A method for predicting the remaining life of bearings based on a dual-weighted degradation model
By fusing bearing vibration characteristics using a dual-weighted degradation model and employing Kalman filtering and iterative weighted estimation techniques, the problems of noise interference and acquisition interval differences were solved, achieving accurate identification of degradation initiation points and prediction of remaining life, thus improving the reliability and accuracy of prediction.
Patent Information
- Application Number
- CN202611472847.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-21
- Publication Date
- 2026-11-17
AI Technical Summary
In existing methods for predicting the remaining life of bearings, noise interference affects the identification of degradation initiation points, degradation parameter estimation struggles to balance recent trends with anomaly suppression, and incremental differences under different acquisition intervals are not adequately considered, all of which affect prediction accuracy.
A dual-weighted degradation model is adopted, which integrates vibration characteristics through principal component analysis, uses Joseph-form Kalman filtering for state estimation and error covariance update, recursively identifies degradation initiation points, and estimates local drift parameters through iterative weighted least squares. The model also incorporates time forgetting and Huber robust weights to mitigate the impact of early data and anomalous increments.
It effectively reduces bearing vibration characteristics, accurately identifies degradation initiation points, reduces false alarms, improves the accuracy and stability of remaining life prediction, and provides a reliable life reference for bearing maintenance.
Smart Images

Figure CN123116167A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of rotating machinery condition monitoring, stochastic degradation modeling, and remaining life prediction, specifically to a bearing remaining life prediction method that combines degradation initiation detection and dual-weighted parameter estimation. Background Technology
[0002] Rolling bearings are crucial components in rotating machinery, bearing loads and supporting rotation. Their operating condition affects the continuous operation and maintenance schedule of the equipment. By acquiring bearing operating data through vibration sensors, identifying degradation states, and predicting remaining life, a basis for maintenance and replacement can be provided.
[0003] Vibration signals during bearing operation typically contain degradation information, operational disturbances, and measurement noise simultaneously. Directly predicting life based on raw vibration amplitudes is susceptible to transient shocks and random fluctuations. Fusing multiple vibration characteristics into a health indicator can compress observational data, but the fused indicator may still contain noise, requiring further condition estimation.
[0004] The entire lifespan of a bearing typically exhibits a phased characteristic of relative stability in the early stages followed by a significant degradation trend in the later stages. If data from both the stable and degradation stages are used together to estimate the degradation rate, a large amount of stable stage data may negatively impact the fitting of the later trend. Furthermore, when using a fixed threshold or a single statistic to detect the degradation initiation point, the choice of threshold and signal noise can also affect the location of the change point.
[0005] Stochastic degradation models can describe degradation trends and their stochastic fluctuations. Model parameters are typically estimated based on observed data and updated with new observations. Parameter estimation needs to consider the degradation state reflected by data at different observation times, as well as the impact of anomalous increments and actual acquisition intervals on the estimation results.
[0006] However, some existing methods for predicting the remaining life of bearings still have the following problems in practical applications:
[0007] 1. Noise interference affects the identification of degradation initiation point: Random fluctuations and instantaneous shocks in health indicators may interfere with the judgment of degradation trend. Determining the degradation initiation point based solely on a single threshold crossing is prone to false alarms, affecting the selection of subsequent modeling data.
[0008] 2. Degradation parameter estimation is difficult to balance recent trends and outlier suppression: All-historical data or equal-weighted estimation is easily affected by early data and cannot reflect recent degradation changes in a timely manner; ordinary least squares is sensitive to large residuals and is easily disturbed by abnormal increments.
[0009] 3. Insufficient consideration of incremental differences under different acquisition intervals: Unequal acquisition intervals or missing data will change the time span of degradation increment. If the changes in its expectation and variance with time interval are not considered, it may affect the estimation of degradation parameters and the calculation of the remaining lifetime prediction interval.
[0010] Therefore, it is necessary to design a bearing remaining life prediction method based on a dual-weighted degradation model to solve the above-mentioned technical problems. Summary of the Invention
[0011] The technical problem to be solved by the present invention is to provide a bearing remaining life prediction method based on a dual-weighted degradation model, which denoises the health index constructed from vibration characteristics, recursively identifies the transition from the stable stage to the degradation stage, and updates the local drift velocity based on recent degradation data, thereby reducing the impact of isolated abnormal increments and early data on the extrapolation of the current life.
[0012] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a bearing remaining life prediction method based on a dual-weighted degradation model, comprising the following steps:
[0013] Step S1: Collect vibration signals of the bearing to be predicted during operation, form vibration data segments according to the collection time, extract the standard deviation, root mean square and peak-to-peak value of each vibration data segment, use principal component analysis to integrate the extracted features into health indicators, and unify the direction of increase of health indicators as degradation worsens.
[0014] Step S2: Establish the state space model of the health index, use Joseph form Kalman filter for state estimation and error covariance update to obtain the denoised health index, and perform natural logarithmic transformation on the denoised health index that meets the positive value condition.
