An EDM and self-attention-biGRU-based rolling bearing residual life prediction method

Abstract

The application discloses a rolling bearing residual life prediction method based on EDM and Self-Attention-BiGRU, and particularly relates to the technical field of bearing residual life prediction, and comprises the following steps: collecting original vibration signals from an experimental platform of accelerated degradation rolling bearings; performing multilayer decomposition on the original vibration signals by using empirical mode decomposition to obtain a plurality of modal function components and corresponding residual components; obtaining an original feature set by calculating dimensionless statistical features based on the modal function components; performing evaluation and screening by using a self-attention mechanism based on the original feature set to obtain an optimized feature data set, and dividing the optimized feature data set into training set data and test set data; constructing and training a model based on bidirectional gate recurrent units and a self-attention mechanism based on the training set data to obtain a rolling bearing residual life prediction model; and verifying the accuracy of the rolling bearing residual life prediction model by using the test set data based on the rolling bearing residual life prediction model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bearing residual life prediction, and particularly relates to a rolling bearing residual life prediction method based on EDM and Self-Attention-BiGRU. BACKGROUND

[0002] With the development of industrial technology, modern mechanical equipment is developing towards large-scale, comprehensive and complex. As an important part of rotating machinery, rolling bearings are widely used in industrial equipment. During the long-term operation of industrial machinery and equipment, it is inevitable to have a series of failure problems such as wear and fatigue due to overload and impact. More seriously, these failures will accelerate the degradation of bearings and have a serious impact on the safe and reliable operation of mechanical equipment. Therefore, how to effectively monitor the running condition and health status of the machine equipment and accurately predict the residual useful life (RUL) of the important parts is the focus of current research. In the current bearing residual life prediction method, the physical model based method needs to establish an accurate simulation degradation model according to the bearing mechanism and operating conditions, but due to the complex and changeable equipment operating environment, it is difficult to predict the residual useful life of the bearing by establishing a general failure model. Therefore, a feature extraction method based on empirical mode decomposition (EMD) and a rolling bearing residual useful life prediction method based on bidirectional gated recurrent unit (BiGRU) and self-attention mechanism (Self-Attention) are proposed. The method can not only extract more identifiable features from the original signal, but also ensure that the local features at different time scales are efficiently utilized. At the same time, the method can easily extract effective features from the shaft motion data, meet the demand for small data dimension, simplify the prediction model, and improve the prediction accuracy. SUMMARY

[0003] The application aims to provide a rolling bearing residual life prediction method based on EDM and Self-Attention-BiGRU, which solves the problems of difficulty in extracting effective features from rolling bearing vibration data, small data dimension, high computational cost and low prediction accuracy caused by complex prediction model.

[0004] In order to achieve the above-mentioned purpose, a rolling bearing residual life prediction method based on EDM and Self-Attention-BiGRU is provided, which comprises the following steps:

[0005] S1, collecting original vibration signals from an accelerated degradation rolling bearing experimental platform;

[0006] S2, performing multi-layer decomposition on the original vibration signal by using empirical mode decomposition to obtain a plurality of modal function components and corresponding residual components;

[0007] S3, obtaining an original feature set by calculating dimensionless statistical features based on the modal function components and the corresponding residual components;

[0008] S4, performing evaluation and screening by using a self-attention mechanism based on the original feature set to obtain an optimized feature dataset, and dividing the optimized feature dataset into training set data and test set data;

[0009] S5, constructing and training a model based on bidirectional gated recurrent units and a self-attention mechanism based on the training set data to obtain a model for predicting the remaining life of a rolling bearing;

[0010] S6, verifying the accuracy of the model for predicting the remaining life of the rolling bearing by using the test set data based on the model for predicting the remaining life of the rolling bearing.

[0011] Further, the method for collecting the original vibration signal from the experimental platform of the accelerated degradation rolling bearing in step S1 specifically comprises:

[0012] The original vibration signal is measured by an acceleration sensor, and the interval time and the sampling time of the measurement are fixed;

[0013] The original vibration signal is subjected to wavelet processing and noise reduction processing to complete preliminary data preprocessing.

