connection layer concatenates the time domain functions of the remaining T time periods of the bearing and obtains the time function of the remaining life of the bearing through a smoothing filter; the product of T and the time length of the window function is the remaining life of the bearing.

[0019] The present invention also provides a computer program product, which is a program code compiled by a computer language according to the above method, and the code can be called and executed by a processor.

[0020] Compared with the prior art, the bearing remaining life prediction system and method based on the large language model provided by the present invention has the following beneficial effects:

[0021] (1) It can avoid collecting a large amount of degradation data from scratch to train a large prediction model, while effectively utilizing the advantages of pre-training strategies to achieve long-term RUL prediction;

[0022] (2) Considering that the pre-trained large language model is trained on the corpus, in order to enable the pre-trained large language model to process the degradation sequence of bearings cross-modally, the present invention converts the time domain signal into a time-frequency matrix through STFT transformation, which can match the format of the information processed by the large language model;

[0023] (3) In order to fully release the general prediction knowledge obtained by the large model in the pre-training stage and activate it for the prediction of bearing RUL, the present invention selectively freezes most of the parameters of the pre-trained large language model and only fine-tunes some parameters, retaining the general prediction knowledge while inheriting the embedded representation learning advantages of the large language model architecture, thereby effectively modeling the complex nonlinear relationship between high-dimensional degradation sequences and RUL sequences;

[0024] (4) The training method provided by the present invention enables a large language model with some layers frozen to simulate the function of the remaining life of a bearing in the presence of noise. Thus, when the input embedding layer of the large language model with some layers frozen inputs the vibration data of the healthy period of the bearing to be measured with noise, the remaining life of the bearing can be predicted. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a block diagram of the composition of the bearing remaining life prediction system based on the large language model provided by the present invention.

[0026] Figure 2 It is a schematic diagram of a structure based on a large language model with some layers frozen provided by the present invention. DETAILED DESCRIPTION

[0027] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0028] In the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance.

[0029] First embodiment

[0030] Figure 1 The present invention provides a block diagram of the bearing remaining life prediction system based on the large language model, such as Figure 1 As shown, the first embodiment of the present invention provides a bearing remaining life prediction system based on a large language model, including a data acquisition module, a transformation module and a remaining life prediction model, wherein:

[0031] The data acquisition module is configured to acquire a vibration signal of a bearing health period to be predicted;

[0032] The transformation module is configured to obtain L short signal sequences by segmenting the vibration signal of the bearing health period obtained by sliding window, and perform STFT transformation on each short signal to obtain a time-frequency feature matrix:

[0033] ;

[0034] In the matrix, The row represents the bearing health period to be predicted. D complex frequency signals of the vibration signal of a sliding window; Indicates The vibration signal of the i-th complex frequency signal of the sliding window; , ;

[0035] The remaining life prediction model is configured to predict the remaining life of the bearing according to the time-frequency feature matrix, and the remaining life prediction model includes a normalized patch layer, an input mapping layer, a large language model with some layers frozen, a denormalization layer, an STFT inverse transformation module, a connection layer and a smoothing filter, wherein the normalized patch layer is used to normalize the time-frequency feature matrix to obtain a normalized time-frequency feature matrix; the input mapping layer is used to perform a convolution operation on the normalized time-frequency feature matrix to align the input embedding layer dimension of the large language model with some layers frozen; the large language model with some layers frozen predicts the normalized remaining life matrix of the bearing according to the matrix input by the input mapping layer, and the denormalization layer denormalizes the normalized remaining life matrix of the bearing to obtain the remaining life matrix; the STFT inverse transformation module performs an STFT inverse transformation on each row of the remaining life matrix to obtain the time domain function of the t-th time period of the remaining life of the bearing, and the connection layer concatenates the time domain functions of the remaining T time periods of the bearing and filters them through a smoothing filter to obtain the time function of the remaining life of the bearing; the product of T and the time length of the window function is the remaining life of the bearing.

[0036] In the first embodiment, the normalization process includes:

[0037] Calculate the mean of each column in the time-frequency feature matrix and variance :

[0038] ; ;

[0039] According to the following formula, any complex frequency signal in the time-frequency characteristic matrix Perform normalization to obtain the normalized time-frequency feature matrix:

[0040] ,

[0041] In the formula, , are learnable affine transformation parameters, is the set value; , is the first in the normalized time-frequency feature matrix The element of the row i-th column, the normalized matrix is ​​represented as:

[0042] .

[0043] In the first embodiment, the large language model with some layers frozen predicts the normalized remaining life matrix of the bearing according to the matrix input by the input mapping layer:

[0044] ;

[0045] In the formula, The row represents the remaining life of the bearing to be predicted. K complex frequency signals of the vibration signal of a sliding window; Indicates The kth normalized complex frequency signal of a sliding window; , .

