CN113947017B

CN113947017B - Method for predicting residual service life of rolling bearing

Info

Publication number: CN113947017B
Application number: CN202111211199.9A
Authority: CN
Inventors: 黄宇; 冯坤; 江志农
Original assignee: Beijing University of Chemical Technology
Current assignee: Beijing University of Chemical Technology
Priority date: 2021-10-18
Filing date: 2021-10-18
Publication date: 2024-07-26
Anticipated expiration: 2041-10-18
Also published as: CN113947017A

Abstract

The invention discloses a method for predicting the residual service life of a rolling bearing, which comprises the following steps: acquiring the trend characteristics of the rolling bearing life data; constructing an LSTM model combined with a self-attention mechanism; training a network by using the trend characteristic data to obtain a model capable of outputting a rolling bearing degradation curve; inputting the trend feature set of the non-full life cycle data of the test bearing into a model to obtain a predicted bearing degradation curve; and performing curve fitting on the predicted degradation curve by using a least square method, calculating the time for the bearing to reach the failure threshold value, obtaining the residual service life of the bearing, and providing support for maintenance personnel to make maintenance decisions. The method can be applied to predicting the residual service life of the rolling bearing running under the complex working condition, accurately predicts the residual time of the rolling bearing reaching the failure point, and provides support for intelligent decision.

Description

Method for predicting residual service life of rolling bearing

Technical Field

The invention relates to a method for predicting the residual service life of a rolling bearing, in particular to a method for predicting the residual service life of a rolling bearing by combining a long and short time memory network (LSTM) and a Self-Attention mechanism (Self-Attention).

Background

In recent years, in the aspect of health monitoring of mechanical equipment, not only the type and the position of the fault can be accurately judged when the equipment has serious faults, but also the residual service life (REMAINING USEFUL LIFE, RUL) of the equipment is predicted after the equipment is diagnosed to be in an early stage of the fault, and an optimal maintenance point is determined, so that technical support is provided for predictive maintenance.

The rolling bearing is the most common and extremely important rotating part in mechanical equipment, and is extremely easy to appear phenomena such as pitting corrosion, abrasion and the like due to factors such as environment, working condition and the like in the running process of the equipment, so that the rolling bearing becomes one of the parts with the worst industrial damage and reliability. Failure of the rolling bearing is likely to cause failure of the entire rotary machine, possibly causing great economic loss to the enterprise. In rotating equipment, about 30% of failures are caused by rolling bearings, the operating state of which has a significant impact on the mechanical equipment. Therefore, the running state of the rolling bearing is monitored, the residual service life of the rolling bearing is predicted, the optimal maintenance point is determined, a maintenance scheme can be established for a decision maker to provide support, the enterprise cost is reduced, huge economic loss is avoided, and even safety accidents caused by equipment failure can be avoided.

With the advent of the big data age, data-driven algorithms are widely applied to various industries. During operation, the rolling bearing generates a large amount of vibration data, which contains abundant bearing operation state information. Therefore, the research of the residual service life prediction method driven by data is of great significance to the health management of mechanical equipment.

In order to make the most efficient use of the mass vibration data collected from the rolling bearing, researchers have proposed using BP artificial neural networks for RUL prediction. According to the method, firstly, fault characteristics are extracted from vibration data, then, bearing fault characteristic data with the whole service life is used as a training sample of the BP neural network, and a mapping relation from the bearing running state to the residual service life is established through the training neural network. Compared with the traditional prediction method based on failure physics, the method is easy to realize, the internal principle is easy to understand, and the range of equipment types capable of being applied is relatively wide. The following problems remain: firstly, the BP neural network can only establish a shallow model, the learned mapping relation is simpler, and the generalization performance is poorer. Because the running environment of the bearing is complex, even if characteristic data are extracted, the characteristics still contain a lot of interference such as environmental noise, and the complexity of the data and the model is unequal, so that the prediction accuracy of the bearing is still provided with a lifting space. And secondly, vibration data acquired by using a sensor is time series data, and the data is characterized in that a certain correlation exists between data at the front and rear continuous time points, and the BP neural network cannot capture the time correlation of the data.

