A method for predicting the service life of a gradient composite coating sliding bearing
By constructing a multi-field coupled physical simulation model of coating degradation and an active learning algorithm, combined with real-time multi-source signal feature extraction and transfer learning framework, the accuracy and efficiency problems of gradient composite coating sliding bearing life prediction were solved, achieving high-precision and low-cost bearing life prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING WANGJIANG IND
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-05
Smart Images
Figure CN122154351A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of bearing life prediction, specifically a method for predicting the life of a gradient composite coated sliding bearing. Background Technology
[0002] As a core support component of rotating machinery, sliding bearings are widely used in wind power, thermal power, aerospace, shipbuilding, and high-end equipment. However, sliding bearings endure heavy loads and high-speed friction over extended periods, leading to wear between their contact surfaces. Excessive wear can cause bearing failure, potentially resulting in equipment downtime, component damage, and even safety accidents. Therefore, predicting bearing life allows for real-time monitoring of bearing degradation, early assessment of potential failures, and timely scheduling of maintenance and replacement cycles. This effectively ensures stable equipment operation, reducing maintenance costs and unplanned downtime losses.
[0003] Currently, bearing life prediction methods are mainly divided into two categories: one is based on physical mechanism models (such as the Archard wear model), which uses fundamental theories such as material friction and wear, and contact mechanics as the core, combining load, friction characteristics, and material parameters to calculate the degree of wear through classical wear models, and then estimate the bearing life. The other is based on data-driven models (such as machine learning methods), which rely on external macroscopic monitoring signals such as vibration and temperature to perform simple data statistics and trend analysis, and then predict the bearing life. However, both of these methods have the following technical problems: Models based on physical mechanisms require precise operating parameters and material properties. To enhance the wear resistance of bearings, industrial applications often involve preparing gradient composite coatings (such as multi-layer structures consisting of hard ceramic layers, transition layers, and self-lubricating layers) on the bearing surface. The wear process of such coatings is complex, involving layer-by-layer failure of the coating materials, making its life prediction more difficult than that of traditional homogeneous bearings. Using this model for bearing life prediction requires a large number of operating parameters and material properties, the methods are cumbersome and inefficient, and this method is more suitable for homogeneous bearings, making it difficult to simulate the nonlinear degradation process under the gradient coating structure.
[0004] Data-driven models require extensive training with historical operating data, wear and failure samples, and necessitate the collection of sufficient early degradation and failure data for model building. Insufficient samples and missing early failure data can lead to severely inaccurate predictions, particularly for novel gradient-coated bearings where a lack of adequate historical failure data is often unavailable for training purely data-driven models. Furthermore, these models only derive lifespan results from data inputs and outputs, failing to reflect the true physical failure mechanisms such as bearing friction, coating wear, and interlayer damage. They cannot clearly explain the intrinsic causal relationship between data changes and bearing aging or coating damage; they can only predict outcomes, not analyze the causes of failure.
[0005] Therefore, existing bearing life prediction methods are difficult to accurately and efficiently predict the life of gradient composite coated sliding bearings, and a new bearing life prediction method is urgently needed to solve the above problems. Summary of the Invention
[0006] This invention provides a method for predicting the life of gradient composite coated sliding bearings, which can solve the problems of poor accuracy and low efficiency of existing bearing life prediction methods when predicting the life of gradient composite coated sliding bearings.
