CN116738868B

CN116738868B - Rolling bearing residual life prediction method

Info

Publication number: CN116738868B
Application number: CN202311028007.XA
Authority: CN
Inventors: 房殿军; 蒋红琰; 蒋铠臣; 郑卓远; 邵佳杰; 杨志浩; 罗尔夫·施密特
Original assignee: Qingdao Sino German Intelligent Technology Research Institute
Current assignee: Qingdao Sino German Intelligent Technology Research Institute
Priority date: 2023-08-16
Filing date: 2023-08-16
Publication date: 2023-11-21
Anticipated expiration: 2043-08-16
Also published as: CN116738868A

Abstract

The invention discloses a method for predicting the residual life of a rolling bearing, which belongs to the technical field of data prediction and comprises the following steps: step 1, collecting vibration signals of a rolling bearing by using an acceleration sensor, and calibrating a degradation starting point of the bearing by adopting a self-adaptive judging method; step 2, carrying out noise reduction treatment on the original vibration signal, extracting characteristics from a time domain, a frequency domain and a time-frequency domain, and carrying out characteristic dimension reduction by using a principal component analysis method; and 3, building a life prediction model based on the attention Bi-GRU neural network and Kalman filtering, and predicting the residual life. According to the invention, feature extraction is performed on a time domain, a frequency domain and a time-frequency domain, and life prediction is performed by adopting a mode of combining an attention bidirectional gate-control cyclic neural network and a Kalman filtering algorithm, so that the method is more in line with the real life cycle of the rolling bearing, and the prediction result is more accurate.

Description

Rolling bearing residual life prediction method

Technical Field

The invention belongs to the technical field of data prediction, and particularly relates to a method for predicting the residual life of a rolling bearing.

Background

With the continuous development of computers, artificial intelligence algorithms and internet communication technologies, the concept of intelligent manufacturing is increasingly paid attention to by industries of various countries, and fault diagnosis and life prediction of industrial equipment parts by using a large amount of data generated in intelligent manufacturing are a great means for preventing production accidents and reducing cost in recent years.

In the intelligent manufacturing age, the integration level, complexity and intelligent level of mechanical equipment are rapidly improved, so that the traditional equipment maintenance and guarantee technology cannot adapt to new requirements. Currently, fault prediction and health management methods (Prognostics and Health Management, PHM) combine modern information technology with neural network technology to enable equipment monitoring, prediction and management in complex projects.

The rolling bearing is used as one of the core basic components of various mechanical equipment, the service life of the rolling bearing directly influences the health condition of related equipment, and whether the rolling bearing works normally or not more directly determines the running stability of the equipment. According to the data, the rolling bearing failure rate is up to 45% -55% in the whole failure accident of the rotating parts. Meanwhile, the running conditions of mechanical equipment in an actual working environment are complex and changeable, the working conditions of different bearings are different, and the problem of predicting the residual life in the degradation process of the mechanical equipment is influenced by a plurality of factors. Therefore, the method for predicting the residual life of the mechanical equipment has the advantages of low cost, high reliability, high prediction precision and engineering practicability, and has positive guiding significance for maintenance, maintenance and timely replacement of core equipment in actual production.

Currently, the main current methods for predicting the remaining life of mechanical equipment are divided into four types, namely a mathematical physical model prediction method, a machine learning-based prediction method, a deep learning prediction method and a combined prediction method.

The deep learning network is the most popular artificial intelligence technology at present, and is widely applied to aspects of image classification, natural language processing, automatic driving and the like. The multi-layer structure of the algorithm can realize the self-adaptive processing of input data through nonlinear functions and linear functions, so that a model can be built more accurately through extracted features. Methods for predicting the remaining life of a mechanical device using deep learning, including Deep Belief Networks (DBNs), deep automatic encoders (SAE), convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), etc., have begun to become popular after the 21 st century.

In the use of deep learning networks for rolling bearing life prediction, models are generally employed: DNN deep neural networks, CNN convolutional neural networks, LSTM long short-term memory networks, and RNN recurrent neural networks. The deep neural network can better fit the mapping relation between input and output through the multi-layer stacked hidden layer compared with the BP neural network of the shallow layer, so that a better training result is achieved. Ren et al have used deep neural networks to study the remaining life of lithium batteries. However, due to the complexity of the network structure of the multi-layer stack, the deep neural network is generally not efficient in training and has great training difficulty. The CNN convolutional neural network has the advantages of stronger analysis capability on high-dimensional data, capability of automatically extracting features and the like because of shared point-by-point convolutional kernels, and is adopted by Li and the like and subjected to related research. However, CNN convolutional neural networks are generally a method of learning and training images, are insensitive to time series, and make subsequent modification of the model difficult due to its encapsulation for feature extraction.

RNN recurrent neural networks offer great advantages over time series processing, and thus RNN recurrent neural network-based studies have had many results in predicting the remaining life of the device. For example, yu et al have predicted the remaining life of a turbofan engine using a two-way RNN model and produced good results. However, the RNN model is too serious to be influenced by global data in training, so that the problems of gradient disappearance, gradient explosion and the like easily occur in long-time sequence training, and the later training of the model is extremely small. Therefore, like Lei, the LSTM long-short-time memory network is used for fault diagnosis of the wind driven generator, and a corresponding diagnosis model is constructed for solving the problem. Training of LSTM long-term memory networks typically consumes significant computational resources and time.

Deep learning networks are capable of learning complex nonlinear relationships by training a multi-layer network. Therefore, they are expected to exhibit good performance in the RUL prediction of mechanical devices. However, it still has its own limitations. In addition to poor interpretability, deep learning networks often require large amounts of high quality training data, which is difficult to collect in industrial applications. In the practical industry, due to the influence of factors such as changes of machine working conditions and interference of environmental noise, distribution differences exist between training and test data generally, so that the RUL prediction performance is reduced. Furthermore, their structure and parameters are generally randomly initialized or manually set, which greatly affects their universal capabilities in different situations.