In response to the last shortcoming of the above-mentioned approach, many researchers have utilized convolutional neural networks for residual life prediction, but convolutional neural networks can only capture time-dependence over the receptive field. Later, some researchers noted the advantages of recurrent neural networks in terms of processing time series data, and utilized recurrent neural networks for residual life prediction. Some researchers have used long and short term memory networks (LSTM) to construct health curves for the evidence bearing decay process and then used a fit prediction method to predict the time remaining for the bearing to reach the point of failure. However, LSTM still has the disadvantage that LSTM can only perform serial operation first, that is, data of each time step is input sequentially, and the calculation is time-consuming. Second, when the sequence length exceeds a certain limit, the LSTM still experiences a gradient extinction phenomenon, resulting in failure to capture time-dependent information of the time step beyond the limit.

In these practices, researchers have investigated this problem using different types of neural networks, but have had their own shortcomings, whether BP neural networks, convolutional neural networks, or long and short term memory networks.

Disclosure of Invention

In order to solve the problems in the prior art, the embodiment of the invention provides a method for predicting the residual service life of a rolling bearing. The technical scheme is as follows:

in one aspect, a method for predicting remaining service life of a rolling bearing is provided, including:

acquiring the trend characteristics of the rolling bearing life data;

Constructing an LSTM model combined with a self-attention mechanism;

Training a network by using the trend characteristic data to obtain a model capable of outputting a rolling bearing degradation curve;

Inputting the trend feature set of the non-full life cycle data of the test bearing into a model to obtain a predicted bearing degradation curve;

And performing curve fitting on the predicted degradation curve by using a least square method, calculating the time for the bearing to reach the failure threshold value, obtaining the residual service life of the bearing, and providing support for maintenance personnel to make maintenance decisions.

Further, the specific steps for acquiring the trend characteristics of the rolling bearing life data are as follows:

(1) An acceleration sensor is arranged on the rolling bearing, and full life cycle monitoring data are collected;

(2) Firstly performing Hilbert transformation on the full life cycle data, forming analysis signals by signals before and after transformation, then solving a module of the analysis signals to obtain envelope signals, and finally performing low-pass filtering and Fourier transformation on the envelope signals to obtain envelope spectrums;

(3) Dividing the envelope spectrum of the vibration signal into n sections according to the average frequency;

(4) Setting a sample at the beginning operation time of the bearing as a standard sample, and respectively calculating pearson correlation coefficients of each sub-frequency band and the frequency band corresponding to the standard sample for all samples;

(5) For all samples, respectively calculating the pearson correlation coefficients of the full-frequency band signal and the standard samples;

(6) And (3) carrying out standardization treatment on n+1 trend characteristics reflecting the bearing degradation process obtained by each sample.

Further, the hilbert transform and fourier transform are specifically:

Hilbert transform produces a positive 90 phase shift to the positive frequency of the original signal, a negative 90 phase shift to the negative frequency, and constructs an analytic signal of the original signal without changing the amplitude of the spectral components; for any time domain signal x _m (t), its Hilbert transform is defined as x _m (t) and Is (are) convolved, i.e

Where x represents a convolution operator and H represents a hilbert operator.

The analytical signal is:

The spectrum z _m (f) of the analysis signal z _m (t) is obtained by a fourier transform method.

Z _m (f) is the envelope spectrum of the original signal.

Further, the calculation formula of the pearson correlation coefficient is as follows:

Where X and Y represent vectors and N represents vector length.

The technical scheme provided by the embodiment of the invention has the beneficial effects that:

the method for predicting the residual service life of the rolling bearing can be applied to predicting the residual service life of the rolling bearing running under complex working conditions, accurately predicts the residual time of the rolling bearing reaching a failure point, and provides support for intelligent decision. The method provided by the invention adopts envelope demodulation analysis instead of traditional spectrum analysis, can effectively demodulate and extract weak fault characteristics, and ensures that the model has strong robustness and generalization. The method provided by the invention segments the envelope demodulation spectrum and calculates the pearson correlation coefficient to measure the degradation degree of the rolling bearing in the running process, and the constructed trend feature set can effectively train the degradation curve prediction model. The method provided by the invention adopts novel neural network structural design, so that the training efficiency of the model is higher, and the prediction of the degradation curve is more accurate.

In conclusion, the method is novel, simple and reliable, has wide application range and is convenient to use in engineering practice.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

Fig. 1 is a complete flow chart of an embodiment of the present invention.

