Rolling bearing service life prediction method based on plane vibration characteristics and degradation history

Abstract

The invention provides a rolling bearing life prediction method based on plane vibration characteristics and degradation history, and the method comprises the steps: constructing a full-life-cycle vibration signal data set of a bearing, and extracting plane vibration characteristics which comprise a principal direction angle and an ellipticity; constructing a degradation historical feature, and performing zero filling on the feature sequence of the whole degradation process of the bearing into a fixed length; training a CNN-LSTM-based hybrid model, and outputting a residual life prediction value after inputting degradation historical features; a weighted mean square error loss function optimization model is adopted, and the prediction precision of a key stage before failure is improved. According to the method provided by the invention, RMSE 0.154 is realized on a PHM2012 data set by fusing multi-directional vibration information and a complete degradation history, the RMSE 0.154 is improved by 47.3% compared with a traditional method, and the method is suitable for bearing health state monitoring and maintenance decision making under variable rotating speed and variable load working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery health monitoring, and specifically to a rolling bearing life prediction method based on planar vibration characteristics and degradation history. Background Art

[0002] Bearings are common key components in mechanical devices, consisting of an inner ring, an outer ring, rolling elements, and a cage, and are used to reduce friction in rotating machinery and support the weight of the bearing components. In major equipment and high-precision devices, high-performance bearings play a key role. Therefore, improving the design and manufacturing level of rolling bearings and perfecting the manufacturing process have become urgent problems to be solved in China's bearing industry.

[0003] As a core component of rotating machinery, rolling bearings are widely used in major equipment such as wind power, high-speed rail, and aeroengines, and their reliability directly affects the safe operation and service life of the equipment. However, under complex working conditions such as high speed, variable load, and strong impact, the fatigue failure of bearings is one of the main reasons for mechanical system failures. According to statistics, about 40% of rotating machinery failures are related to bearing failures, and sudden bearing failures may trigger a chain reaction, causing significant economic losses and even safety accidents. At present, there is insufficient fault warning ability under non-stationary working conditions such as variable speed and variable load, which restricts the independent and controllable development of high-end equipment. Traditional methods for predicting the remaining useful life (RUL) of bearings mainly rely on time-frequency domain feature analysis of vibration signals, but there are the following technical bottlenecks: limitation of single-direction information, existing methods mostly use vibration signals in a single direction (such as vertical acceleration), ignoring the correlation between the load direction and fault evolution in planar vibration, resulting in insufficient extraction of early weak fault features; one-sidedness of degradation modeling, local feature extraction based on a sliding window is difficult to capture the degradation law of the entire life cycle, and traditional statistical features (such as RMS, kurtosis) have poor adaptability to variable speed working conditions; disconnection from physical mechanisms, data-driven models lack the embedding of bearing fault mechanisms (such as the modulation relationship between the inner ring fault characteristic order and speed), and the generalization performance significantly decreases under unknown working conditions.