At present, the research on the residual life of the rolling bearing mainly has the following three problems.

1. Bearing degradation starting point calibration problem: the feature extraction of rolling bearings is generally performed starting from the degradation origin of the rolling bearing. The health status signal accounts for the vast majority of the original time domain root mean square vibration signal of the rolling bearing, and the vibration signal only randomly fluctuates at a low level at this stage, so that the degradation information of the tested bearing is difficult to find from the health status signal. The amplitude of the vibration signal during the degradation phase tends to increase with operating time, which means that the vibration signal during this phase contains rich bearing degradation information. How to find a suitable method to calibrate the degradation starting point of the bearing is a problem, if the health state is not distinguished from the monitoring data of other states, the situation of health state data interference may occur, the occupation of computing resources is increased, and the prediction accuracy is reduced.

2. Feature extraction and data preprocessing problems: the collected vibration signals of the rolling bearing generally contain rich bearing degradation characteristics, such as amplitude characteristics, time domain characteristics, frequency domain characteristics and energy characteristics extracted from the signals, which can be used for representing the health degree of the rolling bearing. Although the abundant signals make the prediction of the residual life of the equipment feasible, too much useless noise is often mixed in the excessively redundant data, and the effective data are often coupled with each other, so that the extraction of fault characteristic information becomes a difficulty. Many residual life prediction methods driven by data are directly input into original time domain vibration signals for research and analysis, but the single original time domain vibration information cannot accurately reflect the degradation property of the rolling bearing, and further influence the prediction accuracy. If the complicated multidimensional input vector extracted from the data is not processed, model training resources are wasted greatly, model training time is increased, and prediction accuracy is reduced.

3. Prediction method model building problem: the current mainstream prediction method is to simply compare the measured value of the monitored parameter with a single static threshold value, so as to realize an alarm function, and lack a reasonable and effective self-adaptive prediction mechanism. The existing rolling bearing service life prediction method based on the deep learning network, such as the influence of global data is excessively emphasized in training of a cyclic neural network (RNN) model, so that the problems of gradient disappearance, gradient explosion and the like easily occur in long-time sequence training, and the later training of the model is extremely low; training of long and short term memory network (LSTM) models, for example, typically consumes significant computational resources and time. In addition, the input and processing of data in long and short term memory networks (LSTM) and gate-controlled cyclic neural networks (GRU) are unidirectional, that is, each training unit only focuses on the historical data before the moment, but under a complex actual scene, the output at the moment may also have a relation with the output at the future moment, and because the early characteristics of the bearing signals in the fault are not obvious, the unidirectional network may miss the characteristic information of part of the fault.

Disclosure of Invention

In order to solve the problems, the invention provides a method for predicting the residual life of a rolling bearing, which extracts characteristics for prediction through original time domain vibration information of the rolling bearing, namely, the characteristics extracted on time domain, frequency domain and time domain are richer, and a principal component analysis method is used for characteristic dimension reduction; on post-processing noise reduction, the attention bidirectional gating circulating neural network and the Kalman filtering algorithm are adopted, so that the effect is better than that of a traditional weighted average filter, the life prediction curve is more in line with the real life cycle of the rolling bearing, the prediction result is more accurate, and the intelligent and accurate production equipment maintenance can be realized.

The technical scheme of the invention is as follows:

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

step 1, collecting vibration signals of a rolling bearing by using an acceleration sensor, and calibrating a degradation starting point of the bearing by adopting a self-adaptive judging method;

step 2, carrying out noise reduction treatment on the original vibration signal, extracting characteristics from a time domain, a frequency domain and a time-frequency domain, and carrying out characteristic dimension reduction by using a principal component analysis method;

and 3, building a life prediction model based on the attention Bi-GRU neural network and Kalman filtering, and predicting the residual life.

Further, in step 1, after the bearing vibration signal is collected, the kurtosis is calculatedThe calculation formula of (2) is as follows:

(1)；

in the method, in the process of the invention,representative signal->Average value of>Representative signal->Center moment of->Representative signal->Standard deviation of (2);

after obtaining the kurtosis value of the bearing vibration signal, adopting 3The self-adaptive judging method of the interval carries out bearing degradation point calibration, and the specific process is as follows: firstly, the average value +_of kurtosis is calculated using the rolling bearing history data of the healthy operating state>Standard deviation of kurtosis->And to determine +.>Interval->The method comprises the steps of carrying out a first treatment on the surface of the Then, use +.>The section is used for identifying the normal state and the abnormal state of the bearing; the newly calculated kurtosis is continuously added during the detection process >And->The intervals are compared, if at a certain moment +.>The kurtosis value of (2) exceeds +.>Judging the intervalAnd is set as the degradation start point of the rolling bearing.

Further, the specific process of step 2 is as follows:

step 2.1, carrying out noise reduction treatment on an original signal, and removing singular values by adopting a wavelet threshold filtering method;

step 2.2, selecting 8 time domain features of a mean value, a standard deviation, a maximum value, a minimum value, kurtosis, a skewness, a root mean square value and a peak-to-peak value of each rectangular window intercepting signal, wherein the mean value, the root mean square value, the variance, the square root amplitude, the skewness, the kurtosis, the maximum value, the absolute average amplitude, the minimum value, the peak-to-peak value, the root mean square root and the center of gravity 12 frequency domain features and the normalized energy value feature obtained by wavelet packet decomposition are time-frequency domain features;

and 2.3, analyzing and dimension-reducing all the features selected in the step 2.3 by adopting a principal component analysis method.

