A bearing vibration suppression method based on local mesh method
By combining local meshing and digital twin technology with deep learning and optimization algorithms, precise suppression of vibration of steel balls in aerospace micro-bearings was achieved. This solved the vibration control problem that traditional design methods could not adapt to in extreme environments, and improved the stability and attitude control accuracy of the bearings in spacecraft.
Patent Information
- Application Number
- CN202611194059.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional bearing design methods cannot effectively suppress the vibration of steel balls in aerospace micro-bearings under extreme space conditions, leading to deviations in attitude control command execution and risks of spacecraft attitude instability. Furthermore, the load-bearing capacity of the lubricating film decreases under high vacuum conditions, the friction coefficient increases, and the contact load distribution becomes uneven. Traditional finite element methods struggle to balance global optimization accuracy and computational efficiency.
The local mesh method is adopted to extract the vibration characteristics of the steel ball through short-time Fourier transform and wavelet packet decomposition. A multi-objective optimization function is established. By combining non-uniform mesh partitioning, Latin hypercube sampling and genetic algorithm, an approximate response surface model is constructed. Gaussian process regression and Bayesian optimization are used to find the globally optimal clearance combination. A digital twin is constructed by combining the lubricating film degradation model and the bearing contact mechanics model. A deep Q-network is used to train an intelligent adjustment strategy for loading parameters to form a closed-loop adaptive vibration suppression.
It achieves precise suppression of steel ball vibration throughout its entire on-orbit lifecycle, adapts to various working conditions and clearance variations, improves the vibration control accuracy and stability of bearings in extreme environments, and reduces the impact of lubricant film degradation on bearing performance.
Smart Images

Figure CN123110285A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of precision transmission support technology, and more specifically, to a bearing vibration suppression method based on the local mesh method. Background Technology
[0002] With the rapid development of the civilian aerospace industry, miniature control moment gyroscopes have become core actuators in satellite attitude control systems. These gyroscopes utilize precision miniature bearings that operate in extreme space environments, typically including high vacuum levels below 10⁻⁵ Pascals, microgravity levels reaching 10⁻⁶ gravitational accelerations, and intense thermal cycling with temperature fluctuations ranging from -50°C to 120°C. The bearings use a one-time-use cured lubricant, making maintenance or replacement impossible during on-orbit operation. Under these conditions, the bearing steel balls continuously generate periodic impact vibrations and high-frequency natural vibrations during their revolution and rotation. This vibrational energy is transmitted via the raceway to the bearing housing and gyroscope shell. This directly affects the stability of the gyroscope's output torque and pointing accuracy. At the same time, the high vacuum environment leads to accelerated evaporation of the lubricating film and an increase in the coefficient of friction. Microgravity conditions cause uneven distribution of contact loads, and temperature fluctuations cause thermal expansion and contraction of materials, thereby changing the assembly gap and contact stress state. The coupling of internal and external factors makes the steel ball exhibit significant nonlinear and time-varying characteristics through vibration. Traditional static optimization schemes under single working conditions are difficult to adapt to the dynamic changes throughout the entire life cycle. When the bearing operates to a specific speed range or experiences temperature fluctuations, the vibration amplitude of the steel ball may suddenly increase and exceed the vibration threshold allowed by the gyroscope control system, which may lead to deviations in the execution of attitude control commands and, in severe cases, the risk of spacecraft attitude instability.
[0003] Currently, aerospace bearing design generally employs overall parameter optimization methods, such as orthogonal experiments or response surface methodology, with the overall vibration intensity or service life of the bearing as the optimization objective. However, it fails to establish a specific objective function for the core vibration source—the steel ball's vibration—causing the optimization direction to deviate from the actual vibration cause. Furthermore, the selection of assembly clearances often relies on recommended tolerance zones in general mechanical design manuals, lacking specific matching tailored to the low-load, high-speed, and high-precision characteristics of aerospace micro-gyroscopes. Loading techniques follow the constant-pressure preload or positioning preload schemes used for terrestrial rolling bearings, failing to consider the attenuation characteristics of the lubricating film's load-bearing capacity under high-vacuum one-time lubrication conditions. Furthermore, the contact dynamics equations between the steel ball and the raceway in the ground simulation model are highly nonlinear, making it difficult for traditional finite element methods to balance global optimization accuracy and computational efficiency. Therefore, there is an urgent need for a structural design method that can address the vibration source and take the vibration of the steel ball as the direct optimization target. This method should combine multi-condition gap test data with space environment-specific loading parameters to form a three-in-one vibration suppression system integrating structural optimization, gap matching, and loading adaptation. It should also have the ability to self-sensitize the on-orbit vibration state and adaptively adjust parameters, thereby achieving precise suppression of vibration of the micro-bearing steel ball under extreme aerospace environments and dynamic adaptation throughout its entire life cycle. Summary of the Invention
[0004] This invention addresses the technical problems existing in the prior art by providing a bearing vibration suppression method based on the local mesh method, thereby solving the problems mentioned in the background art.
[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps:
[0006] Step S1: Acquire broadband vibration signals of the bearing operation, separate the orbital frequency of the steel ball and its harmonic components through short-time Fourier transform and wavelet packet decomposition, extract the vibration amplitude, phase and modulation characteristics of the steel ball, and establish a multi-objective optimization function with the goal of minimizing the vibration amplitude of the steel ball. Calculate the vibration sensitive direction based on the gradient of the multi-objective optimization function, and determine the parameter vector to be optimized and its constraints. The constraints include the lubricating film thickness threshold determined based on the lubricating film degradation model and the bearing contact mechanics model. Step S2: Divide the parameter space corresponding to the parameter vector to be optimized according to the vibration sensitive direction into a non-uniform grid. In the region where the gradient modulus exceeds the preset refinement threshold, refine the grid. Construct an approximate response surface model by Latin hypercube sampling. Then, use a hybrid solver of genetic algorithm and sequential quadratic programming to iteratively search for the optimal parameter combination on the local grid until the steel ball converges to the preset convergence threshold through vibration amplitude. Step S3: Using the optimal parameter combination as the baseline, collect test data of inner and outer ring fit clearance combinations under multiple working conditions to construct a clearance-vibration multidimensional mapping database. Use Gaussian process regression to fit the mapping relationship and search for the global optimal clearance combination through Bayesian optimization to output the standardized assembly clearance recommendation value and temperature-clearance drift model. Step S4: Couple the lubricating film degradation model with the bearing contact mechanics model to construct a digital twin. With the optimal parameter combination and the recommended value of standardized assembly clearance as the initial state, use a deep Q network to train and obtain the intelligent adjustment strategy model of loading parameters and deploy it to the onboard computer of the spacecraft. Based on the on-orbit real-time vibration monitoring data obtained by the onboard accelerometer of the spacecraft, dynamically output the loading parameter adjustment command to form a closed-loop adaptive vibration suppression.
[0007] In a preferred embodiment, the specific operation of extracting the vibration amplitude, phase, and modulation characteristics of the steel ball in step S1 is as follows: A short-time Fourier transform is applied to the broadband vibration signal, with a Hanning window selected as the window function and the window length set to twice the orbital period of the steel ball, to obtain the time spectrum; Simultaneously, the broadband vibration signal is subjected to three-level wavelet packet decomposition, and the Daubechies8 wavelet basis is used to obtain eight sub-band signals generated by wavelet packet decomposition. Based on the pre-calculated theoretical value of the steel ball's orbital frequency, energy ridges with frequencies within the range of the steel ball's orbital frequency plus or minus half a frequency resolution and its integer harmonics plus or minus half a frequency resolution are extracted from the time-frequency spectrum; the time-frequency coefficients corresponding to the energy ridges are mapped to sub-bands, and the pure vibration signal of the steel ball is reconstructed. The vibration amplitude of the steel ball is extracted from the reconstructed vibration signal, and the maximum value of the absolute value of the signal is taken as the vibration amplitude of the steel ball. Extract the phase, which is the arctangent value of the Fourier transform result at the orbital frequency of the steel ball; The modulation features are extracted, and the modulation index is defined as the ratio of the sum of the amplitudes of the sideband components on both sides of the ball's orbital frequency to the amplitude at the ball's orbital frequency, where the sideband spacing is the hold's rotational frequency.