[0046] The denormalization layer denormalizes each element in the normalized residual life matrix according to the following formula to obtain the residual life matrix:

[0047] , β k , γ k Belong to the learnable affine transformation parameters, the remaining life matrix is ​​expressed as:

[0048] .

[0049] Optionally,

[0050] ,

[0051] In the formula, ; ;

[0052] In the formula, , are learnable affine transformation parameters, is a very small value; .

[0053] Figure 2 Schematic diagram of a structure based on a large language model with some layers frozen provided by the present invention. Figure 2 The large language model with some layers frozen as shown includes: an input embedding layer, a position embedding layer, a multi-head attention mechanism layer, a first residual connection layer, a feedforward neural network, and a second residual connection layer. In the present invention, the large language model with some layers frozen is a transformer model. The first embodiment of the present invention trains the large language model with some layers frozen by the following steps:

[0054] Step 1: Extract multiple data sets of bearing life data from the database, each of which includes a pre-trained data vector and the test data vector , the pre-training data vector Input to the input embedding layer of the large language model with some layers frozen, output vector from the second residual connection layer for:

[0055] ,

[0056] in, , , is the current parameter vector of the large language model with some layers frozen; is the kth value of the output vector output from the second residual connection layer, r nk is the test data vector R n The kth value of , k=1,2,...,K;

[0057] Step 2: Calculate the error according to the following formula :

[0058] ;

[0059] Step 3: Update the current parameter vector of the large language model with some layers frozen according to the following formula:

[0060] ,

[0061] In the formula, is the learning rate, Express The gradient of The parameters of the large language model with some layers frozen include: the key value, query value, value of the multi-head attention mechanism layer, and the matrix coefficients and bias of the feedforward neural network;

[0062] Step 4: Determine the error Is it less than the set value? If so, end the training. If not, make the new parameter vector Replace the current parameter vector , and return to step 1.

[0063] Optionally, the error can also be calculated by a loss function.

[0064] The first embodiment of the present invention can use the above-mentioned training method to enable the large language model with some layers frozen to simulate the function of the remaining life of the bearing. Therefore, when the input embedding layer of the large language model with some layers frozen inputs the vibration data of the healthy period of the bearing to be measured, the remaining life can be predicted.

[0065] Second embodiment

[0066] The second embodiment of the present invention only describes the contents that are different from the first embodiment, and the same contents will not be described repeatedly.

[0067] The second embodiment of the present invention is different from the first embodiment only in that the training process of the large language model with some layers frozen is different.

[0068] The first embodiment of the present invention trains a large language model with some layers frozen through the following steps:

[0069] Step 1: Extract multiple data sets of bearing life data from the database, each of which includes a pre-trained data vector and the test data vector , the pre-training data vector Input to the input embedding layer of the large language model with some layers frozen, and output vector from the second residual connection layer :

[0070] ,

[0071] in, , , = , To freeze the current parameter vectors of the large language model of some layers, the parameters of the large language model of some layers are frozen, including: the key value, query value, value of the multi-head attention mechanism layer, and the matrix coefficients and bias of the feedforward neural network; is the random noise generated by the standard normal distribution; is the current mean value, is the current variance; difference; is the kth value of the output vector output from the second residual connection layer, r nk is the test data vector R n The kth value of , k=1,2,...,K;

[0072] Step 2: Let the loss function L be:

[0073] ,

[0074] The gradients of the mean and variance are calculated as follows:

[0075] ,

[0076] ;

[0077] Step 3: Update the current parameter vector of the large language model with some layers frozen according to the following formula:

[0078] ;

[0079] ,

[0080] In the formula, , is the learning rate, Express The gradient of Express The gradient of

[0081] Step 4: Determine whether the loss function value is the minimum. If so, end the training. If not, make the new parameter vector Replace the current parameter vector , the new parameter vector Replace the current parameter vector , and return to step 1.

[0082] The second embodiment of the present invention can use the above-mentioned training method to enable the large language model with some layers frozen to simulate the function of the remaining life of a bearing in the presence of noise. Therefore, when the input embedding layer of the large language model with some layers frozen inputs the vibration data with noise during the healthy period of the bearing to be measured, the remaining life of the bearing can be predicted.

[0083] Third embodiment

[0084] The third embodiment of the present invention only describes the contents that are different from the second embodiment, and the same contents will not be described repeatedly.

