A method for predicting the remaining life of rolling bearings based on square root cubature Kalman filter

Through the square root volume Kalman filtering method, combined with the self-organized mapping neural network and exponential degradation model of the Marbanian distance metric operator, the problem of feature correlation and interval in rolling bearings is solved, and more accurate residual life prediction and Higher prediction reliability.

CN114896861BActive Publication Date: 2024-11-01XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210342484.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-02
Publication Date
2024-11-01
Estimated Expiration
2042-04-02

AI Technical Summary

Technical Problem

The existing rolling bearing residual life prediction method fails to effectively consider the correlation between feature vectors and inconsistent monitoring intervals, resulting in limited prediction effects.

Method used

Using a square root volume Kalman filtering method, the dimensionality reduction is achieved through an self-organized mapping neural network with a Marshallow distance metric operator, health indicators are constructed, and an exponential degradation model considering unequal sampling intervals is established, combining continuous triggering mechanisms and The maximum algorithm is expected to predict the remaining life of rolling bearings.

Benefits of technology

In the process of dimensionality reduction, the topological structure of high-dimensional space is retained, the limitation of equal-space sampling is overcome, the utilization rate of data between different sampling intervals is improved, and the reliability and adaptability of prediction is enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114896861B_ABST
    Figure CN114896861B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering, which extracts multidimensional features from vibration signals of historical failure samples of rolling bearings, constructs a sensitive feature set, trains a self-organizing map neural network with a Mahalanobis distance metric operator based on sensitive feature data of the early stable operation stage of the rolling bearing, performs dimensionality reduction, constructs a health index, and then determines an adaptive degradation threshold of the rolling bearing based on the health index, divides the healthy stage and the degradation stage of the rolling bearing by a continuous trigger mechanism, and predicts the remaining life of the rolling bearing by establishing an exponential degradation model that considers unequal sampling intervals. The method can better retain the topological structure of multidimensional features in high-dimensional space during dimensionality reduction, overcomes the dependence on online monitoring data sampled at equal intervals, and improves the utilization rate of full life data of historical failure samples with different sampling intervals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of bearing remaining life prediction, and in particular relates to a rolling bearing remaining life prediction method based on square root volume Kalman filtering. Background Art

[0002] With the rapid development of modern industrial technology, the safety, stability and reliability of electromechanical equipment are facing higher and higher requirements. As a key component of rotating machinery, rolling bearings are a weak link in equipment reliability. Once a mechanical failure occurs, it will cause equipment failure and economic losses. Therefore, real-time status monitoring and performance degradation assessment of rolling bearings are urgently needed to ensure the reliability of equipment operation. On this basis, the remaining life prediction of rolling bearings can provide a reasonable and feasible basis for the operation and maintenance plan of the equipment.

[0003] At present, in order to realize the prediction of the remaining life of rolling bearings, data-driven methods are mostly used. Since the remaining life of rolling bearings is affected by many factors, a single feature cannot well reflect the health status of rolling bearings. Therefore, many studies are based on the vibration signal of rolling bearings, extracting multidimensional features from them, and then mapping the data to a low-dimensional space to obtain a comprehensive health index. Currently, commonly used dimensionality reduction techniques include principal component analysis, linear discriminant analysis, self-organizing map neural network, etc. Since there is an imbalance between the number of normal samples and the number of failure samples in engineering practice, the self-organizing map neural network that only relies on the normal state feature vector for training is widely used in the health status monitoring of rolling bearings. However, during the training process, the self-organizing map neural network uses the Euclidean distance to calculate the similarity between the input layer feature vector and the output layer neuron node weight, and does not consider the correlation between the elements within the feature vector.

[0004] Secondly, when researchers conduct full-life degradation tests on rolling bearings, they often adopt the method of equal-interval sampling when collecting vibration signals of rolling bearings. However, in engineering practice, due to factors such as changes in monitoring plans, sensor failures, and operator errors, the monitoring data of running rolling bearings may not meet the requirements of equal-interval distribution. On the other hand, historical failure samples of rolling bearings in different periods may also have different monitoring intervals. When dealing with these problems, the effectiveness of this type of remaining life prediction method based on equal-interval sampling will be limited. Summary of the invention

[0005] In view of the shortcomings of the prior art that the correlation between variables and the inconsistency of monitoring intervals are not considered, the purpose of the present invention is to provide a method for predicting the remaining life of a rolling bearing based on square root volume Kalman filtering.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering comprises the following steps:

[0008] Extract multidimensional features from the vibration signal of rolling bearings and perform dimensionality reduction through a self-organizing map neural network with a Mahalanobis distance metric operator to construct health indicators;

[0009] Based on the health index, an adaptive degradation threshold of the rolling bearing is constructed, and a continuous trigger mechanism is used to divide the healthy stage and the degradation stage of the rolling bearing.