FIG. 2 is a full life cycle vibration waveform diagram of a test bearing according to an embodiment of the present invention.

FIG. 3 is a graph of a trend feature visualization of a test bearing according to an embodiment of the present invention.

FIG. 4 is a predicted degradation curve for a test bearing according to an embodiment of the present invention.

FIG. 5 is a graph of the smoothing effect of the predicted degradation curve of the test bearing according to the embodiment of the present invention.

FIG. 6 is a graph showing the fit of the predicted degradation curve of a test bearing according to an embodiment of the present invention.

FIG. 7 is a graph of scoring function of an evaluation model according to an embodiment of the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the embodiments of the present invention will be described in further detail with reference to the accompanying drawings.

The invention discloses a method for predicting the residual service life of a rolling bearing, which comprises the following steps:

acquiring the trend characteristics of the rolling bearing life data;

Constructing an LSTM model combined with a self-attention mechanism;

Specifically, the invention provides a novel rolling bearing life-span data trend feature extraction method, which realizes data reduction and denoising and obtains implicit features of a surface evidence rolling bearing degradation process. The neural network model provided by the invention combines a long-period memory network and an attention mechanism, can realize parallel operation and can capture time correlation information of any time span. The model can return a degradation value indicating the degradation degree of the rolling bearing to a user according to the trend characteristic of the rolling bearing at a given moment. The degradation value returned ranges between 0 and 1, with a value of 0 indicating that the bearing is completely normal and a value of 1 indicating that the bearing has failed. The method provided by the invention uses a least square method for prognosis, and uses the least square method for curve fitting aiming at the degradation curve of the rolling bearing, so as to predict the degradation value and calculate the residual service life of the rolling bearing.

The method mainly comprises the following specific processes in actual application:

(1) And an acceleration sensor is arranged on the rolling bearing, and full life cycle monitoring data are collected.

(6) N+1 trend characteristics reflecting the bearing degradation process obtained by each sample are standardized

(7) Building LSTM model combined with self-attention mechanism

(8) Training a network by using trend characteristic data to obtain a model capable of outputting a rolling bearing degradation curve

(9) Inputting the trend feature set of the non-full life cycle data of the test bearing into a model to obtain a predicted bearing degradation curve

(10) Performing curve fitting on the predicted degradation curve by using a least square method, calculating the time for the bearing to reach the failure threshold value, obtaining the residual service life of the bearing, and providing support for maintenance personnel to make maintenance decisions

Meanwhile, the invention adopts the following technical scheme:

An LSTM model combined with self-attention mechanism trained by using rolling bearing full life cycle data is used for predicting rolling bearing degradation curve, and curve fitting is carried out by using least square method, and residual service life is obtained through calculation.

(1) When a bearing has faults, the fault impact signal and the high-frequency natural frequency oscillation signal of the bearing are modulated, the traditional frequency spectrum analysis can only acquire the simple information of the wide frequency range of the signal, and the fault modulation characteristic frequency is difficult to accurately extract. The envelope demodulation method can effectively demodulate and extract the fault impact signals and clearly display the fault characteristic frequency information in the low frequency band.

(2) The full life cycle data of the rolling bearing required by the acquisition model is used as training data, a sensor is used for acquiring bearing seat acceleration vibration waveform data with the sampling frequency exceeding 20kHz, the fault frequency and the natural frequency of the bearing are covered, and then Hilbert transformation and Fourier transformation are carried out. The Hilbert transform produces a positive 90 phase shift to the positive frequency of the original signal, a negative 90 phase shift to the negative frequency, and constructs an resolved signal of the original signal without altering the amplitude of the spectral components.

For any time domain signal x _m (t), its Hilbert transform is defined as x _m (t) andIs (are) convolved, i.e

Where x represents a convolution operator and H represents a hilbert operator.

The analytical signal is:

Z _m (f) is the envelope spectrum of the original signal.

(3) The magnitude of the envelope spectrum of the normal and fault signals may be quite different, with different ranges of values. The pearson correlation coefficient theory provides a method for solving the problem, which eliminates the difference in the dimensions of different variables, so that no requirement is made on the value range between different variables, and compared with the traditional calculation of the Euclidean distance, the trend is measured by the correlation obtained by the pearson correlation coefficient theory. The calculation formula of the pearson correlation coefficient is as follows:

Where X and Y represent vectors and N represents vector length.