[0085] A third embodiment of the present invention provides a method for predicting the remaining life of a bearing based on a large language model, which comprises the following steps:

[0086] S1-01: Obtain the vibration signal of the bearing health period to be predicted;

[0087] S1-02: The vibration signal of the bearing health period is obtained by sliding window segmentation to obtain L short signal sequences, and each short signal is transformed by STFT to obtain the time-frequency feature matrix:

[0088] ;

[0089] In the matrix, The row represents the bearing health period to be predicted. D complex frequency signals of the vibration signal of a sliding window; Indicates The vibration signal of the i-th complex frequency signal of the sliding window; , ;

[0090] S1-03: Input the time-frequency feature matrix into the remaining life prediction model, the remaining life prediction model predicts the remaining life of the bearing, and the remaining life prediction model includes a normalized patch layer, an input mapping layer, a large language model with some layers frozen, a denormalized layer, an STFT inverse transform module, a connection layer and a smoothing filter, wherein the normalized patch layer is used to normalize the time-frequency feature matrix to obtain a normalized time-frequency feature matrix; the input mapping layer is used to perform a convolution operation on the normalized time-frequency feature matrix to align the input embedding layer of the large language model with some layers frozen dimension; the large language model with some layers frozen predicts the normalized remaining life matrix of the bearing according to the matrix input by the input mapping layer, and the denormalization layer denormalizes the normalized remaining life matrix of the bearing to obtain the remaining life matrix; the STFT inverse transformation module performs STFT inverse transformation on each row of the remaining life matrix to obtain the time domain function of the remaining life of the bearing in the tth time period, and the connection layer concatenates the time domain functions of the remaining T time periods of the bearing and filters them through a smoothing filter to obtain the time function of the remaining life of the bearing; the product of T and the time length of the window function is the remaining life of the bearing.

[0091] Fourth embodiment

[0092] The fourth embodiment of the present invention only describes the contents that are different from the third embodiment, and the same contents are not described again.

[0093] A fourth embodiment of the present invention provides a computer program product, which is a program code compiled in a computer language by the method described in the third embodiment, and the program code can be called and executed by a processor.

[0094] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A bearing remaining life prediction system based on a large language model, It is characterized in that include: A data acquisition module, configured to acquire a vibration signal of a bearing health period to be predicted; The transformation module is configured to obtain L short signal sequences by segmenting the vibration signal of the bearing health period obtained by sliding window, and perform STFT transformation on each short signal to obtain a time-frequency feature matrix: ; In the matrix, The row represents the bearing health period to be predicted. D complex frequency signals of the vibration signal of a sliding window; Indicates The vibration signal of the i-th complex frequency signal of the sliding window; , ; A remaining life prediction model is configured to predict the remaining life of a bearing according to a time-frequency feature matrix. The remaining life prediction model includes a normalized patch layer, an input mapping layer, a large language model with some layers frozen, a denormalized layer, an STFT inverse transformation module, a connection layer and a smoothing filter, wherein the normalized patch layer is used to normalize the time-frequency feature matrix to obtain a normalized time-frequency feature matrix; the input mapping layer is used to perform a convolution operation on the normalized time-frequency feature matrix to align the input embedding layer dimension of the large language model with some layers frozen; the large language model with some layers frozen predicts the normalized remaining life matrix of the bearing according to the matrix input by the input mapping layer, and the denormalized layer denormalizes the normalized remaining life matrix of the bearing to obtain the remaining life matrix; the STFT inverse transformation module performs an STFT inverse transformation on each row of the remaining life matrix to obtain a time domain function of the t-th time period of the remaining life of the bearing, and the connection layer concatenates the time domain functions of the remaining T time periods of the bearing and filters them through a smoothing filter to obtain a time function of the remaining life of the bearing; the product of T and the time length of the window function is the remaining life of the bearing.

2. The system according to claim 1, It is characterized in that Normalization processing includes: Calculate the mean of each column in the time-frequency feature matrix and variance : ; ; According to the following formula, any complex frequency signal in the time-frequency characteristic matrix Perform normalization to obtain the normalized time-frequency feature matrix: , In the formula, , are learnable affine transformation parameters, is a very small value; , is the first in the normalized time-frequency feature matrix The element in the row and column i.

3. The system according to claim 2, It is characterized in that The large language model with some layers frozen predicts the normalized remaining life matrix of the bearing based on the matrix input by the input mapping layer: ; In the formula, The row represents the remaining life of the bearing to be predicted. K complex frequency signals of the vibration signal of a sliding window; Indicates The kth normalized complex frequency signal of a sliding window; , .

4. The system according to claim 3, It is characterized in that The denormalization layer denormalizes each element in the normalized residual life matrix according to the following formula to obtain the residual life matrix: , In the formula, β k , γ k These are learnable affine transformation parameters.

