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

Abstract

The proposed rolling bearing life prediction method based on planar vibration characteristics and degradation history includes: constructing a bearing lifecycle vibration signal dataset and extracting planar vibration characteristics, including principal direction angle and ellipticity; constructing degradation history features by zero-padding the feature sequence of the entire bearing degradation process to a fixed length; training a CNN-LSTM-based hybrid model that inputs the degradation history features and outputs a predicted remaining life value; and optimizing the model using a weighted mean square error loss function to improve prediction accuracy in the critical stage before failure. By integrating multi-directional vibration information with a complete degradation history, the proposed method achieves an RMSE of 0.154 on the PHM2012 dataset, a 47.3% improvement over traditional methods. This method is suitable for bearing health monitoring and maintenance decision-making under variable speed and load conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery health monitoring, and in particular 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. Composed of an inner ring, outer ring, rolling elements, and a cage, they reduce friction and support the weight of rotating machinery. High-performance bearings play a crucial role in major equipment and high-precision machinery. Therefore, improving the design and manufacturing capabilities of rolling bearings and refining their manufacturing processes have become pressing challenges for my country's bearing industry.

[0003] Rolling bearings, as core components of rotating machinery, are widely used in major equipment such as wind turbines, high-speed rail, and aircraft engines. Their reliability directly impacts the safe operation and service life of these equipment. However, under complex operating conditions such as high speed, variable load, and strong impact, bearing fatigue failure is one of the main causes of mechanical system failure. According to statistics, approximately 40% of rotating machinery failures are related to bearing failure, and sudden bearing failures can trigger a chain reaction, causing significant economic losses and even safety accidents. Currently, the lack of fault warning capabilities under non-stationary operating conditions such as variable speed and load has hindered the independent and controllable development of high-end equipment. Traditional bearing remaining life (RUL) prediction methods mainly rely on the time-frequency domain feature analysis of vibration signals, but have the following technical bottlenecks: unidirectional information limitation, existing methods mostly use vibration signals in a single direction (such as vertical acceleration), ignoring the correlation between load direction and fault evolution in plane vibration, resulting in insufficient extraction of early weak fault features; degradation modeling is one-sided, local feature extraction based on sliding windows 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 conditions; physical mechanism is disconnected, and the data-driven model lacks the embedding of bearing fault mechanisms (such as the modulation relationship between the order of inner ring fault features and speed), and the generalization performance is significantly reduced under unknown working conditions.

[0004] The present application is compared with the prior art as follows:

[0005] Technical comparison with application CN202010178652 “Rolling bearing fault diagnosis method based on computational order tracking and spectral kurtosis”;

[0006] 1. Patent application CN202010178652 proposes a rolling bearing fault diagnosis method for variable speed working conditions. Its core technology is to eliminate the influence of speed fluctuations through computational order tracking (COT). This method first uses the speed pulse signal to construct an angle-time quadratic equation (based on the uniform acceleration assumption), and uses cubic spline interpolation to achieve angular domain resampling, converting the non-stationary time domain signal into an angular domain stationary signal; combined with the spectral kurtosis algorithm, it dynamically locates the optimal demodulation frequency band (center frequency) of the impact fault. and bandwidth ) uses an elliptical filter to extract characteristic frequency bands, then performs envelope analysis. The resulting output is an order spectrum (e.g., 18 times the characteristic order of an inner race fault) to pinpoint the fault type under variable speed conditions. This solution's core value lies in resolving the "frequency ambiguity" issue inherent in traditional spectrum analysis. While suitable for dynamic operating conditions such as speed increases and decreases, it can only provide a qualitative assessment of the presence of a fault and cannot quantify remaining life or degradation trends.