[0007] This application provides the following technical solution: a method for predicting the life of a gradient composite coated sliding bearing, comprising the following steps: Step 1: Construct a multi-field coupled physical simulation model of coating degradation; Based on the material properties, layered structure and interface characteristics of gradient composite coatings, establish a multi-physics coupled finite element model that considers hydrodynamic lubrication, contact mechanics and material wear. Step 2: Construct a high-precision proxy model using an active learning algorithm; based on the physical simulation model from Step 1, design an active learning query strategy; Step 3: Online monitoring and multi-domain feature extraction; During actual bearing operation, signals from multiple sensors such as vibration, acoustic emission, and temperature are collected in real time; After preprocessing the signals, multi-domain features that can reflect the coating degradation state are extracted. Step 4: Real-time state mapping and lifetime prediction based on transfer learning; Based on the real-time multi-domain features in Step 3, a transfer learning framework of "teacher model" is established. The online monitoring data is processed using the domain adaptation loss function, and the processed data is input into a lightweight "student prediction network" model to output the estimated value of the cumulative wear of the front coating. Then, the wear data is fitted into a degradation trend curve to infer the remaining service life of the bearing.
[0008] Furthermore, the model established in step one uses bearing operating conditions (load, speed, lubrication conditions) and coating geometric tolerances as random input variables, and outputs a degradation curve of coating thickness over operating time. The hydrodynamic lubrication model is established by using the Reynolds equation to describe the oil film pressure distribution in the sliding bearing, and its simplified form is as follows:
[0009] in, For oil film pressure, Oil film thickness, The viscosity of the lubricating oil. For density, , The tangential velocity of the contact surface; The contact mechanics model is established as follows: for elastic contact between coatings or substrates, Hertzian contact theory is used to calculate contact stress and contact half-width; for contact between two spheres, the maximum contact stress is calculated. and contact radius for:
[0010] in, For normal load, The equivalent radius of curvature, It is the equivalent elastic modulus; The material wear model is established as follows: the reduction in coating thickness is calculated using the Archard wear model; the basic formula describes the wear volume. The relationship between load and sliding distance is as follows:
[0011] in, The wear coefficient is a dimensionless coefficient. For normal load, The sliding distance, The hardness of a relatively soft material is used; however, in finite element analysis, its differential or discrete form is often used to calculate the wear depth at nodes. :
[0012] in, The wear coefficient (unit: 1 / Pa). To contact pressure, This represents the sliding distance.
[0013] Furthermore, in step two, a small number of sample points are initialized within the design space of the input variables for simulation to obtain an initial training set; then the following process is iteratively executed: a Gaussian process regression surrogate model is trained using the current training set; the most valuable new sample points are selected based on the uncertainty measure of the surrogate model for high-fidelity physical simulation, and the new data is added to the training set. This process continues until the prediction accuracy of the surrogate model meets the preset threshold. Gaussian process regression: treating the simulation model as a black box function y=f(x)+ The Gaussian process assumes that the function value follows a multivariate Gaussian distribution, and its predicted distribution is: Posterior mean (predicted value):
[0014] Posterior variance (uncertainty):
[0015] in, It is the kernel function matrix between training samples. It is the kernel function vector between the test points and the training samples. It is the noise variance; Active learning query strategy: Selecting new sample points using expected improvement criteria:
[0016] in, This is the current best observation value. and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. , To balance the parameters of exploration and development.
[0017] Furthermore, in step three, the multi-domain features include time-domain features, frequency-domain features, and nonlinear features; The time-domain features include root mean square value, peak value, and kurtosis. Root mean square value:
[0018] Peak value:
[0019] kurtosis:
[0020] The frequency domain features are obtained by performing a Fourier transform on the signal, analyzing the changes in the amplitude of the characteristic frequencies, and reducing the dimensionality through principal component analysis.
[0021] Extract the amplitude or energy within a specific frequency band. Then, use principal component analysis for dimensionality reduction, the core of which is solving the covariance matrix. eigenvalues and eigenvectors; The nonlinear feature is achieved by using a dynamic time warping algorithm to calculate the nonlinear distance between signal waveforms in order to capture subtle degradation patterns.
[0022] in, It is a regular path. It is the distance between points.
