CN116502049A

CN116502049A - Rolling bearing residual service life prediction method, system, equipment and storage medium

Info

Publication number: CN116502049A
Application number: CN202310744032.1A
Authority: CN
Inventors: 许传诺; 李继明; 杜丰钧; 程学珍
Original assignee: Shandong University of Science and Technology
Current assignee: Shandong University of Science and Technology
Priority date: 2023-06-25
Filing date: 2023-06-25
Publication date: 2023-07-28
Anticipated expiration: 2043-06-25
Also published as: CN116502049B

Abstract

The invention belongs to the technical field of rolling bearings, and particularly discloses a method, a system, equipment and a storage medium for predicting the residual service life of a rolling bearing. The method comprises the following steps: firstly, performing noise reduction processing operation on original vibration signals of a rolling bearing by using a rolling bearing vibration signal data preprocessing method; IMF component selection and reconstruction are carried out according to kurtosis and a correlation coefficient principle, and a rolling bearing fault characteristic reconstruction signal is obtained; step 2, extracting features of the obtained rolling bearing fault feature reconstruction signals to obtain feature sets, selecting the feature sets by using a Spearman-MIR combined method, and constructing low-dimensional sensitive feature sets; and 3, building a DE-BiLSTM prediction model, and inputting the low-dimensional sensitive feature set into the prediction model to predict the residual life of the rolling bearing. The method is beneficial to realizing effective prediction of the residual service life of the rolling bearing, and has high prediction precision.

Description

Rolling bearing residual service life prediction method, system, equipment and storage medium

Technical Field

The invention belongs to the technical field of rolling bearings, and particularly relates to a method, a system, equipment and a storage medium for predicting the residual service life of a rolling bearing.

Background

Rolling bearings are commonly used as "industrial joints" components of equipment, and the safety and reliability of operation of rolling bearings are directly related to the normal operation of the equipment. However, rolling bearings are often operated in complex operating environments, and damage and failure are unavoidable, thereby affecting the operation performance of the apparatus, causing great economic loss and threatening the safety of personnel. The method has the advantages that the residual service life (Remaining Useful Life, RUL) of the rolling bearing is effectively predicted, the fault risk can be avoided, the running stability and the safety are improved, and particularly, after the RUL technology is adopted, the rolling bearing can be timely maintained or replaced before the fault of the rolling bearing occurs, fault early warning can be achieved, the probability of production accidents is reduced, the maintenance cost is reduced, the resource allocation is optimized, the long-term healthy and stable running of equipment is guaranteed, and the method has important practical significance in the aspects of improving the reliability and the safety of the equipment, enhancing the guarantee capability and the like.

The residual service life prediction of the rolling bearing comprises three parts of signal data preprocessing, feature extraction and RUL prediction.

Rolling bearings are usually operated continuously in complex conditions, and the characteristic signals of the rolling bearings are mostly obtained as nonlinear, non-stationary sequence signals, accompanied by a large number of interference signals. In order to accurately acquire the vibration signal of the rolling bearing and reduce signal interference such as noise, data preprocessing is needed. Methods for preprocessing rolling bearing data are currently mainly divided into two categories: firstly, reducing signal noise and optimizing signal-to-noise ratio; and secondly, the fault pulse information is highlighted, and the influence of interference signals is weakened.

Although the data preprocessing has better effects by adopting the method of reducing signal noise and highlighting fault impulse information, if noise is reduced by adopting a decomposition method only, as white noise mixed with an original vibration signal covers the whole frequency bandwidth, each mode inevitably contains some mode noise, which is easy to submerge fault related information, and the vibration signal after the data preprocessing contains a large amount of fault information of the running state of the rolling bearing, however, if the RUL prediction is directly carried out by adopting the vibration signal, inaccurate prediction results may be caused, and therefore, characteristic parameters need to be extracted to intuitively represent the deviation degree, steepness, impact and other fault characteristics of the bearing data. When the rolling bearing is damaged or early-stage fault occurs, state information is changed, characteristic signals are correspondingly changed, and health state assessment is realized by using the changed characteristic parameters.

RUL prediction methods include methods based on mechanism modeling, based on reliability statistical analysis, and based on data driving, however, these methods are not highly accurate in predicting the remaining useful life of a rolling bearing.

Disclosure of Invention

The invention aims to provide a method for predicting the residual service life of a rolling bearing, so that the effective prediction of the residual service life of the rolling bearing is realized, and the prediction precision is high. In order to achieve the above purpose, the invention adopts the following technical scheme:

A method for predicting the residual service life of a rolling bearing comprises the following steps:

firstly, performing data preprocessing operation on original vibration signals of a rolling bearing by using an ISSA-MCKD-ICEEMDAN (integrated circuit-based random access memory) vibration signal data preprocessing method of the rolling bearing so as to highlight periodic signals and reduce signal interference;

IMF component selection and reconstruction are carried out according to kurtosis and a correlation coefficient principle, and a rolling bearing fault characteristic reconstruction signal is obtained;

step 2, extracting features of the rolling bearing fault feature reconstruction signals obtained in the step 1 to obtain feature sets, selecting the feature sets by using a Spearman-MIR combined method, and constructing low-dimensional sensitive feature sets;

and 3, building a DE-BiLSTM prediction model, and inputting the low-dimensional sensitive feature set obtained in the step 2 into the DE-BiLSTM prediction model to predict the residual life of the rolling bearing, so as to obtain a prediction result of the rolling bearing RUL.