[0010] The remaining life of rolling bearings is predicted by considering the exponential degradation model with unequal sampling intervals.

[0011] A further improvement of the present invention is that the method comprises the following steps:

[0012] 1) According to the failure mode of rolling bearings, based on the collected rolling bearing vibration signals, feature extraction and screening are performed to construct a multi-dimensional sensitive feature set;

[0013] 2) Through the trained self-organizing map neural network with Mahalanobis distance metric operator, the multidimensional sensitive features in the multidimensional sensitive feature set are mapped to a low-dimensional space to construct the health index of the rolling bearing;

[0014] 3) Based on the health index of rolling bearings, the dynamic failure threshold of rolling bearings is constructed, and a continuous trigger mechanism is used as a method to judge degradation, so as to realize the division of the healthy stage and the degraded stage of rolling bearings;

[0015] 4) Establish an exponential degradation model considering unequal sampling intervals, and update the degradation model parameters through the expectation maximization algorithm based on the health indicator data of the rolling bearing degradation stage;

[0016] 5) Based on the updated degradation model parameters, the health status of the rolling bearing is updated through the square root volumetric Kalman filter, and the point estimate and probability density estimate of the remaining life of the rolling bearing are calculated.

[0017] A further improvement of the present invention is that the specific steps in step 1) are as follows:

[0018] According to the failure form of rolling bearings and based on the collected historical failure samples of rolling bearings, the collected rolling bearing vibration signals are analyzed, and multiple relevant features in the time domain, frequency domain and time-frequency domain are calculated, including root mean square, peak value and wavelet energy entropy, etc. The Spearman correlation coefficient between each relevant feature and the operating time series is calculated, and the features with Spearman correlation coefficient higher than the average value are screened out as the preliminary sensitive features. Then the covariance matrix of the preliminary sensitive features is calculated, and according to the correlation between the features, redundant features are excluded to determine the multidimensional sensitive feature set.

[0019] A further improvement of the present invention is that the specific steps in step 2) are as follows:

[0020] Select a period of rolling bearing stable operation data as a random sample of the healthy stage, calculate the features in the multidimensional sensitive feature set, normalize them, and calculate the mean and covariance of the sensitive features in the stable operation period;

[0021] Through the sensitive features in the stable operation period, the self-organizing map neural network with Mahalanobis distance metric operator is trained;

[0022] Through the trained self-organizing map neural network with Mahalanobis distance metric operator, the multi-dimensional sensitive features are mapped to low-dimensional space to construct the health index of rolling bearings.

[0023] A further improvement of the present invention is that the specific steps in step 3) are as follows:

[0024] According to the health index of rolling bearing, the dynamic failure threshold z=Q of rolling bearing is calculated. 3 +1.5IQR, IQR=Q 3 -Q 1 , where Q 3 is the third quantile of the health index from the initial moment to the current moment, Q 1 It is the first quantile of the health index from the initial moment to the current moment, and IQR is the difference between the two quantiles;

[0025] The health status of the rolling bearing is judged according to the failure threshold z. When the health index exceeds the failure threshold z for n consecutive times and the n health indexes show an increasing trend, it is judged that the rolling bearing enters the degradation stage, thereby realizing the division of the rolling bearing into the health stage and the degradation stage.

[0026] A further improvement of the present invention is that the specific steps in step 4) are as follows:

[0027] 1) Establish an exponential degradation model considering unequal sampling intervals:

[0028]

[0029] In the formula, xk =(x 1,k ,x 2,k ) T is a hidden variable, y k is the observed value of the health indicator, h(x k )=x 1,k exp(x 2,k t k )+c is an exponential degradation model, is process noise, and the interval between adjacent monitoring times is τ k =t k+1 -t k , r k ~N(0,R) is the measurement noise, c is the offset;

[0030] 2) Through the expectation maximization algorithm, based on the initial time t in the degradation stage 1 To the current time t T The health indicator sequence is used to update the parameters of the degradation model.