[0023] Furthermore, in step four, the real-time multi-domain feature sequence extracted in step three is used as input to construct a transfer learning framework with the surrogate model obtained in step two as the "teacher model." By introducing a domain adaptation loss function, the feature distribution of the online monitoring data is aligned to the feature distribution generated from the simulation data. Then, the aligned features of the online monitoring data are input into a lightweight "student prediction network," which outputs an estimate of the cumulative wear of the pre-coating layer. The predicted wear sequence is compared with the coating failure threshold, and the remaining service life of the bearing is calculated by extrapolating using the fitted degradation trend curve. The transfer learning framework of the "teacher model" is as follows:
[0024] in, It is a feature map that maps data to the reproducing kernel Hilbert space H; The lightweight "student prediction network" employs a Relevance Vector Machine (RVM) model. RVM is a sparse Bayesian model where, for a given input x, its predicted output y follows a Gaussian distribution.
[0025] Among them, weight It is given a sparse prior, that is, the prior is determined by automatic correlation. The hyperparameters are learned by maximizing the marginal likelihood, and finally a sparse weight vector is obtained for prediction. Remaining service life extrapolation: extrapolating the predicted cumulative wear sequence With failure threshold Comparison; commonly used degradation trend fitting models include: Exponential model:
[0026] Linear model:
[0027] Parameters are obtained by fitting historical data. ， Then solve the equation. ， Determine the remaining useful life .
[0028] Furthermore, it also includes step five, prediction result output and model update, outputting the predicted value of the remaining lifespan and its confidence interval. The system can continuously update the feature library based on new monitoring data, and periodically use newly accumulated data to fine-tune the surrogate model and prediction network to achieve online evolution of the prediction model.
[0029] The advantages of this invention are: 1. By understanding the degradation mechanism of the coating, the prediction accuracy is higher.
[0030] This invention establishes a multi-field coupled physical simulation model of coating degradation in step one, combining lubrication characteristics, contact mechanics, and coating wear mechanisms. The aim is to recreate the real failure law of the multi-layer structure, interfacial interaction, and nonlinear layer-by-layer degradation of gradient composite coatings. This overcomes the shortcomings of traditional single physical models, which cannot adapt to the gradient structure of coatings and provide a one-sided characterization of failure mechanisms. It provides reliable physical theoretical support for subsequent life determination. At the same time, it relies on simulation to obtain complete coating degradation evolution laws, laying the foundation for failure threshold delineation and model training. Compared with conventional prediction methods, it effectively improves the prediction accuracy of bearing life.
[0031] 2. Reduce computational costs and improve computational efficiency.
[0032] By introducing an active learning algorithm in step two to construct a high-precision surrogate model, a large number of repetitive simulation calculations are replaced with a small number of high-value simulation samples. While ensuring the accuracy of the model's prediction, the computational cost and time consumption of multiphysics simulation are greatly reduced. This solves the problem of large computational load and low efficiency of conventional prediction models and helps to improve the efficiency of the model in predicting bearing life.
[0033] 3. Acquiring multi-source signals makes the coating degradation characterization state more complete and detailed, which helps to improve the accuracy of bearing life prediction.
[0034] Step three involves real-time acquisition of multi-source monitoring signals such as bearing vibration, acoustic emission, and temperature, reflecting multi-domain characteristics of coating degradation. The aim is to capture subtle deterioration states such as coating wear and interlayer delamination, obtaining richer features and avoiding the limitations of insufficient characterization of single signal features and weak early degradation identification capabilities. This provides comprehensive data support for subsequent state mapping and wear quantification prediction, further improving the accuracy of bearing life prediction.
[0035] 4. Fewer real failure data samples are required, reducing the computational load on the model and ensuring prediction accuracy.
[0036] Step four establishes a transfer learning framework, using a proxy model as the teacher model. The domain adaptation loss function is used to align the feature distributions of the measured monitoring data and the simulation data. The significance of this alignment lies in reducing the domain deviation between the simulation conditions and the actual operating conditions. Then, a lightweight student prediction network is used to quickly output the cumulative wear of the coating. The remaining life is calculated by extrapolating the preset failure threshold and degradation trend. This avoids the difficulty of traditional physical models in simulating the nonlinear degradation of the coating and effectively improves the prediction accuracy.