Further, the specific process of step 2.3 is as follows:

and 2.3.1, carrying out centering treatment on all the characteristic indexes, wherein the following formula is as follows:

(2)；

(3)；

in the method, in the process of the invention, Is the characteristic index vector after centering; />Is->A dimension feature index vector; />Is the original dimension of the data;data sample number, < >>A data sample sequence number; />For a sample matrix, the matrix size is +.>；/>1-dimensional feature index vector centered for 1 st sample, ++>Centered for sample 1 +.>Dimension characteristic index vector, & lt & gt>Is->1-dimensional feature index vector after centering of each sample,/->Is->Sample-centered +.>A dimension feature index vector;

step 2.3.2, calculating to obtain a sample matrixCovariance matrix>And for covariance matrix->And (3) carrying out eigenvalue decomposition to obtain eigenvalues and eigenvectors:

(4)；

in the method, in the process of the invention,as a diagonal matrix, elements on its diagonalIs the eigenvalue obtained by covariance matrix decomposition, orthogonal matrix +.>Each column vector in (a)Feature vectors respectively representing the corresponding feature values;

step 2.3.3, picking out the front having the largest eigenvalueIndividual column vectors->Arranged from large to small according to their characteristic values and arranged in rows from top to bottom to form a new matrix +.>：

(5)；

In the method, in the process of the invention,is the 1 st dimension feature vector, < >>Is->A 1-dimensional feature vector,>is 1 stFeature vector of dimension, ">Is->Personal- >Feature vectors of the dimensions;

step 2.3.4, calculation is reduced toPost-dimensional data->：

*/> (6)。

Further, in step 3, the feature after dimension reduction is input into an attention Bi-GRU neural network, and a Kalman filter is utilized to filter the predicted value at each moment to obtain a final predicted result; the life prediction model built based on the attention Bi-GRU neural network and the Kalman filtering consists of a Bi-directional Bi-GRU learning layer, an attention layer, a Dropout layer, a Flatten layer, a full connection layer and a Kalman filtering post-processing layer.

Further, the Bi-directional Bi-GRU learning layer is composed of a forward GRU and a backward GRU, wherein the forward GRU inputs a forward sequence of fault characteristics, and the backward GRU inputs a backward sequence of fault characteristics; when fault characteristics are transmitted backwards, the model automatically reversely processes a normal sequence to form a backward sequence, and simultaneously transmits the backward sequence and the forward sequence to an output layer for combined calculation to obtain a final value; the Bi-directional Bi-GRU learning layer combines the outputs of the forward GRU and the backward GRU by the following formula, and then calculates and obtains a prediction result by using a Softmax function;

(7)；

in the method, in the process of the invention,is the final output; />Is a forward GRU input; />Is a backward GRU input.

Further, the attention layer uses a SELayer soft attention mechanism to process intermediate layer information, and the calculation formula of the SELayer soft attention mechanism is shown as follows:

(8)；

(9)；

in the method, in the process of the invention,the output of the SELayer soft attention mechanism; />For the sequence number of iteration number, +.>The total iteration times; />Is the weight at the time of adjustment; />For the feature vector before adjustment;

the Dropout layer adopts a Dropout mechanism, and a corresponding calculation formula is shown as follows:

(10)；

(11)；

(12)；

wherein,is provided with->A hidden layer index of a neural network of the layer hidden layer; />Is sparse output; />Is a probability of +.>A vector of bernoulli independent random variables; />Is->An output vector of the layer; />Is the firstLayer->Input vectors for the hidden units; />Is->Layer->Weights of the hidden units; />Is->Layer->Deviation of the individual hidden units; />Is->Layer->Output vectors of the hidden units;is->About->Is a function of (2).

Further, in the flat layer and the full-connection layer, the flat layer tiles the output into 1 dimension, and inputs the output into the full-connection layer for calculation, and the full-connection layer calculation formula is shown as follows:

(13)；

in the method, in the process of the invention,for the input->Data sample,/->Is 1 dimension; />Is- >Data sample and->Connection weights for data samples; />Is->Bias of individual data samples ∈ ->Is->Outputting the full connection layer of the data samples;

the Sigmoid activation function is used at the full connection layer, and the specific formula is as follows:

(14)；

wherein,is->Output of full-connection layer of individual data samples +.>Sigmoid activation function of (a); />Is natural logarithm;

selecting cross entropy as a loss function, cross entropy loss functionThe calculation formula is as follows:

(15)；

in the method, in the process of the invention,is the true remaining life value, +.>Is the model predicted remaining life value.

Further, in the kalman filtering post-processing layer, the final current life prediction value is updated by analyzing the historical prediction data and the prediction value which is not filtered at the moment, and the following formula is a state equation of kalman filtering:

(16)；

(17)；

in the method, in the process of the invention,is->The actual life of the group rolling bearing; />Is->A state transition variable of the group rolling bearing; />Is->The actual life of the group rolling bearing; />Is->Noise of the group rolling bearing obtained by input; />Is the firstPredicted values of the group rolling bearings which are not filtered at the moment; />Is->Measurement value coefficients of the group rolling bearings; />Is->Noise of the group rolling bearing obtained by measurement;

The state update equation of the Kalman filter post-processing layer is shown in the following formula:

(18)；

(19)；

(20)；

(21)；

(22)；

in the method, in the process of the invention,is->An estimated state of the group rolling bearing; />Is->An estimated state of the group rolling bearing; />To act at->Upper->A rank state transformation matrix; />Is->Is->Order control matrix,/->Is a control vector; />Is->Group rolling bearing shape is as follows->Estimating an error covariance matrix a priori; />Is->Group rolling bearing shape is as follows->A posterior estimation error covariance matrix of the order; />Is shaped like +.>A process noise covariance matrix of the order; />Is->Group rolling bearing shape is as follows->A matrix of orders, also known as kalman gain; />Is shaped like +.>An observation matrix of the order; />Transpose the symbol; />Is shaped like +.>A process noise covariance matrix of the order; />Is->A final updated state of the group rolling bearing; />Is->Predicted values of the group rolling bearings which are not filtered at the moment; />Is->Group rolling bearing shape is as follows->A posterior estimation error covariance matrix of the order; />Is->An identity matrix of the order.