[0004] The comparison between the present application and the prior art is as follows: Technical comparison with the application text CN202010178652 "Rolling Bearing Fault Diagnosis Method Based on Computational Order Tracking and Spectral Kurtosis"; 1. The patent of the application text CN202010178652 proposes a rolling bearing fault diagnosis method for variable speed working conditions, and its core technology is to eliminate the influence of speed fluctuations through computational order tracking (COT). This method first constructs an angle-time quadratic equation (based on the assumption of uniform angular acceleration) using the speed pulse signal, and uses cubic spline interpolation to achieve angular resampling, converting the non-stationary time-domain signal into an angular domain stationary signal; combining the spectral kurtosis algorithm to dynamically locate the optimal demodulation frequency band (center frequency With bandwidth ), after extracting the characteristic frequency band through an elliptical filter, envelope analysis is carried out, and finally the order spectrum (such as 18 times the inner ring fault characteristic order) is output to realize the fault type positioning under variable speed. The core value of this scheme lies in solving the "frequency ambiguity" problem of traditional spectrum analysis, which is applicable to dynamic working conditions such as acceleration / deceleration, but can only achieve qualitative judgment of the presence or absence of faults, and cannot quantify the remaining life or degradation trend; 2. There are essential differences in technical objectives and architectures between this application (a rolling bearing life prediction method based on planar vibration characteristics and degradation history) and CN202010178652: This patent focuses on quantitative prediction of remaining life (RUL), rather than qualitative fault diagnosis, belonging to a more advanced stage of equipment health management; It innovatively proposes a dual-track system of planar vibration characteristics (PVF) and degradation history characteristics (DHF) - PVF extracts the principal direction angle ( ) and ellipticity ( ) through covariance matrix decomposition, fuses horizontal / vertical bidirectional vibration physical information to correlate with fault mechanisms, and DHF zero-pads the PVF sequence over the entire life cycle to a fixed length (such as 3000×4) to retain the cumulative damage evolution law; A CNN-LSTM hybrid model is used to learn the mapping relationship from DHF to RUL, and the prediction accuracy of the critical stage before failure is enhanced through a weighted loss function ( ), where is the weighted exponent, is the current time step, is the total life of the bearing, are the true value and predicted value of the remaining life, N is the number of samples, is the sample serial number, and finally a high-precision remaining life percentage is output (such as RMSE 0.154 for the PHM2012 dataset), while the comparative patent only outputs the fault type label.

[0005] There are essential differences between the two in scheduling objectives, system structures, and technical routes.

[0006] Technical comparison with the application text CN202311277320, "A Rolling Bearing Fault Diagnosis Method"; 1. The patent of the application text CN202311277320 aims at early fault detection in high-noise stationary working conditions and proposes a wavelet-EMD joint noise reduction framework. Its technical path is divided into two layers: adaptive noise reduction and anti-aliasing decomposition. In the noise reduction layer, low-correlation nodes are removed through wavelet packet decomposition and cross-correlation coefficient filtering (threshold ), and the global threshold is dynamically optimized based on the principle of minimizing sample entropy., and a new threshold function is designed to retain the weak impact characteristics; at the decomposition layer, the complementary set of high-frequency harmonics added empirical mode decomposition (CEHFHA-EMD) is used to inject high-frequency harmonics to suppress modal aliasing, and the Hilbert marginal spectrum and envelope spectrum (such as the 163Hz inner ring fault fundamental frequency and multiple frequencies) are extracted after decomposing the precise IMF components. This solution achieves high-sensitivity detection of early faults under strong noise, but is limited to current state judgment and cannot predict degradation trends or quantify remaining life; 2. This application and CN202311277320 have made significant leaps in technical dimensions and application depth: In terms of signal dimension, the comparative patent relies on single-point vibration signal analysis, while this patent pioneered the plane vibration vector feature (PVF), which quantifies the load direction and energy distribution (main direction angle and ellipticity) through covariance matrix feature decomposition, directly associates the physical mechanism of bearing fault (such as the order of inner ring fault characteristics), and greatly improves the interpretability of the model; in terms of time scale, the envelope spectrum of the comparative patent only reflects the instantaneous fault state, and this patent constructs a degradation history feature (DHF) covering the degradation trajectory of the entire life cycle (zero-filled fixed sequence), and combines the CNN-LSTM model to capture long-term dependencies, realizing the leap from "fault detection" to "life prediction"; in terms of optimization mechanism, this patent designs a weighted mean square error loss function ( ) Dynamically improve the prediction weight before failure ( ), and the hyperparameters are optimized by adaptive particle swarm algorithm (APSO), while the comparative patent does not involve the prediction accuracy optimization mechanism.

[0007] There are essential differences between the two in terms of system architecture, scheduling goals and technical paths. Summary of the invention

[0008] In view of the problems of insufficient prediction accuracy, low sensitivity to early faults and poor physical interpretability in existing rolling bearing remaining life prediction methods under variable speed and variable load conditions, the present invention proposes a rolling bearing life prediction method based on planar vibration characteristics and degradation history. This method aims to achieve high-precision and highly robust bearing health status assessment and remaining life prediction by integrating multi-directional vibration information with full-cycle degradation laws.