[0007] 2. This application (rolling bearing life prediction method based on plane vibration characteristics and degradation history) has essential differences in technical objectives and architecture from CN202010178652: This patent focuses on quantitative prediction of remaining life (RUL) rather than qualitative diagnosis of faults, and belongs to a higher level of equipment health management; it innovatively proposes a dual-track system of plane vibration characteristics (PVF) and degradation history characteristics (DHF) - PVF extracts the main direction angle ( ) and ellipticity( ), integrating horizontal / vertical bidirectional vibration physical information to associate the fault mechanism, and DHF padding the PVF sequence of the entire life cycle to a fixed length (such as 3000×4) to retain the cumulative damage evolution law; using the CNN-LSTM hybrid model to learn the mapping relationship from DHF to RUL, through the weighted loss function ( ) Enhance the prediction accuracy of the key stage before failure, 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, It is the sample number, and finally outputs the high-precision remaining life percentage (such as RMSE0.154 for the PHM2012 dataset), while the comparative patent only outputs the fault type label.

[0008] There are essential differences between the two in terms of scheduling objectives, system structure and technical routes.

[0009] Technical comparison with application CN202311277320 “A rolling bearing fault diagnosis method”;

[0010] 1. Patent application CN202311277320 proposes a wavelet-EMD joint noise reduction framework for early fault detection in high-noise stable working conditions. Its technical path is divided into two layers: adaptive noise reduction and anti-aliasing decomposition. In the noise reduction layer, wavelet packet decomposition and cross-correlation coefficient filtering (threshold ) Eliminate low-correlation nodes and dynamically optimize the global threshold based on the principle of minimizing sample entropy , and a new threshold function is designed to preserve weak impact characteristics. At the decomposition layer, Empirical Mode Decomposition with Complementary Ensemble High-Frequency Harmonics (CEHFHA-EMD) is used to inject high-frequency harmonics to suppress modal aliasing. After decomposing precise IMF components, the Hilbert marginal spectrum and envelope spectrum (such as the 163Hz inner race fault fundamental frequency and its multiples) are extracted. This solution achieves highly sensitive early fault detection in the presence of strong noise, but is limited to current state judgment and cannot predict degradation trends or quantify remaining life.

[0011] 2. This application and CN202311277320 have made significant breakthroughs in terms of technical dimensions and application depth: in terms of signal dimension, the comparative patent relies on single-point vibration signal analysis, while this patent pioneers 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, while this patent constructs a degradation history feature (DHF) covering the degradation trajectory of the entire life cycle (zero-filled fixed sequence), combines the CNN-LSTM model to capture long-term dependencies, and realizes 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 optimizes hyperparameters through adaptive particle swarm optimization (APSO), while the comparative patent does not involve the prediction accuracy optimization mechanism.

[0012] There are essential differences between the two in terms of system architecture, scheduling goals and technical paths. Summary of the Invention

[0013] In response to the problems of existing rolling bearing remaining life prediction methods under variable speed and variable load conditions, such as insufficient prediction accuracy, low sensitivity to early failures, and poor physical interpretability, 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 robust bearing health status assessment and remaining life prediction by integrating multi-directional vibration information with full-cycle degradation laws.

[0014] To achieve the above object, the technical solution adopted by the present invention is:

[0015] The rolling bearing life prediction method based on plane vibration characteristics and degradation history includes the following specific steps:

[0016] 1) Construct a planar vibration signal dataset for the entire life cycle of rolling bearings and extract the planar vibration feature PVF and degradation history feature DHF;

[0017] 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 the DHF sequence;

[0018] 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.

[0019] As a further improvement of the present invention, the step 1) constructs a planar vibration signal dataset for the entire life cycle of a rolling bearing, specifically comprising the following steps:

[0020] Collect the horizontal and vertical vibration signals of the bearing, with a sampling rate of not less than 25.6 kHz;

[0021] Denoising and normalizing the original signal to obtain a standardized vibration data set;

[0022] Extract time domain, frequency domain and time-frequency domain features and construct a multi-dimensional feature matrix.

[0023] As a further improvement of the present invention, the step 1) extracting the planar vibration feature PVF comprises the following specific steps:

[0024] 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, whose calculation formula is:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] 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.

