Aviation bearing health evolution change point online identification and residual life prediction method
By employing a sliding window mechanism and Welch's t hypothesis testing method, combined with stochastic process modeling, online identification of health evolution change points and prediction of remaining life of aerospace bearings were achieved. This solved the problems of detection delay and prediction bias in multi-stage degradation processes, and improved the accuracy and reliability of predictions.
Patent Information
- Application Number
- CN202511752034.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-17
AI Technical Summary
Existing technologies struggle to accurately identify turning points and effectively predict the remaining life of aerospace bearings during their healthy evolution, especially in the case of detection delays and prediction biases during multi-stage degradation.
We employ Welch's t hypothesis testing under a sliding window mechanism, combined with online change point detection and stochastic process modeling. By constructing a sliding window parameter optimization framework, we achieve online identification of change points in the health evolution of aerospace bearings and prediction of their remaining life.
It improves the accuracy of identifying health evolution change points in aerospace bearings and the reliability of remaining life prediction, reduces prediction uncertainty, and supports the reliability and safety control of aerospace systems.
Smart Images
Figure CN121683206A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for online identification of health evolution change points and prediction of remaining life of aerospace bearings. By designing a hypothesis testing method based on Welch's t under a sliding window mechanism, the method can identify health evolution change points of aerospace bearings online. Furthermore, by constructing three performance indicators to optimize the sliding window parameters, the method uses stochastic process theory to predict the remaining life including change points, thereby helping to achieve effective control of the reliability and safety of aerospace bearings. This invention belongs to the field of equipment reliability engineering. Background Technology
[0002] As a key supporting component of the aero-engine rotor system, aero-engine bearings bear the important functions of supporting the rotor weight, transmitting loads, and ensuring the precise rotation of the rotor system. Their typical structure consists of an inner ring, outer ring, rolling elements, and a cage. Under the continuous operation of complex conditions such as high-speed operation (tens of thousands of revolutions per minute), high-temperature environments (hundreds of degrees Celsius), and variable loads, the health evolution process of aero-engine bearings exhibits a distinct "multi-stage" characteristic. During normal operation, the bearing is in a relatively stable working state, and its health evolution is a slow and steady degradation process. However, when it enters the defect stage such as raceway fatigue, cage wear, or lubrication failure, the rate of health evolution changes significantly due to the cumulative effect of damage, showing an accelerated deterioration trend. If these key turning points in the health evolution process are not identified in a timely and accurate manner, it will not only affect the reliability of the bearing itself but also lead to systematic deviations in the remaining life prediction, thereby affecting the reliability, safety, and maintenance economy of the entire aero-engine and even the whole aircraft.
[0003] Currently, several technical challenges remain in the detection of health evolution and prediction of remaining life for aero-engine bearings. On the one hand, most methods for detecting health evolution change points employ offline analysis, making it difficult to provide real-time condition assessment support for aero-engine health management. On the other hand, existing online detection methods are typically based on traditional hypothesis testing theory, which assumes that the volatility at different stages of health evolution remains constant. However, in practical engineering, due to multiple factors such as changes in operating conditions and material performance degradation, the diffusion characteristics of aero-engine bearing health evolution often change during stage transitions, limiting the applicability of traditional detection methods. Furthermore, most current remaining life prediction methods are based on assumptions about a single health evolution stage, failing to fully consider the changes in degradation characteristics after stage transitions, which may lead to prediction biases in practical engineering applications. Therefore, it is necessary to research a health evolution change point detection and life prediction method that can adapt to the characteristics of actual engineering, providing more reliable technical support for the health status assessment and maintenance decisions of aero-engine bearings. Summary of the Invention
[0004] The purpose of this invention is to provide an online identification method for health evolution change points and remaining life prediction of aerospace bearings. By designing and optimizing the Welch's t hypothesis testing method under the sliding window mechanism, it overcomes the problem of inaccurate identification of evolution change points in the prior art. Furthermore, by integrating online change point detection and stochastic process modeling, it overcomes the problem of insufficient accuracy in remaining life prediction including change points, and provides decision support for the reliability and safety control of aerospace systems.