In addition, on the basis of the method for predicting the residual service life of the rolling bearing, the invention also provides a rolling bearing residual service life prediction system which is suitable for the method, and the system adopts the following technical scheme:

a rolling bearing remaining useful life prediction system, comprising:

the signal data preprocessing module is used for reducing noise of an original vibration signal of the rolling bearing by using an ISSA-MCKD-ICEEMDAN (integrated circuit-based on-chip-based noise) of the vibration signal data preprocessing method of the rolling bearing so as to highlight a periodic signal and reduce signal interference;

the feature extraction module is used for extracting features of the obtained rolling bearing fault feature reconstruction signals to obtain a multi-domain feature parameter set, selecting the feature set by using a Spearman-MIR combined method and constructing a low-dimensional sensitive feature set;

and the RUL prediction module is used for building a DE-BiLSTM prediction model, inputting the obtained low-dimensional sensitive feature set into the DE-BiLSTM prediction model to predict the residual life of the rolling bearing, and obtaining a prediction result of the RUL of the rolling bearing.

The invention also proposes a computer device comprising a memory and one or more processors, based on the method for predicting the remaining life of a rolling bearing.

The memory stores executable code, and the processor is used for realizing the steps of the method for predicting the residual service life of the rolling bearing when executing the executable code.

The invention further provides a computer readable storage medium on which a program is stored, based on the method for predicting the remaining service life of the rolling bearing.

The program, when executed by a processor, is adapted to carry out the steps of the above-mentioned method for predicting the remaining service life of a rolling bearing.

The invention has the following advantages:

as described above, the present invention describes a method, system, apparatus and storage medium for predicting the remaining life of a rolling bearing. The invention provides a preprocessing method of vibration signals, which aims at solving the problems that the vibration signals of the rolling bearing are easy to be polluted by noise and the degradation trend of the bearing cannot be directly and accurately represented in a complex environment, and the preprocessing method firstly adopts the deconvolution (MaximumCorrelation Kurtosis Deconvolution, MCKD) with the maximum correlation kurtosis to highlight the periodic impact component covered by the interference signals, so that the influence of the interference signals is reduced; then, aiming at the problem that the noise reduction effect of the MCKD is influenced by parameter selection, adopting a sparrow search algorithm (Improved Sparrow Search Algorithm, ISSA) to adaptively select the optimal parameter combination of the MCKD so as to avoid subjectivity of manually setting parameters; finally, the improved self-adaptive noise-based complete set empirical mode decomposition (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, ICEEMDAN) is utilized to decompose the vibration signals, the obtained eigenmode functions (Intrinsic Mode Functions, IMF) are screened according to the kurtosis and correlation coefficient principles, a reconstruction signal is obtained, information redundancy is reduced, and the accuracy of degradation information represented by the characteristic parameters extracted later is improved. Secondly, the invention provides a rolling bearing feature extraction and selection method aiming at the problems of incomplete vibration signal feature extraction and information redundancy, wherein the feature extraction and selection method firstly extracts time domain, frequency domain, entropy and fractal dimension features and comprehensively evaluates the running state of the rolling bearing; then, according to the degradation curve of the characteristic parameters, screening characteristic indexes conforming to the degradation trend of the bearing; finally, a combined dimension reduction method of Spearman (Spearman) correlation coefficient and mutual information regression (Mutual Information Regression, MIR) is provided, degradation characteristics with high correlation and low redundancy are selected, and redundant information in characteristic parameters is reduced. In addition, the invention also provides a BiLSTM prediction model based on a differential evolution algorithm (Differential Evolution, DE) aiming at the problem that the ultra-parameter setting of a bidirectional long-short-term memory network (Bidirectional Long Short Term, biLSTM) influences the prediction precision in the RUL prediction of the rolling bearing.

Drawings

FIG. 1 is a flow chart of a method for predicting the remaining service life of a rolling bearing based on DE-BiLSTM in an embodiment of the invention.

FIG. 2 is a schematic diagram of the predicted results of LSTM model on Bearing 1_1 dataset.

FIG. 3 is a schematic diagram of the prediction results of LSTM model on Bearing 2_1 dataset.

FIG. 4 is a diagram showing the prediction results of BiLSTM model on Bearing 1_1 dataset.

FIG. 5 is a diagram showing the prediction results of BiLSTM model on Bearing 2_1 dataset.

FIG. 6 is a diagram showing the predicted results of the DE-LSTM prediction model on the Bearing 1_1 dataset in an embodiment of the present invention.

FIG. 7 is a diagram showing the predicted results of the DE-LSTM prediction model on the Bearing 2_1 dataset in an embodiment of the present invention.

Fig. 8 is a schematic diagram of the prediction result of the BiLSTM model based on the original signal in the dataset Bearing 1_4.

Fig. 9 is a schematic diagram of the prediction result of the BiLSTM model based on the original signal in the data set Bearing 3_3.

FIG. 10 is a diagram showing the prediction results of the DE-BiLSTM model based on the original signal in the data set Bearing 1_4.

FIG. 11 is a diagram showing the prediction results of the DE-BiLSTM model based on the original signal in the data set Bearing 3_3.

Fig. 12 is a schematic diagram of the prediction result of the BiLSTM model based on the preprocessing signal in the dataset Bearing 1_4.

Fig. 13 is a schematic diagram of the prediction result of the BiLSTM model based on the preprocessing signal in the data set Bearing 3_3.

FIG. 14 is a diagram showing the predicted results of the DE-BiLSTM model on the basis of the pre-processed signals in the dataset Bearing 1_4.

FIG. 15 is a diagram showing the predicted results of the DE-BiLSTM model on the basis of the pre-processed signals in the dataset Bearing 3_3.

Detailed Description

Noun interpretation:

ISSA-MCKD-ICEEMDAN, i.e. a combination of ISSA algorithm, MCKD algorithm and ICEEMDAN algorithm;

Spearman-MIR, i.e. a combination of Spearman algorithm and MIR algorithm;

DE-BiLSTM, i.e. a combination of DE algorithm and BiLSTM algorithm.

The invention is described in further detail below with reference to the attached drawings and detailed description:

example 1

The present embodiment 1 describes a method for predicting the remaining service life of a rolling bearing, which is based on an LSTM neural network and is improved in three aspects of vibration signal preprocessing, degradation feature extraction and selection, and RUL prediction.