[0030] As a further improvement of the present invention, the step 1) extracting the degradation history feature DHF comprises the following specific steps:

[0031] Zero-fill the PVF sequence of the bearing's entire life cycle to a fixed length;

[0032] Arrange features in chronological order to preserve the cumulative damage pattern of the degradation process;

[0033] The DHF sequence is standardized to eliminate the dimension effect.

[0034] As a further improvement of the present invention, the CNN-LSTM hybrid model in step 2) is specifically as follows:

[0035] The input layer receives a 3000×4 dimensional DHF feature matrix, where 3000 is the time step and 4 is the feature dimension;

[0036] The convolution layer contains three layers of one-dimensional convolution, and its calculation formula is: , where m is the vibration signal window, y is the feature after convolution, l represents the number of layers, k is the convolution kernel size, w is the weight, and b is the bias. is the sample number, τ is the current global time;

[0037] Using LeakyReLU activation function:

[0038] ;

[0039] Among them, α is 0.01, x is the output of the neuron, is the activation function output;

[0040] The LSTM layer uses a bidirectional LSTM structure, and the calculation formula for each time step t is:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] in, 、 , o represent the forget gate, input gate and output gate respectively, t is the time sequence number, x is the output of the neuron, C is the cell state, is a candidate cell state, is hidden state, and They are the weights and bias coefficients of the forget gate, input gate, output gate, and cell state, and the number of hidden units is set to 128; is the sigmoid activation function, is the tangent activation function.

[0048] The output layer uses the Sigmoid activation function to output the normalized remaining lifespan percentage: , where z is the result of linear transformation, e is the natural coefficient, is the predicted output value.

[0049] As a further improvement of the present invention, the loss function of the weighted mean square error wMSE in step 3) is Defined as:

[0050] ;

[0051] in 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.

[0052] As a further improvement of the present invention, the optimization step of the weighted mean square error wMSE loss function optimization model in step 3) is specifically as follows:

[0053] Adaptive particle swarm algorithm APSO is used to adjust CNN-LSTM hyperparameters;

[0054] Use the validation set RMSE as the fitness function to search for the optimal network configuration;

[0055] The early stopping strategy is used to prevent overfitting and retain the best model weights.

[0056] As a further improvement of the present invention, the speed update formula of the adaptive particle swarm algorithm APSO is:

[0057] , where w is the inertia weight, 、 is the learning factor, 、 is a random number, is the velocity component of particle i in the dth dimension at the k+1th iteration; c is the cognitive learning factor; is the individual historical optimal position component of particle i in the dth dimension at the kth iteration; is the current position component of particle i in the dth dimension at the kth iteration.

[0058] Beneficial effects;

[0059] (1) The proposed remaining life prediction method based on planar vibration features (PVF) and degradation history modeling (DHF) significantly improves the prediction accuracy under variable speed and load conditions by integrating multi-directional vibration information with full-cycle degradation laws. Experiments show that on the PHM2012 standard dataset, the RMSE of this method reaches 0.154, which is 47.3% higher than the traditional single-directional vibration analysis method, and the detection sensitivity of early faults is increased by more than 35%;

[0060] (2) This invention innovatively introduces the plane vibration ellipse feature (main direction angle and ellipticity ) quantifies the load direction and vibration energy distribution through eigendecomposition of the covariance matrix, addressing the loss of physical information in traditional single-direction analysis. This method can directly correlate bearing failure mechanisms (such as the characteristic order of inner race failures), significantly improving the model's interpretability.

[0061] (3) The CNN-LSTM hybrid model designed in this paper combines the spatial feature extraction capability of convolutional neural networks and the temporal modeling advantages of long short-term memory networks. It strengthens the prediction weight of the critical stage before failure through the weighted loss function (wMSE), effectively avoiding the fluctuation problem of prediction results of traditional methods.