[0005] The technical solution of the present invention is as follows: A method for online identification of health evolution change points and prediction of remaining life of aerospace bearings, comprising the following steps: Step 1: Data preprocessing and construction of aircraft bearing health indicators; specifically including the following steps: Figure 1 As shown; 1) Data preprocessing and health indicator construction: Real-time monitoring of raw vibration signals of aircraft bearings using sensors. The original signal is preprocessed (noise reduction, standardization), and then the root mean square (RMS) value is selected as its health indicator. The specific calculation method is as follows: (1) In the formula, It is the vibration signal sampling result within the set time window. This represents the number of sampling points. Based on this, a health index sequence for aerospace bearings is constructed. ,in, This refers to the current moment.
[0006] 2) Degradation increment sequence calculation: Calculate the difference in the RMS (Recovery Status Index) of the aircraft bearing at adjacent detection points to obtain its evolution increment sequence. ,satisfy .in, Indicates the number of times a regular test is conducted. Indicates the first One degradation increment, Indicates the first One degradation increment, and They represent the first Second and third The amount of system degradation during the next periodic inspection. Defined based on engineering experience. .
[0007] Step 2: Detection of health evolution change points based on sliding window hypothesis testing; specifically including the following steps: 1) Real-time setting of sliding window: for any detection time. Define two closely adjacent sliding windows. , Used for dynamically segmenting real-time evolution indicator data streams: Reference Window Define its size as That is, including Each historical evolution increment is primarily used to characterize the evolutionary pattern during the "normal" phase. The reference window for a given time is represented as follows: In the formula and They represent The first and last degradation increment values of the reference window are used at all times. and Based on the current periodic detection time and reference window length and detection window length The window boundaries are calculated together.
[0008] Detection window Define its size as That is, including Each historical evolution increment represents new feature data that needs to be used for testing. The detection window at a given time is represented as follows: In the formula and They represent The first and last degradation increment values of the detection window are checked at all times. and Based on the current periodic detection time and detection window length The window boundaries are calculated together.
[0009] 2) Welch's t-hypothesis testing model construction: The evolution of the aerospace bearing in the "normal" and "defective" stages is characterized by two linear Wiener processes. Based on engineering experience, the diffusion coefficients of the Wiener processes differ between the two stages. At the detection time... Construct statistics : (2) in, and These represent the mean values of the reference window and the detection window, respectively. and Let Variance represent the variance of the reference window and the variance of the detection window, respectively, calculated as follows: (3) Statistic It does not precisely follow a standard t-distribution. Its approximate degrees of freedom can be calculated using the following formula: (4) Hypothesis building: (That is, the mean values of the reference window and the detection window are equal, and there are no evolutionary change points.) (The mean of the detection window is significantly larger than that of the reference window, indicating an evolutionary inflection point.) 3) Evolutionary change point identification rule based on sliding window hypothesis testing: given significance level Calculate the critical values of the quantiles of the t-distribution at this confidence level. And make the following comparison: like If so, then the null hypothesis is rejected, and the aircraft bearing is determined to be in... At that moment, an evolutionary turning point appeared, meaning that after that point, the "defect stage" began; Otherwise, accept the null hypothesis, that is... The time system is still in the "normal phase" and will continue to make judgments at subsequent detection points.
[0010] Step 3: Performance evaluation of the variable point identification model and optimization of sliding window parameters; specifically including the following steps: 1) Offline sample partitioning and parameter space setting: Targeting similar aerospace bearings One sample, using Offline optimization of sliding window parameters was performed on each sample, and the lengths of the reference window and the detection window were set. , .
[0011] 2) Model performance evaluation metrics: using The evolution trajectory of a bearing, and the known true change point position are... Based on the given sliding window parameters ,exist The entire process of detecting health evolution change points based on sliding window hypothesis testing, following "Step Two," is performed on the evolution trajectory of each bearing, and the detection success rate is calculated. Mean absolute error and average detection delay The specific calculation method is as follows: (5) In the formula, and Parameters Down The number of trajectories in the track that successfully detected change points and the number of trajectories that detected change points with delay (success). and The first hThe actual change points and detected change points of the trajectory.