Further, when the firstAfter the data of the group rolling bearing is input into the model, the Kalman filter will be according to the +.>Predictive value and front +.>Calculating the +. >The current optimal life estimate of the group rolling bearing data and thus the +.>The optimized correction value of the group rolling bearing data is updated according to the correction value as the predicted value of the subsequent life, and the predicted life after the correction is added into the basic predicted result sequence;

performing iterative operation on the noise-reduced data, and calculating the first stepPredicted value after noise reduction in iterative process +.>At the time, calculate the complete +.>Predicted value after noise reduction in iterative process +.>To->Predicted value after noise reduction in iterative process +.>Substitute for->Predicted value after noise reduction in iterative process +.>To->Predicted value after noise reduction in iterative processThe method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Is the window size used to reduce noise.

The beneficial technical effects brought by the invention are as follows.

1. Proposes a 3-based methodThe self-adaptive judging method of the interval is used for calibrating the degradation starting point of the bearing. The method utilizes the kurtosis characteristic of the vibration signal and the self-adaptive judging flow to determine the degradation starting point of the rolling bearing, can avoid random noise signals to a great extent from interfering with the calibration of the degradation starting point of the bearing, and reasonably lays a foundation for the subsequent signal characteristic extraction.

2. A bearing characteristic extraction method combining time-frequency analysis and a principal component analysis method is provided. The method integrates the time domain, frequency domain and time-frequency domain feature extraction method, obtains normalized energy value features through wavelet packet transformation, finally uses a principal component analysis method to reduce the dimension of the features to obtain the most suitable input features, and improves the accuracy of prediction while reducing the waste of training resources.

3. Aiming at the problem of neural network structure construction, a combined life prediction model of a Bi-directional gating cycle (Bi-GRU) neural network and Kalman filtering based on an attention mechanism is provided. The method not only improves the defect that the common GRU neural network only considers single-direction information, but also can bidirectionally process the input vector and more comprehensively extract the input information; meanwhile, the attention mechanism is utilized to endow different weights to intermediate variables, so that the resources required by model training are reduced, and the accuracy of life prediction is improved. In addition, a Kalman filter is used as a post-processing method of the prediction result. Compared with experimental results of other methods, the life prediction curve is more fit with the actual life curve and more accurate.

Drawings

Fig. 1 is a flowchart of a method of predicting the remaining life of a rolling bearing according to the present invention.

Detailed Description

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

the invention provides a complete standardized technical scheme for predicting the residual life of the rolling bearing, and a technical scheme flow chart of the complete standardized technical scheme is shown in the attached figure 1. The invention provides a rolling bearing residual life prediction method based on an attention Bi-GRU neural network and Kalman filtering, which comprises the following main steps:

Step 1, collecting vibration signals of the rolling bearing by using an acceleration sensor, and adopting 3And (4) calibrating a bearing degradation starting point by using the section self-adaptive judging method.

To calibrate the degradation onset of a rolling bearing, an appropriate reference is selected to determine its signal state. The kurtosis of the bearing vibration signal is a numerical statistic used for describing the peak degree distribution characteristic of the signal waveform, is also a dimensionless parameter, and is irrelevant to the load and other variables of the bearing. It can also be explained that when the surface of the rolling bearing fails, an impact pulse appears at the failure point on the working surface every time the rolling bearing rotates, and the more serious the surface defect is, the larger the amplitude of the impact response is caused, and the more remarkable the failure phenomenon is.

In general, the kurtosis change has strong correlation with the early failure of the rolling bearing, but is insensitive to the degradation degree of the bearing, so the kurtosis of the original time domain root mean square vibration signal of the rolling bearing can be used for confirming whether the early defect of the bearing occurs or not and calibrating the degradation starting point of the rolling bearing.

Kurtosis ofIs defined as follows：

(1)；

In the method, in the process of the invention,representative signal->Average value of>Representative signal- >Center moment of->Representative signal->Standard deviation of (2).

The vibration signal of the rolling bearing in a healthy state is similar to Gaussian distribution, the kurtosis value is kept at about 3, and when the rolling bearing starts to degrade, the kurtosis value of the signal is rapidly increased, so that the rolling bearing can be used as a representation of the occurrence of the fault of the rolling bearing.

After obtaining the kurtosis value of the bearing vibration signal, a method based on the following is adoptedThe self-adaptive judging method of the interval is used for calibrating the degradation starting point of the bearing. First, the rolling bearing history data of the healthy running state is used to calculate the average value of kurtosisStandard deviation of kurtosis->And to determine +.>Interval of/>. Then, use +.>The interval identifies the normal and abnormal states of the bearing. The newly calculated kurtosis is continuously added during the detection process>And->The intervals are compared, if at a certain moment +.>If the kurtosis value exceeds the interval, the working state of the rolling bearing is considered to be abnormal. Such abnormal conditions may be caused by bearing failure or random noise, and in order to avoid erroneous judgment due to random errors occurring during vibration signal acquisition and feature extraction, it is necessary to introduce a trigger mechanism that is when the successive number of kurtosis values exceeds +. >The interval is only started to consider that the rolling bearing enters the degradation phase, i.e. it is the start of the degradation of the rolling bearing at this time.