[0062] (4) The degradation history feature (DHF) construction method proposed in this paper retains the degradation trajectory of the entire life cycle by zero-filling fixed-length sequences. Compared with the sliding window method, this method is universal and can be extended to other life prediction scenarios of rotating machinery such as gearboxes and motors;

[0063] (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. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of the method of the present invention;

[0065] Figure 2 This is the improved CNN-LSTM model architecture diagram of the present invention;

[0066] Figure 3 The following is a comparison chart of the effects of the present invention under three different load conditions: low, medium and high;

[0067] Figure 3 (a) is a comparison diagram of the effects under low load conditions;

[0068] Figure 3 (b) is a comparison diagram of the effects under medium load conditions;

[0069] Figure 3 (c) is a comparison diagram of the effects under high load conditions. DETAILED DESCRIPTION

[0070] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0071] This invention discloses a rolling bearing life prediction method based on planar vibration characteristics and degradation history. By integrating multi-directional vibration signal characteristics with deep learning technology, this method accurately predicts the remaining service life of rolling bearings. The following describes specific embodiments of the invention in detail with reference to the accompanying drawings.

[0072] As attached Figure 1 As shown, the rolling bearing life prediction method based on plane vibration characteristics and degradation history of the present invention includes the following steps:

[0073] S1. System hardware configuration and data collection, including the following steps:

[0074] The implementation of the present invention first requires the construction of a complete data acquisition system. The system uses the 352C33 ICP acceleration sensor produced by the American PCB company as the core acquisition device. The sensor has a sensitivity of 100mV / g and a frequency response range of 0.5Hz-10kHz, which can accurately capture the vibration signal of the bearing during operation. In terms of installation layout, it is necessary to install a sensor in the horizontal and vertical directions of the bearing seat. The installation position is no more than 50mm away from the outer ring of the bearing, and a magnetic base is used to ensure close contact between the sensor and the bearing seat. Data acquisition uses the NI cDAQ-9178 chassis with the NI 9234 acquisition module, and the sampling frequency is set to 25.6kHz to meet the requirements of the Nyquist sampling theorem. Each time, 10 seconds of vibration signals are continuously collected. To ensure data quality, the acquisition system has a built-in anti-aliasing filter with a cutoff frequency of 10kHz to effectively prevent high-frequency noise interference.

[0075] S2, signal preprocessing and feature extraction;

[0076] The collected raw vibration signal needs to go through a rigorous preprocessing process. First, wavelet threshold denoising is performed, using the db5 wavelet basis function for a 5-layer decomposition. The denoising process can be expressed as follows:

[0077]

[0078] in, represents the wavelet coefficients, is the threshold value, and its value is , is the noise standard deviation, and N is the signal length. After denoising, high-frequency components above 2kHz are retained, which is crucial for capturing early bearing fault characteristics. For outliers in the signal, the Mahalanobis distance is calculated to identify them: .in, is the mean vector, x is the signal strength, is the covariance matrix. A threshold of 3σ is set, and data segments exceeding this threshold are removed. Missing segments are filled using linear interpolation. Finally, MinMaxScaler is used to normalize all features to the range [0, 1].

[0079] S3, plane vibration feature extraction steps:

[0080] The extraction of plane vibration features is one of the core innovations of this invention. First, the covariance matrix of the horizontal and vertical vibration signals is calculated:

[0081] ;

[0082] in, Represents the covariance between two signals, and performs eigendecomposition on the covariance matrix: , where V is the eigenvector matrix and D is the eigenvalue diagonal matrix. This allows us to extract two key features: principal direction angle and ellipticity. To improve the robustness of feature extraction, the MCD (Minimum Covariance Determinant) algorithm is used for iterative calculations until the matrix determinant changes by less than 1%. The specific implementation steps for this iterative calculation are as follows:

[0083] The algorithm first performs an initialization operation: the optimal covariance matrix C is optimally set to the two-dimensional identity matrix, the historical optimal determinant value is initialized to positive infinity, and an initial data subset S is randomly selected as the starting point for the calculation. The algorithm controls the iteration process by continuously monitoring the determinant improvement σ, and terminates the loop when σ is less than a preset threshold ε.