[0012] 3) Determine the optimal window parameters: [This refers to setting the detection success rate...] Mean absolute error and average detection delay Assign weights And calculate the overall score. Sc : (6) in, and Representing parameters respectively Down The maximum mean absolute error and maximum mean detection delay for identifying change points in a trajectory are calculated. The sliding window parameters corresponding to the highest score are determined using an enumeration method. That is, calculate all according to formula (6) respectively. The combined score of the combinations is used to select the window parameters corresponding to the highest-scoring combination as the optimal parameters, which are then fixed as the system parameters for this type of aerospace bearing. These parameters are then used for online identification of bearing evolution change points and prediction of remaining life in subsequent tests.
[0013] Step 4: Remaining lifetime prediction based on online identification of evolutionary change points; specifically including the following steps: 1) Construction of a stochastic process model for the evolution trajectory: The evolution process of the aerospace bearing in the normal and defective stages is modeled using a linear Wiener process model, i.e.: (7) in, and These represent the Wiener process drift coefficient and diffusion coefficient during the normal phase, respectively. and These represent the Wiener process drift coefficient and diffusion coefficient during the defect stage, respectively; and These are all times during the normal phase. and All of these are moments during the defect phase (i.e., after the change point occurs). and These represent two moments in the normal phase (i.e.) The corresponding system degradation amount, and These represent the two moments in the defect stage (i.e.) The corresponding system degradation amount.
[0014] 2) Online identification of bearing evolution change points: based on the optimal window parameters obtained in "Step 3". We conducted online detection of evolutionary change points of test samples to obtain the detected change points and the performance of the model.
[0015] 3) Derivation and prediction of remaining lifetime distribution: detection time Remaining lifespan According to the definition of "first arrival time", Represents random variables The possible values of are: (8) For aerospace bearings in the defect stage (i.e., after the change point occurs), under the evolutionary model framework of this invention, their remaining lifespan... It follows an inverse Gaussian distribution, and its probability density function and distribution function can be expressed as ( ). Indicates the first k (System degradation at each periodic detection time) (9) (10) in, Let be the cumulative distribution function of the standard normal distribution. Using the above formula, calculate... Point estimates and confidence interval estimates of remaining lifetime at time step 1. mean As a point estimate of remaining lifetime: (11) Furthermore, given a confidence level Calculate the prediction interval ,in and They represent The minimum and maximum predicted remaining lifetimes at a given confidence level at time t, satisfy: (12) In terms of prediction error evaluation, point estimation of remaining lifetime mainly considers three error indicators: mean absolute error (MAE), root mean square error (RMSE), and standard deviation of absolute error (SSE). The calculation method is as follows: (13) In the formula, and The first r The actual remaining lifetime and the predicted remaining lifetime at each prediction point The number of prediction points, For the first rThe absolute error of the prediction at each prediction point. Confidence interval estimation mainly considers two error metrics: confidence interval coverage and the standard deviation of the confidence interval width.
[0016] (14) The advantages and beneficial effects of this invention are as follows: ① This invention proposes an online identification method for health evolution change points and remaining life prediction for the multi-stage degradation characteristics of aerospace bearings, which effectively reduces the uncertainty of evolution change points on remaining life prediction; ② Under the condition of "non-homogeneous variance" in the multi-stage evolution model of aerospace bearings, this invention designs a sliding window Welch's t test method. Given a confidence level, the occurrence of change points is accurately determined by analyzing the difference between the means of the reference window and the detection window. ③ This invention constructs a sliding window hypothesis testing and model parameter optimization framework, which significantly improves the accuracy and precision of identifying evolutionary change points; ④ The method described in this invention is scientific, has good processability, and has broad application value. Attached Figure Description
[0017] Figure 1 This is a flowchart of the method described in this invention.
[0018] Figure 2 This is the two-stage evolution trajectory of 10 sample aerospace bearings in this embodiment of the invention.
[0019] Figure 3 The image shows the effect of variable point identification of bearings in the training set (9 samples) under the initial window parameters.
[0020] Figure 4 This is an optimized diagram of the reference window and detection window parameters in an embodiment of the present invention.
[0021] Figure 5 The image shows the effect of variable point identification of bearings in the training set (9 samples) under the optimal window parameters.