[0008] In a preferred embodiment, the specific operations of establishing the multi-objective optimization function and calculating the vibration-sensitive direction in step S1 are as follows: The parameter vector to be optimized includes five parameters: steel ball diameter, inner raceway radius of curvature, outer raceway radius of curvature, retainer window clearance, and fit clearance allowance. A multi-objective optimization function is established with the goal of minimizing the weighted sum of the vibration amplitude and modulation index of the steel ball. The sum of the weight coefficients is one. The steel ball achieves dimensionlessness by dividing its vibration amplitude and modulation index by their respective reference values under the initial parameter combination. The constraints include the bearing's rated dynamic load not exceeding the allowable value, the limiting speed not exceeding the allowable value, and the lubricating film thickness not less than the lubricating film thickness threshold, wherein the lubricating film thickness threshold is obtained by jointly solving the lubricating film degradation model and the bearing contact mechanics model; Multiple sample points are randomly selected in the parameter space to be optimized, and the objective function value of each sample point is calculated. A radial basis function interpolation model is constructed using the Gaussian kernel function as the basis function. The analytical gradient of the radial basis function interpolation model is obtained to obtain the vibration sensitive direction. Each component of the vibration sensitive direction represents the intensity and direction of the comprehensive influence of the corresponding parameter change on the objective function.
[0009] In a preferred embodiment, the specific operation of performing non-uniform mesh generation and constructing an approximate response surface model according to the vibration-sensitive direction in step S2 is as follows: Calculate the gradient magnitude at each sample point based on the vibration-sensitive direction output in step S1, set the preset refinement threshold to 30% of the global maximum gradient magnitude, mark the region with a gradient magnitude greater than or equal to the preset refinement threshold as a vibration-sensitive region, increase the mesh density to four times the initial density in the vibration-sensitive region, and maintain the initial mesh density at the beginning of step S2 in the region not marked as a vibration-sensitive region. Two hundred sample points are extracted at the center of each cell in the non-uniform grid using the Latin hypercube sampling method. For each sample point, step S1 is performed to extract the steel ball and calculate the corresponding objective function value by operating on the vibration amplitude, phase and modulation characteristics. These sample points and their objective function values are used as training data. A radial basis function response surface approximation model is constructed using the Gaussian kernel function as the basis function. The shape parameter of the Gaussian kernel function is taken as the reciprocal of the average Euclidean distance between all pairs of sample points.
[0010] In a preferred embodiment, the specific operation of step S2, which uses a hybrid solver of genetic algorithm and sequential quadratic programming to iteratively search for the optimal parameter combination, is as follows: Using the response surface approximation model as the fitness function, a genetic algorithm is run on all nodes of a non-uniform grid to perform a global coarse search and select the ten individuals with the highest fitness as candidate solutions. The sequential quadratic programming algorithm is started with each candidate solution as the initial point to perform a local fine search. The sequential quadratic programming algorithm gradually approximates the local optimum by constructing a quadratic approximation model of the Lagrange function. The termination condition is that the relative change of the objective function value between two adjacent iterations is less than 10 to the power of negative 6 or the number of iterations exceeds 100. The solution with the smallest objective function value among all local optima is taken as the optimal parameter combination. It is then determined whether the vibration amplitude of the steel ball corresponding to the optimal parameter combination is less than the preset convergence threshold. If the threshold is met, the iteration stops. If not, the mesh near the optimal parameter combination is further refined and the process returns to the Latin hypercube sampling step in step S2 to reconstruct the response surface model and continue iterating.
[0011] In a preferred embodiment, the specific operations of collecting multi-condition test data and constructing a Gaussian process regression model in step S3 are as follows: Using the optimal parameter combination output in step S2 as a fixed baseline, multiple sets of bearing samples are manufactured. The inner ring and shaft fit clearance and the outer ring and bearing housing fit clearance of each set of bearing samples are orthogonally designed within a preset range, with a step size of one micrometer. For each set of gaps, bearing operation tests were conducted under three temperature conditions: -50 degrees Celsius, 25 degrees Celsius, and 120 degrees Celsius. The vibration amplitude, running resistance torque, and coaxiality deviation of the steel ball were measured, and the vibration amplitude, running resistance torque, and coaxiality deviation of the steel ball were used as output vectors. All experimental data were compiled into a dataset. The input vector contained three dimensions: inner ring fit clearance, outer ring fit clearance, and temperature. The output vector contained three dimensions: vibration amplitude, drag torque, and coaxiality. Multi-task Gaussian process regression is used to simultaneously fit the correlation between the three output variables. The joint covariance function is obtained by multiplying the kernel matrix between the output tasks with the Matrn kernel function in the parameter space corresponding to the input vector. The Matrn kernel function is selected as the type with an order of 2 / 5. The length scaling parameter and signal variance of the Matrn kernel function are estimated by maximizing the log marginal likelihood function. The mean function is set as a constant, and all hyperparameters are estimated by maximizing the log marginal likelihood function.
[0012] In a preferred embodiment, the specific operation of searching for the globally optimal gap combination and outputting the temperature-gap drift model through Bayesian optimization in step S3 is as follows: The temperature is fixed at 25 degrees Celsius. The optimization objective is to minimize the predicted mean of the vibration amplitude of the steel ball. The expected improvement is used as the acquisition function. The expression of the expected improvement is calculated based on the minimum vibration amplitude in the currently acquired samples and the predicted distribution of the Gaussian process at the new input, which includes the predicted mean and the predicted standard deviation. Run Bayesian optimization in a two-dimensional gap space. In each iteration, select a new gap combination that maximizes the expected improvement, add it to the dataset and update the Gaussian process model. Repeat the iteration until the maximum number of iterations is reached or the maximum expected improvement is lower than the preset expected improvement stopping threshold, and output the globally optimal gap combination. Using a trained multi-task Gaussian process model, temperature is taken as a variable input at the globally optimal gap combination to calculate the predicted vibration amplitude at different temperatures, and a temperature-vibration amplitude drift model is constructed. This temperature-vibration amplitude drift model adopts a second-order Taylor expansion approximation, and the coefficients are obtained by numerical differentiation of the Gaussian process prediction value in the normal temperature neighborhood. Simultaneously, based on the difference in thermal expansion coefficients between the inner ring and the shaft, and between the outer ring and the bearing housing, the equivalent clearance offset caused by temperature is calculated, and the recommended value for standardized assembly clearance and the temperature-clearance drift model are output.
[0013] In a preferred embodiment, the specific operations of constructing the digital twin and defining the Markov decision process in step S4 are as follows: A digital twin is constructed by coupling the lubricating film degradation model with the bearing contact mechanics model. The lubricating film degradation model describes the exponential decay of the oil film thickness over time. The initial film thickness is determined by the lubricating film thickness threshold constraint in step S1. The degradation rate constant depends on the lubricant evaporation rate and the microgravity environment. The temperature influence factor in the lubricating film degradation model is a linear function. The temperature influence factor is equal to one plus the temperature sensitivity coefficient multiplied by the difference between the current temperature and the room temperature reference temperature. The temperature sensitivity coefficient is used to quantify the degree of correction of the lubricating film degradation rate by temperature. The bearing contact mechanics model calculates the pressure distribution and elastic deformation in the contact area based on Hertz theory, and derives the nonlinear mapping relationship between the vibration amplitude of the steel ball and the preload, film thickness, and temperature. The mapping is based on the response surface approximation model in step S2 and incorporates correction terms for temperature and film thickness. The digital twin uses the optimal parameter combination output in step S2 and the standardized assembly gap recommendation value output in step S3 as fixed structural parameters. With time as the driving variable, it outputs a state vector in real time. The state vector includes the vibration amplitude of the steel ball at the current moment, the current film thickness, the current temperature, the preload at the previous moment, and the equivalent gap offset caused by temperature. The loading parameter adjustment problem is modeled as a discrete-time Markov decision process. The action space is defined to include the preload adjustment and the loading uniformity adjustment. The reward function is defined as the weighted sum of the steel ball passing through vibration amplitude penalty term, lubricating film thickness lower limit violation penalty term, action amplitude penalty term and action frequency penalty term.
[0014] In a preferred embodiment, the specific operation of training the intelligent parameter adjustment strategy model using a deep Q-network in step S4 is as follows: A deep Q-network is constructed. The deep Q-network architecture is a three-layer fully connected neural network. The dimension of the input layer is equal to the dimension of the state vector. The number of neurons in the hidden layer is 128 and 64 respectively. The activation function is a linear rectified function. The dimension of the output layer is equal to the number of actions after the action space is discretized. The state is randomly initialized in the digital twin, the simulation duration is set, and a greedy strategy is used to select actions, with the greedy coefficient decreasing linearly from one to zero. The current state, selected action, immediate reward, and next state obtained from each interaction are stored as an experience tuple in the experience replay buffer. Every fixed number of steps, a batch of experience tuples are randomly sampled from the experience replay buffer, and the temporal difference error is calculated. The temporal difference error adopts the Huber loss function to improve training stability. The target network parameters are copied from the current network parameters every 500 steps. Training continues until the cumulative reward converges, and the network parameters are saved as the loading parameter intelligent adjustment strategy model.