[0015] Step S3: Standardize and limit the adjacent increments of the logarithmic health index based on the health baseline, and recursively calculate the one-sided cumulative sum statistic; when the statistic continuously reaches the discrimination threshold, backtrack to the starting position of the current cumulative segment, determine the first degradation starting point, and record the actual confirmation time.
[0016] Step S4: Extract the most recent degraded data window from the observed data after the initial degradation point, set the exponential time forgetting weight according to the time interval between the observed data and the current time, multiply it with the robust weight corresponding to the Huber loss to form a combined weight, use iterative weighted least squares to estimate the local drift parameters, and estimate the diffusion scale based on the convergence residual.
[0017] Step S5: When the current logarithmic health index has not reached the preset failure threshold and the estimated value of the local drift parameter is positive, the remaining life is calculated using the first arrival time model of local constant drift, starting from the current state; after receiving new data, the degradation data window and model parameters are updated, and the updated bearing remaining life is output.
[0018] A further improvement to the technical solution of the present invention is that step S1 includes the following specific steps:
[0019] Step S11: Obtain bearing operation data through vibration sensors, form vibration data segments according to uniform rules, and record the timestamps corresponding to each data segment.
[0020] Step S12: Assume the k-th vibration data segment contains m sampled values a. k,r Calculate the segment mean, standard deviation, root mean square, and peak-to-peak value:
[0021]
[0022]
[0023]
[0024]
[0025] Where m is greater than 1; the standard deviation, root mean square, and peak-to-peak value are used to form a three-dimensional feature vector z. k =[s k u k v k ] T .
[0026] Step S13: Based on the feature centers μ determined before the prediction begins j and positive scale d j Standardize the matrix; project the first principal component score onto the unit eigenvector w1 corresponding to the largest eigenvalue of the standardized feature covariance matrix.
[0027]
[0028] Step S14: Construct a health index using the orientation coefficient ρ, the positive scale coefficient c, and the offset b.
[0029]
[0030] Where ρ is set to 1 or -1, the health index increases overall as the deterioration worsens; b ensures that the index is positive within the predetermined applicable range. After calibration, the feature standardization parameters, principal component directions, and transformation parameters are kept fixed.
[0031] A further improvement to the technical solution of the present invention is that step S2 includes the following specific steps:
[0032] Step S21, with H k As a potential real health indicator, HI k For the observations, a one-dimensional state-space model is established:
[0033]
[0034] Among them, process noise w k−1 With measurement noise e k They are independent and have a mean of zero, with variances of non-negative Q. k−1 and positive number R k .
[0035] Step S22: Make a prediction based on the state estimate and error variance from the previous time step:
[0036]
[0037] Step S23: Calculate the Kalman gain:
[0038]
[0039] Step S24: Update the error variance using the Joseph form:
[0040]
[0041] Step S25: Using the observed residuals to correct the state estimate, perform a natural logarithmic transformation on the positive denoising index:
[0042]
[0043] Among them, P k This represents the variance of the state estimation error. When the denoising metric does not meet the positive value condition, and the output metric exceeds the applicable range, the construction of the corresponding logarithmic increment is paused.
[0044] A further improvement to the technical solution of the present invention is that step S3 includes the following specific steps:
[0045] Step S31: Construct the incremental Δx using adjacent effective observations. i =x i -x i−1 and actual time difference Δt i =t i -t i−1 >0. Determine baseline drift β0 and positive diffusion scale s0 based on health calibration data, and calculate:
[0046]
[0047]
[0048]
[0049] Where Med represents the median, s min c is the lower limit of the positive scale. M These are positive scaling correction coefficients predetermined based on a reference noise distribution or independently calibrated data. When using a normal reference distribution and converting the median absolute deviation to a standard deviation scale, take:
[0050]
[0051] Where Φ is the standard normal distribution function. When using other reference distributions, correction coefficients that match the distribution are predetermined, and the detection parameters and residual discrimination parameters are recalibrated on the same scale.
[0052] Step S32: Set the detection startup index n0 and G n0 =0, with c G Standardize the increment limit and recursively calculate the one-sided cumulative sum:
[0053]
[0054] Where ν is the reference offset, satisfying 0 < ν < c G .
[0055] Step S33, when G i The number of consecutive valid observations is not less than the threshold h. G At that time, confirm the degradation; denote the index of the first threshold crossing of this continuous segment as k. c Let the actual confirmed index be k. a =k c +q-1. Set h G >c G -ν, limits the triggering effect of a single limit increment starting from zero.
[0056] Step S34, put k c The index of the most recent accumulation and zeroing is used as the starting point η for the first degradation:
[0057]
[0058] Record the degradation initiation estimate t respectively η With actual confirmation time t ka After confirmation, keep η fixed and pass the effective increment after the starting point to step S4.
[0059] The dual-weighted degradation model adopts a two-stage structure, including a stable stage model and a local prediction model for the degradation stage:
[0060]
[0061]
[0062] Where t0 and x0 are the initial time and initial logarithmic index, respectively, and σ1 is taken as s0 during the steady-state phase; B1 and B2 represent standard Brownian motion; the local prediction model is used after the degradation initiation point has been identified, and the local drift estimate β̂ is updated with new observations. k and diffusion scale σ̂ 2,k .