5. The system according to claim 4, It is characterized in that Methods for training a large language model with some layers frozen include: Step 1: Extract multiple data sets of bearing life data from the database, each of which includes a pre-trained data vector and the test data vector , the pre-training data vector Input to the input embedding layer of the large language model with some layers frozen, output vector from the second residual connection layer for: , in, , , = , To freeze the current parameter vector of the large language model with some layers, the parameters of the large language model with some layers frozen include: the key value, query value, value of the multi-head attention mechanism layer, and the matrix coefficient and bias of the feedforward neural network; is the random noise generated by the standard normal distribution; is the current mean value, is the current variance; is the kth value of the output vector output from the second residual connection layer, r nk is the test data vector R n The kth value of , k=1,2,...,K; Step 2: Let the loss function L be: , The gradients of the mean and variance are calculated as follows: , ; Step 3: Update the current parameter vector of the large language model with some layers frozen according to the following formula: ; , In the formula, , is the learning rate, Express The gradient of Express The gradient of Step 4: Determine whether the loss function value is the minimum. If so, end the training. If not, make the new parameter vector Replace the current parameter vector , the new parameter vector Replace the current parameter vector , and return to step 1.

6. A method for predicting the remaining life of bearings based on a large language model. It is characterized in that The steps include: S1-01: Obtain the vibration signal of the bearing health period to be predicted; S1-02: The vibration signal of the bearing health period is obtained by sliding window segmentation to obtain L short signal sequences, and each short signal is transformed by STFT to obtain the time-frequency feature matrix: ; In the matrix, The row represents the bearing health period to be predicted. D complex frequency signals of the vibration signal of a sliding window; Indicates The vibration signal of the i-th complex frequency signal of the sliding window; , ; S1-03: Input the time-frequency feature matrix into the remaining life prediction model, the remaining life prediction model predicts the remaining life of the bearing, the remaining life prediction model includes a normalized patch layer, an input mapping layer, a large language model with some layers frozen, a denormalized layer, an STFT inverse transform module, a connection layer and a smoothing filter, wherein the normalized patch layer is used to normalize the time-frequency feature matrix to obtain a normalized time-frequency feature matrix; the input mapping layer is used to perform a convolution operation on the normalized time-frequency feature matrix to align the input embedding layer dimension of the large language model with some layers frozen degree; the large language model with some layers frozen predicts the normalized remaining life matrix of the bearing according to the matrix input by the input mapping layer, and the denormalization layer denormalizes the normalized remaining life matrix of the bearing to obtain the remaining life matrix; the STFT inverse transformation module performs STFT inverse transformation on each row of the remaining life matrix to obtain the time domain function of the remaining life of the bearing in the tth time period, and the connection layer concatenates the time domain functions of the remaining T time periods of the bearing and obtains the time function of the remaining life of the bearing through a smoothing filter; the product of T and the time length of the window function is the remaining life of the bearing.

7. The method according to claim 6, It is characterized in that Normalization processing includes: Calculate the mean of each column in the time-frequency feature matrix and variance : ; ; According to the following formula, any complex frequency signal in the time-frequency characteristic matrix Perform normalization to obtain the normalized time-frequency feature matrix: , In the formula, , are learnable affine transformation parameters, is a very small value; , is the first in the normalized time-frequency feature matrix The element in the row and column i.

8. The method according to claim 7, It is characterized in that The large language model with some layers frozen predicts the normalized remaining life matrix of the bearing based on the matrix input by the input mapping layer: ; In the formula, The row represents the remaining life of the bearing to be predicted. K complex frequency signals of the vibration signal of a sliding window; Indicates The kth normalized complex frequency signal of a sliding window; , .

9. The method according to claim 8, It is characterized in that The denormalization layer denormalizes each element in the normalized residual life matrix according to the following formula to obtain the residual life matrix: , In the formula, β k , γ k These are learnable affine transformation parameters.

10. The method according to claim 9, It is characterized in that Methods for training a large language model with some layers frozen include: Step 1: Extract multiple data sets of bearing life data from the database, each of which includes a pre-trained data vector and the test data vector , the pre-training data vector Input to the input embedding layer of the large language model with some layers frozen, output vector from the second residual connection layer for: , in, , , = , To freeze the current parameter vector of the large language model with some layers, the parameters of the large language model with some layers frozen include: the key value, query value, value of the multi-head attention mechanism layer, and the matrix coefficient and bias of the feedforward neural network; is the random noise generated by the standard normal distribution; is the current mean value, is the current variance; is the kth value of the output vector output from the second residual connection layer, r nk is the test data vector R n The kth value of , k=1,2,...,K; Step 2: Let the loss function L be: , The gradients of the mean and variance are calculated as follows: , ; Step 3: Update the current parameter vector of the large language model with some layers frozen according to the following formula: ; , In the formula, , is the learning rate, Express The gradient of Express The gradient of Step 4: Determine whether the loss function value is the minimum. If so, end the training. If not, make the new parameter vector Replace the current parameter vector , the new parameter vector Replace the current parameter vector , and return to step 1.