[0015] In a preferred embodiment, the specific operation of deploying the intelligent adjustment strategy model of loading parameters to the onboard computer and forming a closed-loop adaptive vibration suppression in step S4 is as follows: The trained deep Q-network is converted into a greedy policy and deployed to the onboard computer carried by the spacecraft. During in-orbit operation, the bearing vibration signal is collected by the onboard accelerometer in each preset control cycle, and the current vibration amplitude of the steel ball is extracted by time-frequency analysis in step S1. Read the temperature sensor data and calculate the equivalent gap offset using the temperature-gap drift model from step S3; read the preload force from the memory of the onboard computer at the previous moment and call the lubricating film degradation model to estimate the current film thickness; The state vector is formed and input into the greedy policy network, and the optimal action is output, including the preload adjustment amount and the loading uniformity adjustment amount; the preload adjustment is performed by the piezoelectric actuator, and the loading uniformity adjustment is performed by the fine-tuning spring shims; Simultaneously, the vibration amplitude of the steel ball measured in orbit is compared with the predicted vibration amplitude of the steel ball output by the digital twin. If the deviation exceeds the preset threshold, the degradation rate constant and temperature sensitivity coefficient in the digital twin model are corrected online. The degradation rate constant and temperature sensitivity coefficient are updated by recursive Bayesian estimation to ensure that the digital twin is always aligned with the actual degradation trajectory.
[0016] The beneficial effects of this invention are as follows: Vibration characteristics of the steel ball are accurately extracted through short-time Fourier transform and wavelet packet decomposition; vibration-sensitive directions are locked using multi-objective optimization functions and gradient calculations; structural parameters are efficiently optimized using non-uniform mesh generation, Latin hypercube sampling, response surface approximation models, and a hybrid solver combining genetic algorithms and sequential quadratic programming; globally optimal gap combinations and temperature-gap drift models are obtained using Gaussian process regression and Bayesian optimization; finally, a digital twin is constructed by coupling the lubricating film degradation model with the bearing contact mechanics model, and a deep Q-network is used to train an intelligent adjustment strategy model for loading parameters. Combined with real-time data from the onboard accelerometer, a closed-loop adaptive vibration suppression is formed, achieving precise control from the vibration source to the entire on-orbit lifecycle. Attached Figure Description
[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0019] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0020] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.
[0021] Example 1: This example provides the following... Figure 1 The method for suppressing bearing vibration based on the local mesh method is shown, and specifically includes the following steps: Step S1: Acquire broadband vibration signals of the bearing operation, separate the orbital frequency of the steel ball and its harmonic components through short-time Fourier transform and wavelet packet decomposition, extract the vibration amplitude, phase and modulation characteristics of the steel ball, and establish a multi-objective optimization function with the goal of minimizing the vibration amplitude of the steel ball. Calculate the vibration sensitive direction based on the gradient of the multi-objective optimization function, and determine the parameter vector to be optimized and its constraints. The constraints include the lubricating film thickness threshold determined based on the lubricating film degradation model and the bearing contact mechanics model. Step S2: Divide the parameter space corresponding to the parameter vector to be optimized according to the vibration sensitive direction into a non-uniform grid. In the region where the gradient modulus exceeds the preset refinement threshold, refine the grid. Construct an approximate response surface model by Latin hypercube sampling. Then, use a hybrid solver of genetic algorithm and sequential quadratic programming to iteratively search for the optimal parameter combination on the local grid until the steel ball converges to the preset convergence threshold through vibration amplitude. Step S3: Using the optimal parameter combination as the baseline, collect test data of inner and outer ring fit clearance combinations under multiple working conditions to construct a clearance-vibration multidimensional mapping database. Use Gaussian process regression to fit the mapping relationship and search for the global optimal clearance combination through Bayesian optimization to output the standardized assembly clearance recommendation value and temperature-clearance drift model. Step S4: Couple the lubricating film degradation model with the bearing contact mechanics model to construct a digital twin. With the optimal parameter combination and the recommended value of standardized assembly clearance as the initial state, use a deep Q network to train and obtain the intelligent adjustment strategy model of loading parameters and deploy it to the onboard computer of the spacecraft. Based on the on-orbit real-time vibration monitoring data obtained by the onboard accelerometer of the spacecraft, dynamically output the loading parameter adjustment command to form a closed-loop adaptive vibration suppression.
[0022] In this embodiment, the specific operation of extracting the vibration amplitude, phase, and modulation characteristics of the steel ball in step S1 is as follows: A short-time Fourier transform was applied to the broadband vibration signal, with a Hanning window selected as the window function and the window length set to twice the orbital period of the steel ball to obtain the time spectrum. The orbital period of the steel ball was calculated using the bearing pitch circle diameter, the steel ball diameter, and the spindle speed. Under typical operating conditions, the orbital period of the steel ball is approximately 0.01 seconds. Therefore, the window length was set to 0.02 seconds to ensure that the frequency resolution was sufficient to distinguish the vibration of the steel ball from adjacent interference components. Simultaneously, the broadband vibration signal is subjected to three-level wavelet packet decomposition using the Daubechies8 wavelet basis to obtain eight sub-band signals generated by wavelet packet decomposition. The three-level wavelet packet decomposition divides the signal frequency band into eight equal-width sub-bands. The frequency range covered by each sub-band is determined according to the sampling frequency. For example, when the sampling frequency is 100 kHz, the bandwidth of each sub-band is 12.5 kHz, which is sufficient to accommodate the vibration of the steel ball and its harmonic components. Based on the pre-calculated theoretical value of the steel ball's orbital frequency, energy ridges are extracted from the time spectrum whose frequencies fall within the range of the steel ball's orbital frequency plus or minus half a frequency resolution and its integer harmonics plus or minus half a frequency resolution. The frequency resolution is determined by the window length of the short-time Fourier transform; when the window length is 0.02 seconds, the frequency resolution is 50 Hz. Therefore, the extraction range is the steel ball's orbital frequency plus or minus 25 Hz, and its second and third harmonics plus or minus 25 Hz. The method for extracting energy ridges is to search for local maxima at each frequency point along the time axis in the time spectrum. The energy ridges are formed by connecting the maximum values at consecutive time points that satisfy the amplitude being higher than the background noise threshold (usually set to 10% of the maximum amplitude). The time-frequency coefficients corresponding to the energy ridges are mapped to sub-bands to reconstruct the pure steel ball passing vibration signal. The specific mapping method is as follows: the time-frequency coefficients at the location of the energy ridges on the time spectrum are assigned to the wavelet packet coefficients of the corresponding time point and sub-band index, and the coefficients at other positions are set to zero. Then, the wavelet packet coefficients are subjected to inverse wavelet packet transform to recover the time domain signal, which is the pure steel ball passing vibration signal. The vibration amplitude of the steel ball is extracted from the reconstructed vibration signal. The maximum absolute value of the signal is taken as the vibration amplitude. Taking the absolute value is to eliminate the phase effect. The maximum value reflects the peak impact intensity of the steel ball vibration, and the unit is gravitational acceleration g. The phase is extracted by taking the arctangent value of the Fourier transform result at the orbital frequency of the steel ball. The specific calculation process is as follows: perform a Fourier transform on the reconstructed signal, extract the complex component corresponding to the orbital frequency of the steel ball, denoted as R in real part and I in imaginary part. The phase is equal to the arctangent function value, i.e., I divided by the arctangent of R. The calculation result is expressed in radians and ranges from negative π to π. Modulation features are extracted, and the modulation index is defined as the ratio of the sum of the amplitudes of the sideband components on both sides of the ball's orbital frequency to the amplitude at the ball's orbital frequency. The sideband interval is the cage rotational frequency. The cage rotational frequency is calculated from the cage rotational speed, and a typical value is 0.4 times the ball's orbital frequency. The sideband components appear at the positions where the ball's orbital frequency is plus or minus the cage rotational frequency. The modulation index is dimensionless; the smaller the value, the weaker the nonlinear modulation of the ball's vibration, and the closer the vibration waveform is to a sine wave. The specific steps in step S1 to establish the multi-objective optimization function and calculate the vibration-sensitive direction are as follows: The parameter vector to be optimized is defined as including five parameters: steel ball diameter, inner raceway radius of curvature, outer raceway radius of curvature, cage window clearance, and fit clearance allowance. The typical value ranges for the five parameters are as follows: steel ball diameter 0.5 mm to 2 mm, inner raceway radius of curvature 0.3 mm to 1.5 mm, outer raceway radius of curvature 0.3 mm to 1.5 mm, cage window clearance 0.01 mm to 0.1 mm, and fit clearance allowance 0.001 mm to 0.02 mm. A multi-objective optimization function is established with the goal of minimizing the weighted sum of the vibration amplitude and modulation index of the steel ball. The sum of the weight coefficients is one. The vibration amplitude and modulation index of the steel ball are divided by their respective baseline values under the initial parameter combination to achieve dimensionlessness. The initial parameter combination is the initial design parameters of the bearing at the beginning of step S1. The initial parameter combination is given by the bearing design drawings, such as a steel ball diameter of 1 mm and an inner ring raceway curvature radius of 0.8 mm. The baseline value is calculated by substituting the initial parameter