[0014] Further, the method for performing multi-layer decomposition on the original vibration signal by using empirical mode decomposition to obtain a plurality of modal function components and corresponding residual components in step S2 specifically comprises:

[0015] For a given original time series signal x(t), the empirical mode decomposition process is as follows:

[0016] (1) Find the local maximum and minimum points of x(t), and use cubic spline interpolation to fit the maximum and minimum value point sequences respectively to obtain the upper envelope line and the lower envelope line; take the average of the upper and lower envelope lines to obtain m1(t);

[0017] (2) Calculate the difference p1(t) between the original time series data x(t) and the average envelope line m1(t), that is:

[0018] p1(t) = x(t) - m1(t)

[0019] (3) If p1(t) satisfies the condition of the modal function component, it is the first modal function component; otherwise, repeat steps (1) and (2) with p1(t) as the new original time series data until the condition of the modal function component is satisfied;

[0020] (4) After obtaining the first modal function component p1(t), it is decomposed from the original time series x(t) to obtain the residual component u1(t), that is:

[0021] u1(t) = x(t) - p1(t)

[0022] (5) The residual component u1(t) is taken as new data input, and steps (1) to (5) are re-executed to obtain a new residual component u2(t) and a second modal function component p2(t); and so on, until the residual component u k (t) of the kth modal function component p k (t) is constant or a monotonic function and cannot be decomposed any more, and the entire empirical mode decomposition process is completed; at this time, the original time series x(t) can be represented as:

[0023]

[0024] In the formula, u k (t) can be regarded as the trend or mean value of x(t); q1(t), q2(t), …, q k (t) are modal function components of x(t) and represent high-frequency components to low-frequency components of the original time series data.

[0025] That is, the original signal is subjected to multi-layer empirical mode decomposition, thereby obtaining the corresponding intrinsic modal function components and the corresponding residual components.

[0026] Further, the method for obtaining the original feature set by calculating dimensionless statistical features based on the modal function components and the corresponding residual components in step S3 specifically includes:

[0027] A plurality of dimensionless statistical features are used to represent each modal function component obtained after empirical mode decomposition, and these dimensionless statistical features include root mean square value, energy, Shannon entropy, peak factor, skewness, variance, mean value, standard deviation, and kurtosis.

[0028] Based on statistical analysis of a plurality of dimensionless statistical features, key features of the fault data are captured to obtain a multi-dimensional original feature set.

[0029] Further, the method of step S4, based on the original feature set, using self-attention mechanism for evaluation and screening to obtain the optimized feature data set, and dividing the optimized feature data set into training set data and test set data, specifically includes:

[0030] The calculation of the self-attention mechanism is divided into two steps:

[0031] First step: calculate the attention weight between any two vectors in the input sequence;

[0032] Second step: calculate the weighted average of the input sequence according to the attention weight;

[0033] The specific operation method is as follows:

[0034] Q = XW q

[0035] K = XW k

[0036] V = XW ν

[0037]

[0038] In the formula: Q is the query matrix; K is the key matrix; V is the value matrix; dim is the dimension of Q, K and V.

[0039] The three evaluation features used for screening the original feature set are correlation (Corr(f, t)), monotonicity (Mon(f)) and variance (Var(f)) to effectively reflect the features of the degradation process.

[0040] The center moving method is used to regard the feature f as a random process, which is divided into a trend part, f T represents the average trend and a random part, f R represents the residual, as shown in the following formula:

[0041] f(t k ) = f T (t k ) + f R (t k )

[0042] Where: f(t k ) is the degradation feature at time t k ;

[0043] The three evaluation indexes are as follows:

[0044]

[0045] Wherein, k is the total number of observations; h(t) is a step function;

[0046] The screening equation is as follows:

[0047] Ce=0.2Var(f)+0.5Mon(f)+0.3Corr(f,t)

[0048] The weight value before each feature evaluation represents the importance of the index, and the weight value ranges within (0, 1);

[0049] The screened feature data set is finally divided into a training set and a test set.

[0050] Further, the method for constructing and training the model based on the bidirectional gate recurrent unit and the self-attention mechanism based on the training set data in step S5 specifically comprises:

[0051] (1) a 14-layer network structure of the bidirectional gate recurrent unit and the self-attention mechanism is designed;

[0052] (2) the network structure of the bidirectional gate recurrent unit and the self-attention mechanism is added with the self-attention mechanism after the bidirectional gate recurrent unit layer;

[0053] (3) the mean square error is used as a loss function in the training process of the model based on the bidirectional gate recurrent unit and the self-attention mechanism, and the Adam optimizer optimization algorithm is selected to optimize the model training;

[0054] (4) the features with high scores screened by the feature screening equation are used as inputs of the rolling bearing residual life prediction model to obtain the rolling bearing residual life prediction model.