[0022] Figure 6 This is a diagram showing the effect of testing the evolution change point detection and remaining life prediction of the bearing in an embodiment of the present invention. Detailed Implementation
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] This invention provides a method for online identification of health evolution change points and prediction of remaining life of aerospace bearings, which is implemented as follows; Step 1: Data preprocessing and construction of aircraft bearing health indicators; specifically including the following steps: This step generated a sequence of RMS health indicators for 10 aerospace bearings. , For the current moment, such as Figure 2 As shown. Furthermore, the evolutionary increment sequence was calculated. .
[0025] Step 2: Detection of health evolution change points based on sliding window hypothesis testing; specifically including the following steps: 1) Real-time setting of sliding window: for any detection time. Set the initial length of the sliding window. and According to the formula and Real-time construction of reference and detection windows.
[0026] 2) Welch's t-hypothesis testing model construction: at any detection time... Based on the constructed reference window and detection window, the hypothesis test statistic at this moment is calculated according to formulas (2) and (3). Further, the approximate degrees of freedom of this statistic are calculated using formula (4). Construct a hypothesis testing model.
[0027] 3) Evolutionary change point identification rule based on sliding window hypothesis testing: In the embodiments of this invention, a significance level is given. Use MATLAB to calculate the critical value at this confidence level. and compare and Size relationship: If This indicates the presence of an evolutionary change point; otherwise, it indicates the absence of an evolutionary change point. Based on the initial window parameters, the change point identification results for the nine sample bearings are obtained, such as... Figure 3 As shown. According to Figure 3 Get, in and Under the given parameter settings, the variable point recognition effect was not ideal, so subsequent window parameter optimization steps were performed.
[0028] Step 3: Performance evaluation of the variable point identification model and optimization of sliding window parameters; specifically including the following steps: 1) Offline sample segmentation and parameter space setting: For 10 samples of similar aerospace bearings, 9 samples were used for offline optimization of sliding window parameters, and the length range of the reference window and the detection window was set. , The step size for each step is 2, which is used for subsequent parameter optimization.
[0029] 2) Model performance evaluation metrics: Using the evolution trajectories of 9 sample bearings, the true change point positions are known to be 38. Based on the given sliding window parameters... The entire process of detecting health evolution change points based on sliding window hypothesis testing, which runs on the evolution trajectory of 9 sample bearings in "Step Two", is performed. The parameter combination is calculated according to formula (5). The success rate of detection Mean absolute error and average detection delay .
[0030] 3) Determine the optimal window parameters: allocate detection success rate Mean absolute error and average detection delay The weights are respectively And calculate the comprehensive score according to formula (6). Sc :
[0031] The sliding window parameters corresponding to the highest score are determined through calculation. As the optimal parameter (see parameter optimization process), Figure 4 As shown in the figure, it is fixed as the system parameter of this type of aircraft bearing, and used to test the online identification of bearing evolution change points and the prediction of remaining life. Figure 5 The results of online variable point detection for nine sample bearings under optimal window parameters are shown in Table 1.
[0032] Table 1. Detection effect of variable point test on bearing set (under optimal window parameters)
[0033] Step 4: Remaining lifetime prediction based on online identification of evolutionary change points; specifically including the following steps: 1) Construction of a stochastic process model for the evolution trajectory: Using the linear Wiener process, i.e., formula (7), the evolution process of the aerospace bearing in the normal and defective stages is constructed for predicting the remaining life of the bearing when it is in the defective stage. In this embodiment of the invention, the step of estimating the Wiener process parameters is simplified. The linear Wiener process parameters of this type of bearing in the defective stage are known to be... The evolutionary failure threshold for this type of bearing is .
[0034] 2) Online identification of evolutionary change points of the test bearing: For the sample bearing used for testing (i.e., bearing 10), under the optimal window parameters obtained in "Step 3", online identification of the evolutionary change points of the test bearing is carried out, and the detected change points are... The change point recognition effect is as follows Figure 6 As shown in (a), the success rate and accuracy of change point recognition for this test sample are both 100%.