[0009] To achieve the above object, the technical solution adopted by the present invention is: The rolling bearing life prediction method based on plane vibration characteristics and degradation history includes the following specific steps: 1) Construct a planar vibration signal dataset of rolling bearings throughout their life cycle and extract the planar vibration feature PVF and degradation history feature DHF; 2) Constructing and training a CNN-LSTM hybrid model, which is used to output a predicted value of the remaining useful life (RUL) after inputting a DHF sequence; 3) Based on the weighted mean square error (wMSE) loss function, the model is optimized to improve the prediction accuracy in the critical stage before failure and output the optimal RUL prediction result.

[0010] As a further improvement of the present invention, the step 1) constructs a planar vibration signal data set for the entire life cycle of a rolling bearing, and the specific steps are as follows, including: Collect the vibration signals of the bearing in the horizontal and vertical directions, with a sampling rate of not less than 25.6 kHz; De-noising and normalizing the original signal to obtain a standardized vibration data set; Extract time domain, frequency domain and time-frequency domain features and construct a multi-dimensional feature matrix.

[0011] As a further improvement of the present invention, the step 1) extracting the planar vibration feature PVF comprises the following specific steps, including: Calculate the covariance matrix of the vibration signal , obtain the main direction angle through feature decomposition and ellipticity ; Use the robust covariance estimation algorithm to remove outliers until the matrix determinant converges; expand the 2D statistical features, including the plane RMS and Crest coefficient, and the calculation formula is: ; ; ; ; in, is the polar radian of the main direction of vibration, are the major and minor axes of the fitted ellipse; N is the number of sampling points in the vibration signal, is the sample serial number, It is the amplitude value of the sampling point in the x / y direction.

[0012] As a further improvement of the present invention, the step 1) extracting the degradation history feature DHF comprises the following specific steps, including: Zero-fill the PVF sequence of the bearing's entire life cycle to a fixed length; Arrange features in chronological order to preserve the cumulative damage pattern of the degradation process; The DHF sequence is standardized to eliminate the dimension effect.

[0013] As a further improvement of the present invention, the CNN-LSTM hybrid model in step 2) is specifically as follows: The input layer receives a 3000×4 dimensional DHF feature matrix, where 3000 is the time step and 4 is the feature dimension; Convolutional layer, including three one-dimensional convolutions, and its operation formula is: , where m is the vibration signal window, y is the feature after convolution, l represents the layer number, k is the convolution kernel size, w is the weight, b is the bias, is the sample serial number, and τ is the current global time; The LeakyReLU activation function is adopted: ;

[0014] where α is taken as 0.01, x is the output of the neuron, is the output of the activation function; LSTM layer, adopting a bidirectional LSTM structure, and the calculation formula for each time step t is: ; ; ; ; ; ;

[0015] where , , o represent the forget gate, input gate and output gate respectively, t is the time ordinal number in the sequence, x is the output of the neuron, C is the cell state, is the candidate cell state, is the hidden state, and are the weight and bias coefficients of the forget gate, input gate, output gate and cell state respectively, and the number of hidden units is set to 128; is the sigmoid activation function, is the tangent activation function.

[0016] Output layer, adopting the Sigmoid activation function, and outputting the normalized remaining life percentage: , where z is the result of linear transformation, e is the natural coefficient, is the predicted output value.

[0017] As a further improvement of the present invention, the loss function of the weighted mean square error wMSE in step 3) is defined as: ;

[0018] where is the weighted exponent, is the current time step, is the total life of the bearing, are the true value and predicted value of the remaining useful life, N is the number of samples, is the sample serial number.

[0019] As a further improvement of the present invention, the optimization steps of optimizing the model based on the weighted mean square error wMSE loss function in step 3) are specifically as follows: Adopt the adaptive particle swarm optimization algorithm APSO to adjust the CNN-LSTM hyperparameters; Use the validation set RMSE as the fitness function to search for the optimal network configuration; Prevent overfitting through the early stopping strategy and retain the best model weights.