[0055] Further, the method for verifying the accuracy of the rolling bearing residual life prediction model based on the rolling bearing residual life prediction model and using the test set data in step S6 specifically comprises:

[0056] The test set data is verified by using a univariate experimental method to verify different parameters.

[0057] Compared with the prior art, the beneficial effects of the present application are:

[0058] The rolling bearing residual life prediction method based on EDM and Self-Attention-BiGRU of the present application has the advantages of: 1. solving the problem of model information overload; 2. improving the accuracy and robustness; 3. the prediction model is simple and low in cost; and 4. facilitating the extraction of effective features of bearing vibration signals.

[0059] On the basis of the implementation manners of the above aspects, the application can be further combined to provide more implementation manners. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0061] Figure 1 Flow chart for bearing residual life prediction of the present application;

[0062] Figure 2 Structure diagram of self-attention mechanism of the present application;

[0063] Figure 3 Self-Attention-BiGRU network structure of the present application. DETAILED DESCRIPTION

[0064] In order to make the technical personnel in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort should be within the scope of protection of the present application. If not specifically stated, the technical means used in the embodiments are conventional means familiar to those skilled in the art.

[0065] The embodiments of the present application provide a rolling bearing residual life prediction method based on EDM and Self-Attention-BiGRU. Referring to Figure 1 , Figure 1 The embodiments of the present application provide a flow chart of a rolling bearing residual life prediction method based on EDM and Self-Attention-BiGRU. The prediction method comprises the following steps:

[0066] S1, collecting original vibration signals from an accelerated degradation rolling bearing experimental platform;

[0067] S2, using empirical mode decomposition to perform multi-layer decomposition on the original vibration signals to obtain a plurality of modal function components and corresponding residual components;

[0068] S3, based on the modal function components and corresponding residual components, obtaining an original feature set by calculating dimensionless statistical features;

[0069] S4, based on the original feature set, using self-attention mechanism for evaluation and screening, obtaining the optimized feature data set, and dividing the optimized feature data set into training set data and test set data;

[0070] S5, based on the training set data, constructing and training a model based on bidirectional gated recurrent unit and self-attention mechanism, obtaining a model for predicting the remaining life of the rolling bearing;

[0071] S6, based on the model for predicting the remaining life of the rolling bearing, using the test set data to verify the accuracy of the model for predicting the remaining life of the rolling bearing.

[0072] Based on the EDM and Self-Attention-BiGRU rolling bearing residual life prediction method provided by the embodiment of the application, first, the original vibration signal is collected from the accelerated degradation rolling bearing experimental platform, and the empirical mode decomposition (EMD) is used for multi-layer decomposition to separate multiple modal function components and residual components, thereby improving the effectiveness of feature extraction. Then, dimensionless statistical features are calculated based on these components to form an original feature set, and a self-attention mechanism is used to evaluate and screen an optimized feature data set, thereby enhancing the accuracy of feature selection. Then, the optimized feature data set is divided into a training set and a test set, a model based on bidirectional gated recurrent unit and self-attention mechanism is constructed and trained based on the training set, and finally a model for predicting the remaining life of the rolling bearing is obtained. Finally, the accuracy of the model is verified by the test set data to ensure its stability and reliability in practical application. This method not only improves the accuracy of feature extraction and selection, but also optimizes the model training process, thereby realizing more accurate and reliable prediction of the remaining life of the rolling bearing.

[0073] In one embodiment, the method of collecting original vibration signals from the accelerated degradation rolling bearing experimental platform in step S1 specifically includes:

[0074] The original vibration signal is measured by an acceleration sensor, and the interval time and sampling time of the measurement are fixed. The original vibration signal is processed by wavelet and noise reduction to complete preliminary data preprocessing.

[0075] Specifically, the collected data is based on the experimental platform of accelerated aging of rolling bearings, and the vibration data of the bearings in the horizontal direction and the vertical direction in the whole life cycle is collected by using an acceleration sensor, the data is collected once every 10s, the sampling time is 0.1s, the sampling frequency is 25.6kHz, and the length of the data collected each time in the horizontal direction and the vertical direction is 2560 points. At the same time, it has been proved through experiments that the collected bearing degradation data can be better used for RUL prediction when the equipment is in a light load condition, that is, the collected data for model establishment in this method is under the light load working condition of the equipment; the horizontal vibration signal is used as the original data input of the EDM; and the collected data is preliminarily processed by using wavelet processing and noise reduction processing.

