CN112036547A - Rolling bearing residual life prediction method combining automatic feature extraction with LSTM - Google Patents

Rolling bearing residual life prediction method combining automatic feature extraction with LSTM Download PDF

Info

Publication number
CN112036547A
CN112036547A CN202010881816.5A CN202010881816A CN112036547A CN 112036547 A CN112036547 A CN 112036547A CN 202010881816 A CN202010881816 A CN 202010881816A CN 112036547 A CN112036547 A CN 112036547A
Authority
CN
China
Prior art keywords
time
model
rolling bearing
bearing
residual life
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010881816.5A
Other languages
Chinese (zh)
Other versions
CN112036547B (en
Inventor
翟怡萌
张启亮
姜丽萍
刘振
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jiangsu Xugong Information Technology Co ltd
Original Assignee
Jiangsu Xugong Information Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Jiangsu Xugong Information Technology Co ltd filed Critical Jiangsu Xugong Information Technology Co ltd
Priority to CN202010881816.5A priority Critical patent/CN112036547B/en
Publication of CN112036547A publication Critical patent/CN112036547A/en
Application granted granted Critical
Publication of CN112036547B publication Critical patent/CN112036547B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Computing Systems (AREA)
  • Mathematical Optimization (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Pure & Applied Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Human Resources & Organizations (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Operations Research (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)

Abstract

The invention discloses a rolling bearing residual service life prediction method combining automatic feature extraction and LSTM, which can predict the residual service life of a rolling bearing according to a vibration signal. Firstly, carrying out data cleaning on vibration data of a rolling bearing in a full life cycle, and eliminating abnormal values; then, performing time-frequency analysis on the vibration signal to obtain a PWVD time-frequency image data set representing the degradation state of the bearing; then based on transfer learning, using a pre-trained VGG16 model to perform automatic feature extraction; and finally, the extracted features are sent to an LSTM network to realize the residual life prediction. The predicted value of the residual life has small mean square error, can monitor the degradation state of the bearing in real time, prevent major accidents from happening, and provide reference opinions for predictive maintenance.

Description

Rolling bearing residual life prediction method combining automatic feature extraction with LSTM
Technical Field
The invention relates to a rolling bearing residual life prediction method combining automatic feature extraction and LSTM, and belongs to the technical field of intelligent fault diagnosis.
Background
The main bearing is a large number of common basic units present in rotary machines. The main bearing is generally a rolling bearing and plays a role in supporting and torque transmission. According to statistics, 30% of rotating machinery faults are caused by rolling bearings, and the doubly-fed wind turbine generators of multiple domestic wind turbine generator manufacturing plants have thousands of rolling bearings or gear failures. When the bearing fails and is not repaired or replaced in time, the gearbox which is tens of times higher than the bearing value or other parts are damaged. If the residual service life of the rolling bearing can be accurately predicted, and the fault part is isolated and replaced in time, the resource waste and the economic loss can be greatly reduced.
The method for predicting the residual life of the rolling bearing is divided into three categories: a mechanism model based approach, a statistical based approach and an artificial intelligence based approach. The advent of the big data and artificial intelligence era has promoted the rapid development of data-driven and artificial intelligence-based methods for predicting remaining life. At present, many scholars realize the residual life prediction method based on a Long Short Term Memory (LSTM) recurrent neural network, but the method needs to manually extract a large number of features and perform feature screening, and has artificial subjectivity. There are also scholars who combine Convolutional Neural Network (CNN) and LSTM to realize automatic feature extraction, but use the frequency spectrum as input, only consider the frequency domain of the vibration signal, do not consider the non-stationary characteristic of the vibration signal, and cannot accurately reflect the degradation state of the bearing. Therefore, automatic feature extraction is carried out on a Pseudo Wigner VilleDistribution (PWVD) time-frequency image of the bearing vibration signal based on transfer learning, and the prediction of the residual life is realized by combining with the LSTM, which is beneficial to improving the prediction precision.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a rolling bearing residual life prediction method combining automatic feature extraction and LSTM.