[0031] A further improvement of the present invention is that the specific steps of step 5) are as follows:

[0032] 5.1) Substituting the updated parameters of the degradation model into the degradation model to achieve recursive update of the health status of the rolling bearing;

[0033] 5.2) Extrapolate the health indicator series to obtain the point estimate and probability distribution estimate of remaining life expectancy.

[0034] A further improvement of the present invention is that the specific steps of 5.1) are as follows:

[0035] a) Remember m k-1 is time t k-1 Hidden variable x k-1 The expected estimate of S P,k-1 is time t k-1 Hidden variable x k-1 The square root estimate of the covariance of Q,k-1 is time t k-1 Process noise q k-1 The square root of the covariance of k Expected value of latent variable prediction and the square root of the covariance

[0036] Establishing augmented latent variables S R,k is time t k Measurement noise k The square root of the variance of is obtained by square root volume Kalman filtering at time t kExpected Estimation of Augmented Latent Variables and the square root estimate of the covariance The expected estimate m of the latent variable is obtained by matrix decomposition k and the square root estimate of the covariance S P,k ;

[0037] b) Whenever a new monitoring value is obtained, the current time t is obtained by performing the prediction update and measurement update steps of step a). T estimates of health indicators.

[0038] A further improvement of the present invention is that the specific steps of 5.2) are as follows:

[0039] Point Estimate of Remaining Life

[0040] Using the Monte Carlo method, we obtain the particle The remaining life The probability distribution of the remaining life is estimated to be

[0041] Where w is the set failure threshold, t T is the current moment, m T =(m 1,T ,m 2,T ) T is the expected estimate of the latent variable at the current moment, P T is the covariance estimate of the latent variables at the current moment.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] The present invention constructs health indicators based on the characteristics of rolling bearing vibration signals, realizes the mapping of feature vectors from high-dimensional space to low-dimensional space through a self-organizing map neural network with a Mahalanobis distance metric operator, and considers the correlation between variables in the dimensionality reduction process, so that the topological structure in the high-dimensional space is better preserved. On the one hand, this method better preserves the topological structure of multidimensional features in the high-dimensional space during the dimensionality reduction process, and on the other hand, overcomes the dependence on online monitoring data sampled at equal intervals, while improving the utilization rate of the full life data of historical failure samples with different sampling intervals.

[0044] Furthermore, the present invention adopts a combination of quantiles as an adaptive dynamic threshold and determines the degradation stage through a continuous trigger mechanism, which is less disturbed by noise and can better adapt to different rolling bearings with individual differences;

[0045] Furthermore, the exponential degradation model adopted by the present invention takes into account the situation of unequal sampling intervals, and can better adapt to the degradation assessment in the actual engineering environment;

[0046] Furthermore, the augmented form of the square root cubature Kalman filter adopted in the present invention has better numerical stability than the particle filter and the like, and adopts the expectation maximization algorithm to update the parameters of the model, making the results of long-term online operation more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flowchart of a method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering proposed for the invention;

[0048] Figure 2 This is the full life vibration waveform of the rolling bearing;

[0049] Figure 3 This is the full life health index value diagram of the rolling bearing;

[0050] Figure 4 This is the remaining life prediction result diagram of the rolling bearing in the later stage of the prediction period. DETAILED DESCRIPTION

[0051] The present invention is described in detail below with reference to the accompanying drawings.

[0052] Reference Figure 1 , a method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering, comprising the following steps:

[0053] (1) According to the main failure modes of rolling bearings and based on the historical failure samples of rolling bearings, the collected rolling bearing vibration signals are analyzed, and multiple related features in the time domain, frequency domain, and time-frequency domain are calculated, including root mean square, peak value, and wavelet energy entropy.

[0054] Calculate the Spearman correlation coefficient between each relevant feature and the running time series, and select the features with Spearman correlation coefficient higher than the average value as the preliminary sensitive features. Then, calculate the covariance matrix of the preliminary sensitive features, exclude redundant features based on the correlation between the features, and finally determine the streamlined multidimensional sensitive feature set.