combination into the feature extraction process of step S1 to calculate the corresponding vibration amplitude and modulation index of the steel ball, which are used as the denominators. The typical weight coefficient is 0.7 for the vibration amplitude and 0.3 for the modulation index, prioritizing the suppression of amplitude while taking into account the modulation effect. The constraints include that the rated dynamic load of the bearing does not exceed the allowable value, the limiting speed does not exceed the allowable value, and the lubricating film thickness is not less than the lubricating film thickness threshold. The lubricating film thickness threshold is obtained by jointly solving the lubricating film degradation model and the bearing contact mechanics model. The lubricating film degradation model describes the process of the lubricating film thickness decaying exponentially with time in a high vacuum environment. Its mathematical expression is: the lubricating film thickness equals the initial thickness multiplied by the negative degradation rate of the natural constant multiplied by the time power and then multiplied by the temperature influence factor. The bearing contact mechanics model calculates the pressure distribution and elastic deformation of the contact area based on Hertzian contact theory. The steps for solving both together are as follows: First, calculate the equivalent radius of curvature of the contact area and the load per unit length based on the current parameter combination. Then, substitute them into the elastohydrodynamic lubrication film thickness formula to obtain the minimum film thickness value. Compare this minimum film thickness value with the critical film thickness threshold (usually taken as 0.1 micrometers). If it is less than the threshold, it is determined that the constraint is not satisfied. The critical film thickness threshold is determined based on the lubricant characteristics and surface roughness. In this embodiment, it is taken as 0.15 micrometers. Multiple sample points are randomly selected within the parameter space to be optimized. The objective function value of each sample point is calculated. A radial basis function interpolation model is constructed using the Gaussian kernel function as the basis function. The analytical gradient of the radial basis function interpolation model is calculated to obtain the vibration-sensitive direction. Each component of the vibration-sensitive direction represents the intensity and direction of the comprehensive influence of the corresponding parameter change on the objective function. The number of randomly selected sample points is set to 50, and the Latin hypercube sampling method is used to ensure that the sample points are uniformly distributed in the five-dimensional parameter space. The shape parameter of the Gaussian kernel function is determined based on the average distance between the sample points, with a typical value being the reciprocal of the average distance. The process of constructing the radial basis function interpolation model is as follows: using the coordinates of 50 sample points and the corresponding objective function values as training data, solving the linear equation system to obtain the interpolation coefficients for each sample point, and then linearly combining the interpolation coefficients with the Gaussian kernel function to obtain a continuously differentiable surrogate model; The process of finding the analytical gradient of the surrogate model is as follows: calculate the partial derivative for each parameter component. Since the Gaussian kernel function is differentiable, the partial derivatives can be calculated directly, resulting in a five-dimensional gradient vector. The absolute value of each component represents the sensitivity of the parameter to the objective function, and the positive or negative sign indicates the direction of the parameter's influence on the objective function.
[0023] In this embodiment, it is specifically necessary to explain the following steps in step S2: performing non-uniform mesh generation based on the vibration-sensitive direction and constructing an approximate response surface model. The gradient magnitude at each sample point is calculated based on the vibration-sensitive direction output in step S1. A preset refinement threshold is set to 30% of the global maximum gradient magnitude. Regions with gradient magnitudes greater than or equal to the preset refinement threshold are marked as vibration-sensitive regions. Within these regions, the mesh density is increased to four times the initial density. In regions not marked as vibration-sensitive regions, the initial mesh density at the beginning of step S2 is maintained. The gradient magnitude calculation process is as follows: the square root of the sum of the squares of the five-dimensional vibration-sensitive direction vector output in step S1 is taken to obtain the gradient magnitude at that point. The global maximum gradient magnitude refers to the maximum gradient magnitude among all sample points in the entire parameter space to be optimized. The preset refinement threshold of 30% of the global maximum gradient magnitude is a balance point derived from empirical statistics, ensuring that the sensitive region is sufficiently refined while avoiding excessive refinement that leads to a surge in computational load. The specific implementation of increasing the mesh density to four times is as follows: the number of mesh nodes in each dimension of the vibration-sensitive region is doubled. For example, if each dimension was originally divided into 10 equal parts, it is now divided into 20 equal parts, thus making the total number of mesh units in the region 16 times the original. Two hundred sample points are extracted at the center of each cell in a non-uniform grid using the Latin hypercube sampling method. For each sample point, step S1 is performed to extract the steel ball and calculate the corresponding objective function value through operations on vibration amplitude, phase, and modulation characteristics. These sample points and their objective function values are used as training data. A radial basis function response surface approximation model is constructed using the Gaussian kernel function as the basis function. The shape parameter of the Gaussian kernel function is taken as the reciprocal of the average Euclidean distance between all pairwise sample points. The specific process of Latin hypercube sampling is as follows: First, the value range of each parameter is divided into 200 intervals. Then, a value is randomly selected in each interval. The values of each parameter are then randomly paired and combined to ensure that all intervals of each parameter are uniformly covered, and finally 200 sample points are generated. The shape parameter is calculated as follows: calculate the Euclidean distance between each pair of all 200 sample points, resulting in 19,900 distance values. Calculate the arithmetic mean of these distance values, and then take its reciprocal as the shape parameter of the Gaussian kernel function. This shape parameter determines the width of the kernel function; the smaller the average distance, the sharper the kernel function, and the more sensitive the interpolation model is to local changes. The specific operation of step S2, which uses a hybrid solver combining genetic algorithm and sequential quadratic programming to iteratively search for the optimal parameter combination, is as follows: Using the response surface approximation model as the fitness function, a genetic algorithm is run on all nodes of a non-uniform grid for a global coarse search. The population size is set to 100, the crossover probability is 0.8, the mutation probability is 0.05, and the maximum number of generations is 50. The ten individuals with the highest fitness are selected as candidate solutions from the final population. The specific execution steps of the genetic algorithm are as follows: First, initialize the population by randomly generating 100 individuals, each corresponding to a set of parameter vectors to be optimized; Second, calculate the fitness value of each individual, which is the output value of the response surface approximation model at that individual; Third, the selection operation uses roulette wheel selection, where individuals with smaller fitness values have a higher probability of being selected; Fourth, the crossover operation simulates binary crossover on the selected parent individuals with a probability of 0.8 to generate offspring individuals; Fifth, the mutation operation performs polynomial mutation on the offspring individuals with a probability of 0.05 to introduce random perturbation; Sixth, repeat steps two through five until the number of generations reaches 50. Finally, the ten individuals with the smallest fitness values are selected as candidate solutions from the last generation population. A sequential quadratic programming algorithm is initiated with each candidate solution as the initial point to perform a local fine search. The sequential quadratic programming algorithm gradually approximates the local optimum by constructing a quadratic approximation model of the Lagrange function. The termination condition is that the relative change in the objective function value between two consecutive iterations is less than 10 to the power of -6, or the number of iterations exceeds 100. The specific execution steps of the sequential quadratic programming algorithm are as follows: First, given the initial point, calculate the gradient vector and approximate Hessian matrix of the objective function at the current point; Second, construct a quadratic programming subproblem, that is, approximate the objective function with a quadratic function and the constraint conditions with a linear function at the current point; Third, solve the quadratic programming subproblem to obtain the search direction and Lagrange multipliers; Fourth, perform a one-dimensional line search along the search direction, determine the step size, and update the current point; Fifth, check the termination condition: if the relative change in the objective function value between two consecutive iterations is less than 10 to the power of -6, or the number of iterations exceeds 100, stop; otherwise, return to step 2. This algorithm can quickly converge to a local optimum and has low requirements for the quality of the initial point. The solution with the smallest objective function value among all local optima is taken as the optimal parameter combination. It is then determined whether the vibration amplitude of the steel ball corresponding to the optimal parameter combination is less than a preset convergence threshold. If this threshold is met, the iteration stops. If not, the mesh near the optimal parameter combination is further refined, and the iteration returns to the Latin hypercube sampling step in step S2 to rebuild the response surface model and continue iterating. In this embodiment, the preset convergence threshold is set to 0.1 micrometers, which corresponds to the upper limit of the engineering allowable displacement amplitude of the steel ball through vibration. The specific method for further refining the mesh is as follows: Using the optimal parameter combination as the center, the range is expanded by 10% in each dimension, doubling the mesh density of this local area, i.e., doubling the number of nodes in each dimension. Then, Latin hypercube sampling is re-executed, sample points are extracted on the newly refined mesh, and the response surface model is built. The hybrid solver is run again to search for the optimal solution. This process is repeated until the convergence condition is met, typically requiring two to three rounds of refinement to achieve convergence.