[0076] In one embodiment, the method of using empirical mode decomposition to perform multi-layer decomposition on the original vibration signal in step S2 to obtain a plurality of modal function components and corresponding residual components, specifically includes:

[0077] Finding the extreme value points of the original vibration signal; fitting the extreme value points with envelope lines, fitting the maximum value with an upper envelope line and fitting the minimum value with a lower envelope line, calculating the mean value of the envelope lines; calculating the intermediate signal, which is the difference between the original signal and the mean value; judging whether the intermediate signal meets the modal function component condition, if not, returning to the first step; if the intermediate signal meets the modal function component condition, repeating the above steps to continue the decomposition as the original signal.

[0078] Specifically, empirical mode decomposition (EDM) is an adaptive data processing method; it has obvious advantages in processing non-stationary and nonlinear data; the basic idea of EMD is mainly to decompose the complex original signal into a group of independent, nearly periodic intrinsic modal functions (IMF); among them, the number of zero points and extreme points of IMF must be equal, or the difference between them must not exceed 1; the upper and lower envelope lines composed of the local maximum and minimum values of IMF, respectively, have a mean value of 0 at any time.

[0079] For a given original time series signal x(t), the EMD process is as follows:

[0080] 1) Find the local maximum and minimum points of x(t), and use cubic spline interpolation to fit the maximum and minimum value point sequences respectively to obtain the upper and lower envelope lines; take the average of the upper and lower envelope lines to obtain m1(t).

[0081] 2) Calculate the difference p1(t) between the original time series data x(t) and the average envelope line m1(t), that is:

[0082] p1(t)=x(t)-m1(t)

[0083] 3) If p1(t) satisfies the condition of IMF component, it is the first IMF component; otherwise, repeat step 1) and step 2) with p1(t) as the new original time series data until the condition of IMF component is satisfied.

[0084] 4) After obtaining the first IMF component p1(t), it is decomposed from the original time series x(t) to obtain the residual component u1(t), that is:

[0085] u1(t) = x(t) - p1(t)

[0086] 5) The residual component u1(t) is taken as the new data input, and steps 1) to 5) are re-executed to obtain a new residual component u2(t) and a second IMF component p2(t); and so on, until the residual component u k (t) of the kth IMF component p k (t) is a constant or a monotonic function and cannot be decomposed any more, the entire EMD process is completed; at this time, the original time series x(t) can be represented as:

[0087]

[0088] wherein, u k (t) can be regarded as the trend or mean value of x(t); q1(t), q2(t), …, q k (t) are IMF components of x(t) and represent high-frequency components to low-frequency components of the original time series data.

[0089] According to the above method, the input original signal is subjected to multi-layer EMD, so as to obtain corresponding intrinsic mode functions (IMF) and corresponding residual components.

[0090] In an embodiment, the method for obtaining the original feature set by calculating dimensionless statistical features based on the modal function components and corresponding residual components in step S3 specifically includes:

[0091] A plurality of dimensionless statistical features are used to represent each modal function component obtained after the empirical mode decomposition processing, and these dimensionless statistical features include but are not limited to: root mean square value, energy, Shannon entropy, peak factor, skewness, variance, mean value, standard deviation, and kurtosis, etc.; the processed signal is the component obtained after the EDM processing, and the features of the statistical fault data are obtained from these components, so that a multi-dimensional original feature set can be obtained; the main formula of feature extraction is as follows:

[0092] Root mean square:

[0093]

[0094] Energy: Used to measure the total energy of a signal in the time domain.

[0095]

[0096] Where, x i is the value of each sample of the signal, and N is the number of samples in the signal.

[0097] Shannon entropy: In information theory, it is used to measure the uncertainty or amount of information.

[0098]

[0099] Where p(x) i () is a random variable X taking the value x. i The probability is log2, which is a logarithmic operation with base 2; the unit of Shannon entropy is usually bit, and when the base is 2, the unit of Shannon entropy can be expressed as bit.

[0100] Trigonometric functions: involve the periodicity and oscillation characteristics of signals;

[0101] Peak value: The peak value is the maximum value that a signal reaches over a period of time.