[0055] (2) In the early operation stage of the rolling bearing, a period of stable operation data is selected as a random sample of the healthy stage, the sensitive features in the multidimensional sensitive feature set are calculated, the sensitive features are normalized, and the mean μ and covariance Σ of the sensitive features in the selected stable operation period are calculated. Figure 2 As shown in the figure, due to the fluctuations caused by the presence of noise, the mean and covariance of sensitive features are easily affected by outliers. Therefore, in order to obtain more robust mean and covariance estimates, a subset with a sample size of h is determined by the minimum covariance determinant, and the mean μ and covariance Σ of the sensitive features are calculated through this subset.

[0056] Then, the self-organizing map neural network with Mahalanobis distance metric operator is trained by using the sensitive features in the selected stable operation period. The specific process is as follows:

[0057] 1) Initialize the network: determine the number of neuron nodes in the output layer n, and initialize the weight vector m of each neuron node in the output layer j , m j ∈R n ;

[0058] 2) In the k+1th iteration, for the i-th input sample x i , x i ∈R n , calculate x i And the weight vector of each neuron node in the output layer The similarity: In the formula ||·|| M represents the Mahalanobis distance;

[0059] 3) Determine the output layer neuron nodes that correspond to sample x i The corresponding best matching unit c(x i ), whose weight vector satisfy

[0060] 4) Update the best matching unit c(x i )The weight of the neuron node in the neighborhood: In the formula, is the sample x i The corresponding best matching unit c(x i ) is the weight vector of the neuron nodes in the neighborhood, is the neighborhood function, α is the learning rate, 0<α<1, and it decreases monotonically with the increase of the number of iterations. σ controls the range of the neighborhood and also decreases monotonically with the increase of the number of iterations.

[0061] 5) Repeat steps 2) to 4) above until the required number of iterations is met.

[0062] When the self-organizing map neural network is trained, the sensitive feature vector x of the rolling bearing monitored in real time is i Input to the self-organizing map neural network to determine the best matching unit c(x i ), calculate the minimum quantization error, that is This enables the mapping of sensitive feature vectors from high-dimensional space to low-dimensional space. Figure 3 As shown, the health index of the rolling bearing is constructed using the minimum quantization error obtained.

[0063] (3) Calculate the dynamic failure threshold z = Q based on the health index of the rolling bearing 3 +1.5IQR, IQR=Q 3 -Q 1 , where Q 3 is the third quantile of the health index from the initial moment to the current moment, Q 1 It is the first quantile of the health index from the initial moment to the current moment, and IQR is the difference between the two quantiles.

[0064] Then the health status of the rolling bearing is judged according to the dynamic failure threshold z. In order to reduce the influence of noise, Figure 3 As shown, it is set that when the health index exceeds the dynamic failure threshold z for n consecutive times and these n health indicators show an increasing trend, it is judged that the rolling bearing enters the degradation stage, thereby realizing the division of the rolling bearing into the healthy stage and the degradation stage.

[0065] (4) Establish a degradation model and realize online update of degradation model parameters.

[0066] 1) Establish an exponential degradation model considering unequal sampling intervals:

[0067]

[0068] In the formula, x k =(x 1,k ,x 2,k ) T is a hidden variable, y k is the observed value of the health indicator, h(x k )=x 1,k exp(x 2,k t k )+c is an exponential degradation model, is process noise, and the interval between adjacent monitoring times is τ k =t k+1 -t k , r k ~N(0,R) is the measurement noise, c is the offset;

[0069] In the degradation model, the process noise is related to time. As the sampling interval increases, the covariance of the process noise increases and the uncertainty of the latent variables increases.

[0070] 2) Through the expectation maximization algorithm, based on the initial time t in the degradation stage 1 To the current time t T The health indicator sequence of the exponential degradation model is updated

[0071] Step E:

[0072]

[0073] In the formula, Θ (n) is the parameter obtained in the previous iteration, m 1 is the initial time t 1 Hidden variable x 1 The expected estimate, P 1 is the initial time t 1 Hidden variable x 1 The covariance estimate of is the initial time t of the smoothed estimate 1 Hidden variable x 1 expectations, is the initial time t of the smoothed estimate 1 Hidden variable x 1 The covariance of .

[0074] M-step:

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] In the formula, the expected solution can be obtained by RTS smoothing and volume transformation. The full life data of historical failure samples can be used to train the degradation model and determine a reasonable initial value for the maximum expectation algorithm.