As shown in fig. 1, the method for predicting the remaining service life of the rolling bearing comprises the following steps:

and step 1, firstly, performing data preprocessing operation on original vibration signals of the rolling bearing by using an ISSA-MCKD-ICEEMDAN (integrated circuit based on the information) vibration signal data preprocessing method of the rolling bearing so as to highlight periodic signals and reduce signal interference.

And selecting and reconstructing IMF components according to kurtosis and a correlation coefficient principle to obtain a rolling bearing fault characteristic reconstruction signal.

In a system with more interference factors or complex system, the degradation characteristic information of the rolling bearing is weak and severely affected by noise interference, the signal-to-noise ratio is reduced, and since the noise mixed with the original vibration signal covers the whole frequency bandwidth, each mode of the ICEEMDAN inevitably contains some mode noise, which easily floods the fault related information.

Therefore, before ICEEMDAN is decomposed, the vibration signal needs to be preprocessed, namely, the MCKD algorithm is adopted, the original periodic pulse signal is increased as much as possible, the interference influence is reduced, and a feature set with obvious bearing degradation trend is obtained, so that the novel method based on ISSA-MCKD-ICEEMDAN data preprocessing is provided on the basis of solving the problems that the rolling bearing is easy to be interfered by signals such as noise and the like in the running process, namely:

firstly, parameters L and T of MCKD are optimized by ISSA, an original signal is preprocessed to highlight a periodic pulse signal, and the signal is decomposed and reconstructed by ICEEMDAN algorithm, so that the defect of the ICEEMDAN algorithm is overcome, effective information filtering is reduced, and information detail identification degree is improved while denoising.

The ISSA-MCKD-ICEEMDAN data preprocessing process is as follows:

initializing an ISSA population, and setting the population scale and the maximum iteration number.

The range of the parameter filter length L and the deconvolution period T of the MCKD is defined as a spatial domain of ISSA algorithm population search, the maximum relevant kurtosis value of the MCKD is used as a fitness function, and (L, T) is used as a parameter combination to be optimized.

The optimal parameter combination of the maximum correlation kurtosis deconvolution MCKD is selected through ISSA self-adaption.

And 1.2, performing enhanced pretreatment on the original vibration signal of the rolling bearing with noise by using the MCKD after the parameters are adaptively selected so as to highlight fault pulses and reduce noise interference, thereby obtaining a noise reduction signal.

And step 1.3, ICEEMDAN decomposition is carried out on the noise reduction signal to obtain a plurality of IMF component signals.

And step 1.4, selecting an IMF component from the IMF component signals in step 1.3 for reconstruction based on a correlation coefficient and a kurtosis principle, and obtaining a rolling bearing fault feature reconstruction signal with obvious fault features.

MCKD is sensitive to signal periodicity so it is easy to extract a faulty signal, but it is critical to and greatly affected by input parameters. The filter length L determines the number of samples required by each iteration, and the iteration time is influenced excessively, and the calculation accuracy is influenced excessively; the parameter T is affected by errors caused by the running wear of the rolling bearing and is difficult to manually determine.

Therefore, aiming at the problem of the input parameters of the MCKD, the embodiment of the invention optimizes the parameters of the MCKD by using the ISSA. The specific procedure for optimizing MCKD parameters using ISSA is as follows:

setting the population quantity X as 100, searching the number of the users as 20%, and setting the maximum correlation kurtosis value of the signals of the MCKD as a fitness function according to the filter length L epsilon [1,400], the period T epsilon [1,400], and the step 1.1.1.

Step 1.1.2. Setting upper and lower bounds of the region, initializing the sparrow position in the space domain according to the improved Circle chaotic map, and setting the objective function as parameters of MCKD, namely the filter length L and the deconvolution period T.

Step 1.1.3. Introducing self-adaptive weight omega, updating the position of the searcher, and enabling the follower to find food around the optimal area found by the searcher, possibly competing for the food, so that the follower becomes the searcher. If the sparrow individual is in a very starved state, then the sparrow individual needs to find food elsewhere, and the location update function of the follower is used for location update.

Step 1.1.4. When the alerter encounters a hazard, an update is made using an alerter location update function. Sparrows at the periphery of the population are gathered towards the safety area, and the sparrows at the center of the population randomly walk to be close to other sparrows.

And 1.1.5, introducing a reverse learning strategy, and performing reverse solution treatment on the sparrow positions, so that the algorithm better jumps out of local optimum, and the optimizing precision of the algorithm is improved.

And 1.1.6, finding the optimal position in the space domain in iteration, stopping iteration and outputting the optimal position, namely the optimal filter length L and the deconvolution period T, when the maximum iteration times are reached, otherwise, turning to the step 1.1.3.

In addition, in order to verify the effectiveness of the ISSE-MCKD-ICEEMDAN pretreatment method provided by the method, an XJTU-SY rolling bearing accelerated life test data set is used for verification in the embodiment of the invention.

The time domain waveform diagram fault pulse of the original signal in the XJTU-SY rolling bearing accelerated life test data set is not prominent, and the envelope spectrum fault characteristic frequency is also not obvious, because the signal contains more interference and is subjected to pretreatment.

Therefore, the ISSA-MCKD is adopted to preprocess the original signal, and the impact component is highlighted, and the specific process is as follows:

firstly, optimizing parameters of the MCKD by using an ISSA algorithm, setting the value range of the parameter L to be [1,400], setting the value range of the parameter T to be [1,400], and setting the optimal parameter combination (L, T) = (339,185) of the MCKD;

the original signal is then preprocessed using the MCKD after the parameters have been adaptively selected.

ICEEMDAN decomposition is performed on the ISSA-MCKD processed signal for subsequent degradation feature extraction. The ICEEMDAN decomposition yields 13 IMF components, with the useful information gradually decreasing as the components increase.