[0024] In this embodiment, the specific operation of collecting multi-condition test data and constructing a Gaussian process regression model in step S3 is as follows: Using the optimal parameter combination output in step S2 as a fixed baseline, multiple sets of bearing samples were manufactured. The inner ring-shaft fit clearance and outer ring-bearing housing fit clearance of each set of bearing samples were orthogonally designed within a preset range, with a step size of one micrometer. The specific values of the preset range were: inner ring-shaft fit clearance from -5 micrometers to +15 micrometers, and outer ring-bearing housing fit clearance from -3 micrometers to +10 micrometers. The orthogonal design adopted a full factorial design scheme, generating a total of 441 combinations of 21 clearance combinations multiplied by 21 clearance combinations. However, considering the experimental cost, a partial factorial design was actually adopted, and 150 representative combinations were selected for the experiment. For each clearance combination, bearing operation tests were conducted under three temperature conditions: -50°C, 25°C, and 120°C. The vibration amplitude, operating resistance torque, and coaxiality deviation of the steel ball were measured, and these parameters were used as output vectors. The specific process of the operation test was as follows: the bearing was mounted on a dedicated test bench and operated at a constant speed of 3,000 revolutions per minute. After the temperature stabilized, the vibration signal measured by the accelerometer was recorded. The vibration amplitude of the steel ball was extracted using the time-frequency analysis method in step S1. The operating resistance torque was measured using a torque sensor, and the coaxiality deviation was measured using a laser displacement sensor. Each condition was measured three times, and the average value was taken to reduce random errors. All experimental data were compiled into a dataset. The input vector contained three dimensions: inner ring fit clearance, outer ring fit clearance, and temperature. The output vector contained three dimensions: vibration amplitude, drag torque, and coaxiality. The total number of samples in the dataset was 150 gap combinations multiplied by three temperature conditions, for a total of 450 samples. Each sample had a three-dimensional input vector and a three-dimensional output vector. A multi-task Gaussian process regression is employed to simultaneously fit the correlation between three output variables. The joint covariance function is obtained by multiplying the kernel matrix between the output tasks with the Matrn kernel function in the parameter space corresponding to the input vectors. A 2 / 5 order Matrn kernel function is selected. The length scale parameter and signal variance of the Matrn kernel function are estimated by maximizing the log-marginal likelihood function. The mean function is set as a constant, and all hyperparameters are estimated by maximizing the log-marginal likelihood function. The kernel matrix between the output tasks is a 3x3 positive definite matrix, where the diagonal elements represent the variance of each output variable itself, and the off-diagonal elements represent the covariance between different output variables, such as the correlation between vibration amplitude and drag torque. The elements of this matrix are estimated during the optimization process, with the initial value set as a diagonally dominant identity matrix. The specific form of the Matrn kernel function is as follows: the Euclidean distance between two input vectors is used as the independent variable, multiplied by the square root of five and divided by the length scale parameter, and then substituted into the product of a polynomial containing first-order and second-order terms and an exponential function; the signal variance controls the overall fluctuation of the objective function; the maximization of the log-marginal likelihood function adopts a quasi-Newton optimization algorithm, with the upper limit of the number of iterations set to two hundred and the gradient tolerance set to ten to the power of negative five; after optimization, the estimated values of all hyperparameters are obtained, including the length scale parameter, signal variance, kernel matrix elements, and constant mean. The specific steps in step S3, which involve searching for the globally optimal gap combination using Bayesian optimization and outputting the temperature-gap drift model, are as follows: The temperature is fixed at 25 degrees Celsius. The optimization objective is to minimize the predicted mean of the vibration amplitude of the steel ball. The expected improvement is used as the acquisition function. The expression for the expected improvement is calculated based on the minimum vibration amplitude in the currently acquired samples and the predicted distribution of the Gaussian process at the new input, including the predicted mean and predicted standard deviation. The specific calculation process for the expected improvement is as follows: First, calculate the minimum vibration amplitude in the currently acquired samples, called the current optimal value. Then, for candidate gap combinations, use the Gaussian process model to obtain their predicted mean and predicted standard deviation. Calculate the difference between the current optimal value and the predicted mean, and divide this difference by the predicted standard deviation to obtain the standardized deviation. Find the cumulative distribution function of the standard normal distribution at this standardized deviation, as well as the probability density function at this standardized deviation. The expected improvement equals this difference multiplied by the cumulative distribution function value plus the predicted standard deviation multiplied by the probability density function value. This acquisition function automatically balances exploration and utilization, favoring utilization when the predicted mean is much smaller than the current optimal value, and favoring exploration when the predicted standard deviation is large. Bayesian optimization is performed in a two-dimensional gap space. In each iteration, a new gap combination that maximizes the expected improvement is selected, added to the dataset, and the Gaussian process model is updated. This iteration is repeated until the maximum number of iterations is reached or the maximum expected improvement is lower than a preset expected improvement stopping threshold. The globally optimal gap combination is then output. The preset expected improvement stopping threshold is set to 10 to the power of -3 micrometers. That is, when the maximum expected improvement is less than this value, it is considered that the benefit of continuing optimization is negligible, and the iteration is stopped. The maximum number of iterations is set to 50. In each iteration, the maximization of the expected improvement is achieved by generating 1,000 candidate points using Latin hypercube sampling. The expected improvement value of each candidate point is calculated, and the gap combination corresponding to the maximum value is taken as a new sample point. After adding the new sample point to the dataset, the Gaussian process model is retrained, and all hyperparameters are updated. Using a trained multi-task Gaussian process model, temperature is taken as a variable input at the globally optimal gap combination to calculate the predicted vibration amplitude at different temperatures, thus constructing a temperature-vibration amplitude drift model. This model employs a second-order Taylor expansion approximation, and the coefficients are obtained by numerically differentiating the Gaussian process prediction values in the neighborhood of room temperature. The specific process of numerical differentiation is as follows: at the globally optimal gap combination, the predicted vibration amplitudes are calculated at 25 degrees Celsius (room temperature), 26 degrees Celsius (room temperature plus 1 degree), and 24 degrees Celsius (room temperature minus 1 degree). The first-order coefficient is equal to the predicted value at 1 degree Celsius (room temperature plus 1 degree) minus the predicted value at 1 degree Celsius (room temperature minus 1 degree) and then divided by two. The second-order coefficient is equal to the predicted value at 1 degree Celsius (room temperature plus 1 degree) minus twice the predicted value at room temperature plus the predicted value at 1 degree Celsius (room temperature minus 1 degree) and then divided by one. This yields the two coefficients of the temperature-vibration amplitude drift model, which can be used to predict the change in vibration amplitude relative to room temperature at any temperature. Simultaneously, based on the difference in thermal expansion coefficients between the inner ring and the shaft, and between the outer ring and the bearing housing, the equivalent clearance offset caused by temperature is calculated, and the recommended value for standardized assembly clearance and the temperature-clearance drift model are output. The difference in thermal expansion coefficients between the inner ring and the shaft is taken as 1.2 times 10 to the power of negative 5 per degree Celsius, and the difference in thermal expansion coefficients between the outer ring and the bearing housing is taken as 0.8 times 10 to the power of negative 5 per degree Celsius. The calculation process for the equivalent clearance offset is as follows: multiply the difference in thermal expansion coefficients by the temperature change to obtain the offset of the inner and outer ring clearances, in micrometers; the recommended value for standardized assembly clearance is to round the globally optimal clearance combination obtained by Bayesian optimization to the closest standard tolerance level, for example, to the nearest integer multiple of one micrometer, and mark the corresponding tolerance zone code.