[0020] As a further improvement of the present invention, the velocity update formula of the adaptive particle swarm optimization algorithm APSO is: , where w is the inertia weight, , are the learning factors, , are random numbers, is the velocity component of particle i in the d-th dimension at the (k + 1)-th iteration; c is the cognitive learning factor; is the individual historical optimal position component of particle i in the d-th dimension at the k-th iteration; is the current position component of particle i in the d-th dimension at the k-th iteration.

[0021] Beneficial effects;

[0022] (1) The remaining useful life prediction method based on planar vibration features (PVF) and degradation history modeling (DHF) proposed by the present invention significantly improves the prediction accuracy under variable speed and variable load conditions by fusing multi-directional vibration information and full-cycle degradation laws. Experiments show that on the PHM2012 standard data set, the RMSE of this method reaches 0.154, which is 47.3% higher than that of traditional single-direction vibration analysis methods, and the detection sensitivity to early faults is increased by more than 35%; (2) The present invention innovatively introduces planar vibration ellipse features (principal direction angle and ellipticity ), and quantifies the load direction and vibration energy distribution through covariance matrix eigenvalue decomposition, solving the problem of losing physical information in traditional single-direction analysis. This method can directly correlate with the bearing fault mechanism (such as the characteristic order of inner race faults), greatly improving the interpretability of the model; (3) The CNN-LSTM hybrid model designed in the present invention combines the spatial feature extraction ability of the convolutional neural network and the temporal modeling advantage of the long short-term memory network. By strengthening the prediction weight of the critical stage before failure through the weighted loss function (wMSE), it effectively avoids the problem of prediction result fluctuations in traditional methods; (4) The method for constructing the degradation history feature (DHF) proposed in the present invention preserves the full-life cycle degradation trajectory by zero-padding the fixed-length sequence. Compared with the sliding window method, this method has universality and can be extended and applied to the life prediction scenarios of other rotating machinery such as gearboxes and motors; (5) The adaptive optimization strategy (APSO) of the present invention can automatically adjust the model hyperparameters, significantly reducing the dependence on expert experience in engineering implementation and providing an out-of-the-box intelligent prediction solution for industrial sites. Description of the Drawings

[0023] Figure 1 is the flowchart of the method of the present invention; Figure 2 is the architecture diagram of the improved CNN-LSTM model of the present invention; Figure 3 is the comparison diagram of the effects of the present invention under three different load conditions of low, medium, and high; Figure 3 in which (a) is the comparison diagram of the effects under low load conditions; Figure 3 in which (b) is the comparison diagram of the effects under medium load conditions; Figure 3 in which (c) is the comparison diagram of the effects under high load conditions. Detailed Embodiment

[0024] The following further describes the present invention in detail in conjunction with the drawings and specific embodiments: The present invention discloses a rolling bearing life prediction method based on planar vibration characteristics and degradation history. This method realizes the accurate prediction of the remaining service life of rolling bearings by fusing multi-directional vibration signal characteristics and deep learning technology. The following details the specific embodiments of the present invention in conjunction with the drawings.

[0025] As shown in the attached Figure 1 figures, the rolling bearing life prediction method based on planar vibration characteristics and degradation history of the present invention includes the following steps: S1. System hardware configuration and data collection, including the following steps: The implementation of the present invention first requires building a complete data acquisition system. This system uses the 352C33 type ICP acceleration sensor produced by PCB Company in the United States as the core acquisition device. This sensor has a sensitivity of 100mV / g and a frequency response range of 0.5Hz - 10kHz, and can accurately capture the vibration signals during the operation of the bearing. In terms of installation and layout, one sensor needs to be installed in each of the horizontal and vertical directions of the bearing housing, and the installation position is no more than 50mm away from the outer ring of the bearing. A magnetic base is used to ensure the close contact between the sensor and the bearing housing. Data acquisition uses the NI cDAQ - 9178 chassis in cooperation with the NI 9234 acquisition module, and the sampling frequency is set to 25.6kHz, meeting the requirements of the Nyquist sampling theorem. Each time, 10 - second vibration signals are continuously acquired. To ensure data quality, the acquisition system is built - in with an anti - aliasing filter with a cut - off frequency of 10kHz, effectively preventing high - frequency noise interference.