In order to reduce the redundancy of information and improve the accuracy of extracting a degradation characteristic set, the invention applies a correlation coefficient and kurtosis rule to calculate the correlation coefficient and kurtosis value of an IMF component obtained after ICEEMDAN decomposition.

The periodic pulse of the time domain waveform diagram is more prominent after the signal is decomposed and reconstructed by ICEEMDAN, the fault characteristic frequency and the frequency multiplication thereof can be clearly identified, the fault characteristic extraction effect is more excellent, and more accurate bearing degradation information can be obtained.

In addition, in order to further illustrate the effectiveness of the ISSE-MCKD-ICEEMDAN pretreatment method provided by the invention, the signal processed by the custom parameter MCKD, the signal processed by the ISSA parameter optimization and selection MCKD, and the signal processed by the ISSA parameter optimization and selection MCKD and ICEEMDAN are compared with the original signal.

By calculating the first 5 characteristic fault frequenciesf _g And the ratio of the pre-processed conflict signal to the sum of the amplitude values of 0-1000 Hz, wherein the larger the amplitude ratio is, the more obvious the pre-processed conflict signal is. The calculation results are shown in table 1 below.

Table 1 amplitude ratio after noise reduction in three methods

Signal signal	Amplitude ratio
		MCKD dropNoisy signal	0.0012
ISSA-MCKD noise-reduced signal	0.0033
		ISSA-MCKD-ICEEMDAN noise-reduced signal	0.0087

From table 1, it can be seen that, according to the comparison analysis of the amplitude ratios of the three signals, the result obtained by combining MCKD and icemdan is the optimal difference, so that it is proved that the ISSA-MCKD-icemdan data preprocessing method provided by the invention can better remove noise interference in the original signal, and the pulse component is highlighted under the condition that the information contained in the original signal is unchanged, so that the interference suppression effect is better, and the accuracy is improved for extracting degradation characteristics later.

The method comprises the steps of firstly improving an initialization process of SSA, improving optimizing efficiency, optimizing parameter setting of MCKD by using ISSA, then providing an ISSA-MCKD-ICEEMDAN data preprocessing method on the basis, designing a data preprocessing flow, finally carrying out verification analysis by using an XJTU-SY test data set, and comparing amplitude ratio results.

And 2, extracting features of the rolling bearing fault feature reconstruction signals obtained in the step 1 to obtain feature sets, selecting the feature sets by using a Spearman-MIR combined method, and constructing low-dimensional sensitive feature sets.

The rolling bearing performance degradation curves under different working conditions are different, so that a large difference exists between a rolling bearing RUL prediction model obtained through data set training and the service life of an actual rolling bearing.

Therefore, it is necessary to extract features reflecting the consistency of the degradation curves of the respective rolling bearings.

According to the invention, 33 degradation characteristic parameters based on time domain, frequency domain, entropy and fractal dimension are extracted from the vibration signal after data preprocessing, and degradation trend graphs of the 33 degradation characteristic parameters are analyzed and screened, so that a Spearman-MIR combined method is designed, and the characteristic parameters with low redundancy and high correlation are selected as the health index of the rolling bearing.

The step 2 specifically comprises the following steps:

and 2.1, extracting time domain, frequency domain, entropy and fractal dimension characteristics as much as possible from the obtained rolling bearing fault characteristic reconstruction signals obtained in the step 1, thereby comprehensively evaluating the running state of the rolling bearing.

Unlike conventional feature parameters, degradation features require higher sensitivity to identify each performance degradation stage, while vibration signals of rolling bearings are very complex, extracting only a single feature, and may result in information loss.

In this embodiment, in order to comprehensively evaluate the running state of the rolling bearing, 33 feature indexes are extracted, wherein the feature indexes include 16 time domain features, 13 frequency domain features, 3 entropy features and 1 fractal dimension feature.

Time domain analysis is the simplest and most direct signal analysis method adopted in most rolling bearing online monitoring systems at present. According to different working conditions, the time domain characteristic value can be changed correspondingly, and the working environment has a larger influence on the time domain characteristic value.

The disadvantage of the dimensional time domain features is unstable performance, the performance is complex in actual engineering, the dimensionless time domain features are not interfered by load and rotation speed changes basically, and the running state can be described relatively intuitively.

Thus, the dimensionless feature is typically combined with the dimensionless feature to collectively characterize the bearing degradation state.

The dimensional time domain parameter describes the bearing state and can reflect partial fault information, including maximum value, minimum value, mean value, variance, square root amplitude, absolute average amplitude, peak-to-peak value, root mean square value, skewness and kurtosis.

The dimensionless characteristic parameter indexes mainly comprise a peak value factor, a waveform factor, a pulse factor, a margin factor, a skewness factor and a kurtosis factor, wherein the peak value factor reflects the peak value degree in the waveform; the pulse factor is sensitive to the impact signal; form factor = pulse factor/peak factor; the margin factor may then represent the fatigue wear level of the component; the skewness factor is used to describe the distribution of the variables; the kurtosis factor reflects the shock characteristics of the vibration signal.

The spectrum analysis is to analyze the time domain signal after fourier transform decomposition into the frequency domain. When the health state of the rolling bearing is changed, frequency components in the frequency spectrum are correspondingly changed, so that signal frequency spectrum information can be accurately represented by analyzing frequency domain features of the vibration signal, and then the running state of the rolling bearing under different working conditions is detected.

For fourier transformed vibration signalsS(k) And (3) representing.

In this embodiment, 13 frequency domain features are extracted, using F respectively ₁ -F ₁₃ Representation, wherein F ₁ Reflecting the vibration energy of the frequency domain, F ₂ -F ₅ Reflecting the change of the position of the main frequency band, feature F ₆ -F ₁₃ Reflecting the degree of dispersion or concentration of the spectrum.