(4) The long-period memory network combined with the self-attention mechanism can perform serial-parallel operation, and compared with the traditional long-period memory network which can only perform serial operation, the training speed is greatly improved. Under the condition that the length of an input sequence is long, the gradient vanishing phenomenon still exists in the long-short-period memory network, hidden information in the long input sequence is difficult to mine, and a self-attention mechanism focuses limited attention on important information by calculating weights among different time points of the input sequence. Therefore, the self-attention mechanism can capture time correlation information of an input sequence with any length, and the prediction performance of the degradation curve model is improved. Let h= [ H ₁,h₂,…,h_T ] be the output matrix of the long-short-term memory network, T denote the length of the input sequence, then the implicit representation r of the input sequence obtained by the self-attention mechanism is obtained by:

M＝tanh(H)

α＝softmax(w^TM)

r＝Hα^T

And (3) implicitly representing r to perform attention operation again on the high-order features extracted by the long-period memory network, mining an implicit evidence capable of reflecting trend features deeper, and finally inputting a full-connection layer to map the hidden evidence into a degradation value of the bearing.

(5) The degradation curve predicted by the model has errors, the traditional interpolation fitting method requires the interpolation curve to pass through all observation points with errors, when the data quantity is relatively large, the curve fitting effect is not ideal due to overhigh polynomial times, the least square method does not require the fitting curve to pass through known points, and the best function matching of the data is searched by minimizing the square sum of the errors, so that the accuracy of curve fitting is improved, and the prediction accuracy of the residual service life is improved. Assuming a given function f (x; α ₁,α₂,…,α_n) and its measured values y ₁,y₂,…,y_N of N different observation points x ₁,x₂,…,x_N, the least squares fit requires that the unknown parameter set α ₁,α₂,…,α_n be determined that holds the following equation.

Specifically, the degradation of the bearing is a slow process, and the typical full life cycle data acquisition time is quite long, so that the residual life prediction model combining the long-period memory network and the attention mechanism designed by the invention is applied to the public data set, and the whole flow of the method is shown in figure 1.

In this example, the test adds radial load to accelerate the bearing decay while the bearing is running and acquires life cycle data for the bearing over several hours. In the test, two unidirectional acceleration sensors are adopted to collect vibration data of the bearing in the horizontal and vertical directions, the sampling frequency of the test is 25.6kHz, the sampling interval is 1 minute, the sampling time is 1.28 seconds each time, and therefore the length of each data sample is 32768. In the present embodiment, the vibration acceleration signal in the horizontal direction is used for verification. In the test, 3 kinds of working conditions are designed, the working condition description of the used public data set is shown in table 1, 5 bearings are arranged under each kind of working condition, and the bearing data description of the used public data set is shown in table 2. The training set should contain all kinds of faults, so in this embodiment, bearings 1_1, 1_2, 1_3, 2_1 and 3_2 are selected as training bearing data, and the remaining 10 bearings randomly intercept part of the data as a test set.

TABLE 1

TABLE 2

In applying the present invention, a trend feature set is first constructed. Taking the bearing 1_4 as an example, table 1 is vibration data of the whole life cycle of the bearing, and the vibration amplitude of the whole degradation process of the bearing is gradually increased from the table, so that it is meaningful to construct a trend feature set of the degradation of the bearing. The method is an original method and mainly comprises the following two steps: firstly, performing Hilbert transformation on an original signal, solving a module of an analysis signal to obtain an envelope signal, and performing Fourier transformation on the envelope signal to obtain an envelope spectrum; the second step equally divides the envelope spectrum into n segments (n=4 in this example) according to the frequency ranges, resulting in 4 sub-bands, the frequency ranges being 0-6.4kHz, 6.4kHz-12.8kHz, 12.8kHz-19.2kHz and 19.2kHz-25.6kHz, respectively. And taking the first time sample as a standard sample, calculating the pearson correlation coefficient of each sub-frequency segment of the full life cycle sample and the sub-frequency segment corresponding to the standard sample, then calculating the pearson correlation coefficient of the full frequency band envelope spectrum and the standard sample, and finally obtaining 5 trend characteristics. Taking the bearing 3_2 as an example, fig. 3 is a visual result of the trend feature set of the bearing 3_2 after the dimension reduction by using t-sne, and it can be seen from the graph that the trend feature set presents a certain trend, which illustrates that the method of the present invention is effective.