[0026] S2. Signal pre - processing and feature extraction; The acquired original vibration signals need to go through a strict pre - processing process. First, wavelet threshold denoising is performed. The db5 wavelet basis function is selected for 5 - layer decomposition, and its denoising process can be expressed by the following formula:

[0027] where, represents the wavelet coefficient, is the threshold, and its value is , is the noise standard deviation, and N is the signal length. After denoising, the high - frequency components above 2kHz are retained, which is crucial for capturing the early fault characteristics of the bearing. For the outliers in the signal, they are identified by calculating the Mahalanobis distance: . Where, is the mean vector, x is the signal intensity, is the covariance matrix. The threshold is set to 3σ, and the data segments exceeding this threshold will be removed. The missing parts are filled by linear interpolation. Finally, MinMaxScaler is used to normalize all features to the [0, 1] interval.

[0028] S3. Steps for extracting planar vibration characteristics: The extraction of planar vibration characteristics is one of the core innovation points of the present invention. First, calculate the covariance matrix of the vibration signals in the horizontal and vertical directions: ;

[0029] where, represents the covariance between the two signals. Perform eigenvalue decomposition on the covariance matrix: , where V is the feature vector matrix and D is the eigenvalue diagonal matrix. Two key features can be extracted from this: the main direction angle and the ellipticity. To improve the robustness of feature extraction, the MCD (Minimum Covariance Determinant) algorithm is used for iterative calculation until the change rate of the matrix determinant is less than 1%. The steps of its iterative calculation and the specific implementation method are as follows: This algorithm first performs an initialization operation: set the optimal covariance matrix C_best as the two-dimensional identity matrix, initialize the historical optimal determinant value as positive infinity, and randomly select an initial data subset S as the starting point for calculation. The algorithm controls the iterative process by continuously monitoring the determinant improvement amount σ, and terminates the loop when σ is less than the preset threshold ε.

[0030] Each iteration contains four core steps: Statistic calculation: Calculate the covariance matrix C_S and its determinant value based on the current subset S.

[0031] Optimal solution update: If the current determinant is less than the historical optimal value, update the optimal covariance matrix C_best with C_S, calculate the difference between the new and old determinants and assign it to σ, and record the new determinant as the optimal value. This design ensures that the algorithm continuously approaches a more compact data distribution.

[0032] Outlier identification: Calculate the Mahalanobis distance for all data points. This distance is constructed through the current subset mean μ_S and the covariance inverse matrix C_S^(-1), quantifying the deviation degree of each point from the subset.

[0033] Dynamic subset screening: Set an adaptive threshold θ based on the standard deviation, and only retain the points with a Mahalanobis distance lower than θ to form a new subset S. This mechanism effectively excludes outliers with large statistical deviations.

[0034] When the iteration terminates, the algorithm outputs the finally optimized covariance matrix C_best. Its robustness comes from a dual closed-loop design: driving the optimization of the covariance matrix through the determinant minimization objective, and at the same time using the Mahalanobis distance threshold to achieve the dynamic purification of the data subset, making the result insensitive to outliers.

[0035] S31. In the data preparation stage, prepare a dataset X containing n two-dimensional data points, ensuring the correct data format. Initialize the parameters required for calculation, including setting the initial covariance matrix as a 2nd-order identity matrix, the reference determinant value as positive infinity, and randomly select an initial data subset S; S32. Enter the main loop and calculate the covariance matrix of the current subset S; S33. Calculate the determinant value of this covariance matrix; S34. Compare the current determinant value with the historical optimal value. If it is better, update the optimal estimate and the reference value; S35. Calculate the Mahalanobis distance of all data points relative to the current subset; S36. Screen data points according to a preset threshold and update subset S; S37. Determine whether the termination condition is met. If so, end; if not, return to S32 to regenerate the offspring. The termination condition is that when the difference in determinants between two consecutive iterations is less than the preset convergence threshold, the calculation process is terminated. At the same time, set the maximum number of iterations as a safety limit to prevent infinite loops. In some embodiments of the present invention, the maximum number of iterations is set to 100. Return the finally calculated robust covariance matrix, which is robust to outliers and can accurately reflect the distribution characteristics of the main body of the data.