[0063] The dual weighting refers to the process of multiplying a time-forgetting weight determined based on the observation time interval and a Huber robust weight determined based on the standardized residuals in the estimation of local drift parameters during the degradation phase. This results in a combined weight, which is then used for iterative weighted estimation. The time-forgetting weight is used to reduce the influence of earlier observations, while the Huber robust weight is used to reduce the influence of larger standardized residuals.
[0064] A further improvement to the technical solution of the present invention is that step S4 includes the following specific steps:
[0065] Step S41: Construct the nearest window I from the observed data after the degradation starting point. k ={i: max(η+1, k-W+1)≤i≤k}, where the exponential forgetting weight is determined based on actual time:
[0066]
[0067] Where W is the maximum window increment, and T f It is a positive forgetting time constant; the number of effective increments is less than n. min Continue collecting data as needed.
[0068] Step S42: Calculate the initial drift and normalized initial residuals using time weighting.
[0069]
[0070] The fixed positive scale ζ used in the current iteration is determined by the weighted median absolute deviation. k :
[0071]
[0072] Med v This means sorting the values from smallest to largest and then pressing v. i,k The weighted median obtained by accumulating the weights to half of the total weights; c M Let these be the scale correction coefficients. Minimize the following Huber objective function:
[0073]
[0074] Huber loss is defined as follows:
[0075]
[0076]
[0077] κ is a positive adjustment constant used to define the boundary between squared loss and linear loss.
[0078] Step S43: Calculate the robust weight u based on the standardized residual of the a-th iteration. i and combined weight ω i And update the drift estimate:
[0079]
[0080]
[0081]
[0082] After convergence, the updated value will be used as β̂ k When the maximum number of iterations is reached and convergence is still not achieved, the output parameters are in a non-convergent state.
[0083] Step S44: Estimate the local diffusion scale based on the converged normalized residuals, and calculate the effective sample size using the final combined weights:
[0084]
[0085]
[0086]
[0087] Under finite window and time-weighted conditions, the diffusion scale is used as a local approximation; when the effective sample size is lower than the preset condition, the lifetime is temporarily postponed and data collection continues.
[0088] A further improvement to the technical solution of the present invention is that step S5 includes the following specific steps:
[0089] Step S51: Determine the logarithmic space failure threshold D based on the engineering failure criteria or independently calibrated bearings that can be obtained before prediction, ensuring that the calibration and prediction use the same feature processing, filtering and scaling rules.
[0090] Step S52, calculate d k =D-x kIt checks whether the current indicator has exceeded the threshold, whether the local drift is positive, whether the iteration has converged, and whether the amount of effective data meets the requirements. When the threshold is exceeded, it outputs zero remaining lifetime and remains in the exceeded-threshold state until maintenance or recalibration confirmation; for other cases that do not meet the extrapolation conditions, it outputs the corresponding status.
[0091] Step S53: For bearings that satisfy the extrapolation conditions, starting from the current state, adopt the local constant drift first arrival time model:
[0092]
[0093]
[0094]
[0095] Where l represents the future time calculated from the current moment, L k Let L be a random variable representing the remaining lifetime. k This is a lifetime point estimate under a local model. For f k Integral distribution F is obtained k The conditional prediction interval for a given local parameter is calculated based on its quantile; this interval is not the same as the interval that already covers all parameter errors and model errors.
[0096] Step S54: Output remaining lifetime, condition interval, degradation start point, actual confirmation time, effective sample size, and calculation status; after receiving new observations, update the window and parameters, and recalculate the lifetime.
[0097] By employing the above technical solutions, the technical effects of this invention are as follows: It integrates and denoises bearing vibration characteristics to characterize degradation with a unified index; it recursively detects degradation initiation points using cumulative statistics, reducing redundant candidate searches; it limits the triggering effect of single extreme increments through amplitude limiting and continuous confirmation; it multiplies the time-forgotten weight with the Huber robust weight to form a combined weight, mitigating the impact of earlier observation data and abnormal increments on the estimation of local degradation rates; and it provides a life reference for bearing maintenance after clarifying the failure threshold and extrapolation conditions. The window and forgetting settings need to balance change response and estimation stability, and actual performance is evaluated using independent bearing data. Attached Figure Description
[0098] Figure 1 This is a flowchart illustrating the overall process of predicting the remaining life of bearings according to the present invention.
[0099] Figure 2 This is a flowchart of the degradation initiation detection and local parameter update process of the present invention.
[0100] Figure 3 This is a flowchart of the dual-weighted construction and local drift parameter update of the present invention. Detailed Implementation
[0101] The present invention will be further described in detail below with reference to embodiments:
[0102] like Figure 1 As shown, a method for predicting the remaining life of bearings based on a dual-weighted degradation model includes the following steps:
[0103] Step S1: Collect bearing vibration signals, extract standard deviation, root mean square and peak-to-peak values, construct health indicators using principal component analysis, and unify the degradation direction of the indicators.