[0037] 5. The model iterates dynamically, resulting in more stable long-term prediction accuracy.
[0038] Step five outputs the remaining service life prediction results and corresponding confidence intervals, objectively characterizing the uncertainty and reliability range of the prediction results, making the service life assessment results more rigorous and reliable. At the same time, the surrogate model and lightweight prediction network are dynamically fine-tuned and optimized to realize the online iterative evolution of the model, effectively improving the problem of decreased adaptability to operating conditions and reduced prediction accuracy after long-term use of fixed models. It can adapt to fluctuations in actual operating conditions and long-term aging changes in coatings, ensuring the stability and long-term effectiveness of service life prediction throughout the entire life cycle of the bearing. Attached Figure Description
[0039] Figure 1 This is an overall flowchart of the method of the present invention.
[0040] Figure 2 This is a flowchart of a first embodiment of the method of the present invention. Detailed Implementation
[0041] The following detailed description illustrates the specific implementation method: Example 1 A method for predicting the life of a gradient composite coated sliding bearing includes the following steps: Step 1: Construct a physical simulation model of coating degradation with multi-field coupling.
[0042] Based on the material properties, layered structure and interface characteristics of gradient composite coatings, a multi-physics coupled finite element model considering hydrodynamic lubrication, contact mechanics and material wear is established. The established model uses bearing operating conditions (load, speed, lubrication conditions) and coating geometric tolerances as random input variables, and outputs the degradation curve of coating thickness over operating time. The hydrodynamic lubrication model is established by using the Reynolds equation to describe the oil film pressure distribution in the sliding bearing, and its simplified form is as follows:
[0043] in, For oil film pressure, Oil film thickness, The viscosity of the lubricating oil. For density, , The tangential velocity of the contact surface; The contact mechanics model is established as follows: for elastic contact between coatings or substrates, Hertzian contact theory is used to calculate contact stress and contact half-width; for contact between two spheres, the maximum contact stress is calculated. and contact radius for:
[0044] in, For normal load, The equivalent radius of curvature, It is the equivalent elastic modulus; The material wear model is established as follows: the reduction in coating thickness is calculated using the Archard wear model; the basic formula describes the wear volume. The relationship between load and sliding distance is as follows:
[0045] in, The wear coefficient is a dimensionless coefficient. For normal load, The sliding distance, The hardness of a relatively soft material is used; however, in finite element analysis, its differential or discrete form is often used to calculate the wear depth at nodes. :
[0046] in, The wear coefficient (unit: 1 / Pa). To contact pressure, This represents the sliding distance.
[0047] Step 2: Construct a high-precision proxy model using an active learning algorithm.
[0048] Based on the physical simulation model in step one, an active learning query strategy is designed; a small number of sample points are initialized in the design space of the input variables for simulation to obtain an initial training set; then the following process is executed iteratively: a Gaussian process regression surrogate model is trained using the current training set; the most valuable new sample points are selected according to the uncertainty measure of the surrogate model for high-fidelity physical simulation, and the new data is added to the training set. This process continues until the prediction accuracy of the surrogate model meets the preset threshold. Gaussian process regression: treating the simulation model as a black box function y=f(x)+ The Gaussian process assumes that the function value follows a multivariate Gaussian distribution, and its predicted distribution is: Posterior mean (predicted value):
[0049] Posterior variance (uncertainty):
[0050] in, It is the kernel function matrix between training samples. It is the kernel function vector between the test points and the training samples. It is the noise variance; Active learning query strategy: Selecting new sample points using expected improvement criteria:
[0051] in, This is the current best observation value. and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. , To balance the parameters of exploration and development.