[0102] Peak = max(x1, x2, ..., x N )

[0103] Variance: measures the degree of dispersion of signal data.

[0104]

[0105] Where μ is the mean of the signal, x i is the value of each sample of the signal, and N is the number of samples in the signal.

[0106] Mean:

[0107]

[0108] Where, x i is the sample value of the signal, and N is the number of samples.

[0109] Standard deviation: the square root of the variance, representing the degree of dispersion of the signal data.

[0110]

[0111] Where μ is the mean of the signal, x i is the value of each sample of the signal, and N is the number of samples in the signal.

[0112] Skewness: describes the degree of skewness in the distribution of signal data.

[0113]

[0114] where μ is the mean of the signal, σ is the standard deviation of the signal, x i is each sample value of the signal, and N is the number of samples of the signal.

[0115] Kurtosis: describes the degree of peak of the signal data distribution

[0116]

[0117] where μ is the mean of the signal, σ is the standard deviation of the signal, x i is each sample value of the signal, and N is the number of samples of the signal.

[0118] In one embodiment, the method of step S4, based on the original feature set, using self-attention mechanism for evaluation and screening, obtaining the optimized feature data set, and dividing the optimized feature data set into training set data and test set data, specifically includes:

[0119] The calculation of the self-attention mechanism is divided into two steps:

[0120] First step: Calculate the attention weight between any two vectors in the input sequence;

[0121] Second step: Calculate the weighted average value of the input sequence according to the attention weight;

[0122] The three evaluation feature indicators used for screening the original feature set are correlation (Corr(f, t)), monotonicity (Mon(f)), and variance (Var(f));

[0123] The selected screening equation is as follows:

[0124] Ce = 0.2Var(f) + 0.5Mon(f) + 0.3Corr(f, t).

[0125] Specifically, the method uses self-attention mechanism (Self-Attention) for data feature screening. This mechanism is improved on the basis of attention mechanism (Attention). At the same time, this mechanism not only can quickly screen out key information and reduce attention to other irrelevant information, but also can reduce the dependence on external information. By introducing the self-attention mechanism, not only the problem of model information overload is solved, but also the accuracy and stability of the network are improved. The structure of the self-attention mechanism is as shown in Figure 2 .

[0126] Figure 2 where x i is the input sequence, i = 1, 2, 3,... t; v i(i=1,2,3…t) is the value vector generated by the input sequence; x ti (i=1,2,3…t) is the result of the input sequence operating on the respective q and k vectors and passing through the SoftMax function; b i (i=1,2,3…t) is the result of the i-th position information in the input sequence operating on all position information through the attention mechanism.

[0127] The calculation of Self-Attention is divided into two steps, the first step: calculate the attention weight between any two vectors in the input sequence; the second step: calculate the weighted average value of the input sequence according to the attention weight.

[0128] The specific operation method is as follows:

[0129] Q=XW q

[0130] K=XW k

[0131] V=XW ν

[0132]

[0133] In the formula, Q is the query matrix; K is the key matrix; V is the value matrix; dim is the dimension of Q, K and V.

[0134] The three evaluation features used by the application are: correlation (Corr(f,t)), monotonicity (Mon(f)) and variance (Var(f)) to screen features that can effectively reflect the degradation process.

[0135] The center moving method is adopted to regard the feature f as a random process, which is divided into a trend part, f T represents the average trend and the random part, f R represents the residual, as shown in the following formula:

[0136] f(t k )=f T (t k )+f R (t k )

[0137] Wherein, f(t k ) is the degradation feature of time t k

[0138] The three evaluation indexes are as follows:

[0139]

[0140] ​Wherein, k is the total number of observations; h(t) is a step function;

[0141] The screening equation is as follows:

[0142] Ce=0.2Var(f)+0.5Mon(f)+0.3Corr(f,t)

[0143] The weight value of each feature before evaluation represents the importance of the index, and the weight value ranges within (0, 1).

[0144] The feature data set after screening is finally divided into: training set and test set.

[0145] In one embodiment, the method for constructing and training a model based on bidirectional gate recurrent unit and self-attention mechanism based on the training set data in step S5 to obtain the model for predicting the residual life of the rolling bearing specifically comprises:

[0146] (1) A 14-layer bidirectional gate recurrent unit and self-attention mechanism network structure is designed as shown in Figure 3

[0147] (2) The bidirectional gate recurrent unit and self-attention mechanism network structure is added with a self-attention mechanism after the bidirectional gate recurrent unit layer.