[0104] Step S1 specifically includes the following steps:
[0105] Step S11: Collect bearing operation data using vibration sensors installed on the rotating machinery, form vibration data segments according to a unified rule, and record the timestamps corresponding to each data segment. The sampling frequency within a data segment and the update interval between adjacent data segments are set separately; the timestamps are strictly incremented, and when a missing data segment occurs, the actual time difference between adjacent valid observations is used.
[0106] Step S12: For the k-th vibration data segment, let the sampled value be a. k,r For each r range from 1 to m, where m is greater than 1, calculate the mean, standard deviation, root mean square, and peak-to-peak value of the range.
[0107]
[0108]
[0109]
[0110]
[0111] Among them, the standard deviation represents the degree of fluctuation within the data segment, the root mean square represents the overall amplitude of the fluctuation, and the peak-to-peak value represents the amplitude span. These three features are combined to form z. k =[s k u k v k ] T .
[0112] Step S13: Determine the feature centers μ using the calibration data available before the prediction begins. j and positive scale d j The features are standardized, and the unit eigenvector w1 corresponding to the largest eigenvalue of the standardized feature covariance matrix is calculated. The feature standardization and the first principal component score are:
[0113]
[0114]
[0115] Wherein, in this embodiment d j Take the standard deviation of the calibration features. If zero-scale features exist, re-select valid calibration data or positive scaling parameters before processing.
[0116] Step S14: Determine the orientation coefficient ρ, positive scale coefficient c, and offset b based on the calibration data to obtain the health indicators:
[0117]
[0118] Where ρ is 1 or -1, so that the index increases overall as degradation worsens; b is selected according to the predetermined applicable range, so that HI k The result is positive. After calibration, the feature center, scale, projection direction, and transformation parameters of each dimension remain fixed; when recalibration is required, the corresponding historical data and failure thresholds are reprocessed using the same new scale.
[0119] Step S2: Perform Joseph-style Kalman filtering on the health indicators obtained in step S1, and convert the positive denoising indicators to logarithmic space.
[0120] Step S2 specifically includes the following steps:
[0121] Step S21, with H k As a potential real health indicator, HI k For the observations, a one-dimensional state-space model is established:
[0122]
[0123] Process noise w k−1 With measurement noise e k They are independent and have a mean of zero, with variances Q and Q', respectively. k−1 and R k The state is initialized with the first valid health indicator, and the initial error variance is non-negative. Q k−1 Non-negative, R k It is positive; the noise parameter is determined based on the calibration data of the corresponding scale.
[0124] Step S22: Based on the previous state estimate, perform state prediction and error variance prediction:
[0125]
[0126] Step S23: Calculate the Kalman gain:
[0127]
[0128] Step S24: Update the error variance according to the Joseph form:
[0129]
[0130] The updated formula includes contributions from both prediction and measurement errors, which helps maintain the non-negativity of the error variance.
[0131] Step S25: Correct the state estimate using the observed residuals and perform a natural logarithmic transformation on the positive denoising index:
[0132]
[0133]
[0134] If the denoising metric does not meet the positive value condition, the report metric is deemed to be out of applicable range, and the construction of the corresponding logarithmic increment is paused. After obtaining a new effective metric, the increment is constructed based on its actual time difference with the previous effective metric.
[0135] Step S3, as follows Figure 2 As shown, the logarithmic health index increment is standardized and limited by the health baseline, the one-sided cumulative sum statistic is recursively calculated, and the starting point of the first degradation is determined by combining the continuous confirmation condition.
[0136] Step S3 specifically includes the following steps:
[0137] Step S31: Define the index increment and time increment for adjacent valid observations:
[0138]
[0139] In health calibration data, Δx i / Δt i The median is used as the baseline drift β0. For b i =(Δx) i -β0Δt i ) / √Δt i Calculate the median absolute deviation to obtain the health diffusion scale and standardized increment:
[0140]
[0141]
[0142] Where Med represents the median, s min To prevent a positive lower limit where the scale is zero, its unit is consistent with the diffusion scale; c M This is the scaling correction factor. In this embodiment, a normal reference distribution is used to perform standard deviation scaling on the absolute deviation of the median, taking:
[0143]
[0144] Where Φ is the standard normal distribution function. The median of the absolute value of a standard normal variable is Φ. −1 (0.75), approximately 0.67449, so its reciprocal is approximately 1.4826. σ1 of the stable phase model is taken as s0. The baseline calibration data were obtained through the same feature processing, filtering, and logarithmic transformation, and have matching operating conditions and update intervals with the bearing to be predicted.
[0145] Step S32: Set the detection startup index n0 and G n0 =0, limit the amplitude of the standardized increment, and recursively calculate the statistic:
[0146]
[0147]
[0148] Where c G ν is the limiting constant, and ν is the reference offset, 0 < ν < c G The statistic only retains cumulative evidence of upward shifts. G The thresholds, ν, and subsequent consecutive counts are determined before prediction and remain fixed during detection.