[0035] 3) Prediction of remaining life of test bearings based on change point detection: from the detection change point of bearing 10 back Starting from a given time point, select 26 prediction points, and assign a confidence level. The remaining life point estimate and confidence interval estimate of the aircraft bearing are performed using formulas (11) and (12) respectively, and the results are as follows: Figure 6 As shown in (b). The prediction effect was evaluated according to formulas (13) and (14), and the results are shown in Table 2. Based on the data in the charts, it can be seen that after accurately identifying the health evolution change point of the bearing, its remaining life prediction effect is excellent, and all of them are "conservative" prediction results, which is of great reference for subsequent predictive maintenance and other control links.
[0036] Table 2: Prediction effect of remaining life of test bearings
[0037] In summary, this invention provides an online identification method for health evolution change points and remaining life prediction of aerospace bearings. The core innovations of this method are: First, by employing the Welch's t hypothesis testing method under a sliding window mechanism, accurate online identification of health evolution change points in aerospace bearings is achieved, effectively overcoming the problem of change point detection delay in traditional methods. Second, a degradation model based on a multi-stage linear Wiener process is constructed, accurately describing the evolutionary characteristics of aerospace bearings from the normal stage to the defect stage. Furthermore, a sliding window parameter optimization framework is established; by comprehensively evaluating the detection success rate, mean absolute error, and mean detection delay, the optimal window parameters are determined, significantly improving the accuracy and timeliness of change point detection. Finally, based on the inverse Gaussian distribution, the complete probability distribution of remaining life is derived, enabling point and interval estimation of remaining life, providing a reliable basis for predictive maintenance of aerospace bearings. Compared with existing technologies, this invention not only identifies health evolution change points in a timely and accurate manner but also improves the reliability of remaining life prediction, providing effective technical support for safety control and maintenance decisions in aerospace systems, and has significant application value.
Claims
1. An aero-bearing health evolution change point online identification and residual life prediction method, characterized in that: The steps are as follows: Step one: data preprocessing and health index construction of aero-bearing; including data preprocessing and health index construction and degradation increment sequence calculation; Step two: health evolution change point detection based on sliding window hypothesis testing; Including sliding window real-time setting, Welch's t hypothesis testing model construction and evolution change point identification rule based on sliding window hypothesis testing; Step three: performance evaluation of change point identification model and sliding window parameter optimization; Including offline sample division and parameter space setting, model performance evaluation index and determination of optimal window parameter; Step four: residual life prediction based on online identification of evolution change point; Including the construction of the random process model of the evolution trajectory, Online identification of evolution change points of test bearings and derivation and prediction of residual life distribution.
2. The method according to claim 1, characterized in that: In step one, data preprocessing and health index construction, specifically: using sensors to monitor the original vibration signal of the aviation bearing in real time The original signal is preprocessed, and the root mean square value RMS is selected as the health index, and the calculation method is: (1) In the formula, is a vibration signal sampling result in a set time window, is a sampling point number; a health index sequence of the aviation bearing is constructed as , wherein, is a current time.
3. The method according to claim 1 or 2, characterized in that: In step one, the degradation increment sequence is calculated, specifically by calculating the difference in the health index of the aircraft bearing at adjacent detection points to obtain the evolution increment sequence. ,satisfy ;in, Indicates the number of times a regular test is conducted. Indicates the first One degradation increment, Indicates the first One degradation increment, and They represent the first Second and third The amount of system degradation during the next periodic inspection; defined based on engineering experience. .
4. The method of claim 1, wherein the method further comprises: In step two, the sliding window is set in real time, specifically: for any detection time , two closely adjacent sliding windows , are defined for dynamic segmentation of real-time evolving index data streams: Reference window : defining size of , i.e. containing history evolution increments, for representing evolution pattern of normal stage; Reference window at time t is represented as: , where and represent the first and last degradation increment values of reference window at time t, and are calculated based on current periodic detection time and reference window length and detection window length and are jointly calculated window boundaries. Detection window Define size as That is, including Each historical evolution increment represents new feature data that needs to be used for testing; The detection window at a given time is represented as follows: In the formula and They represent The first and last degradation increment values of the detection window are checked at all times. and Based on the current periodic detection time and detection window length The window boundaries are calculated together.