The specific judgment logic is as follows: first, the iteration number is to be calculatedInitializing to 0, finding that kurtosis value exceeds the value of kurtosis for the first timeTime of section->The method comprises the steps of carrying out a first treatment on the surface of the Then let->Find a moment +.>At the moment +.>Post-continuum->Kurtosis of individualValue->Satisfy->This time is taken as->Record->Is->The method comprises the steps of carrying out a first treatment on the surface of the Finally, let->From->Begin to enlarge until->The value of (2) satisfies->Namely, from the moment +.>The start of bearing failure starts to degrade at this point +.>Can be marked as the start of degradation of the bearing, at which point the remaining life of the bearing is also defined by +.>And starts to gradually descend.

The theory behind the adaptive judging method of the invention is that in general, the abnormal state caused by random noise in the normal operation stage is almost impossible to appear continuously for a plurality of times. When selectedAlong with->This may mean that a fault has occurred when the increase in (c) starts to remain stable, so that the adaptive judgment method can largely avoid the interference of random noise signals with the calibration of the bearing degradation start point.

And step 2, carrying out noise reduction treatment on the original vibration signal, extracting characteristics from a time domain, a frequency domain and a time-frequency domain, and carrying out characteristic dimension reduction by using a principal component analysis method. The specific process is as follows:

Step 2.1, signal noise reduction;

the noise of the original signal is firstly reduced before the feature extraction is carried out by the interference of the sensor or other external factors, and the singular values are removed. The method adopts wavelet threshold filtering, namely, a threshold value is set by using a statistical method to remove noise;

when the data is processed, a truncation method is adopted, and the process of eliminating the singular value of the signal is called truncation processing. In the original time domain signal, the upper and lower thresholds of the signal amplitude are calculated correspondingly based on a normal distribution 3sigma principle, and singular values exceeding the range are replaced by the thresholds;

step 2.2, extracting time domain, frequency domain and time-frequency domain characteristics;

the method comprises the steps of selecting 8 time domain features of an average value, a standard deviation, a maximum value, a minimum value, kurtosis, a deflection, a root mean square value and a peak-to-peak value of each rectangular window intercepting signal (namely, each time step), wherein the average value, the mean square value, the root mean square value, the variance, the square root amplitude, the deflection, the kurtosis, the maximum value, the absolute average amplitude, the minimum value, the peak-to-peak value, the root mean square and the center of gravity 12 of spectrum frequency and the normalized energy value feature obtained through wavelet packet decomposition, so that 21 features are adopted to predict the residual life of the rolling bearing;

Step 2.3, main component analysis feature dimension reduction;

and analyzing and dimension-reducing the 21 selected and extracted features by adopting a principal component analysis method. The method can convert the multidimensional feature into a plurality of key features without reducing the information content of the original data, thereby achieving the aim of minimum coupling between the features. The essence of the principal component analysis method is that the feature vectors in the original feature set are spatially transformed so that they are orthogonal to each other by using a spatial transformation method, so that the largest feature difference can be represented by fewer feature components. Redundant information in the feature set can be effectively removed through a principal component analysis method.

The main component analysis method comprises the following specific steps:

(2)；

(3)；

in the method, in the process of the invention,is the characteristic index vector after centering; />Is->A dimension feature index vector; />Is the original dimension of the data;data sample number, < >>A data sample sequence number; />For a sample matrix, the matrix size is +.>；/>1-dimensional feature index vector centered for 1 st sample, ++>Centered for sample 1 +.>Dimension characteristic index vector, & lt & gt>Is->1-dimensional feature index vector after centering of each sample,/- >Is->Sample-centered +.>A dimension feature index vector;

(4)；

step 2.3.4, calculation is reduced toPost-dimensional data->：

*/> (6)；

When the feature dimension reduction of the principal component analysis is carried out, the number of principal components is very important, and the detection performance of the principal component analysis process is directly determined. The method selects CPV indexes in the principal component analysis in fault diagnosis analysis, wherein the CPV indexes determine principal component numbers according to the cumulative sum percentage of principal component variances. The characteristic values obtained in the step 2.3.2 are arranged according to the order of magnitude, and then the accumulated values are calculated, the curve formed by the accumulated values is analyzed, and the corresponding abscissa value of the inflection point in the curve is the number of principal components to be selected.

When the number of principal elements reaches 15, 99.73% of degradation information in the data set is already included, and the data set can be regarded as containing all information according to the 3sigma principle, so that the method determines the number of principal elements to be 15. At this time, the data characteristic fusion based on the principal component analysis can obtain a better bearing degradation performance index.

And 3, building a life prediction model based on the attention Bi-GRU neural network and Kalman filtering, and predicting the residual life. And inputting the feature after dimension reduction into an attention Bi-GRU neural network, and filtering the predicted value at each moment by using a Kalman filter to obtain a final predicted result.

The life prediction model built based on the attention Bi-GRU neural network and the Kalman filtering consists of a Bi-directional Bi-GRU learning layer, an attention layer, a Dropout layer, a Flatten layer, a full connection layer and a Kalman filtering post-processing layer. The Bi-directional Bi-GRU learning layer has an output dimension of 128, the attention layer has an output dimension of 256, the dropout layer has an output dimension of 256 (23.3% of neurons sleep), the flat layer has an output dimension of 1 dimension, the full connection layer uses a Sigmoid activation function, and the output dimension is 1 dimension. The principles and structures of the layers are described in detail below.

Bi-directional Bi-GRU learning layer: the Bi-directional Bi-GRU learning layer provided by the invention also comprises a forward hiding layer and a backward hiding layer, can jointly process an input sequence in two directions and obtain a final result, and can just solve the problem that a common GRU neural network can only extract unidirectional information. The bidirectional Bi-GRU learning layer consists of forward GRU and backward GRU, and the forward and backward processes are traditional GRU neural network, but the directions are opposite, wherein the forward process extracts the characteristics of history information before the intermediate layer point, the backward process extracts the characteristics of information behind the intermediate layer point, and the two directions jointly determine the prediction result. The GRU is a gated recurrent neural network.