[0025] In this embodiment, the specific operation of constructing the digital twin and defining the Markov decision process in step S4 is as follows: A digital twin is constructed by coupling the lubricating film degradation model with the bearing contact mechanics model. The lubricating film degradation model describes the exponential decay of the oil film thickness over time. The initial film thickness is determined by the lubricating film thickness threshold constraint in step S1. The degradation rate constant depends on the lubricant evaporation rate and the microgravity environment. The temperature influence factor in the lubricating film degradation model is a linear function, equal to 1 plus a temperature sensitivity coefficient multiplied by the difference between the current temperature and the ambient reference temperature. The temperature sensitivity coefficient quantifies the degree to which temperature corrects the lubricating film degradation rate; a typical value for the temperature sensitivity coefficient is 0.02 degrees Celsius, and the ambient reference temperature is 25 degrees Celsius. The lubricating film degradation model... The specific calculation process is as follows: First, determine the initial film thickness, which is determined by the lubricating film thickness threshold constraint in step S1, and is usually taken as 0.15 micrometers; then, calculate the time variable based on the running time recorded by the on-orbit timer; the degradation rate constant is determined based on the lubricant evaporation test data, and in this embodiment, it is taken as 0.001 per hour; the temperature influence factor is calculated based on the current temperature sensor reading. For example, when the current temperature is 80 degrees Celsius, the temperature influence factor is equal to 1 plus 0.02 multiplied by the difference between 80 and 25, and the calculation result is 2.1; finally, multiply the initial film thickness by the negative degradation rate constant of the natural constant multiplied by the time power and then multiplied by the temperature influence factor to obtain the current film thickness; The bearing contact mechanics model calculates the pressure distribution and elastic deformation in the contact area based on Hertzian theory, and derives the nonlinear mapping relationship between the vibration amplitude of the steel ball and the preload, film thickness, and temperature. This mapping is based on the response surface approximation model from step S2, incorporating correction terms for temperature and film thickness. The specific construction process of the nonlinear mapping relationship is as follows: using the response surface approximation model constructed in step S2 as the baseline model, this model takes structural parameters such as the steel ball diameter and raceway curvature radius as input, and outputs the vibration amplitude of the steel ball. Based on this, two correction factors are added: the first correction factor is the film thickness correction coefficient, etc. The ratio of the current film thickness to the initial film thickness reflects the amplification effect of lubricant film degradation on vibration amplitude. The second correction factor is the temperature correction factor, which is equal to the sum of the temperature sensitivity factor and the temperature change. This factor reflects the influence of the material stiffness change caused by temperature change on vibration amplitude. The vibration amplitude output by the reference model is multiplied by the film thickness correction factor and then by the temperature correction factor to obtain the corrected vibration amplitude. This mapping relationship is implemented in the digital twin by a lookup table. The table resolution is one Newton per interval of preload, 0.01 micrometer per interval of film thickness, and ten degrees Celsius per interval of temperature. The digital twin uses the optimal parameter combination output in step S2 and the standardized assembly clearance recommendation value output in step S3 as fixed structural parameters. With time as the driving variable, it outputs a state vector in real time. The state vector includes the current vibration amplitude of the steel ball, the current film thickness, the current temperature, the preload from the previous moment, and the equivalent clearance offset caused by temperature. The six components of the state vector are arranged in the digital twin in the following order: the first component is the current vibration amplitude of the steel ball, calculated by a nonlinear mapping relationship; the second component is the current film thickness, calculated by the lubricating film degradation model; the third component is the current temperature, read in real time by a temperature sensor; the fourth component is the preload from the previous moment, read from the onboard computer's memory; the fifth component is the equivalent clearance offset of the inner ring, calculated by the temperature-clearance drift model in step S3; and the sixth component is the equivalent clearance offset of the outer ring, also calculated by the temperature-clearance drift model. The loading parameter adjustment problem is modeled as a discrete-time Markov decision process. The action space is defined to include the preload adjustment and the loading uniformity adjustment. The reward function is defined as the weighted sum of the steel ball passing vibration amplitude penalty, the lubricating film thickness lower limit violation penalty, the action amplitude penalty, and the action frequency penalty. The weight coefficients of each penalty in the reward function are as follows: the weight coefficient of the steel ball passing vibration amplitude penalty is 0.5, the weight coefficient of the lubricating film thickness lower limit violation penalty is 0.3, the weight coefficient of the action amplitude penalty is 0.15, and the weight coefficient of the action frequency penalty is 0.05. The calculation method for the steel ball vibration amplitude penalty term is as follows: divide the current vibration amplitude by the optimal vibration amplitude after convergence in step S2, and then multiply by the weighting coefficient; the calculation method for the lubricating film thickness lower limit violation penalty term is as follows: if the current film thickness is less than the lubricating film thickness safety threshold (taken as 0.1 micrometer), then the penalty term is equal to the safety threshold minus the current film thickness and then multiplied by the weighting coefficient, otherwise it is zero; the motion amplitude penalty term is equal to the absolute value of the preload adjustment multiplied by the weighting coefficient; the motion frequency penalty term is equal to the value of the indicator function multiplied by the weighting coefficient, where the indicator function is one when the motion vector is not zero, and zero otherwise; The specific operation of training the intelligent parameter adjustment strategy model using a deep Q-network in step S4 is as follows: A deep Q-network was constructed, consisting of a three-layer fully connected neural network. The input layer dimension equals the state vector dimension, and the number of neurons in the hidden layers are 128 and 64, respectively. The activation function is a linear rectified function, and the output layer dimension equals the number of actions after discretization of the action space. The action space is discretized as follows: the preload adjustment ranges from -5 Newtons to +5 Newtons, with each Newton interval representing a discrete action, resulting in 11 discrete values; the loading uniformity adjustment takes three discrete values: -0.1, 0, and +0.1; the combination of these two values results in 33 discrete actions. The neural network weights are initialized using the Kaiming He initialization method, and the biases are initialized to zero. The optimizer uses an adaptive moment estimation algorithm, with the learning rate set to 0.001, the first-order moment decay coefficient set to 0.9, and the second-order moment decay coefficient set to 0.999. In the digital twin, the state is randomly initialized, the simulation duration is set, and a greedy strategy is used to select actions. The greedy coefficient decreases linearly from 1 to 0.01. The simulation duration is set to 5,000 steps, with each step corresponding to one hour of on-orbit operation. The specific method of the greedy coefficient decay is as follows: the decay amount per step is equal to the difference between 1 and 0.01 divided by 2,000, that is, the decay amount per step is 0.00495. After 2,000 steps, the greedy coefficient drops to 0.01 and then remains constant. When the state is randomly initialized, each component of the state vector is randomly sampled from the distribution of multi-condition test data in step S3. Among them, the vibration amplitude is uniformly sampled from 0 to 1 micrometer, the film thickness is uniformly sampled from 0.1 to 0.3 micrometers, the temperature is uniformly sampled from -50 degrees Celsius to 120 degrees Celsius, the preload is uniformly sampled from 0 to 50 Newtons, and the gap offset is uniformly sampled from -5 micrometers to +5 micrometers. Each interaction generates a current state, selected action, immediate reward, and next state, which are stored as an experience tuple in the experience replay buffer. Every fixed number of steps, a batch of experience tuples is randomly sampled from the experience replay buffer, and the temporal difference error is calculated. The Huber loss function is used to improve training stability, with a discount factor set to 0.99. The target network parameters are copied from the current network parameters every 500 steps. Training continues until the cumulative reward converges, and the network parameters are saved as a loading parameter intelligent adjustment strategy model. The capacity of the experience replay buffer is set to 10,000 experience tuples. When the buffer is full, new tuples overwrite the oldest tuples. Every 100 steps, 256 experience tuples are randomly sampled from the buffer as a batch. The threshold of the Huber loss function is set to 1; when the absolute value of the temporal difference error is less than 1, squared error is used; otherwise, linear error is used. The criterion for cumulative reward convergence is: the average reward change is less than 1% over 100 consecutive rounds, at which point training stops. The average reward is recorded every 10 rounds during training. If the average reward continues to rise, training continues, up to a maximum of 1,000 rounds. The specific operation in step S4 of deploying the intelligent adjustment strategy model of loading parameters to the onboard computer and forming a closed-loop adaptive vibration suppression is as follows: The trained deep Q-network is converted into a greedy policy and deployed to the onboard computer of the spacecraft. The greedy policy means that for any input state vector, the action that maximizes the output value of the deep Q-network is selected as the action to be executed, and random exploration is no longer performed. During deployment, the trained network parameters are burned into the read-only memory of the onboard computer in the form of floating-point arrays. The processor of the onboard computer is a radiation-hardened field-programmable gate array with parallel computing capabilities, which can complete a forward inference within ten milliseconds. During on-orbit operation, the bearing vibration signal is collected by the onboard accelerometer in each preset control cycle. The current vibration amplitude of the steel ball is extracted by the time-frequency analysis in step S1. The preset control cycle is set to ten minutes, that is, the status acquisition and action decision are performed once every ten minutes. The sampling frequency of the onboard accelerometer is set to 10 kHz, the acquisition time is 0.1 seconds, and a total of 1,000 data points are collected. The specific process of time-frequency analysis is the same as that in step S1, including short-time Fourier transform and wavelet packet decomposition to extract the vibration amplitude of the steel ball. The system reads temperature sensor data and calculates the equivalent gap offset using