[0148] (3) The model based on bidirectional gate recurrent unit and self-attention mechanism adopts mean square error as the loss function in the training process, and selects Adam optimizer optimization algorithm for model training.

[0149] (4) The features with high scores are screened out by the feature screening equation as the input of the rolling bearing residual life prediction model to obtain the model for predicting the residual life of the rolling bearing.

[0150] Specifically, the model is established, the Self-Attention-BiGRU network structure is as shown in Figure 3 , which is 14 layers in total; the features with high scores are screened out by the feature screening equation as the input of the RUL prediction model to predict the residual service life of the bearing; in the network structure Figure 3 , a self-attention mechanism is added after the BiGRU layer, which reduces the attention to irrelevant information and reduces the dependence on external information, solves the problem of information overload of the model, and helps to improve the accuracy of RUL prediction.

[0151] ​We can roughly divide the life cycle of the bearing into four stages: normal working stage, in this stage, the amplitude of the bearing vibration signal is low; early degradation stage, with the passage of time, the amplitude of the vibration signal gradually increases, in this stage, the prediction of RUL can be carried out; in the middle stage of degradation, the amplitude of the vibration signal will continue to increase; and in the later stage of degradation, when the amplitude of the signal further significantly increases, the bearing is considered to be completely failed; the data in the training set is input into the model for prediction, and the corresponding remaining useful life curve (predicted value curve) is drawn.

[0152] In the training process, the model uses mean squared error (MSE) as the loss function, and uses Adam optimizer optimization algorithm for model training.

[0153] In one embodiment, the method for verifying the accuracy of the rolling bearing remaining life prediction model in step S6 uses test set data to verify the accuracy of the rolling bearing remaining life prediction model, which specifically includes:

[0154] The test set data is verified by using a single variable experimental method to verify different parameters.

[0155] Specifically, the test set is verified. In order to verify the accuracy of the EDM and Self-Attention BiGRU method, the RUL prediction model established in step 5 is used to verify the bearing remaining life through the data in the test set. According to the data in the test set, the data is input into the prediction model, and the corresponding remaining useful life curve (true value curve) is drawn. Finally, the predicted value curve and the true value curve are compared to see if they can be overlapped, thereby verifying.

[0156] The beneficial effects of the present application are: in the RUL prediction of the bearing, different hyperparameter values need to be set for verification, so the optimizer, learning rate, sample number and iteration parameter need to be considered; the method uses Adam optimizer optimization algorithm for model training, because it has the characteristics of simple algorithm implementation, high calculation efficiency and fast convergence speed; the learning rate and sample parameter need to be selected according to the data demand, because the learning rate and sample number that is too large or too small will affect the accuracy of RUL prediction, therefore, in order to determine the optimal parameter value of the learning rate, sample number and iteration number, the present application uses a single variable experimental method to verify different parameters.

[0157] The above describes one embodiment of the present application in detail, but the content is only the preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A method for predicting the remaining life of rolling bearings based on EDM and Self-Attention-BiGRU, characterized in that, The prediction method includes the following steps: S1. Collect the original vibration signal from the experimental platform of the accelerated degradation rolling bearing; S2. The original vibration signal is decomposed into multiple layers using empirical mode decomposition to obtain multiple modal function components and corresponding residual components; S3. Based on the modal function components and the corresponding residual components, the original feature set is obtained by calculating the dimensionless statistical features; S4. Based on the original feature set, an evaluation and screening process is performed using a self-attention mechanism to obtain an optimized feature dataset, and the optimized feature dataset is divided into training set data and test set data. S5. Based on the training set data, construct and train a model based on bidirectional gated recurrent units and self-attention mechanism to obtain a model for predicting the remaining life of rolling bearings. S6. Based on the rolling bearing remaining life prediction model, use test set data to verify the accuracy of the rolling bearing remaining life prediction model. The method described in step S5 for constructing and training a model based on a bidirectional gated recurrent unit and a self-attention mechanism to obtain a model for predicting the remaining life of rolling bearings specifically includes: (1) Design a 14-layer bidirectional gated recurrent unit and self-attention mechanism network structure; (2) For the network structure of the bidirectional gated recurrent unit and the self-attention mechanism, a self-attention mechanism is added after the bidirectional gated recurrent unit layer; (3) The mean square error is used as the loss function for the model based on the bidirectional gated recurrent unit and self-attention mechanism during the training process, and the Adam optimizer is selected to optimize the model training. (4) The features with high scores are selected by the feature selection equation and used as the input of the rolling bearing remaining life prediction model to obtain the rolling bearing remaining life prediction model.