The information entropy is one of the effective factors for quantitatively revealing uncertain information in the signal and reflecting the characteristics contained in the signal, and the basic idea is as follows: events with low probability of occurrence contain a higher amount of information.

The entropy features selected in the embodiment include power spectrum entropy, singular spectrum entropy and energy entropy.

The power spectral entropy is a quantitative description of the complexity of the energy distribution of a signal in the frequency domain. The closer the signal is to white noise, the more dispersed the power spectrum and the greater the power spectrum entropy.

In order to reveal the details of the internal energy variation of the complex signal, singular spectrum entropy is also needed, and the basic idea is to acquire the intrinsic complexity characteristics of the complex signal by carrying out phase space reconstruction and singular value decomposition on the time domain signal sequence of the system.

There are various approaches to energy entropy, typically combined with signal decomposition algorithms, such as EMD, CEEMD, VMD.

The fractal dimension is an important parameter in fractal theory.

The fractal theory is one of three nonlinear theories, can express the nonlinearity, self-similarity and complexity of signals, and the fractal dimension can be adopted to represent the irregularity of the signals under a certain time domain length of the rolling bearing vibration signals.

Each feature analysis method has advantages and disadvantages, can comprehensively consider, extract multi-domain features to comprehensively represent the degradation trend of the rolling bearing, and select all extracted feature parameters by using a proper feature selection method to obtain an optimal feature set.

Feature selection becomes key to data dimension reduction in order to handle high-dimensional data. Therefore, before the rolling bearing performs the remaining life prediction work, feature selection needs to be performed on the extracted 33 features.

Firstly, aiming at a large number of features extracted for comprehensively detecting the state of a bearing in the earlier stage, firstly screening out features which have the same degradation trend with the bearing and are less affected by interference; the redundancy of the feature set is then reduced based on the Spearman-MIR method. The method selects the filtering method to select the characteristics, directly selects a small amount of sensitive characteristics, and reduces the running time of the RUL prediction model.

And 2.2, screening characteristic indexes which accord with the degradation trend of the bearing as a characteristic set according to the degradation curve of the characteristic parameters, wherein the step 2.2 is one-step screening, and firstly screening out the characteristics which are the same as the degradation trend of the bearing and are less affected by interference.

Aiming at the time domain, frequency domain, entropy and fractal dimension characteristics extracted in the step 2.1, the running state of the bearing is stable in the early running stage of the bearing, and the characteristic parameters are not changed obviously; when the life cycle of the bearing enters the later stage, the bearing has a degradation trend, and the characteristic parameters can be increased or decreased excessively. In the aspect of time domain characteristics, for dimensional time domain characteristics, variance, square root amplitude, absolute average amplitude, root mean square value, skewness and kurtosis are gradually changed and become larger in value along with time increment, so that the response to the bearing state change is sensitive. And the minimum value and the burrs of the peak value are too much, are easy to interfere, and cannot provide accurate indexes for predictive analysis.

For dimensionless time domain features, peak factors, pulse factors and margin factors continuously and severely fluctuate throughout the life cycle, while waveform factors, skewness factors and kurtosis factors suddenly fluctuate in a larger range although degradation starting time is suddenly changed, and do not accord with degradation trend, so 7 indexes are selected according to the time domain features: variance, square root amplitude, absolute average amplitude, maximum, root mean square value, skewness, and kurtosis. In the frequency domain characteristic indexes, the invention selects the characteristic 1, the characteristic 6, the characteristic 9, the characteristic 10 and the characteristic 13 as characteristic values, and the 5 indexes are similar to the variation trend of the bearing degradation curve, so that the characteristic values are reserved. The entropy and fractal dimension are not used because there are too many burrs and there is no significant tendency to rise or fall.

The calculation formulas of the feature 1, the feature 6, the feature 9, the feature 10 and the feature 13 are respectively as follows:

；

wherein, the liquid crystal display device comprises a liquid crystal display device,；

。

wherein, the liquid crystal display device comprises a liquid crystal display device,Kthe length of the signal is indicated and,f _k representing the spectrum. 7 indexes are selected in the time domain, 5 indexes are selected in the frequency domain through one-time screening, and 12 characteristic indexes are combined to form 12 dimensionsFeature vectors.

And 2.3, performing secondary screening, namely, further selecting degradation features with high correlation and low redundancy by utilizing a combined dimension reduction method of the Spearman correlation coefficient and mutual information regression MIR, and constructing a low-dimension sensitive feature set.

Spearman (Spearman) correlation coefficient, also known as Spearman rank correlation coefficient, is a non-parametric measure of rank-based, reflecting the correlation between direction of variation trend and intensity of two variables.

For Spearman correlation coefficientP _s If (3)P _s If positive, the two variables are positively correlated; if it isP _s Negative, then the two variables are inversely related; if it isP _s Equal to 0, then the two variables are independent of their variation; when (when)P _s When 1 or-1, it is indicated that the two variables have a strictly monotonic functional relationship. The correlation strength ranking table of Spearman correlation coefficients is as follows:

TABLE 2 correlation intensity

Spearman correlation coefficientP _S	Correlation strength
		-1~-0.5, 0.5~1	Strong correlation
-0.5~-0.3, 0.3~0.5	Correlation in
		-0.3~-0.1, 0.1~0.3	Weak correlation
-0.1~0.1	No correlation

P _s Is a non-parametric or non-distributed rank statistic measure of the intensity and direction of any monotonic association between two ordering variables or one ordering variable and one measured variable. In the process of calculationP _s Previously, the two variables were first comparedX _i 、Y _i And then the rank is obtained by arranging the bit times according to the data sizeX ^’ _i 、Y ^’ _i Calculating the differenced _i ，nAnd finally, carrying out formula (2) to solve the result for the number of data contained in the variable.