And secondly, building an LSTM model used by the invention and combined with a self-attention mechanism. The model in the invention is a supervised learning model, so that corresponding labels are required to be set when the model is trained, and the labels are degradation values. The preset degradation value is a ratio, the whole life cycle of a bearing is assumed to be T, and when the bearing works and runs to a time T, the label value of the bearing is T/T. The loss function L of the neural network is designed as a mean square error, and the formula is as follows:

wherein n is the number of training samples, y _i is the label, To predict the output of the network.

After the prediction network is built and the loss function is determined, the network parameters are learned by using a method of backward propagation gradient descent, and in the embodiment, an Adam optimization algorithm based on self-adaptive gradient and momentum is used.

And training a neural network by utilizing the trend characteristics of the training bearing, testing on a testing set after training a model, and outputting a degradation curve of the bearing. Taking test bearing 1_4 as an example, FIG. 4 is a predicted degradation curve for bearing 1_4.

The degradation curve directly output by the model locally has fluctuation with different degrees, and the prediction of the subsequent residual service life is influenced. The curve is smoothed by means of a sliding average method, setting the sliding window size to 10, the sliding average operation being as follows:

Where w is the length of the sliding window, d _i is the degradation value of the ith sample of the current bearing, and N degradation values are used, corresponding to N samples. The smoothing effect of the degradation curve of the bearing 1_4 is shown in fig. 5.

And finally, fitting the smooth degradation curve by using a least square method. Taking test bearing 1_4 as an example, fig. 6 is a fit of the degradation curve of the bearing. From the graph, the fitted curve can be well fitted with the degradation curve, and the development trend of the degradation curve is fitted. The degradation value of the bearing is the ratio of the current time T to the full life time T, so the failure threshold of the bearing is set to be 1. The moment when the value of the fitted curve is equal to the failure threshold value is the predicted service life end point t _{Life span}, the current running moment of the bearing is t _{Currently, the method is that}, and the calculation formula of the residual service life RUL of the bearing is as follows:

RUL＝t_{Life span}-t_{Currently, the method is that}

Two indicators are set up in this embodiment to evaluate the method. First, the prediction error percentage% Er _i is as follows:

Wherein ActRUL _i denotes the actual remaining service life of the bearing, predRUL _i denotes the predicted remaining service life of the method according to the invention.

The next is the model average Score. When% Er _i >0, it indicates that the predicted remaining service life is smaller than the actual remaining service life, i.e. the failure point of the bearing is predicted in advance, and the model is valid, the score a _i of the model on the bearing is shown as follows:

when%er _i < = 0, it indicates that the predicted remaining service life is greater than the actual remaining service life, that is, the failure point of the bearing cannot be predicted, and it indicates that the prediction of the model is not ideal, and at this time, the score a _i of the model on the bearing is as follows:

The total score of the model is the average score on each test bearing:

Meanwhile, in the embodiment, fig. 7 is also provided, and a score function diagram of the evaluation model is provided; and table 3, error comparison and score plot.

TABLE 3 Table 3

The foregoing is only illustrative of the present invention and is not to be construed as limiting thereof, but rather as various modifications, equivalent arrangements, improvements, etc., within the spirit and principles of the present invention.

Claims

1. A method of predicting remaining useful life of a rolling bearing, comprising:

acquiring the trend characteristics of the rolling bearing life data;

Constructing an LSTM model combined with a self-attention mechanism;

performing curve fitting on the predicted degradation curve by using a least square method, calculating the time of the bearing reaching the failure threshold value, obtaining the residual service life of the bearing, and providing support for maintenance personnel to make maintenance decisions;

the specific steps for acquiring the trend characteristics of the rolling bearing life data are as follows:

2. The method according to claim 1, wherein the hilbert transform and fourier transform are specifically:

Wherein, represents a convolution operator, and H represents a hilbert operator;

The analytical signal is:

Z _m (f) is the envelope spectrum of the original signal.