The structural characteristics and the working principle of the forward GRU and the backward GRU are the same. The forward GRU inputs a forward sequence of fault signatures and the backward GRU inputs a backward sequence of fault signatures. When fault characteristics are propagated backwards, the model automatically reversely processes the normal sequence to form a backward sequence, and simultaneously transmits the backward sequence and the forward sequence to an output layer to be combined and calculated to obtain a final value. The Bi-directional Bi-GRU learning layer combines the outputs of the forward and backward GRUs through the following formulas and then calculates the prediction result using the Softmax function.

(7)；

In the method, in the process of the invention,is the final output; />Is a forward GRU input; />Input for backward GRU;

compared with other neural networks, the Bi-GRU information extraction capability is more comprehensive by a Bi-directional propagation mechanism, and the effect of predicting the residual life of the bearing is better. The network structure model can combine the past and future information of the predicted point, so that the accuracy of prediction can be further improved.

The invention adopts the attention mechanism to give different weights to all the output intermediate variables, and solves the problem of difficult processing of information in two directions, thereby improving the prediction accuracy. The soft attention mechanism mainly refers to that when information processing is performed, instead of selecting one or a plurality of pieces of information to be processed, weight distribution is performed on the information to be processed, and then the input sequence is input into a neural network for calculation after weighted average processing is performed on the input sequence. The attention mechanism can allocate the resources to more critical parts under the condition of limited computer resources, and focus on input information which is more important to the current task, so that the information overload problem of an algorithm can be solved while the processing efficiency and accuracy of the prediction task are improved.

Attention layer: the invention uses a SELayer soft attention mechanism for processing intermediate layer information, wherein the SELayer soft attention mechanism consists of a full connection layer, a Dropout layer for avoiding overfitting, a batch regularization layer for facilitating better training, a ReLU activation function, another full connection layer, another Dropout layer, another batch regularization layer and a Sigmoid activation function. The calculation formula of the SELayer soft attention mechanism is shown as follows:

(8)；

(9)；

in the method, in the process of the invention,the output of the SELayer soft attention mechanism; />For the sequence number of iteration number, +.>The total iteration times; />Is the weight at the time of adjustment; />To adjust the feature vector before.

The SELayer soft attention mechanism is mainly used for highlighting fault information of rolling bearing features in a Bi-GRU model, reducing attention to irrelevant information, fusing information in different directions and realizing higher residual life prediction accuracy.

Dropout layer: in deep learning model training, the situations of insufficient data set samples, excessive training times, high model complexity and the like may occur, and the trained model may excessively extract training set features, so that test data or actual situations cannot be accurately predicted, which is called an overfitting phenomenon. Overfitting can greatly impact the generalization ability and robustness of the final model, and to address this problem, methods such as adding additional information in the dataset, noise, and changing the network structure of the training model, which are used to compensate for overfitting, are called regularization methods, can be employed.

The Dropout mechanism is one of the most effective methods to solve the over-fitting problem of neural networks. The Dropout mechanism means that the neural units of the hidden layer are randomly dormant according to a set proportion in the training of the neural network, and the super parameters of the dormant neural units are not changed in the iterative updating, and the corresponding formulas are as follows:

(10)；

(11)；

(12)；

wherein,is provided with->A hidden layer index of a neural network of the layer hidden layer; />Is sparse output; />Is a probability of +.>A vector of bernoulli independent random variables; />Is->An output vector of the layer; />Is the firstLayer->Input vectors for the hidden units; />Is->Layer->Weights of the hidden units; />Is->Layer->Deviation of the individual hidden units; />Is->Layer->Output vectors of the hidden units;is->About->Is a function of (2);

the Dropout mechanism is to add a Bernoulli distribution when data is transmitted, and randomly cancel the influence of some nerve units in the upper layer on the model. The Dropout mechanism can effectively reduce the occurrence of the over-fitting phenomenon of the Bi-GRU network, and the calculated amount of the Bi-GRU network is also greatly reduced due to the deletion of part of neurons, so that the related resources required by training are reduced, and the learning efficiency and the prediction accuracy of the Bi-GRU network are further improved.

The flat layer and the full connection layer: since the remaining life of the rolling bearing output by the model is one-dimensional number, a flat layer is needed to tile the output into 1 dimension, and then the output is input into a full-connection layer for calculation, and the full-connection layer calculation formula is shown as follows:

(13)；

in the method, in the process of the invention,for the input->Data sample,/->Is 1 dimension; />Is->Data sample and->Connection weights for data samples; />Is->Bias of individual data samples ∈ ->Is->And outputting the full connection layer of the data samples.

The Sigmoid activation function is used in the full connection layer, the value range is between 0 and 1, the value range is exactly the same as the calibrated remaining life value range, and the better effect is achieved in training, and the specific formula is as follows:

(14)；

wherein,is->Output of full-connection layer of individual data samples +.>Sigmoid activation function of (a); />Is natural logarithm.

The invention selects the cross entropy as the loss function, because the cross entropy loss function has better effect than the mean square error loss function when the sigmoid is used as the activation function of the last layer, and has the advantages that: the error is large, and the parameter update is fast; the error is small, the parameter updating is slow, so that the problem of too slow weight updating speed caused by the mean square error loss function is solved, and the entropy loss function is crossed The calculation formula is as follows: />

(15)；

In the method, in the process of the invention,is the true remaining life value, +.>The residual life value obtained by model prediction reflects the difference value between the residual life of the real sample and the predicted residual life.