[0036] S4. Degradation history feature construction method: The construction of degradation history features adopts an innovative zero-padding fixed-length sequence method. Suppose m groups of PVF features are collected during the entire life cycle of the bearing, and each group of features contains n parameters. Then the DHF matrix at time t is constructed as:

[0037] Among them, is the feature vector, is the maximum life constant, c is the number of channels, is the real number space. When m < 3000, it is filled with zero vectors. This processing method not only retains the complete trajectory of the degradation process but also solves the problem of local information fragmentation caused by the traditional sliding window method.

[0038] S5. Construction of CNN-LSTM hybrid model: The CNN-LSTM hybrid neural network architecture designed in the present invention is as Figure 2 shown and specifically includes the following layers: S51. Input layer, receiving a 3000×4-dimensional DHF feature matrix, where 3000 is the time step and 4 is the feature dimension; S52. Convolutional layer, including 3 layers of one-dimensional convolution, and its operation formula is: , where m is the vibration signal window, y is the feature after convolution, l represents the layer number, k is the convolution kernel size, w is the weight, b is the bias, i is the sample number, and τ is the current global time. The number of channels of the three layers of convolution is set to 16, 32, and 64 respectively, and the LeakyReLU activation function is adopted: ;

[0039] Among them, α takes 0.01, x is the output of the neuron, and f(x) is the output of the activation function; S53. LSTM layer, adopting a bidirectional LSTM structure, and the calculation formula for each time step t is as follows: ; ; ; ; ; ;

[0040] Among them, f, i, and o respectively represent the forget gate, input gate, and output gate. t is the time ordinal number within the sequence, x is the output of the neuron, C is the cell state, is the candidate cell state, h is the hidden state, and are respectively the weights and bias coefficients of the forget gate, input gate, output gate, and cell state, is the sigmoid activation function, and tanh is the tangent activation function. Set the number of hidden units to 128; S54. Output layer, adopting the sigmoid activation function, and outputting the normalized remaining life percentage: , where z is the result of the linear transformation, e is the natural coefficient, is the predicted output value.

[0041] S6. Model training and optimization steps: The model training adopts a weighted mean square error loss function: , where is the weighted exponent, is the current time step, is the total life of the bearing, are the true value and predicted value of the remaining life, N is the number of samples, is the sample serial number. The optimization adopts the Adam optimizer, with the initial learning rate set to 0.001, decaying by 10% every 20 epochs. For the optimization hyperparameters, the adaptive particle swarm optimization (APSO) is adopted. Set the population size to 50 and the maximum number of iterations to 100. Among them, the effect comparison diagram under the low load condition is as shown in Figure 3 (a) in Figure 3 , the effect comparison diagram under the medium load condition is as shown in Figure 3 (b) in

[0042] The above is only a preferred embodiment of the present invention, and it is not any other form of limitation to the present invention. Any modification or equivalent change made according to the technical essence of the present invention still belongs to the scope protected by the present invention.

Claims

1. A rolling bearing life prediction method based on planar vibration characteristics and degradation history, characterized in that: It includes the following specific steps: 1) Construct a plane vibration signal dataset for the whole life cycle of a rolling bearing, and extract the plane vibration feature PVF and the degradation history feature DHF; 2) Construct and train a CNN-LSTM hybrid model, which is used to output the remaining useful life (RUL) prediction value after inputting the DHF sequence; 3) Optimize the model based on the weighted mean square error (wMSE) loss function to improve the prediction accuracy in the critical stage before failure, and output the optimal RUL prediction result.