[0052] Step 3: Online Monitoring and Multi-Domain Feature Extraction; During actual bearing operation, signals from multiple sensors, including vibration, acoustic emission, and temperature, are collected in real time; after preprocessing the signals, multi-domain features reflecting the coating degradation state are extracted; these multi-domain features include time-domain features, frequency-domain features, and nonlinear features; among which, time-domain features include root mean square value, peak value, and kurtosis. Root mean square value:
[0053] Peak value:
[0054] kurtosis:
[0055] Frequency domain characteristics: Perform Fourier transform on the signal, analyze the changes in characteristic frequency amplitudes, and perform dimensionality reduction through principal component analysis;
[0056] Extract the amplitude or energy within a specific frequency band. Then, use principal component analysis for dimensionality reduction, the core of which is solving the covariance matrix. eigenvalues and eigenvectors; Nonlinear characteristics: A dynamic time warping algorithm is used to calculate the nonlinear distance between signal waveforms in order to capture subtle degradation patterns;
[0057] in, It is a regular path. It is the distance between points.
[0058] Step 4: Real-time state mapping and lifetime prediction based on transfer learning; Based on the real-time multi-domain features in Step 3, a transfer learning framework of "teacher model" is established. The online monitoring data is processed using the domain adaptation loss function, and the processed data is input into a lightweight "student prediction network" model to output the estimated value of the cumulative wear of the front coating. Then, the wear data is fitted into a degradation trend curve to infer the remaining service life of the bearing.
[0059] Specifically, the real-time multi-domain feature sequence extracted in step three is used as input to construct a transfer learning framework with the surrogate model obtained in step two as the "teacher model." A domain adaptation loss function is introduced; in this example, MMD (Maximum Mean Difference) is chosen to align the feature distribution of the online monitoring data with the feature distribution generated from the simulation data. The aligned features of the online monitoring data are then input into a lightweight "student prediction network," which outputs an estimate of the cumulative wear of the pre-coating layer. The predicted wear sequence is compared with the coating failure threshold, and the remaining service life of the bearing is calculated by extrapolating using the fitted degradation trend curve. The transfer learning framework incorporating MMD (Maximum Mean Difference) is as follows:
[0060] in, It is a feature map that maps data to the reproducing kernel Hilbert space H; The lightweight "student prediction network" employs a Relevance Vector Machine (RVM) model. RVM is a sparse Bayesian model where, for a given input x, its predicted output y follows a Gaussian distribution.
[0061] Among them, weight It is given a sparse prior, that is, the prior is determined by automatic correlation. The hyperparameters are learned by maximizing the marginal likelihood, and finally a sparse weight vector is obtained for prediction. Remaining service life extrapolation: extrapolating the predicted cumulative wear sequence With failure threshold Comparison; commonly used degradation trend fitting models include: Exponential model:
[0062] Linear model:
[0063] Parameters are obtained by fitting historical data. ， Then solve the equation. ， Determine the remaining useful life .
[0064] Step 5: Output prediction results and update the model. Output the predicted value of the remaining lifespan and its confidence interval. The system can continuously update the feature library based on new monitoring data and periodically fine-tune the surrogate model and prediction network using newly accumulated data to achieve online evolution of the prediction model.
[0065] Confidence interval output: For Bayesian frameworks (such as GPR, RVM), the prediction results can directly provide the mean and variance, thus allowing the calculation of confidence intervals (e.g., 95% confidence interval). ).
[0066] Online model updates: Bayesian updates or incremental learning can be used. For example, for the GPR model, when new data pairs are obtained... Then, it is added to the training set and the posterior distribution is recalculated. For neural networks, gradient-based fine-tuning can be used, with the loss function being:
[0067] in, For model parameters, and This is a weighting factor.
[0068] The following is a bearing life prediction for a gradient-coated sliding bearing in a certain type of wind turbine gearbox, based on the method described above.