Kalman filter post-processing layer: the Kalman filtering belongs to a software filtering method, and the basic principle is as follows: taking the minimum mean square error as an optimization target, constructing a noise-signal state space model, and updating a final current life predicted value by analyzing historical predicted data and a predicted value which is not filtered at the moment, wherein the following formula is a Kalman filtering state equation:

(16)；

(17)；

in the method, in the process of the invention,is->The actual life of the group rolling bearing; />Is->A state transition variable of the group rolling bearing; />Is->The actual life of the group rolling bearing; />Is->Input-derived noise of group rolling bearingsSound; />Is the firstPredicted values of the group rolling bearings which are not filtered at the moment; />Is->Measurement value coefficients of the group rolling bearings; />Is->Noise of the group rolling bearing obtained by measurement;

(18)；

(19)；

(20)；

(21)；

(22)；

in the method, in the process of the invention,is->An estimated state of the group rolling bearing; />Is->An estimated state of the group rolling bearing; / >To act at->Upper->A rank state transformation matrix; />Is->Is->Order control matrix,/->Is a control vector; />Is->Group rolling bearing shape is as follows->Estimating an error covariance matrix a priori; />Is->Group rolling bearing shape is as follows->A posterior estimation error covariance matrix of the order; />Is shaped like +.>A process noise covariance matrix of the order; />Is->Group rolling bearing shape is as follows->A matrix of orders, also known as kalman gain; />Is shaped like +.>An observation matrix of order, which has the function of converting the real state space into a new observation space; />Transpose the symbol; />Is shaped like +.>A process noise covariance matrix of the order;is->A final updated state of the group rolling bearing; />Is->Predicted values of the group rolling bearings which are not filtered at the moment; />Is->Group rolling bearing shape is as follows->A posterior estimation error covariance matrix of the order; />Is->An identity matrix of the order.

When the first isAfter the data of the group rolling bearing is input into the model, the Kalman filter will be according to the +.>Predictive value and front +.>Calculating the +.>The current optimal life estimate of the group rolling bearing data and thus the +.>And (3) optimizing the correction value of the group rolling bearing data, sequentially updating the predicted value of the subsequent life according to the correction value, and adding the predicted life after the correction into a basic predicted result sequence.

The invention performs iterative operation on the noise-reduced data to strengthen the noise reduction effect, and calculates the first stepPredicted value after noise reduction in iterative process +.>At the time, calculate the complete +.>Predicted value after noise reduction in iterative process +.>To->Predicted value after noise reduction in iterative process +.>Substitute for->Predicted value after noise reduction in iterative process +.>To->Predicted value after noise reduction in iterative process +.>The method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Is the window size used to reduce noise. When a certain predicted value is noise-reduced and corrected, the predicted value is added to the next noise-reducing link immediately, so that iteration is continued, and a better predicted effect is obtained.

In order to prove the feasibility and superiority of the invention, experiments are carried out, wherein an XJTU-SY rolling bearing accelerated life test data set and a 2012PHM challenge match data set are utilized in the experiments, and meanwhile, two evaluation indexes, namely root mean square error and predictive score, which are mainly used in most related researches in the field are utilized for comparison and evaluation.

The form of the root mean square error index is very similar to the standard deviation in the statistical concept, and is a commonly used prediction accuracy evaluation index, and the calculation formula is as follows:

(23)；

wherein,is the root mean square error value; / >Is indicated at->The actual rolling bearing remaining life at the moment;is indicated at->Predicting the residual life of the rolling bearing at the moment; />Representing the number of predicted data points in a prediction period, +.>Representing the sequence number of the predicted data point within the prediction period.

The magnitude of the root mean square error can reflect the accuracy of the predicted result, and the greater the magnitude of the root mean square error, the lower the accuracy of the predicted result.

The prediction score index is from IEEE 2012PHM data challenge, and when the predicted life is longer than the actual life, the penalty coefficient is larger. The prediction accuracy of the estimated bearing life can be intuitively represented, and the calculation formula is as follows:

(24)；

(25)；

wherein,a value that is a predictive score; />Is an intermediate variable.

The method has the advantages that the influence of different parameters on the network prediction performance is detected by the same prediction performance index, and is compared and analyzed with other neural network prediction methods, and on the XJTU-SY rolling bearing accelerated life test data set, the prediction score index of each final prediction method is obtained, so that the prediction result of the method is improved by 23% compared with the best GRU prediction network in other prediction methods, and is improved by 70.2% compared with the average value of other methods. On the root mean square error index, the prediction result of the method is improved by 41.7% compared with the best GRU prediction network in other prediction methods, and is improved by 67.7% compared with the average value of other methods. On 2012PHM challenge match data set, the method of the invention is improved by 63.8% compared with STFT-CNN on root mean square error index and 55.23% compared with CNN model based on wavelet transformation and global pooling. From the data, the method can better fit the life degradation trend of the rolling bearing under the complex working condition, and compared with other neural network models, the method has better fitting effect and more accurate prediction result of the residual life.

It should be understood that the above description is not intended to limit the invention to the particular embodiments disclosed, but to limit the invention to the particular embodiments disclosed, and that the invention is not limited to the particular embodiments disclosed, but is intended to cover modifications, adaptations, additions and alternatives falling within the spirit and scope of the invention.