[0149] Step S33, when G i Not less than h G When h is reached, the count increments by one; when h is less than h, the count increments by one. G When the consecutive count reaches q, the continuous count is reset to zero. Degeneration is confirmed when the continuous count reaches q. Let k be the index of the first threshold crossing of the current continuous segment. c The actual confirmed index is k. a =k c +q-1. Using h G >c G -ν ensures that when the statistic starts from zero, a single limited increment cannot independently reach the discrimination threshold.
[0150] In this embodiment, the detection parameter is set to c. G =4, ν=0.5, h G =8, q=3. When selecting the threshold, the complete detection process is run on independent health data for a predetermined monitoring duration, and the parameters are determined by combining the health false alarm situation and the detection delay of independent degradation calibration data. The above values are the parameters used in this embodiment. The actual cumulative false alarm rate depends on the observation correlation, monitoring duration, and operating conditions, and needs to be evaluated separately.
[0151] Step S34, when n0≤j<k c Within the range, trace back to the most recent G j The index = 0 determines the starting point of the first degradation:
[0152]
[0153] Record t η and x η At the same time, save the actual confirmation time t. ka . t η t is the degradation starting point for backtracking estimation. ka This is the moment when the system obtains sufficient evidence; both are output separately. After confirmation, η remains fixed, and subsequent valid increments are sent to step S4.
[0154] In the stable phase, the baseline trend and stochastic fluctuations are characterized by X(t) = x0 + β0(t - t0) + σ1B1(t - t0); when β0 is zero, it degenerates into a zero-drift model. In the degradation phase, the local drift and diffusion scale are estimated from recent data and updated with observations, thus allowing different local degradation rates at different prediction times.
[0155] Step S4, as follows Figure 3 As shown, the nearest window is established from the observed data after the initial degradation point. The exponential time forgetting weight is multiplied by the Huber robust weight to form a combined weight. Iterative weighted least squares is used to estimate the current local drift parameters, and the diffusion scale is estimated based on the convergence residual.
[0156] Step S4 specifically includes the following steps:
[0157] Step S41: Set the maximum number of window increments W and the minimum number of valid increments n. min and the forgetting time constant T f The window index satisfies max(η+1, k-W+1)≤i≤k, denoted as I. k If the effective increment within the window is less than n min Continue accumulating observations; otherwise, calculate the time weight:
[0158]
[0159] In this embodiment, the parameters are set to W = 20, n min =8, T f =10Δt ref , where Δt ref This is the reference update interval determined during the calibration phase. During equal-interval data collection, the weights of adjacent old data are calculated according to exp(-Δt). ref / T f Attenuation; when collecting data at unequal intervals, the actual time difference is used directly.
[0160] Step S42: Calculate the initial drift estimate using time weights and calculate the normalized initial residuals:
[0161]
[0162]
[0163] With time weight v i,k Define the weighted median Med v Sort the values to be calculated in ascending order, and take the value where the cumulative weight first reaches half of the total weight. Determine the fixed iteration scale based on the initial residual.
[0164]
[0165] Keep ζ in this parameter iteration k Fixed, minimize the following objectives:
[0166]
[0167] The Huber loss used in this embodiment is:
[0168]
[0169]
[0170] The small residuals use squared loss, while the large residuals use linearly increasing loss. In this embodiment, κ = 1.345 is chosen to limit the impact of large residuals on drift estimation.
[0171] Step S43: Using 'a' as the iteration number, calculate the standardized residual, robust weights, and combined weights based on the current drift estimate:
[0172]
[0173]
[0174]
[0175] Where, |r i When |≤κ, the robust weight is 1; when |>κ, the weight decreases as the absolute value of the residual increases; the actual squared residual weight is equivalent to ω. i / Δt i Perform a maximum of 20 iterations, where the absolute difference between two consecutive drift estimates does not exceed 10. −6 ×max(β ref ,|β (a) Stop when |), β ref As a positive reference scale with drift units, this embodiment takes s0 / √Δt. ref After convergence, we obtain β̂. k If convergence fails, record the corresponding state and temporarily postpone the output of new lifetime extrapolation values.
[0176] Step S44: Calculate the normalized residual e after convergence. iThe local diffusion scale is estimated by its weighted median absolute deviation:
[0177]
[0178]
[0179] In this embodiment, c is used in the scaling of the healthy baseline, initial residual, and convergence residual. M =1.4826; finite window, time weighting, and noise deviating from the normal distribution affect this approximation. This result is used as a local diffusion scale. Meanwhile, ω is recalculated from the converged residuals. i And calculate the effective sample size:
[0180]
[0181] This embodiment uses N eff A value of at least 5 is used as one of the conditions for output lifetime extrapolation. If the value is less than this, an insufficient amount of valid data is reported and acquisition continues; after acquiring new observations, the sliding window is activated and steps S41 to S44 are re-executed. The data used for estimation are all from current and historical valid observations.
[0182] Step S5: Calculate the remaining lifetime based on the current state, local degradation parameters, and preset failure threshold, and update the prediction results with new observations.