(1)

（2）

Mutual information (Mutual information, MI) is used to analyze the linear or nonlinear relationship between each feature and the tag, which can be either regressive or classifying, thus containing two types—mutual information classification (Mutual Information Classification, MIC) and mutual information regression (MutualInformation Regression, MIR).

MI is used for analyzing the association degree between each characteristic index and the label, the value of the obtained result MI is between [0,1], and the larger the value is, the larger the correlation between the two variables is indicated, and the larger the association degree between the characteristic index and the service life prediction is indicated.

Formally, have a joint distributionp(x,y) Random variable of (2)XAndYmutual information between them is defined by the relation (3):

（3）

if the distribution density of the random variable is p(x,y) The mutual information can be expressed as:

（4）

and carrying out Spearman correlation coefficient calculation on 12 features (variance, square root amplitude, absolute average amplitude, maximum value, root mean square value, skewness, kurtosis, frequency domain feature 1, frequency domain feature 6, frequency domain feature 9, frequency domain feature 10 and frequency domain feature 13) extracted by one-time screening. The variance, the absolute average amplitude, the frequency domain feature 1 and the frequency domain feature 6 are all in strong correlation, and the kurtosis and the root mean square value are in strong correlation, so that the redundancy degree of the feature set is increased.

The correlation among the indexes is analyzed by adopting the Spearman correlation coefficient, the redundancy of the feature vector data set can be reduced only, and the influence degree of each feature on the residual service life of the bearing is not considered, so that the MIR is adopted for further feature selection. The variance, square root amplitude, absolute average amplitude, maximum value, root mean square value, skewness, kurtosis, frequency domain feature 1, frequency domain feature 6, frequency domain feature 9, frequency domain feature 10 and frequency domain feature 13 are respectively expressed as feature numbers 1-12.

Specific values for the extent of influence of each feature of the MIR calculation are shown in table 3.

TABLE 3 mutual information

Characteristic index	MIR value
		Variance of	0.862016
Square rootAmplitude value	0.333808
		Absolute average amplitude	0.553854
Maximum value	0.421730
		Root mean square value	0.757629
Skewness of inclination	0.498260
		Kurtosis of	0.761403
Frequency domain feature 1	0.632087
		Frequency domain features 6	0.237652
Frequency domain features 9	0.617941
		Frequency domain features 10	0.059549
Frequency domain features 13	0.732193

As can be seen from table 3, the square root amplitude, absolute average amplitude, maximum, skewness, frequency domain feature 6 and frequency domain feature 10 have smaller MIR values of 0.333808, 0.553854, 0.421730, 0.498260, 0.237652 and 0.059549, respectively, with the 6 feature indices contributing the least to bearing life. And (3) combining the Spearman correlation coefficient and the MIR, and finally reserving 4 characteristic indexes of variance, kurtosis, frequency domain characteristics 9 and frequency domain characteristics 13 as an input characteristic set of the life prediction model.

It is clear from step 2.3 that the secondary screening is based on the Spearman-MIR combination method, and that sensitive features with a relatively large correlation can be selected as degradation indicators for predicting the remaining service life of the rolling bearing.

The step 2 is used for extracting and selecting the characteristics of the preprocessed vibration signals, namely, the 33 characteristics of the 4 types of time domain, frequency domain, entropy and fractal dimension are extracted respectively, and aiming at the characteristic set with huge quantity, the characteristics are analyzed based on correlation and monotonicity indexes, a Spearman-MIR combined dimension reduction method is provided, the characteristic parameters with high correlation and low redundancy are screened out, and finally the characteristic set containing 4 characteristics is obtained.

The DE algorithm is a simple, efficient and easy-to-implement global optimization algorithm in continuous space, so that the super-parameters of BiLSTM are optimized by the DE algorithm in the embodiment.

BiLSTM has strong memory and data processing capabilities, and has great advantages for dealing with nonlinear sequence problems. Therefore, regression modeling is carried out on the input feature set by utilizing the LSTM, and a final prediction result is obtained.

It follows that the combination of the DE algorithm and the BiLSTM can effectively solve the RUL prediction problem of the rolling bearing.

The BiLSTM neural network not only can mine the spatial and temporal correlation between output variables and related input variables, but also has been greatly developed in the field of complex condition prediction.

The method has good prediction precision on the health index extracted by the rolling bearing, however, the determination of model parameters is difficult, the internal parameters of BiLSTM are too many, the model training time is long, and the model is easy to excessively fit.

In the BiLSTM neural network model parameters, the learning rate is used as an over-parameter for controlling the learning speed of the network, the network cannot be converged due to the fact that too large oscillation occurs, and the convergence complexity of the network is increased and the convergence speed is influenced if too small oscillation occurs; the number of hidden layer nodes is a direct cause of over fitting, and the more the number of layers is, the smaller the error of the whole network is, but the whole network is complicated, the training time of the network is increased, and the over fitting is also possible to occur; l2 regularization can smooth the parameters, preventing overfitting. Learning rate, number of hidden layer nodes, and L2 regularized values can affect the performance of the network training.

Therefore, the super-parameters of the BiLSTM are trained by using the DE algorithm, and finally, the optimized super-parameters are used for constructing the DE-BiLSTM prediction model, so that the workload required by manual parameter adjustment can be reduced, and the prediction accuracy can be improved.

The step 3 specifically comprises the following steps:

and 3.1, training the super-parameters of the two-way long-short-term memory network BiLSTM by using a differential evolution method DE, and constructing a DE-BiLSTM prediction model by using the optimized super-parameters.

And 3.2, inputting the low-dimensional sensitive feature set into a DE-BiLSTM prediction model to predict the rolling bearing RUL.