[0084] Each iteration consists of four core steps:

[0085] Statistics calculation: Calculate the covariance matrix CS and its determinant value based on the current subset S.

[0086] Optimal Solution Update: If the current determinant is less than the historical optimal value, the optimal covariance matrix C is updated using CS. The difference between the old and new determinants is calculated and assigned to σ, and the new determinant is recorded as the optimal value. This design ensures that the algorithm continues to approach a more compact data distribution.

[0087] Outlier identification: The Mahalanobis distance is calculated for all data points. This distance is constructed by the current subset mean μS and the inverse covariance matrix CS−1 to quantify the degree of deviation of each point from the subset.

[0088] Dynamic subset screening: An adaptive threshold θ is set based on the standard deviation, and only points with a Mahalanobis distance lower than θ are retained to form a new subset S. This mechanism effectively excludes outliers with large statistical deviations.

[0089] When the iteration terminates, the algorithm outputs the final optimized covariance matrix C. Its robustness comes from the double closed-loop design: the covariance matrix optimization is driven by the determinant minimization objective, and the Mahalanobis distance threshold is used to achieve dynamic purification of data subsets, making the results insensitive to outliers.

[0090] S31. Data preparation: Prepare a dataset X containing n two-dimensional data points, ensuring the data format is correct. Initialize the parameters required for the calculation, including setting the initial covariance matrix to the second-order identity matrix, the base determinant value to positive infinity, and randomly select an initial data subset S.

[0091] S32, enter the main loop, calculate the covariance matrix of the current subset S;

[0092] S33, calculating the determinant value of the covariance matrix;

[0093] S34. Compare the current determinant value with the historical optimal value, and if it is better, update the optimal estimate and the benchmark value;

[0094] S35, calculating the Mahalanobis distance of all data points relative to the current subset;

[0095] S36. Filter data points according to a preset threshold and update the subset S;

[0096] S37. Determine whether the termination condition is met. If so, terminate the process. If not, return to S32 and regenerate the offspring. The termination condition is that the calculation process is terminated when the difference in the determinant of two consecutive iterations is less than a preset convergence threshold. A maximum number of iterations is also set as a safety limit to prevent infinite loops. In some embodiments of the present invention, the maximum number of iterations is set to 100. The final calculated robust covariance matrix is ​​returned. This matrix is ​​robust to outliers and can accurately reflect the distribution characteristics of the data.

[0097] S4. Degradation history feature construction method:

[0098] The construction of degradation history features adopts an innovative zero-filled fixed-length sequence method. Assuming that m groups of PVF features are collected throughout the entire life cycle of the bearing, and each group of features contains n parameters, the constructed DHF matrix at time t is:

[0099]

[0100] in, is the eigenvector, is the maximum lifetime constant, c is the number of channels, is a real number space, and when m<3000, it is filled with zero vectors. This processing method not only preserves the complete trajectory of the degradation process, but also solves the problem of local information fragmentation caused by the traditional sliding window method.

[0101] S5. CNN-LSTM hybrid model construction:

[0102] The CNN-LSTM hybrid neural network architecture designed by the present invention is as follows Figure 2 As shown, it specifically includes the following levels:

[0103] S51, input layer, receives the 3000×4-dimensional DHF feature matrix, where 3000 is the time step and 4 is the feature dimension;

[0104] S52, convolution layer, contains 3 layers of one-dimensional convolution, and its calculation formula is: , where m is the vibration signal window, y is the convolution feature, l represents the number of layers, 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 in the three layers of convolution is set to 16, 32, and 64 respectively, and the LeakyReLU activation function is used:

[0105] ;

[0106] Among them, α is 0.01, x is the output of the neuron, and f(x) is the output of the activation function;

[0107] S53, LSTM layer, uses a bidirectional LSTM structure. The calculation formula for each time step t is:

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] Among them, f, i, and o represent the forget gate, input gate, and output gate respectively, t is the time sequence number, x is the output of the neuron, and C is the cell state. is the candidate cell state, h is the hidden state, and They are 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;

[0115] S54, the output layer uses the Sigmoid activation function to output the normalized remaining lifespan percentage: , where z is the result of linear transformation, e is the natural coefficient, is the predicted output value.