[0081] (5) Figure 4 As shown, the condition of the rolling bearing is estimated, and the remaining life of the rolling bearing is predicted.

[0082] 1) Using the augmented form of square root cubature Kalman filtering, the updated degradation model parameter θ is substituted into the degradation model to achieve recursive update of the rolling bearing health status:

[0083] Prediction update: remember m k-1 is time t k-1 Hidden variable x k-1 The expected estimate of S P,k-1 is time t k-1 Hidden variable x k-1 The square root estimate of the covariance of Q,k-1 is time t k-1 Process noise q k-1 The square root of the covariance of , and then the degradation model can be obtained at time t kExpected value of latent variable prediction and the square root of the covariance

[0084] Measurement Update: Building Augmented Latent Variables S R,k is time t k Measurement noise k The square root of the variance of , obtained by square root volume Kalman filtering at time t k Expected Estimation of Augmented Latent Variables and the square root estimate of the covariance By matrix decomposition, we can further obtain the expected estimate m of the latent variable k and the square root estimate of the covariance S P,k .

[0085] Whenever a new monitoring value is obtained, the current time t can be obtained by performing the above prediction update and measurement update steps. T estimates of health indicators.

[0086] 2) By extrapolating the health indicator series, we can obtain the point estimate and probability distribution estimate of remaining life expectancy.

[0087] The point estimate of remaining life is w is the set failure threshold.

[0088] Using the Monte Carlo method, we obtain the particle The remaining life The probability distribution of the remaining life is estimated to be

[0089] Where w is the set failure threshold, t T is the current moment, m T =(m 1,T ,m 2,T ) T is the expected estimate of the latent variable at the current moment, P T is the covariance estimate of the latent variables at the current moment.

[0090] In order to demonstrate the effectiveness of the method of the present invention, a full-life sample of a rolling bearing is used to carry out performance degradation assessment and remaining life prediction using the method of the present invention:

[0091] like Figure 2 As shown, the vibration signal of the rolling bearing is relatively stable in the early stage, with random fluctuations, and enters the degradation stage after a period of stability.

[0092] like Figure 3As shown in the figure, after the sensitive features are determined, the dimension reduction is performed through the self-organizing map neural network with the Mahalanobis distance metric operator. The comprehensive health index of the rolling bearing can be further constructed based on the minimum quantization error. Based on this health index, the rolling bearing can be divided into the healthy stage and the degradation stage.

[0093] like Figure 4 As shown in the figure, the remaining life prediction of rolling bearings is realized by using square root volume Kalman filtering and extrapolating the sequence. As the monitoring data increases, the remaining life prediction value is getting closer and closer to the true value.

[0094] The present invention can better preserve the topological structure of multidimensional features in high-dimensional space during the dimensionality reduction process, overcome the dependence on online monitoring data with equal interval sampling, and at the same time improve the utilization rate of the full life data of historical failure samples with different sampling intervals, and can better adapt to the actual engineering environment.