The step 3.1 specifically comprises the following steps:

step 3.1.1, reading a feature vector set, dividing a training set, a verification set and a test set, and carrying out normalization processing;

step 3.1.2, building a BiLSTM model, and determining learning rate, hidden layer node number and L2 regularization as optimizing parameters;

step 3.1.3, initializing current algebra, maximum iteration times, population scale, scaling factors and cross probability parameters;

step 3.1.4. Generating next generation individuals according to the mutation operation, the crossover operation and the selection operation of the DE;

step 3.1.5, repeatedly executing the step 3.1.4 to obtain a population of the next generation; evaluating the fitness value of the next generation population, wherein the minimum fitness value is a global minimum value, and the corresponding individual is a global optimal individual;

Step 3.1.6, judging whether an iteration ending condition is reached, namely, whether the maximum iteration times are reached; outputting an optimal individual if the condition of stopping iteration is reached, otherwise, turning to step 3.1.4;

and 3.1.7, transmitting the optimized learning rate, hidden layer node number and L2 regularization into a BiLSTM neural network structure to construct a DE-BiLSTM prediction model.

The step 3.1.5 specifically comprises the following steps:

step 3.1.5.1, determining DE parameters including maximum iteration times, population size, crossover probability parameters, scaling factors and value range; initializing a population, and enabling the population to cover the range of the DE algorithm;

3.1.5.2, endowing the values of population individuals with parameters and weights of BiLSTM, inputting training, performing bidirectional propagation on variables, and taking the error between the obtained predicted value and the actual value as the fitness of the individuals to obtain the fitness of all the individuals;

step 3.1.5.3, updating the value of each individual in the population, wherein the updated vector is used as a mutation vector;

3.1.5.4 after generating the mutant vector, performing cross operation on the source vector and the corresponding mutant vector to generate a cross vector;

step 3.1.5.5, determining a vector reserved in the next iteration according to the adaptation value; if the adaptive value cross vector is higher than the target vector, reserving the cross vector; if the adaptation value crossing vector is smaller than the target vector, the target vector is reserved;

3.1.5.6, judging whether the jump-out loop condition is met, wherein the condition is defined as the accuracy requirement required by prediction and the limited maximum iteration number; if the condition is met, reserving an optimal solution, and continuing to perform the next step;

if the condition is not satisfied, executing to step 3.1.5.2, and performing the next iteration;

step 3.1.5.7, taking the optimal solution of DE as the super-parameter learning rate, the hidden layer node number and the L2 regularization of BiLSTM.

The step 4 is to divide the full life cycle of the rolling bearing into a stable operation period and a wear period according to the RMS value, determine the degradation starting time of entering the wear period, determine the index of the degradation period node and the predictable interval, and improve the accuracy of the RUL prediction model. Then, aiming at the problem of difficult hyper-parameter adjustment of BiLSTM, the DE algorithm is utilized for optimization. On the basis, a rolling bearing RUL prediction model of the DE-BiLSTM is provided.

In addition, the invention also demonstrates the proposed DE-BiLSTM predictive model on the disclosed XJTU-SY bearing dataset, and the model has the advantages of accuracy and fitting by comparing with the LSTM and BiLSTM models before optimization.

The data adopted in the embodiment are based on XJTU-SY rolling bearing accelerated life test data sets, and degradation characteristic sets obtained after data preprocessing and characteristic extraction are carried out to carry out bearing residual service life prediction experiments.

The data set has 3 working conditions, 5 working conditions are used for each working condition, bearing 1_1 and Bearing2_1 are used as test sets, bearing 2_5 and Bearing 3_3 are used as verification sets, and the rest are used as training sets. The prediction results are shown in fig. 2 to 7.

In the LSTM model, the error of the predicted value and the actual value is larger, and the predicted value and the actual value basically coincide at the end stage of degeneration; in the BiLSTM prediction model, the error of Bearing2_1 is larger than that of Bearing 1_1, the prediction curve is far away from the actual life degradation curve, and larger deviation still exists in the later stage; the DE-LSTM prediction model is still Bearing2_1 with larger error, but gradually fits in the middle and later stages of prediction, approaching the actual life. Evaluation index calculation is carried out on the 4 prediction models, and the following table is obtained:

table 4 evaluation index of predictive model

As can be seen from Table 4, the 4 models of Bearing 1_1 are better than Bearing2_1 in overall evaluation index, the errors RMSE and MAE of DE-BiLSTM are smaller than those of the other 3 models, and the fitting effect is also better.

In addition, in order to prove the effectiveness of the data preprocessing flow and the DE-BiLSTM prediction model adopted by the invention, comparison experiments are carried out, namely a BiLSTM model based on an original signal, a DE-BiLSTM model based on the original signal, a BiLSTM model based on a denoising preprocessing signal and a DE-BiLSTM model based on the denoising preprocessing signal.

The comparison results are shown in fig. 8 to 15. As can be seen from fig. 8 to 15:

the prediction model before and after preprocessing can achieve a better fitting effect only by a certain data sample, the prior data is less, the error is larger, the fluctuation is reduced and the deviation is reduced along with the increase of the input feature set, and the difference is that the deviation between the model prediction degradation curve directly adopting the original signal and the actual degradation curve is larger.

The evaluation index was calculated for the above four cases, and the results were obtained as follows:

table 5 evaluation index of predictive model

As can be seen from table 5: the RUL prediction is directly carried out on the original signal, the RMSE and MAE values of the BiLSTM prediction model are larger than 10, the error is larger, and the fitting effect is poor; when the DE algorithm is used for optimizing and selecting the super-parameters of the BiLSTM, the error is reduced to be less than 10, but the error is still larger, which indicates that signal interference such as noise in the actually collected vibration signals has a larger influence on the rolling bearing RUL prediction. In addition, the original signal is preprocessed and then RUL prediction is carried out, the error is reduced to below 10, R2 also reaches above 0.96, and the fitting effect is good; when the DE-BiLSTM model is adopted, the error is reduced compared with the error when the DE algorithm is not used for optimization, and the fitting effect is better.