5. The method of claim 1, wherein: In step two, Welch's t hypothesis testing model is constructed, specifically: at the detection time Constructing the statistic : (2) wherein, and and denote the mean of the reference window and the detection window, respectively, and and denote the variance of the reference window and the detection window, respectively, calculated as follows: (3) statistic does not follow the standard t distribution, representing its approximate degrees of freedom, calculated according to the following formula: (4) Hypothesis construction: i.e. the mean of the reference window is equal to the mean of the detection window, no evolving change point; The mean of the detection window is significantly greater than the reference window, and there is an evolutionary turning point.
6. The method of claim 1 or 4 or 5, wherein: In step two, the evolutionary change point detection rule based on sliding window hypothesis testing is as follows: given a significance level , compute the quantile critical value of t-distribution at this confidence level , and make the following comparison: If , the original hypothesis is rejected, and it is determined that the aero-bearing has evolved a defect at the instant in time, i.e., it has entered a defect phase after that point in time. Otherwise, the null hypothesis is accepted, i.e. If the system is still in the normal phase at the time, the decision continues at the next detection point.
7. The method of claim 1, wherein: In step three, offline sample division and parameter space setting, specifically: for the same kind of aviation bearing samples, use samples to perform sliding window parameter offline optimization, and set the length of the reference window and the detection window , , ; model performance evaluation index, specifically: use the evolution track of bearings, and the known true change point position is ; based on the given sliding window parameters , run the health evolution change point detection full process based on sliding window hypothesis testing on the evolution track of bearings, and calculate the detection success rate , the mean absolute error and the average detection delay ; the specific calculation method is: (5) wherein and are parameters The number of trajectories in which a change point was successfully detected and the number of trajectories in which a change point was detected with delay; and are the h true change point position and the detected change point position for the i-th trajectory.
8. The method according to claim 1 or 7, characterized in that: In step three, the optimal window parameters are determined, specifically: the detection success rate is... Mean absolute error and average detection delay Assign weights And calculate the overall score. Sc : (6) wherein, and respectively represent the parameters Under The maximum average absolute error and the maximum average detection delay of the variable point recognition in the track are determined by enumeration method , for the subsequent evolution of bearing variable point online identification and residual life prediction.
9. The method of claim 1, wherein: In step four, the random process model of the evolution trajectory is constructed, specifically: the linear Wiener process model is used to model the evolution process of aero-bearing in the normal and defect stages, that is: (7) wherein, and Wiener process drift coefficient and diffusion coefficient of the normal phase, respectively; and Wiener process drift coefficient and diffusion coefficient of the defect phase, respectively; and are the times of the normal phase, and are the times of the defect phase, and denote the system degradation quantities corresponding to the two times of the normal phase, respectively, and denote the system degradation quantities corresponding to the two times of the defect phase, respectively.
10. The method of claim 1 or 9, wherein: In step four, the evolution change point of the bearing is identified online, specifically: based on the obtained optimal window parameter , the evolution change point of the test sample is detected online, and the performance of the detected change point and model is obtained; Derivation and prediction of residual life distribution, specifically: (8) wherein, is the detection time, is the remaining life, represents the value of the remaining life . Residual life for aero bearing in defect stage Subject to inverse Gaussian distribution, the probability density function and distribution function are represented as: (9) (10) wherein, represents the degradation amount of the system at the k nth periodic inspection time, is the cumulative distribution function of the standard normal distribution; calculates the point estimate and the confidence interval estimate of the remaining life at the nth periodic inspection time; and the mean of the distribution of the degradation amount is used as the point estimate of the remaining life. (11) Further, given a confidence level , the prediction interval is computed where denote the minimum and maximum values, respectively, of the remaining lifetime at time t predicted at a given confidence level, satisfying: (12) In terms of prediction error evaluation, the point estimate of the remaining life considers three error indicators: mean absolute error (MAE), root mean square error (RMSE), and standard deviation of absolute error The calculation method is: (13) wherein, and are the true and predicted residual life of the r th prediction point, respectively, is the number of prediction points, is the r th prediction point, and the confidence interval estimation considers two error indicators: the coverage rate of the confidence interval and the standard deviation of the confidence interval width. (14)。