the temperature-gap drift model from step S3. It reads the preload force from the onboard computer's memory at the previous moment and calls the lubricating film degradation model to estimate the current film thickness. The temperature sensor is a platinum resistance temperature sensor with an accuracy of ±0.5 degrees Celsius. The calculation process for the equivalent gap offset is as follows: subtract 25 degrees Celsius from the temperature sensor reading to obtain the temperature change; multiply the difference in the inner ring thermal expansion coefficients (1.2 times 10 to the power of -5 per degree Celsius) by the temperature change to obtain the inner ring equivalent gap offset in micrometers; multiply the difference in the outer ring thermal expansion coefficients (0.8 times 10 to the power of -5 per degree Celsius) by the temperature change to obtain the outer ring equivalent gap offset. The process of estimating the current film thickness using the lubricating film degradation model is as follows: read the initial film thickness (0.15 micrometers) and degradation rate constant (0.001 per hour) from the onboard computer's memory, read the cumulative running time from the on-orbit timer, and substitute them into the exponential decay formula to calculate the current film thickness. A state vector is formed and input into a greedy policy network. The network outputs the optimal action, including the preload adjustment and the loading uniformity adjustment. The six components of the state vector are concatenated into a one-dimensional array in sequence and input into the greedy policy network. The network outputs thirty-three action values. The action index corresponding to the maximum value is taken and decoded to obtain the preload adjustment and the loading uniformity adjustment. The unit of the preload adjustment is Newton, and the unit of the loading uniformity adjustment is a dimensionless proportionality coefficient. Preload adjustment is performed via a piezoelectric actuator, and load uniformity adjustment is performed via a fine-tuning spring shim. The piezoelectric actuator has a stroke range of -10 micrometers to +10 micrometers, corresponding to a preload adjustment range of -5 Newtons to +5 Newtons. The control voltage is 0 to 100 volts, and the preload adjustment amount is converted into a voltage command via a digital-to-analog converter. The fine-tuning spring shim is adjusted via a stepper motor with a step angle of 1.8 degrees. Each step corresponds to a load uniformity adjustment of 0.01, and the stepper motor rotation angle is controlled by a pulse width modulation signal. Simultaneously, the vibration amplitude of the steel ball measured in orbit is compared with the predicted vibration amplitude of the steel ball output by the digital twin. If the deviation exceeds a preset threshold, the degradation rate constant and temperature sensitivity coefficient in the digital twin model are corrected online. The degradation rate constant and temperature sensitivity coefficient are updated using recursive Bayesian estimation to ensure that the digital twin is always aligned with the actual degradation trajectory. The preset threshold is set to 10%, meaning that correction is triggered when the absolute value of the difference between the measured value and the predicted value divided by the predicted value is greater than 0.1. The specific process of recursive Bayesian estimation is as follows: it is assumed that the degradation rate constant and temperature sensitivity coefficient follow a Gaussian distribution, with initial means of 0.001 per hour and 0.02 per degree Celsius, and initial variances of 0.001. Each time correction is triggered, the difference between the current measured vibration amplitude and the predicted vibration amplitude is used as the observation value, the Kalman gain is calculated, and the means and variances of the two parameters are updated. After correction, the updated parameters are written into the digital twin model for prediction in the next cycle. This correction mechanism ensures that the digital twin remains synchronized with the actual state during the degradation of the lubricating film.
[0026] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0027] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0028] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0029] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0030] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0031] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0032] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A bearing vibration suppression method based on the local mesh method, characterized in that, Specifically, the steps include the following: Step S1: Acquire broadband vibration signals of the bearing operation, separate the orbital frequency of the steel ball and its harmonic components through short-time Fourier transform and wavelet packet decomposition, extract the vibration amplitude, phase and modulation characteristics of the steel ball, and establish a multi-objective optimization function with the goal of minimizing the vibration amplitude of the steel ball. Calculate the vibration sensitive direction based on the gradient of the multi-objective optimization function, and determine the parameter vector to be optimized and its constraints. The constraints include the lubricating film thickness threshold determined based on the lubricating film degradation model and the bearing contact mechanics model. Step S2: Divide the parameter space corresponding to the parameter vector to be optimized according to the vibration sensitive direction into a non-uniform grid. In the region where the gradient modulus exceeds the preset refinement threshold, refine the grid. Construct an approximate response surface model by Latin hypercube sampling. Then, use a hybrid solver of genetic algorithm and sequential quadratic programming to iteratively search for the optimal parameter combination on the local grid until the steel ball converges to the preset convergence threshold through vibration amplitude. Step S3: Using the optimal parameter combination as the baseline, collect test data of inner and outer ring fit clearance combinations under multiple working conditions to construct a clearance-vibration multidimensional mapping database. Use Gaussian process regression to fit the mapping relationship and search for the global optimal clearance combination through Bayesian optimization to output the standardized assembly clearance recommendation value and temperature-clearance drift model. Step S4: Couple the lubricating film degradation model with the bearing contact mechanics model to construct a digital twin. With the optimal parameter combination and the recommended value of standardized assembly clearance as the initial state, use a deep Q network to train and obtain the intelligent adjustment strategy model of loading parameters and deploy it to the onboard computer of the spacecraft. Based on the on-orbit real-time vibration monitoring data obtained by the onboard accelerometer of the spacecraft, dynamically output the loading parameter adjustment command to form a closed-loop adaptive vibration suppression.
2. The bearing vibration suppression method according to claim 1, characterized in that: The specific steps for extracting the vibration amplitude, phase, and modulation characteristics of the steel ball in step S1 are as follows: A short-time Fourier transform is applied to the broadband vibration signal, with a Hanning window selected as the window function and the window length set to twice the orbital period of the steel ball, to obtain the time spectrum; Simultaneously, the broadband vibration signal is subjected to three-level wavelet packet decomposition, and the Daubechies8 wavelet basis is used to obtain eight sub-band signals generated by wavelet packet decomposition. Based on the pre-calculated theoretical value of the steel ball's orbital frequency, energy ridges with frequencies within the range of the steel ball's orbital frequency plus or minus half a frequency resolution and its integer harmonics plus or minus half a frequency resolution are extracted from the time-frequency spectrum; the time-frequency coefficients corresponding to the energy ridges are mapped to sub-bands, and the pure vibration signal of the steel ball is reconstructed. The vibration amplitude of the steel ball is extracted from the reconstructed vibration signal, and the maximum value of the absolute value of the signal is taken as the vibration amplitude of the steel ball. Extract the phase, which is the arctangent value of the Fourier transform result at the orbital frequency of the steel ball; The modulation features are extracted, and the modulation index is defined as the ratio of the sum of the amplitudes of the sideband components on both sides of the ball's orbital frequency to the amplitude at the ball's orbital frequency, where the sideband spacing is the hold's rotational frequency.
3. The bearing vibration suppression method according to claim 2, characterized in that: The specific steps in step S1 to establish the multi-objective optimization function and calculate the vibration-sensitive direction are as follows: The parameter vector to be optimized includes five parameters: steel ball diameter, inner raceway radius of curvature, outer raceway radius of curvature, retainer window clearance, and fit clearance allowance. A multi-objective optimization function is established with the goal of minimizing the weighted sum of the vibration amplitude and modulation index of the steel ball. The sum of the weight coefficients is one. The steel ball achieves dimensionlessness by dividing its vibration amplitude and modulation index by their respective reference values under the initial parameter combination. The constraints include the bearing's rated dynamic load not exceeding the allowable value, the limiting speed not exceeding the allowable value, and the lubricating film thickness not less than the lubricating film thickness threshold, wherein the lubricating film thickness threshold is obtained by jointly solving the lubricating film degradation model and the bearing contact mechanics model; Multiple sample points are randomly selected in the parameter space to be optimized, and the objective function value of each sample point is calculated. A radial basis function interpolation model is constructed using the Gaussian kernel function as the basis function. The analytical gradient of the radial basis function interpolation model is obtained to obtain the vibration sensitive direction. Each component of the vibration sensitive direction represents the intensity and direction of the comprehensive influence of the corresponding parameter change on the objective function.
4. The bearing vibration suppression method according to claim 3, characterized in that: The specific steps in step S2, which involve dividing the mesh into non-uniform grids based on the vibration-sensitive direction and constructing an approximate response surface model, are as follows: Calculate the gradient magnitude at each sample point based on the vibration-sensitive direction output in step S1, set the preset refinement threshold to 30% of the global maximum gradient magnitude, mark the region with a gradient magnitude greater than or equal to the preset refinement threshold as a vibration-sensitive region, increase the mesh density to four times the initial density in the vibration-sensitive region, and maintain the initial mesh density at the beginning of step S2 in the region not marked as a vibration-sensitive region. Two hundred sample points are extracted at the center of each cell in the non-uniform grid using the Latin hypercube sampling method. For each sample point, step S1 is performed to extract the steel ball and calculate the corresponding objective function value by operating on the vibration amplitude, phase and modulation characteristics. These sample points and their objective function values are used as training data. A radial basis function response surface approximation model is constructed using the Gaussian kernel function as the basis function. The shape parameter of the Gaussian kernel function is taken as the reciprocal of the average Euclidean distance between all pairs of sample points.