Claims

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

in the step 1, after the bearing vibration signal is collected, the kurtosis is calculatedThe calculation formula of (2) is as follows:

(1)；

after obtaining the kurtosis value of the bearing vibration signal, adopting 3The self-adaptive judging method of the interval carries out bearing degradation point calibration, and the specific process is as follows: firstly, the average value +_of kurtosis is calculated using the rolling bearing history data of the healthy operating state>Standard deviation of kurtosis->And to determine +.>Interval->The method comprises the steps of carrying out a first treatment on the surface of the Then, use +.>The section is used for identifying the normal state and the abnormal state of the bearing; the newly calculated kurtosis is continuously added during the detection process >And->The intervals are compared, if at a certain moment +.>The kurtosis value of (2) exceeds +.>The section is determined to be the degradation starting point of the rolling bearing at the time;

step 3, building a life prediction model based on the attention Bi-GRU neural network and Kalman filtering, and predicting the residual life;

in the step 3, the feature after dimension reduction is input into an attention Bi-GRU neural network, and a Kalman filter is utilized to filter the predicted value at each moment to obtain a final predicted result; the life prediction model built based on the attention Bi-GRU neural network and the Kalman filtering consists of a Bi-directional Bi-GRU learning layer, an attention layer, a Dropout layer, a Flatten layer, a full connection layer and a Kalman filtering post-processing layer;

the bidirectional Bi-GRU learning layer consists of a forward GRU and a backward GRU, wherein the forward GRU inputs a forward sequence of fault characteristics, and the backward GRU inputs a backward sequence of fault characteristics; when fault characteristics are transmitted backwards, the model automatically reversely processes a normal sequence to form a backward sequence, and simultaneously transmits the backward sequence and the forward sequence to an output layer for combined calculation to obtain a final value; the Bi-directional Bi-GRU learning layer combines the outputs of the forward GRU and the backward GRU by the following formula, and then calculates and obtains a prediction result by using a Softmax function;

(7)；

the attention layer uses a SELayer soft attention mechanism to process intermediate layer information, and the calculation formula of the SELayer soft attention mechanism is shown as follows:

(8)；

(9)；

(10)；

(11)；

(12)；

wherein,is provided with->A hidden layer index of a neural network of the layer hidden layer; />Is sparse output;is a probability of +.>A vector of bernoulli independent random variables; />Is->An output vector of the layer; />Is the firstLayer->Input vectors for the hidden units; />Is->Layer->Weights of the hidden units; />Is->Layer->Deviation of the individual hidden units; />Is->Layer->Output vectors of the hidden units;is->About->Is a function of (2);

in the flat layer and the full-connection layer, the flat layer tiles output into 1 dimension, and inputs the output into the full-connection layer for calculation, and the full-connection layer calculation formula is shown as follows:

(13)；

in the method, in the process of the invention, For the input->Data sample,/->Is 1 dimension; />Is->Data sample and->Connection weights for data samples; />Is->Bias of individual data samples ∈ ->Is->Outputting the full connection layer of the data samples;

(14)；

(15)；

in the method, in the process of the invention,is the true remaining life value, +.>The residual life value obtained by model prediction;

in the Kalman filtering post-processing layer, the final current life predicted value is updated by analyzing historical predicted data and the predicted value which is not filtered at the moment, and the following formula is a Kalman filtering state equation:

(16)；

(17)；

in the method, in the process of the invention,is->The actual life of the group rolling bearing; />Is->A state transition variable of the group rolling bearing;is->The actual life of the group rolling bearing; />Is->Noise of the group rolling bearing obtained by input; />Is->Predicted values of the group rolling bearings which are not filtered at the moment; />Is->Measurement value coefficients of the group rolling bearings; / >Is the firstNoise of the group rolling bearing obtained by measurement;

(18)；

(19)；

(20)；

(21)；

(22)；

in the method, in the process of the invention,is->An estimated state of the group rolling bearing; />Is->An estimated state of the group rolling bearing; />To act at->Upper->A rank state transformation matrix; />Is->Is->Order control matrix,/->Is a control vector; />Is->Group rolling bearing shape is as follows->Prior estimation of orderAn error covariance matrix; />Is->Group rolling bearing shape is as follows->A posterior estimation error covariance matrix of the order; />Is shaped like +.>A process noise covariance matrix of the order; />Is->Group rolling bearing shape is as follows->A matrix of orders, also known as kalman gain; />Is shaped like +.>An observation matrix of the order; />Transpose the symbol; />Is shaped like +.>Process noise of the orderA covariance matrix; />Is->A final updated state of the group rolling bearing; />Is->Predicted values of the group rolling bearings which are not filtered at the moment; />Is->Group rolling bearing shape is as follows->A posterior estimation error covariance matrix of the order; />Is->A unit matrix of the order;

when the first isAfter the data of the group rolling bearing is input into the model, the Kalman filter will be according to the +.>Predictive value and front +. >Calculating the +.>The current optimal life estimate of the group rolling bearing data and thus the +.>The optimized correction value of the group rolling bearing data is updated according to the correction value as the predicted value of the subsequent life, and the predicted life after the correction is added into the basic predicted result sequence;

performing iterative operation on the noise-reduced data, and calculating the first stepPredicted value after noise reduction in iterative process +.>At the time, calculate the complete +.>Predicted value after noise reduction in iterative process +.>To->Predicted value after noise reduction in iterative processSubstitute for->Predicted value after noise reduction in iterative process +.>To->Predicted value after noise reduction in iterative process/>The method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Is the window size used to reduce noise.

2. The method for predicting the residual life of a rolling bearing according to claim 1, wherein the specific process of step 2 is as follows:

3. The method for predicting the residual life of a rolling bearing according to claim 2, wherein the specific process of step 2.3 is as follows:

(2)；

(3)；

in the method, in the process of the invention,is the characteristic index vector after centering; />Is->A dimension feature index vector; />Is the original dimension of the data; />Data sample number, < >>A data sample sequence number; />For a sample matrix, the matrix size is +.>；/>1-dimensional feature index vector centered for 1 st sample, ++>Centered for sample 1 +.>Dimension characteristic index vector, & lt & gt>Is->1-dimensional feature index vector after centering of each sample,/->Is->Sample-centered +.>A dimension feature index vector;

(4)；

(5)；

In the method, in the process of the invention,is the 1 st dimension feature vector, < >>Is->A 1-dimensional feature vector,>is 1 +.>Feature vector of dimension, ">Is->Personal->Feature vectors of the dimensions;

step (a)2.3.4, calculation is reduced toPost-dimensional data->：

*/> (6)。