[0069] First, based on step one, determine the parameters such as the thickness, elastic modulus, hardness, and wear coefficient of each layer of the coating for a certain type of bearing. Then, establish a multi-field coupled wear model in ANSYS simulation software, setting the fluctuation range of rotational speed and load as random inputs, and pre-setting the complete depletion of the self-lubricating layer as the coating failure threshold, for example, setting the remaining coating thickness ≤ 2μm as the failure threshold.
[0070] Then, several sample points are initialized within the design space of rotational speed and load for high-fidelity simulation to obtain an initial training set. Subsequently, the process of "Gaussian process regression model training → uncertainty measurement to screen high-value samples → high-fidelity simulation to supplement data → updating training set" is executed iteratively. After about 40 iterations, the prediction error of the Gaussian process surrogate model drops below 5%, meeting the preset accuracy threshold. The iteration is then stopped, and a high-precision surrogate model is obtained. This significantly reduces the computational cost and time consumption of multiphysics simulation, achieving a balance between high accuracy and computational efficiency.
[0071] Vibration sensors are installed on the actual pump body of the wind turbine gearbox to collect signals online. Every hour of operation, a multi-domain feature vector, including time-domain, frequency-domain, and nonlinear features, is extracted. Using a trained transfer learning framework model, this feature vector is mapped to the current total wear depth of the coating. When the prediction indicates that the self-lubricating layer thickness is about to fall below a preset threshold, the system issues a warning and predicts the remaining bearing life to be approximately 50 hours based on the slope of the degradation curve. Verification shows that the actual failure time and prediction error are within 10%, proving that the present invention can effectively improve the accuracy and computational efficiency of predicting the life of gradient composite coating sliding bearings, accurately reflect the layer-by-layer wear degradation process of the coating, and achieve reliable prediction of the remaining service life.
[0072] The above are merely embodiments of the present invention, and the invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A method for predicting the life of a gradient composite coated sliding bearing, characterized in that: Step 1: Construct a multi-field coupled physical simulation model of coating degradation; Based on the material properties, layered structure and interface characteristics of gradient composite coatings, establish a multi-physics coupled finite element model that considers hydrodynamic lubrication, contact mechanics and material wear. Step 2: Construct a high-precision proxy model using an active learning algorithm; Based on the physical simulation model in step one, an active learning query strategy is designed. Step 3: Online monitoring and multi-domain feature extraction; During actual bearing operation, signals from multiple sensors such as vibration, acoustic emission, and temperature are collected in real time; After preprocessing the signals, multi-domain features that can reflect the coating degradation state are extracted. Step 4: Real-time state mapping and lifetime prediction based on transfer learning; Based on the real-time multi-domain features in step three, a transfer learning framework for the "teacher model" is established. The online monitoring data is processed using the domain adaptation loss function, and the processed data is input into a lightweight "student prediction network" model. The model outputs an estimate of the cumulative wear of the front coating. Then, the wear data is fitted into a degradation trend curve to infer the remaining service life of the bearing.
2. The life prediction method for a gradient composite coating sliding bearing according to claim 1, characterized in that: The model established in step one uses bearing operating conditions (load, speed, lubrication conditions) and coating geometric tolerances as random input variables, and outputs the degradation curve of coating thickness over operating time. The hydrodynamic lubrication model is established by using the Reynolds equation to describe the oil film pressure distribution in the sliding bearing, and its simplified form is as follows: in, For oil film pressure, Oil film thickness, The viscosity of the lubricating oil. For density, , The tangential velocity of the contact surface; The contact mechanics model is established as follows: for elastic contact between coatings or substrates, Hertzian contact theory is used to calculate contact stress and contact half-width; for contact between two spheres, the maximum contact stress is calculated. and contact radius for: in, For normal load, The equivalent radius of curvature, It is the equivalent elastic modulus; The material wear model is established as follows: the reduction in coating thickness is calculated using the Archard wear model; the basic formula describes the wear volume. The relationship between load and sliding distance is as follows: in, The wear coefficient is a dimensionless coefficient. For normal load, The sliding distance, The hardness of a relatively soft material is used; however, in finite element analysis, its differential or discrete form is often used to calculate the wear depth at nodes. : in, The wear coefficient (unit: 1 / Pa). To contact pressure, This represents the sliding distance.