The invention is realized by the following technical scheme: a method for predicting the residual life of a rolling bearing by combining automatic feature extraction and LSTM is characterized by comprising the following steps:
step 1: acquiring full-life-cycle vibration data of a rolling bearing, and performing time-frequency analysis on a vibration signal to obtain a two-dimensional time-frequency image data set;
step 2: migrating parameters of a VGG16 pre-training model to a target model of an automatic feature extraction module, fully utilizing the universal feature learning capability of the pre-training model in a natural image data set, then taking the two-dimensional time-frequency image data set obtained in the step 1 as the input of a deep convolution network, taking the residual life ratio as the output of the model, setting a parameter updating strategy, and finely adjusting the target model of the automatic feature extraction module;
and step 3: taking the output of the last full-connection layer of the target model of the automatic feature extraction module as a feature vector, setting a time step, generating a feature vector set of the sample number, the time step and the feature number, taking the feature vector set in the time step as the input of an LSTM (least squares metric) residual life prediction model at a certain moment, taking the residual life proportion at the current moment as the model output, and training the residual life prediction model;
and 4, step 4: repeating the step 1 to the step 3, collecting vibration signals of the current moment and the first 29 moments of the bearing to be tested to obtain a two-dimensional time-frequency image of the bearing to be tested, inputting the two-dimensional time-frequency image of the bearing to be tested into the automatic feature extraction deep convolution network trained in the step 3 to obtain a feature vector of the bearing to be tested, and sending the feature vector to the residual life prediction model trained in the step 4 to obtain a residual life proportion of the bearing to be tested;
and 5: and calculating the final residual life according to the residual life proportion and the running time of the rolling bearing.
In the step 1, the time frequency analysis is PWV analysis, and the two-dimensional time frequency image is a PWV time frequency image.
The PWV analysis is a time-frequency analysis method, and for a signal s (t), its WVD is defined as:
Figure BDA0002654315790000021
in the formula (1), x (t) is an analytic signal obtained by Hilbert transform, x (t) is complex conjugate of x (t), the bandwidth of the obtained analytic signal is halved, distortion influence of negative frequency in the signal is avoided, and cross term interference can be effectively reduced, wherein x (t + tau/2) x (t-tau/2) is a transient autocorrelation function of the signal x (t), so that W is the instantaneous autocorrelation function of the signal x (t), and the signal x (t) is a complex conjugate of x (t) and ts(t, Ω) is the local power spectrum of signal x (t);
considering further reducing the influence of cross interference terms, windowing the signal in the time dimension, and combining with a Hamming window, the side lobe intensity of the signal can be effectively reduced, and the improved WVD distribution can be expressed as:
Figure BDA0002654315790000022
discretizing the formula (2) by making t equal to nTsWherein T issFor sampling interval time, segmenting a signal according to frequency conversion multiples, taking a time window w (k) with the length of D being 2M-1, dividing x (t) into P sections in a time dimension, respectively solving improved WVD distribution in each section, and then adding and solving an average value to obtain a time-frequency spectrum characteristic after time domain synchronous averaging, thus obtaining:
Figure BDA0002654315790000031
normalization, let Ts＝1，ω＝ΩTsThe following can be obtained:
Figure BDA0002654315790000032
in the step 2, the automatic feature extraction module is realized based on transfer learning, namely, the knowledge learned from a source data set is transferred to a target data set, a pre-training model is defined as a network structure which is fully trained on a large data set and obtains optimal model parameters, the feature learning capability of the pre-training model in a natural image data set is transferred to an automatic feature extraction task, and a pre-training model is used for learning a time-frequency image and performing automatic feature extraction;
the parameter updating strategy mainly comprises the following 4 steps:
(1) training a source data set to obtain a neural network model, namely a source model;
(2) creating a new neural network model with the same structure as the source model, namely a target model, and copying all parameters except the output layer of the pre-trained model into the target model;
(3) adding an output layer with the output size being the number of the types of the target data set for the target model, and randomly initializing the model parameters of the layer;
(4) training a target model on a target data set, training an output layer from the beginning, and finely adjusting parameters of other layers based on parameters of a source model;
the residual life ratio is defined as the ratio of the residual life of the rolling bearing at a certain moment to the full life cycle of the rolling bearing, and the formula is as follows:
Figure BDA0002654315790000033
in the formula, ytRUL being the remaining life ratio of a certain bearing at time ttThe FLC is the remaining life of a bearing at time t, and the FLC is the full life cycle of the bearing.