2. The rolling bearing life prediction method based on plane vibration features and degradation history according to claim 1, wherein: The step 1) of constructing a plane vibration signal dataset for the whole life cycle of a rolling bearing is specifically as follows, including: Collect the vibration signals in the horizontal and vertical directions of the bearing, and the sampling rate is not less than 25.6 kHz; Denoise and normalize the original signal to obtain a standardized vibration dataset; Extract time-domain, frequency-domain and time-frequency-domain features to construct a multi-dimensional feature matrix.

3. The rolling bearing life prediction method based on plane vibration features and degradation history according to claim 1, wherein: The step 1) of extracting the plane vibration feature PVF is specifically as follows, including: Calculate the covariance matrix of the vibration signal , and obtain the principal direction angle through eigen decomposition and ellipticity ; Use a robust covariance estimation algorithm to remove outliers until the matrix determinant converges; Expand 2D statistical features, including planar RMS and Crest factor, and their calculation formulas are as follows: ； ； ; ； Among them, is the polar coordinate radian of the main vibration direction, are the major and minor axes of the fitted ellipse; N is the number of sampling points in the vibration signal, is the sample serial number, are the amplitude values of the sampling points in the x / y directions.

4. The rolling bearing life prediction method based on plane vibration features and degradation history according to claim 1, wherein: The step 1) of extracting the degradation history feature DHF is specifically as follows, including: Zero-pad the PVF sequence of the whole life cycle of the bearing to a fixed length; Arrange the features in chronological order to retain the cumulative damage law of the degradation process; Standardize the DHF sequence to eliminate the influence of dimension.

5. The rolling bearing life prediction method based on plane vibration features and degradation history according to claim 1, wherein: The CNN-LSTM hybrid model in the step 2) is specifically as follows: Input layer, receiving a 3000×4-dimensional DHF feature matrix, where 3000 is the time step and 4 is the feature dimension; Convolutional layer, including three one-dimensional convolutions, and its operation formula is: , where m is the vibration signal window, y is the feature after convolution, l represents the layer number, k is the convolution kernel size, w is the weight, and b is the bias, is the sample serial number, and τ is the current global time; Adopt the LeakyReLU activation function: ； Among them, α takes 0.01, x is the output of the neuron, is the output of the activation function; LSTM layer, adopting a bidirectional LSTM structure, and the calculation formula for each time step t is: ； Among them, , , and o respectively represent the forget gate, input gate, and output gate. t is the time ordinal number within the sequence, x is the output of the neuron, C is the cell state, is the candidate cell state, is the hidden state, and are respectively the weights and bias coefficients of the forget gate, input gate, output gate, and cell state. Set the number of hidden units to 128; is the sigmoid activation function, is the tangent activation function; The output layer uses the Sigmoid activation function to output the normalized remaining life percentage: , where z is the result of linear transformation, and e is the natural coefficient, is the predicted output value.

6. The rolling bearing life prediction method based on plane vibration features and degradation history according to claim 1, wherein: The loss function of the weighted mean square error wMSE in step 3) is defined as: ； where is the weighted index, is the current time step, is the total bearing life, are the true and predicted values of the remaining life, N is the number of samples, is the sample serial number.

7. The rolling bearing life prediction method based on planar vibration characteristics and degradation history according to claim 6, wherein: The optimization steps of optimizing the model based on the weighted mean square error (wMSE) loss function in the step 3) are specifically as follows: Adopt the adaptive particle swarm optimization (APSO) algorithm to adjust the CNN-LSTM hyperparameters; Use the RMSE of the validation set as the fitness function to search for the optimal network configuration; Prevent overfitting through the early stopping strategy and retain the best model weights.

8. The rolling bearing life prediction method based on planar vibration characteristics and degradation history according to claim 7, wherein: The velocity update formula of the adaptive particle swarm optimization (APSO) algorithm is: , where w is the inertia weight, , are learning factors, , are random numbers, is the velocity component of particle i in the d-th dimension at the (k + 1)-th iteration; c is the cognitive learning factor; is the individual historical best position component of particle i in the d-th dimension at the k-th iteration; is the current position component of particle i in the d-th dimension at the k-th iteration.

Images