[0183] Step S5 specifically includes the following steps:
[0184] Step S51: Before prediction begins, determine the failure threshold D based on the engineering failure criteria or independently calibrated bearings. Calibration and prediction use the same feature standardization, principal component direction, positive value transformation, filtering, and logarithmic scaling; D remains constant throughout the prediction process.
[0185] Step S52, let d k =D-x k The system checks whether the current metric has exceeded the threshold, whether the local drift is positive, whether the iteration has converged, and whether the effective data volume meets the requirements. When the current metric has exceeded the threshold, the system saves the threshold-exceeding state and outputs zero remaining lifetime; this state remains valid until it is maintained or recalibrated and confirmed. If the initial degradation starting point has not yet been confirmed, the data volume is insufficient, the drift is not positive, or the parameters have not converged, the system outputs the corresponding state and continues data collection.
[0186] Step S53: For bearings that satisfy the extrapolation conditions, using the current time as the prediction starting point, the local constant drift approximation is adopted:
[0187]
[0188] Where B(l) represents the standard Brownian motion increment process starting from the current moment. Under the condition that the threshold has not yet been exceeded, the remaining lifetime is defined as:
[0189]
[0190] The conditional first arrival time density and lifetime point estimate under the local model are as follows:
[0191]
[0192]
[0193] After applying windowing and robust weights, β̂ k It is usually not equal to the cumulative increment of the self-degradation starting point divided by the cumulative time. Therefore, the remaining lifetime is directly estimated by the current distance from the threshold and the local drift.
[0194] Further calculations were performed on the cumulative distribution and the 95% conditional prediction interval under fixed local parameters:
[0195]
[0196]
[0197] The intervals described reflect the randomness of future processes given current parameters. Uncertainties introduced by local drift continuing to change over future time, parameter estimation errors, and starting point detection errors require further verification or modeling; the above intervals are not considered as intervals where these errors have been calibrated.
[0198] Step S54: Output the remaining lifetime, condition interval, first degradation start point, actual confirmation time, effective sample size, and calculation status to the bearing condition monitoring interface. After obtaining new observations, update the window and model parameters and re-execute steps S52 and S53. The current parameters remain unchanged in a single future extrapolation and are updated after the next observation arrives.
[0199] To illustrate the calculation process for degradation initiation detection, this embodiment uses a set of given standardized increments for calculation. The input sequence is 0.2, 0.4, 1.4, 1.5, 1.6, 1.7, 2.0, 2.2, 2.5, 2.4, 2.4, with an initial G0 = 0. Using c... G =4, ν=0.5, h G When η = 8 and q = 3, the cumulative sums are 0, 0, 0.9, 1.9, 3.0, 4.2, 5.7, 7.4, 9.4, 11.3, and 13.2, respectively. The 9th observation first crosses the threshold, the 11th observation completes continuous confirmation, and the most recent zeroing position is the 2nd observation. Therefore, η = 2 and k = 3. a =11.
[0200] like Figure 3 As shown, steps S41 to S43 divide the observations in the recent window into two processing branches: time weight calculation and standardized residual calculation. The time branch calculates v based on the time difference between the observation and the current time. i,k The residual branch is calculated based on the current drift estimate r. i And obtain u according to Huber rules i Multiplying the two branches yields the combined weight ω. i The combined weights are used to update the drift parameters, and the new parameters are then used to calculate the residuals until the convergence condition is met. The attached figure illustrates the data relationships involved in the parameter updates.
[0201] In this embodiment, the remaining lifetime assessment uses Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). Using r... j Let r̂ represent the actual remaining lifetime at the j-th evaluation time. j Represents the corresponding predicted value, denoted by N. e This indicates the number of moments involved in the evaluation. Calculation:
[0202]
[0203]
[0204]
[0205] MAE and RMSE use the same time unit as the remaining life, listed in minutes in Table 1; MAPE is listed as a percentage and is calculated only for evaluation points where the actual remaining life is greater than zero. Different methods should be compared using the same bearing, prediction time, failure criteria, and time scale.
[0206] The remaining bearing life assessment selected Bearing 1_1, Bearing 2_2, and Bearing3_2 from the PRONOSTIA data. For each bearing, the predicted remaining life value at each assessment time was paired with the actual remaining life, and MAE, RMSE, and MAPE were calculated according to the above formulas. The evaluation index recording format is shown in Table 1.
[0207] Table 1. Record of Bearing Remaining Life Evaluation Indicators
[0208] bearing number MAE / min RMSE / min MAPE / % Bearing 1_1 7.40 12.28 25.66 Bearing 2_2 6.55 10.87 28.27 Bearing 3_2 2.09 3.09 16.68
[0209] When performing the above bearing verification, calibration data and evaluation data are divided according to complete bearings. Before evaluation, the health index construction parameters, filtering parameters, detection parameters, and failure thresholds are fixed, and predictions are made using only current and historical observations. At the same time, the degradation start point, actual confirmation time, effective sample size, and states that cannot be extrapolated are recorded. False alarms, detection delays, and interval coverage are statistically analyzed. The role of each component is evaluated through control experiments that remove amplitude limiting, continuous confirmation, time forgetting, or robust weights.