In the step 3, the feature vector F extracted by the automatic feature extraction module is expressed as F e RN：
F＝[f1,f2,…,fN] (15)
In the formula (f)1～fNRepresenting an N-dimensional characteristic of the rolling bearing at a certain moment;
the input of the LSTM residual life prediction model is specifically defined as xt∈RM×N：
Figure BDA0002654315790000041
In the formula, xtCharacteristic matrix for rolling bearing used for residual life prediction at time t, wherein
Figure BDA0002654315790000042
For the characteristic vector of the rolling bearing at time t, likewise, Ft-M+1And Ft-M+2Respectively representing the characteristic vectors of the rolling bearing at the moment t-M +1 and the moment t-M + 2;
the output of the LSTM residual life prediction model is the residual life ratio y of the rolling bearing at the moment tt。
In the step 5, the remaining life calculating step is:
(1) establishing a linear equation between the predicted residual life ratio and the operated time through linear regression;
t＝a·yt+b (17)
(2) calculating the full life cycle duration of the rolling bearing when ytWhen the life time is 0, the residual service life of the bearing is 0, and the running time, namely the full life cycle FLC is a.yt+b＝a·0+b＝b；
(3) The remaining service life of the bearing at any time t:
RULt＝FLC-t (18)
where t is the running time of the rolling bearing, ytFor the remaining life ratio of the rolling bearing after running to time t, FLC is the full life cycle, RUL, of the bearingtThe remaining life of the bearing after running to time t.
The invention has the beneficial effects that: from the combined view of time and frequency, the change of the vibration non-stationary signal frequency of the rolling bearing along with the time is described, the degradation state of the rolling bearing is represented by using a two-dimensional time-frequency image, expert experience is not needed, and meanwhile the subjectivity of manual feature extraction is avoided.
Based on deep transfer learning, a pre-training model which is completely trained on a natural image data set is used as a transfer learning object, and an automatic feature extraction module is initialized by using parameters of the pre-training model. Meanwhile, a certain parameter fine-tuning strategy is set in consideration of the difference between the vibration data and the natural image, so that the network can extract the abstract characteristics of the time-frequency image. Therefore, the automatic feature extraction model with fast convergence and high precision is obtained through training.
And constructing a residual life prediction model by using an LSTM network, setting a certain time step, and taking a characteristic matrix of a period of time as model input to fully utilize the advantage of the LSTM in time sequence data prediction.
Drawings
The invention is further illustrated below with reference to the figures and examples.
FIG. 1 is a functional block diagram of the present invention;
FIG. 2 is a waveform diagram of a full life cycle vibration signal of the bearing 1 of the present invention;
FIG. 3 is a waveform of the bearing vibration during the smooth running phase of the present invention;
FIG. 4 is a bearing vibration time-frequency diagram at the smooth-running stage of the present invention;
FIG. 5 is a waveform of the bearing vibration at the amplitude increase stage of the present invention;
FIG. 6 is a bearing vibration time-frequency plot at the amplitude increase stage of the present invention;
FIG. 7 is a waveform of the bearing vibration at the amplitude ramp of the present invention;
FIG. 8 is a bearing vibration time-frequency diagram at the amplitude surge stage of the present invention;
fig. 9 is a predicted value of the remaining life ratio of the bearing 3 of the present invention.