[0116] S6. Model training and optimization steps:

[0117] The model training uses the weighted mean square error loss function: ,in 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 number. The Adam optimizer was used for optimization, with an initial learning rate of 0.001 and a 10% decay every 20 epochs. The adaptive particle swarm optimization (APSO) algorithm was used to optimize hyperparameters. The population size was set to 50, and the maximum number of iterations was 100. The comparison of the results under low-load conditions is shown in the figure below. Figure 3 (a) in the figure shows the effect comparison under medium load conditions. Figure 3 (b) in the figure shows the effect comparison under high load conditions. Figure 3 (c) in the.

[0118] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A rolling bearing life prediction method based on plane vibration characteristics and degradation history, characterized by: The specific steps include: 1) Construct a planar vibration signal dataset for the entire life cycle of rolling bearings and extract the planar vibration feature PVF and degradation history feature DHF; The step 1) extracting the planar vibration feature PVF comprises the following specific steps: 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, whose 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, is the amplitude value of the sampling point in the x / y direction; 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 the DHF sequence; The CNN-LSTM hybrid model in step 2) is 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; The convolution layer contains three layers of one-dimensional convolution, and its calculation formula is: , where m is the vibration signal window, y is the feature after convolution, l represents the number of layers, k is the convolution kernel size, w is the weight, and b is the bias. is the sample number, τ is the current global time; Using LeakyReLU activation function: ； Among them, α is 0.01, x is the output of the neuron, is the activation function output; The LSTM layer uses a bidirectional LSTM structure, and the calculation formula for each time step t is: ； in, 、 , o represent the forget gate, input gate and output gate respectively, t is the time sequence number, x is the output of the neuron, C is the cell state, is a candidate cell state, is hidden state, and They are the weights and bias coefficients of the forget gate, input gate, output gate, and cell state, and the number of hidden units is set 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 lifespan percentage: , where z is the result of linear transformation, e is the natural coefficient, is the predicted output value; 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.

2. The rolling bearing life prediction method based on plane vibration characteristics and degradation history according to claim 1 is characterized in that: The step 1) constructs a planar vibration signal dataset for the entire life cycle of a rolling bearing. The specific steps are as follows: Collect the horizontal and vertical vibration signals of the bearing, with a sampling rate of not less than 25.6 kHz; Denoising 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.

3. The method for predicting the life of a rolling bearing based on plane vibration characteristics and degradation history according to claim 1, characterized in that: The step 1) extracting the degradation history feature DHF includes the following specific steps: 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.

4. The method for predicting the life of a rolling bearing based on plane vibration characteristics and degradation history according to claim 1, wherein: The loss function of weighted mean square error wMSE in step 3) Defined as: ； in 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.

5. The method for predicting the life of a rolling bearing based on plane vibration characteristics and degradation history according to claim 4, characterized in that: The optimization steps of the weighted mean square error (WMSE) loss function optimization model in step 3) are as follows: Adaptive particle swarm algorithm APSO is used to adjust CNN-LSTM hyperparameters; Use the validation set RMSE as the fitness function to search for the optimal network configuration; The early stopping strategy is used to prevent overfitting and retain the best model weights.

6. The method for predicting the life of a rolling bearing based on plane vibration characteristics and degradation history according to claim 5, characterized in that: The speed update formula of the adaptive particle swarm algorithm APSO is: , where w is the inertia weight, 、 is the learning factor, 、 is a random number, is the velocity component of particle i in the dth dimension at the k+1th iteration; c is the cognitive learning factor; is the individual historical optimal position component of particle i in the dth dimension at the kth iteration; is the current position component of particle i in the dth dimension at the kth iteration.

Images