[0210] The method described in this embodiment can be executed by a computing device to process bearing vibration information into current state, degradation start point, and remaining life, providing data basis for bearing operation monitoring and maintenance arrangements.
Claims
1. A method for predicting the remaining life of bearings based on a dual-weighted degradation model, characterized in that, Includes the following steps: Step S1: Collect vibration signals of the bearing to be predicted during operation, form vibration data segments according to the collection time, extract the standard deviation, root mean square and peak-to-peak value of each vibration data segment, use principal component analysis to integrate the extracted features into health indicators, and unify the direction of increase of health indicators as degradation worsens. Step S2: Establish the state space model of the health index, use Joseph form Kalman filter for state estimation and error covariance update to obtain the denoised health index, and perform natural logarithmic transformation on the denoised health index that meets the positive value condition. Step S3: Standardize and limit the adjacent increments of the logarithmic health index based on the health baseline, and recursively calculate the one-sided cumulative sum statistic; when the statistic continuously reaches the discrimination threshold, backtrack to the starting position of the current cumulative segment, determine the first degradation starting point, and record the actual confirmation time. Step S4: Extract the most recent degraded data window from the observed data after the initial degradation point, set the exponential time forgetting weight according to the time interval between the observed data and the current time, multiply it with the robust weight corresponding to the Huber loss to form a combined weight, use iterative weighted least squares to estimate the local drift parameters, and estimate the diffusion scale based on the convergence residual. Step S5: When the current logarithmic health index has not reached the preset failure threshold and the estimated value of the local drift parameter is positive, the remaining life is calculated using the first arrival time model of local constant drift, starting from the current state; after receiving new data, the degradation data window and model parameters are updated, and the updated bearing remaining life is output.
2. The bearing remaining life prediction method according to claim 1, characterized in that, Step S1 includes the following specific steps: Step S11: Obtain vibration data through vibration sensors installed on the rotating machinery where the bearing is located, record the timestamps corresponding to each vibration data segment, and keep the sampling and segmentation rules consistent between the calibration stage and the prediction stage; Step S12: For the k-th vibration data segment containing m vibration sample values, calculate the standard deviation, root mean square, and peak-to-peak value respectively to form a three-dimensional feature vector; Step S13: Based on the feature center value and scale parameter determined before the prediction begins, the three-dimensional feature vector is standardized, and the standardized feature vector is projected onto the calibrated first principal component direction to obtain the first principal component score. Step S14: Transform the score of the first principal component using a fixed direction coefficient, positive scaling coefficient, and offset to form a health index that is positive and has a consistent degradation direction within a predetermined applicable range; keep the standardization parameters, projection direction, and transformation parameters unchanged in the same prediction process.
3. The bearing remaining life prediction method according to claim 2, characterized in that, In step S12, the vibration data segment is denoted as a. k,1 To a k,m The mean is denoted as ā k The standard deviation, root mean square, and peak-to-peak value are calculated according to the following formulas: The three-dimensional feature vector is z k =[s k u k v k ] T Where m is greater than 1, u k v is the root mean square. k The peak-to-peak value; in step S13, the first principal component direction is the unit eigenvector corresponding to the largest eigenvalue of the standardized feature covariance matrix after calibration.
4. The bearing remaining life prediction method according to claim 1, characterized in that, Step S2 includes the following specific steps: Step S21: Using real health indicators as the state and the health indicators obtained in step S1 as the observations, establish a state-space model: Among them, w k−1 The mean is zero and the variance is Q. k−1 Process noise, e k The mean is zero and the variance is R. k The measurement noise is independent of each other; Step S22: Make a prediction based on the state estimate and error variance from the previous time step: Step S23: Calculate the Kalman gain: Step S24: Update the error variance using the Joseph form: Step S25: Update the estimated values of the true health indicators using the observed residuals and perform a natural logarithmic transformation: Among them, P k The state estimation error variance is initialized to be non-negative, the process noise variance is non-negative, and the measurement noise variance is positive. When the denoising health index does not meet the positive value condition, the output index exceeds the applicable range and no logarithmic transformation is performed on the value.
5. The bearing remaining life prediction method according to claim 1, characterized in that, The dual-weighted degradation model adopts a two-stage structure of a stable stage and a degradation stage. The local drift parameters in the degradation stage are estimated using a dual-weighted method according to step S4. The stable stage model and the local prediction model in the degradation stage are respectively expressed as follows: Where X(t) is the logarithmic health index, t0 is the initial time, x0 is the initial index, β0 is the health baseline drift parameter, σ1 is the diffusion coefficient in the stable phase, and t η x is the starting point of the first degradation. η =X(t) η );t k and x k These represent the current moment and the current logarithmic health indicator, β̂. k and σ̂ 2,k B1 and B2 are the local drift parameters and diffusion scale estimated for the degraded data window, respectively, and represent standard Brownian motion. The local prediction model uses the current parameters for each prediction and re-estimates them after obtaining new observations.