Detailed Description
The method for predicting the residual life of the rolling bearing by combining automatic feature extraction and LSTM, as shown in figures 1 to 3, is characterized by comprising the following steps:
step 1: acquiring the vibration Data of the rolling bearing in the full life cycle and time-frequency analysis, and applying an extra load or increasing the rotating speed to the bearing by using the vibration signal of the rolling bearing in the full life cycle from an IEEE PHM 2012 Data ChallengePRONOSTIA test bed to achieve the purpose of accelerating failure. The experimental bearings 1-7 are operated under the working conditions that the rotating speed is 1800r/min and the load is 4000N, the acceleration sensor collects data every 10s, the time length of data collection every time is 0.1s, and 2560 data points are collected every time. When the acceleration amplitude exceeds 20g, the test is considered to be failed, and the test is ended. All 7 bearings in this condition run from normal to failure. As shown in Table 1, the number of samples of each bearing is shown, wherein the bearing 3 is taken as a test bearing, and the rest are taken as training bearings. As shown in fig. 1, a full life cycle waveform of the bearing 1 is shown. It can be seen that in the early stage of the experiment, the vibration amplitude steadily fluctuates, the fluctuation buoyancy is increased, the amplitude is increased sharply until the final stage, and the whole degradation trend accords with the operation degradation process of the equipment.
The vibration signals are subjected to time-frequency analysis, wherein the time-frequency analysis expresses one-dimensional time sequence signals in a two-dimensional time-frequency combined distribution mode, so that not only are various frequency components of the signals obtained, but also the corresponding relation of the frequency components in the time dimension can be established. WVD is a nonlinear time-frequency analysis method, performs Fourier transform on a signal instantaneous autocorrelation function, and has the advantages of simple calculation, good time-frequency focusing property, strong direct impression and the like. The autocorrelation function brings about a serious cross term interference problem, and various cross coupling term inhibition methods are proposed successively, wherein PWVD has high time frequency resolution, and can inhibit cross interference terms while inheriting good time frequency focusing of the WVD.
Wherein, for signal s (t), its WVD is defined as:
Figure BDA0002654315790000061
in formula (1), x (t) is the analytic signal obtained by Hilbert transform, and x (t) is the complex conjugate of x (t). The bandwidth of the obtained analytic signal is halved, the distortion influence of negative frequency in the signal is avoided, and meanwhile, the cross term interference can be effectively reduced. x (t + τ/2) x (t- τ/2) is the instantaneous autocorrelation function of signal x (t), thus Ws(t, Ω) is the local power spectrum of the signal x (t).
Considering further reducing the influence of cross interference terms, windowing the signal in the time dimension, and combining with a Hamming window, the side lobe intensity of the signal can be effectively reduced, and the improved WVD distribution can be expressed as:
Figure BDA0002654315790000062
discretizing the formula (2) by making t equal to nTsWherein T issIs the sampling interval time. And (3) segmenting signals according to frequency conversion multiples, taking a time window w (k) with the length D of 2M-1, dividing x (t) into P segments in the time dimension, respectively solving improved WVD distribution in each segment, and then adding the segments to obtain an average value to obtain time-frequency spectrum characteristics after time domain synchronous averaging. The following can be obtained:
Figure BDA0002654315790000063
normalization, let Ts＝1，ω＝ΩTsThe following can be obtained:
Figure BDA0002654315790000064
fig. 2 shows a waveform diagram of a vibration signal and a PWVD time-frequency image of the bearing 1 at different degradation stages, where 3 diagrams respectively represent that the bearing 1 is at 3 different degradation stages, and as the amplitude of the vibration signal increases, the frequency component of the signal changes, and the amplitude increases from about 4000Hz to about 0-2000Hz during normal and smooth operation, and in addition, an obvious impact occurs in the vibration mode, and the vibration mode is also reflected in the time-frequency image. By constructing the time-frequency characteristic image, the fault identification is converted into the image identification problem. The deep convolutional neural network can automatically learn the global and local characteristics of the time-frequency image through the action of the convolutional layer filter.