3. The life prediction method for a gradient composite coating sliding bearing according to claim 2, characterized in that: In step two, a small number of sample points are initialized in the design space of the input variables for simulation to obtain an initial training set. Then, the following process is executed iteratively: a Gaussian process regression surrogate model is trained using the current training set; the most valuable new sample points are selected based on the uncertainty measure of the surrogate model for high-fidelity physical simulation, and the new data is added to the training set. This process continues until the prediction accuracy of the surrogate model meets the preset threshold. Gaussian process regression: treating the simulation model as a black box function y=f(x)+ The Gaussian process assumes that the function value follows a multivariate Gaussian distribution, and its predicted distribution is: Posterior mean (predicted value): Posterior variance (uncertainty): in, It is the kernel function matrix between training samples. It is the kernel function vector between the test points and the training samples. It is the noise variance; Active learning query strategy: Selecting new sample points using expected improvement criteria: in, This is the current best observation value. and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. , To balance the parameters of exploration and development.
4. The life prediction method for a gradient composite coating sliding bearing according to claim 3, characterized in that: In step three, the multi-domain features include time-domain features, frequency-domain features, and nonlinear features; The time-domain features include root mean square value, peak value, and kurtosis. Root mean square value: Peak value: kurtosis: The frequency domain features are obtained by performing a Fourier transform on the signal, analyzing the changes in the amplitude of the characteristic frequencies, and reducing the dimensionality through principal component analysis. Extract the amplitude or energy within a specific frequency band. Then, use principal component analysis for dimensionality reduction, the core of which is solving the covariance matrix. eigenvalues and eigenvectors; The nonlinearity Features: Employs a dynamic time warping algorithm to calculate the nonlinear distance between signal waveforms in order to capture subtle degradation patterns; in, It is a regular path. It is the distance between points.
5. The life prediction method for a gradient composite coating sliding bearing according to claim 4, characterized in that: In step four, the real-time multi-domain feature sequence extracted in step three is used as input to construct a transfer learning framework with the surrogate model obtained in step two as the "teacher model." By introducing a domain adaptation loss function, the feature distribution of the online monitoring data is aligned to the feature distribution generated from the simulation data. Then, the aligned features of the online monitoring data are input into a lightweight "student prediction network," which outputs an estimate of the cumulative wear of the pre-coating layer. The predicted wear sequence is compared with the coating failure threshold, and the remaining service life of the bearing is calculated by extrapolating using the fitted degradation trend curve. The transfer learning framework of the "teacher model" is as follows: in, It is a feature map that maps data to the reproducing kernel Hilbert space H; The lightweight "student prediction network" employs a Relevance Vector Machine (RVM) model. RVM is a sparse Bayesian model where, for a given input x, its predicted output y follows a Gaussian distribution. Among them, weight It is given a sparse prior, that is, the prior is determined by automatic correlation. The hyperparameters are learned by maximizing the marginal likelihood, and finally a sparse weight vector is obtained for prediction. Remaining service life extrapolation: extrapolating the predicted cumulative wear sequence With failure threshold Comparison; commonly used degradation trend fitting models include: Exponential model: Linear model: Parameters are obtained by fitting historical data. ， Then solve the equation. ， Get the remaining useful life .
6. The life prediction method for a gradient composite coating sliding bearing according to claim 1, characterized in that: It also includes step five, prediction result output and model update, outputting the predicted value of the remaining lifespan and its confidence interval. The system can continuously update the feature library based on new monitoring data, and periodically use newly accumulated data to fine-tune the surrogate model and prediction network to achieve online evolution of the prediction model.