6. The bearing remaining life prediction method according to claim 5, characterized in that, Step S3 includes the following specific steps: Step S31: Define adjacent increments Δx i =x i -x i−1 and positive time increment Δt i =t i -t i−1 β0 and the positive diffusion scale s0 are determined using pre-prediction health calibration data, and the standardized increment is calculated: Step S32: Using the limiting constant c G For standardized increment limit, and recursively calculate the one-sided cumulative sum statistic according to the reference offset ν: Step S33, when G i The number of consecutive valid observations is not less than the discrimination threshold h. G At that time, confirm the starting point of degradation; denote the first observation index of this continuous sequence as k. c Let the actual confirmed index be k. a =k c +q-1; Step S34, put k c The most recent G j The index = 0 is used as η, and t is recorded. η and t ka The alarm confirmation time is recorded; after the initial degradation start point is confirmed, η is kept fixed, and the subsequent effective increments are passed to step S4; when the new amount does not meet the continuous confirmation condition, the continuous count is cleared and data collection continues. Among them, G is set at the startup index n0. n0 =0, the backtracking index is located at n0≤j<k c ν is greater than zero and less than c G q is a positive integer, h G Greater than c G -ν.
7. The bearing remaining life prediction method according to claim 6, characterized in that, β0 is taken from the health calibration data Δx i / Δt i The median; let b i =(Δx) i -β0Δt i ) / √Δt i , with b i The median absolute deviation multiplied by the positive scaling correction factor c M and the lower limit of positive numbers s min Take the maximum value to obtain s0. M Predetermined based on reference noise distribution or independent calibration data; when using a normal reference distribution for standard deviation scaling, take: Where Φ is the standard normal distribution function. The c G 、ν、h G q is determined based on independent health calibration data and a preset monitoring duration and remains fixed during the testing period; a set of implementation parameters is c. G =4, ν=0.5, h G =8 and q=3.
8. The bearing remaining life prediction method according to claim 1, characterized in that, Step S4 includes the following specific steps: Step S41: Construct the degradation data window I k ={i: max(η+1, k-W+1)≤i≤k}, where W is the maximum number of increments that the window can contain, according to the forgetting time constant T. f Determine time weights: Step S42: Perform initial weighted estimation with time weights, and determine the positive scale ζ using the initial residuals. k and maintain ζ in the current iteration. k Fixed, minimize the Huber objective function: Where, ρ κ (u) Take u when |u|≤κ 2 / 2, when |u|>κ, take κ|u|-κ 2 / 2, κ is a positive adjustment constant; Step S43: In the a-th iteration, calculate the robust weights and combined weights based on the standardized residuals, and update the drift estimate: When the preset convergence condition is met, the update result will be used as β̂. k If the convergence condition is not met even after reaching the upper limit of the number of iterations, the output parameter is in an unconverged state and data continues to be received. Step S44, using e i =(Δx) i -β̂ k Δt i ) / √Δt i For the residual, calculate its value according to v. i,k The weighted median and median absolute deviation are used to obtain the diffusion scale according to the following formula: Among them, Med v s represents the weighted median determined by sorting numerically and accumulating time weights. min c is the lower limit of the positive scale. M The positive scaling correction coefficients are pre-determined based on the reference noise distribution or independently calibrated data; when using a normal reference distribution for standard deviation scaling, c M Take 1.4826; if the effective increment in the degraded data window is less than the preset minimum number, continue collecting data.
9. The bearing remaining life prediction method according to claim 8, characterized in that, Step S5 includes the following specific steps: Step S51: Determine the logarithmic space failure threshold D based on the failure criteria available before prediction or the independently calibrated bearing. The threshold is constructed using the same health index and scaling rules as the bearing to be predicted. Step S52: Calculate the distance d from the current state to the failure threshold. k =D-x k Check the convergence state and positive value condition of the local drift estimation; Step S53, when d k >0, local drift estimation has converged and β̂ k When the value is greater than 0, the remaining lifetime point estimate under the local model is calculated according to the following formula: When the current indicator has reached or exceeded D, the remaining lifetime is set to zero and the over-threshold state is retained; when D has not yet been reached and the drift estimate is not positive or the parameters have not converged, the state that does not meet the lifetime extrapolation condition is output. Step S54: Update the degradation data window, local drift parameters, and diffusion scale with new observations, and recalculate the remaining lifetime; the threshold-crossing state remains valid until it is maintained or recalibrated and confirmed.
10. The bearing remaining life prediction method according to claim 9, characterized in that, Given the current positive drift estimate and a fixed diffusion scale in local predictions, the remaining lifetime distribution is calculated using the following first arrival time density: According to its cumulative distribution F k The quantile calculation conditional lifespan prediction interval is used to output the effective sample size N corresponding to the combined weights. eff When N eff When the value is below a preset threshold, the output parameter estimate is based on the insufficient state and the lifespan extrapolation is temporarily suspended: The conditional lifetime prediction interval reflects the random fluctuations of the process under given local parameters, N eff Used to characterize the effective amount of data used for estimation.