Step 2: and (4) constructing an automatic feature extraction module based on transfer learning. The process comprises the following steps:
(1) and training on the source data set to obtain a neural network model, namely the source model. The pre-trained model of VGG16 on ImageNet dataset in the mxnet framework is used here directly.
(2) And creating a new neural network model which has the same structure as the source model, namely the target model. And copying all parameters except the output layer of the pre-training model into the target model.
(3) Because the output of the automatic feature extraction module is the residual life ratio and the dimension is 1, an output layer with the output size of 1 is added to the target model, and the model parameters of the layer are initialized randomly.
(4) And training a target model on a target data set, namely a rolling bearing two-dimensional time-frequency image data set. The parameters of the first 4 convolution modules are fixed, the parameters of the last convolution module and its following fully-connected layer are fine-tuned with a small learning rate (0.001), while the parameters of the output layer are trained ab initio with a learning rate of 10 times (0.01). Therefore, the extraction capability of the pre-training model to the general features is reserved, the extraction capability of the training model to the abstract features of the two-dimensional time-frequency image data set is simultaneously reserved, and the parameter migration from the pre-training model which is completely trained on the natural image data set to the automatic feature extraction convolution model is realized.
The model output is the remaining life ratio defined in equation (5), using the L2 loss as a loss function for model training.
And step 3: constructing an LSTM residual life prediction model, wherein the process comprises the following steps:
(1) obtaining the output of the last full connection layer of the automatic feature extraction module as a feature vector F ═ F1,f2,…,fN],F∈RNWhere N is 4096.
(2) Reconstructing the feature vectors of all samples to generate a feature matrix (number of samples, time step, feature dimension), i.e. a feature matrix x composed of the feature vectors in the time step as input to the LSTM residual life prediction modelt∈RM ×N：
Figure BDA0002654315790000071
The output is the remaining life ratio defined in equation (5).
(3) Creating a single hidden layer networkAnd the number of the hidden layer neurons is 30, and network parameters are initialized randomly. Setting the learning rate to be 0.1, the time step size to be 30, the batch _ size to be 100 and the number of iteration steps to be 20000; using a training set { xt,yt}T t＝1And training the model, wherein the mean square error is an evaluation index.
And 4, step 4: for a test bearing, time-frequency image characteristics are obtained through time-frequency analysis, a characteristic vector is obtained through an automatic characteristic extraction module, and the characteristic vector is sent to an LSTM residual life prediction model through reconstruction, so that the residual life proportion is predicted. The predicted results are shown in the figure. It can be seen that the proposed automatic feature extraction + LSTM method has smaller mean square error and higher prediction precision when predicting the proportion of the remaining life.
And 5: the remaining life calculating step is as follows:
(1) establishing a linear equation between the predicted residual life ratio and the operated time through linear regression; for the test bearing 3, the remaining life as required by the IEEE2012 PHM tournament at run-up to 18010s was calculated. The linear equation t-27045.2 · y is obtained using the remaining life ratio before operation to 18010s and the elapsed timet+24815；
(2) Calculating the full life cycle duration of the rolling bearing when ytWhen the service life of the bearing is 0, the residual service life of the bearing is 0, and the running time of the bearing, namely the full life cycle FLC is 24815 s;
(3) remaining service life of the bearing 3 running to 18010s at this time: RUL1801024815-.