Through the analysis, the method provided by the invention has better applicability and effectiveness for the rolling bearing.

Example 2

This embodiment 2 describes a rolling bearing remaining life prediction system based on the same inventive concept as the rolling bearing remaining life prediction method described in embodiment 1 above.

Specifically, the rolling bearing remaining service life prediction system in this embodiment includes:

It should be noted that, in the rolling bearing residual service life prediction system, the implementation process of the functions and roles of each functional module is specifically shown in the implementation process of the corresponding steps in the method in the above embodiment 1, and will not be described herein again.

Example 3

Embodiment 3 describes a computer apparatus for implementing the steps of the rolling bearing remaining life prediction method described in embodiment 1 above.

The computer device includes a memory and one or more processors. Executable code is stored in the memory for implementing the steps of the rolling bearing remaining life prediction method described above when the processor executes the executable code.

In this embodiment, the computer device is any device or apparatus having data processing capability, which is not described herein.

Example 4

This embodiment 4 describes a computer-readable storage medium for implementing the steps of the rolling bearing remaining life prediction method described in embodiment 1 above.

The computer-readable storage medium in this embodiment 4 has stored thereon a program for implementing the steps of the above-described rolling bearing remaining service life prediction method when executed by a processor.

The computer readable storage medium may be an internal storage unit of any device or apparatus having data processing capability, such as a hard disk or a memory, or may be an external storage device of any device having data processing capability, such as a plug-in hard disk, a Smart Media Card (SMC), an SD Card, a Flash memory Card (Flash Card), or the like, which are provided on the device.

The foregoing description is, of course, merely illustrative of preferred embodiments of the present invention, and it should be understood that the present invention is not limited to the above-described embodiments, but is intended to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.

Claims

1. The method for predicting the residual service life of the rolling bearing is characterized by comprising the following steps of:

firstly, performing data noise reduction operation on original vibration signals of a rolling bearing by using an ISSA-MCKD-ICEEMDAN (integrated circuit-based random access memory) vibration signal data preprocessing method of the rolling bearing so as to highlight periodic signals and reduce signal interference;

2. The method for predicting remaining life of a rolling bearing according to claim 1, wherein,

the step 1 specifically comprises the following steps:

initializing an ISSA population, and setting the population scale and the maximum iteration number;

defining the range of the length L and deconvolution period T of a parameter filter of the MCKD as a space domain for searching an ISSA algorithm population, taking the maximum relevant kurtosis value of the signal of the MCKD as a fitness function, and taking (L, T) as a parameter combination to be optimized;

the optimal parameter combination of the deconvolution MCKD with the maximum correlation kurtosis is adaptively selected through an ISSA algorithm;

step 1.2, performing enhanced pretreatment on an original vibration signal of the rolling bearing with noise by using the MCKD after the parameters are adaptively selected so as to highlight fault pulses and reduce noise interference, thereby obtaining a noise reduction signal;

Step 1.3, ICEEMDAN decomposition is carried out on the noise reduction signal to obtain a plurality of IMF component signals;

3. The method for predicting remaining life of a rolling bearing according to claim 2, wherein,

the step 1.1 specifically comprises the following steps:

step 1.1.1, setting the population number X as 100, the number of searchers as 20%, the filter length L epsilon [1,400], the period T epsilon [1,400], and taking the signal maximum correlation kurtosis value of MCKD as a fitness function;

step 1.1.2, setting upper and lower bounds of a sparrow population region, initializing the position of sparrows in a spatial domain according to improved Circle chaotic mapping, and setting an objective function as a filter length L of MCKD and a deconvolution period T;

step 1.1.3, introducing self-adaptive weight to update the position of the searcher, and allowing the follower to find food around the optimal area found by the searcher to compete for the searcher; if the sparrow individual is in a very starved state, seeking food in other places, and updating the position by using a position updating function of the follower;

Step 1.1.4, when the alerter encounters danger, updating by using an alerter position updating function; sparrows at the periphery of the population are gathered towards a safety area, and the sparrows at the center of the population randomly walk to be close to other sparrows;

step 1.1.5, introducing a reverse learning strategy, and performing reverse solution treatment on the sparrow positions;

4. The method for predicting remaining life of a rolling bearing according to claim 1, wherein,

the step 2 specifically comprises the following steps:

step 2.1, extracting time domain, frequency domain, entropy and fractal dimension characteristics from the obtained rolling bearing fault characteristic reconstruction signals obtained in the step 1, thereby comprehensively evaluating the running state of the rolling bearing;

step 2.2, screening characteristic indexes which accord with the bearing degradation trend as a characteristic set according to the degradation curve of the characteristic parameters;

and 2.3, selecting degradation features with high correlation and low redundancy from the feature set by utilizing a combined dimension reduction method of the Spearman correlation coefficient and mutual information regression MIR, and constructing a low-dimension sensitive feature set.

5. The method for predicting remaining life of a rolling bearing according to claim 1, wherein,

the step 3 specifically comprises the following steps:

step 3.1, training hyper-parameters of a two-way long-short-term memory network BiLSTM by using a differential evolution method DE, and constructing a DE-BiLSTM prediction model by using optimized hyper-parameters;

6. The method for predicting remaining life of a rolling bearing according to claim 5, wherein,

the step 3.1 specifically comprises the following steps:

7. The method for predicting remaining life of a rolling bearing according to claim 6, wherein,

the step 3.1.5 specifically comprises the following steps:

8. A rolling bearing remaining useful life prediction system, comprising:

9. A computer device comprising a memory and one or more processors, the memory having executable code stored therein, wherein the processor, when executing the executable code,

a method of predicting the remaining life of a rolling bearing according to any one of claims 1 to 7 is achieved.

10. A computer-readable storage medium having a program stored thereon, which when executed by a processor, implements the rolling bearing remaining life prediction method according to any one of claims 1 to 7.