2. The method for predicting the remaining life of rolling bearings based on EDM and Self-Attention-BiGRU according to claim 1, characterized in that, The method for acquiring raw vibration signals from the experimental platform of the accelerated degradation rolling bearing described in step S1 specifically includes: The original vibration signal is measured by an accelerometer, and the measurement interval and sampling time are fixed. The original vibration signal is subjected to wavelet processing and noise reduction to complete the initial data preprocessing.

3. The method for predicting the remaining life of rolling bearings based on EDM and Self-Attention-BiGRU according to claim 1, characterized in that, The method described in step S2 for performing multi-level decomposition of the original vibration signal using empirical mode decomposition to obtain multiple modal function components and corresponding residual components specifically includes: For a given original time series signal The empirical mode decomposition process is as follows: (1) Find For local maxima and minima, cubic spline interpolation is used to fit the sequences of maxima and minima respectively, obtaining upper and lower envelopes. The average of the upper and lower envelopes is then calculated. ; (2) Calculate the original time series data With the average envelope difference ,Right now: ； (3) If If the conditions for modal function components are met, then it is the first modal function component; otherwise, it will be... Repeat steps (1) and (2) as new raw time series data until the conditions of the modal function components are met; (4) After obtaining the first modal function component , from the original time series The residual components are decomposed from the original value. ,Right now: ； (5) The residual components As new data input, repeat steps (1) to (5) to obtain new residual components. and the second modal function component And so on, until the k-th modal function component. residual components If the function is a constant or monotonic function and cannot be further decomposed, the entire empirical mode decomposition process is complete; at this point, the original time series... It can be represented as: ； In the formula, It can be regarded as The trend or mean; q1(t), q2(t), ..., q n (t) is the modal function component of x(t), representing the high-frequency to low-frequency components of the original time series data; That is, multi-level empirical mode decomposition is performed on the input original signal to obtain the corresponding intrinsic mode function components and the corresponding residual components.

4. The method for predicting the remaining life of rolling bearings based on EDM and Self-Attention-BiGRU according to claim 1, characterized in that, The method for obtaining the original feature set by calculating dimensionless statistical features based on the modal function components and the corresponding residual components in step S3 specifically includes: Various dimensionless statistical features were used to characterize the modal function components obtained after empirical mode decomposition. These dimensionless statistical features include: root mean square value, energy, Shannon entropy, peak factor, skewness, variance, mean, standard deviation, and kurtosis. Based on statistical analysis of various dimensionless statistical features, key features of fault data are captured, resulting in a multi-dimensional original feature set.

5. The method for predicting the remaining life of rolling bearings based on EDM and Self-Attention-BiGRU according to claim 1, characterized in that, The method described in step S4, which involves evaluating and filtering based on the original feature set using a self-attention mechanism to obtain an optimized feature dataset, and then dividing the optimized feature dataset into training and testing sets, specifically includes: The computation of the self-attention mechanism consists of two steps: Step 1: Calculate the attention weights between any two vectors in the input sequence; Step 2: Calculate the weighted average of the input sequence based on these attention weights; The specific calculation method is as follows: ； In the formula: For query matrix; The key matrix; It is a value matrix; for , and dimensionality; The three evaluation metrics used to screen the original feature set are: Corr(f,t), Monotonicity Mon(f), and Variance Var(f) to screen features that can effectively reflect the degradation process. Using the center-movement method to extract features Treating it as a stochastic process, we divide it into a trend component. Indicates the average trend and the random component. The residual is represented by the following formula: ； in: For time Degenerative characteristics; The three evaluation indicators are as follows: ； ； ； in: The total number of observations; It is a step function; The screening equation is as follows: ； The weight value before each feature evaluation represents the importance of that indicator, and the weight value ranges from (0,1); After filtering, the feature dataset is finally divided into a training set and a test set.

6. The method for predicting the remaining life of rolling bearings based on EDM and Self-Attention-BiGRU according to claim 1, characterized in that, The method for verifying the accuracy of the rolling bearing remaining life prediction model based on the test set data in step S6 is specifically to verify the test set data by using a univariate experimental method to verify different parameters.

Images