5. The bearing vibration suppression method according to claim 4, characterized in that: The specific operation of step S2, which uses a hybrid solver combining genetic algorithm and sequential quadratic programming to iteratively search for the optimal parameter combination, is as follows: Using the response surface approximation model as the fitness function, a genetic algorithm is run on all nodes of a non-uniform grid to perform a global coarse search and select the ten individuals with the highest fitness as candidate solutions. The sequential quadratic programming algorithm is started with each candidate solution as the initial point to perform a local fine search. The sequential quadratic programming algorithm gradually approximates the local optimum by constructing a quadratic approximation model of the Lagrange function. The termination condition is that the relative change of the objective function value between two adjacent iterations is less than 10 to the power of negative 6 or the number of iterations exceeds 100. The solution with the smallest objective function value among all local optima is taken as the optimal parameter combination. It is then determined whether the vibration amplitude of the steel ball corresponding to the optimal parameter combination is less than the preset convergence threshold. If the threshold is met, the iteration stops. If not, the mesh near the optimal parameter combination is further refined and the process returns to the Latin hypercube sampling step in step S2 to reconstruct the response surface model and continue iterating.
6. The bearing vibration suppression method according to claim 5, characterized in that: The specific steps for collecting multi-condition test data and constructing a Gaussian process regression model in step S3 are as follows: Using the optimal parameter combination output in step S2 as a fixed baseline, multiple sets of bearing samples are manufactured. The inner ring and shaft fit clearance and the outer ring and bearing housing fit clearance of each set of bearing samples are orthogonally designed within a preset range, with a step size of one micrometer. For each set of gaps, bearing operation tests were conducted under three temperature conditions: -50 degrees Celsius, 25 degrees Celsius, and 120 degrees Celsius. The vibration amplitude, running resistance torque, and coaxiality deviation of the steel ball were measured, and the vibration amplitude, running resistance torque, and coaxiality deviation of the steel ball were used as output vectors. All experimental data were compiled into a dataset. The input vector contained three dimensions: inner ring fit clearance, outer ring fit clearance, and temperature. The output vector contained three dimensions: vibration amplitude, drag torque, and coaxiality. Multi-task Gaussian process regression is used to simultaneously fit the correlation between the three output variables. The joint covariance function is obtained by multiplying the kernel matrix between the output tasks with the Matrn kernel function in the parameter space corresponding to the input vector. The Matrn kernel function is selected as the type with an order of 2 / 5. The length scaling parameter and signal variance of the Matrn kernel function are estimated by maximizing the log marginal likelihood function. The mean function is set as a constant, and all hyperparameters are estimated by maximizing the log marginal likelihood function.
7. The bearing vibration suppression method according to claim 6, characterized in that: The specific steps in step S3, which involve searching for the globally optimal gap combination using Bayesian optimization and outputting the temperature-gap drift model, are as follows: The temperature is fixed at 25 degrees Celsius. The optimization objective is to minimize the predicted mean of the vibration amplitude of the steel ball. The expected improvement is used as the acquisition function. The expression of the expected improvement is calculated based on the minimum vibration amplitude in the currently acquired samples and the predicted distribution of the Gaussian process at the new input, which includes the predicted mean and the predicted standard deviation. Run Bayesian optimization in a two-dimensional gap space. In each iteration, select a new gap combination that maximizes the expected improvement, add it to the dataset and update the Gaussian process model. Repeat the iteration until the maximum number of iterations is reached or the maximum expected improvement is lower than the preset expected improvement stopping threshold, and output the globally optimal gap combination. Using a trained multi-task Gaussian process model, temperature is taken as a variable input at the globally optimal gap combination to calculate the predicted vibration amplitude at different temperatures, and a temperature-vibration amplitude drift model is constructed. This temperature-vibration amplitude drift model adopts a second-order Taylor expansion approximation, and the coefficients are obtained by numerical differentiation of the Gaussian process prediction value in the normal temperature neighborhood. Simultaneously, based on the difference in thermal expansion coefficients between the inner ring and the shaft, and between the outer ring and the bearing housing, the equivalent clearance offset caused by temperature is calculated, and the recommended value for standardized assembly clearance and the temperature-clearance drift model are output.
8. The bearing vibration suppression method according to claim 7, characterized in that: The specific operations for constructing the digital twin and defining the Markov decision process in step S4 are as follows: A digital twin is constructed by coupling the lubricating film degradation model with the bearing contact mechanics model. The lubricating film degradation model describes the exponential decay of the oil film thickness over time. The initial film thickness is determined by the lubricating film thickness threshold constraint in step S1. The degradation rate constant depends on the lubricant evaporation rate and the microgravity environment. The temperature influence factor in the lubricating film degradation model is a linear function. The temperature influence factor is equal to one plus the temperature sensitivity coefficient multiplied by the difference between the current temperature and the room temperature reference temperature. The temperature sensitivity coefficient is used to quantify the degree of correction of the lubricating film degradation rate by temperature. The bearing contact mechanics model calculates the pressure distribution and elastic deformation in the contact area based on Hertz theory, and derives the nonlinear mapping relationship between the vibration amplitude of the steel ball and the preload, film thickness, and temperature. The mapping is based on the response surface approximation model in step S2 and incorporates correction terms for temperature and film thickness. The digital twin uses the optimal parameter combination output in step S2 and the standardized assembly gap recommendation value output in step S3 as fixed structural parameters. With time as the driving variable, it outputs a state vector in real time. The state vector includes the vibration amplitude of the steel ball at the current moment, the current film thickness, the current temperature, the preload at the previous moment, and the equivalent gap offset caused by temperature. The loading parameter adjustment problem is modeled as a discrete-time Markov decision process. The action space is defined to include the preload adjustment and the loading uniformity adjustment. The reward function is defined as the weighted sum of the steel ball passing through vibration amplitude penalty term, lubricating film thickness lower limit violation penalty term, action amplitude penalty term and action frequency penalty term.
9. The bearing vibration suppression method according to claim 8, characterized in that: The specific operation of training the intelligent parameter adjustment strategy model using a deep Q-network in step S4 is as follows: A deep Q-network is constructed. The deep Q-network architecture is a three-layer fully connected neural network. The dimension of the input layer is equal to the dimension of the state vector. The number of neurons in the hidden layer is 128 and 64 respectively. The activation function is a linear rectified function. The dimension of the output layer is equal to the number of actions after the action space is discretized. The state is randomly initialized in the digital twin, the simulation duration is set, and a greedy strategy is used to select actions, with the greedy coefficient decreasing linearly from one to zero. The current state, selected action, immediate reward and next state obtained from each interaction are stored as an experience tuple in the experience replay buffer. Every fixed number of steps, a batch of experience tuples are randomly sampled from the experience replay buffer, and the temporal difference error is calculated. The temporal difference error adopts the Huber loss function to improve training stability. The target network parameters are copied from the current network parameters every 500 steps. Training continues until the cumulative reward converges, and the network parameters are saved as loading parameters for intelligent adjustment of the strategy model.
10. The bearing vibration suppression method according to claim 9, characterized in that: The specific operation in step S4 of deploying the intelligent adjustment strategy model of loading parameters to the onboard computer and forming a closed-loop adaptive vibration suppression is as follows: The trained deep Q-network is converted into a greedy policy and deployed to the onboard computer carried by the spacecraft. During in-orbit operation, the bearing vibration signal is collected by the onboard accelerometer in each preset control cycle, and the current vibration amplitude of the steel ball is extracted by time-frequency analysis in step S1. Read the temperature sensor data and calculate the equivalent gap offset using the temperature-gap drift model from step S3; read the preload force from the memory of the onboard computer at the previous moment and call the lubricating film degradation model to estimate the current film thickness; The state vector is formed and input into the greedy policy network, and the optimal action is output, including the preload adjustment amount and the loading uniformity adjustment amount; the preload adjustment is performed by the piezoelectric actuator, and the loading uniformity adjustment is performed by the fine-tuning spring shims; Simultaneously, the vibration amplitude of the steel ball measured in orbit is compared with the predicted vibration amplitude of the steel ball output by the digital twin. If the deviation exceeds the preset threshold, the degradation rate constant and temperature sensitivity coefficient in the digital twin model are corrected online. The degradation rate constant and temperature sensitivity coefficient are updated by recursive Bayesian estimation to ensure that the digital twin is always aligned with the actual degradation trajectory.