Claims

1. A method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering, characterized in that: The following steps are involved: Extract multidimensional features from the vibration signal of rolling bearings and perform dimensionality reduction through self-organizing map neural network with Mahalanobis distance metric operator to construct health indicators; An adaptive degradation threshold for rolling bearings is constructed based on health indicators, and a continuous triggering mechanism is used to divide rolling bearings into healthy and degraded stages. The remaining life of rolling bearings is predicted by considering the exponential degradation model with unequal sampling intervals. The specific steps include: 1) According to the failure mode of rolling bearings, based on the collected rolling bearing vibration signals, feature extraction and screening are performed to construct a multi-dimensional sensitive feature set; 2) Using a trained self-organizing map neural network with a Mahalanobis distance metric operator, the multidimensional sensitive features in the multidimensional sensitive feature set are mapped to a low-dimensional space to construct a health indicator for the rolling bearing; 3) Based on the health indicators of rolling bearings, a dynamic failure threshold for rolling bearings is constructed, and a continuous trigger mechanism is used as a method to judge degradation, thereby realizing the division of rolling bearings into healthy and degraded stages; 4) Establish an exponential degradation model that takes into account unequal sampling intervals. Based on the health indicator data of the rolling bearing degradation stage, the degradation model parameters are updated using the expectation maximization algorithm. 5) Based on the updated degradation model parameters, the rolling bearing health status is updated through square root cubature Kalman filtering, and the point estimate and probability density estimate of the remaining life of the rolling bearing are calculated; The specific steps in step 4) are as follows: 1) Establish an exponential degradation model considering unequal sampling intervals: Where x k =(x 1,k ,x 2,k ) T is a hidden variable, y k is the observed value of the health indicator, h(x k )=x 1,k exp(x 2,k t k )+c is an exponential degradation model, is process noise, and the interval between adjacent monitoring times is τ k =t k+1 -t k , r k ~N(0,R) is the measurement noise, c is the offset; 2) Through the expectation maximization algorithm, based on the initial time t1 to the current time t in the degradation stage T The health indicator sequence is used to update the parameters of the degradation model; The specific steps of step 5) are as follows: 5.1) Substituting the updated degradation model parameters into the degradation model to achieve recursive update of the rolling bearing health status; 5.2) Extrapolate the health indicator series to obtain point estimates and probability distribution estimates of remaining life expectancy; The specific steps of 5.1) are as follows: a) Remember m k-1 is time t k-1 Hidden variable x k-1 The expected estimate of S P,k-1 is time t k-1 Hidden variable x k-1 The square root estimate of the covariance of Q,k-1 is time t k-1 Process noise q k-1 The square root of the covariance of , the next moment t is obtained by the degradation model k Expected value of latent variable prediction and the square root of the covariance Establishing augmented latent variables S R,k is time t k Measurement noise r k The square root of the variance of , obtained by square root volume Kalman filtering at time t k Expected Estimation of Augmented Latent Variables and the square root estimate of the covariance The expected estimate m of the latent variable is obtained by matrix decomposition k and the square root estimate of the covariance S P,k ; b) Whenever a new monitoring value is obtained, the current time t is obtained by performing the prediction update and measurement update steps of step a). T estimates of health indicators; The specific steps of 5.2) are as follows: Point estimate of remaining life Using the Monte Carlo method, we obtain the particle The remaining life The probability distribution of remaining life is estimated to be Where w is the set failure threshold, t T is the current moment, m T =(m 1,T ,m 2,T ) T is the expected estimate of the latent variable at the current moment, P T is the covariance estimate of the latent variables at the current moment.

2. The method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering according to claim 1, characterized in that: The specific steps in step 1) are as follows: According to the failure mode of rolling bearings and based on the collected historical failure samples of rolling bearings, the collected rolling bearing vibration signals are analyzed, and multiple relevant features in the time domain, frequency domain and time-frequency domain are calculated, including root mean square, peak value and wavelet energy entropy, etc. The Spearman correlation coefficient between each relevant feature and the running time series is calculated, and the features with Spearman correlation coefficient higher than the average value are screened out as the preliminary sensitive features. Then, the covariance matrix of the preliminary sensitive features is calculated. According to the correlation between the features, redundant features are excluded and the multidimensional sensitive feature set is determined.

3. The method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering according to claim 1, characterized in that: The specific steps in step 2) are as follows: Data from a period of stable operation of a rolling bearing is selected as a random sample of the healthy stage. The features in the multidimensional sensitive feature set are calculated and normalized. The mean and covariance of the sensitive features during the stable operation period are also calculated. The self-organizing map neural network with Mahalanobis distance metric operator is trained by using sensitive features during the stable operation period. Through the trained self-organizing map neural network with Mahalanobis distance metric operator, multi-dimensional sensitive features are mapped to low-dimensional space to construct the health index of rolling bearings.

4. The method for predicting the remaining life of a rolling bearing based on square root volumetric Kalman filtering according to claim 1, characterized in that: The specific steps in step 3) are as follows: According to the health index of the rolling bearing, the dynamic failure threshold of the rolling bearing is calculated as z = Q3 + 1.5IQR, IQR = Q3 - Q1, where Q3 is the third quantile of the health index from the initial moment to the current moment, Q1 is the first quantile of the health index from the initial moment to the current moment, and IQR is the difference between the two quantiles; According to the failure threshold z Determine the health status of the rolling bearing and set the failure threshold when the health indicator exceeds the failure threshold for n consecutive times z When the n-time health index shows an increasing trend, it is judged that the rolling bearing has entered the degradation stage, thereby realizing the division of the rolling bearing into the healthy stage and the degradation stage.