TABLE 1 Condition 1 number of samples of each bearing
Figure BDA0002654315790000081
TABLE 2 prediction error of residual life at 318010s bearing
Figure BDA0002654315790000082

Claims (6)

1. A method for predicting the residual life of a rolling bearing by combining automatic feature extraction and LSTM is characterized by comprising the following steps:
step 1: acquiring full-life-cycle vibration data of a rolling bearing, and performing time-frequency analysis on a vibration signal to obtain a two-dimensional time-frequency image data set;
step 2: migrating parameters of a VGG16 pre-training model to a target model of an automatic feature extraction module, fully utilizing the universal feature learning capability of the pre-training model in a natural image data set, then taking the two-dimensional time-frequency image data set obtained in the step 1 as the input of a deep convolution network, taking the residual life ratio as the output of the model, setting a parameter updating strategy, and finely adjusting the target model of the automatic feature extraction module;
and step 3: taking the output of the last full-connection layer of the target model of the automatic feature extraction module as a feature vector, setting a time step, generating a feature vector set of the sample number, the time step and the feature number, taking the feature vector set in the time step as the input of an LSTM (least squares metric) residual life prediction model at a certain moment, taking the residual life proportion at the current moment as the model output, and training the residual life prediction model;
and 4, step 4: repeating the step 1 to the step 3, collecting vibration signals of the current moment and the first 29 moments of the bearing to be tested to obtain a two-dimensional time-frequency image of the bearing to be tested, inputting the two-dimensional time-frequency image of the bearing to be tested into the automatic feature extraction deep convolution network trained in the step 3 to obtain a feature vector of the bearing to be tested, and sending the feature vector to the residual life prediction model trained in the step 4 to obtain a residual life proportion of the bearing to be tested;
and 5: and calculating the final residual life according to the residual life proportion and the running time of the rolling bearing.
2. The method for predicting the residual life of a rolling bearing by combining automatic feature extraction and LSTM according to claim 1, wherein:
in the step 1, the time frequency analysis is PWV analysis, and the two-dimensional time frequency image is a PWV time frequency image.
3. The method for predicting the residual life of a rolling bearing by combining automatic feature extraction and LSTM according to claim 2, wherein:
the PWV analysis is a time-frequency analysis method, and for a signal s (t), its WVD is defined as:
Figure FDA0002654315780000011
in the formula (1), x (t) is an analytic signal obtained by Hilbert transform, x (t) is complex conjugate of x (t), the bandwidth of the obtained analytic signal is halved, distortion influence of negative frequency in the signal is avoided, and cross term interference can be effectively reduced, wherein x (t + tau/2) x (t-tau/2) is a transient autocorrelation function of the signal x (t), so that W is the instantaneous autocorrelation function of the signal x (t), and the signal x (t) is a complex conjugate of x (t) and ts(t, Ω) is the local power spectrum of signal x (t);
considering further reducing the influence of cross interference terms, windowing the signal in the time dimension, and combining with a Hamming window, the side lobe intensity of the signal can be effectively reduced, and the improved WVD distribution can be expressed as:
Figure FDA0002654315780000021
discretizing the formula (2) by making t equal to nTsWherein T issFor sampling interval time, segmenting a signal according to frequency conversion multiples, taking a time window w (k) with the length of D being 2M-1, dividing x (t) into P sections in a time dimension, respectively solving improved WVD distribution in each section, and then adding and solving an average value to obtain a time-frequency spectrum characteristic after time domain synchronous averaging, thus obtaining:
Figure FDA0002654315780000022
normalization, let Ts＝1，ω＝ΩTsThe following can be obtained:
Figure FDA0002654315780000023
4. the method for predicting the residual life of a rolling bearing by combining automatic feature extraction and LSTM according to claim 1, wherein:
in the step 2, the automatic feature extraction module is realized based on transfer learning, namely, the knowledge learned from a source data set is transferred to a target data set, a pre-training model is defined as a network structure which is fully trained on a large data set and obtains optimal model parameters, the feature learning capability of the pre-training model in a natural image data set is transferred to an automatic feature extraction task, and a pre-training model is used for learning a time-frequency image and performing automatic feature extraction;
the parameter updating strategy mainly comprises the following 4 steps:
(1) training a source data set to obtain a neural network model, namely a source model;
(2) creating a new neural network model with the same structure as the source model, namely a target model, and copying all parameters except the output layer of the pre-trained model into the target model;
(3) adding an output layer with the output size being the number of the types of the target data set for the target model, and randomly initializing the model parameters of the layer;
(4) training a target model on a target data set, training an output layer from the beginning, and finely adjusting parameters of other layers based on parameters of a source model;
the residual life ratio is defined as the ratio of the residual life of the rolling bearing at a certain moment to the full life cycle of the rolling bearing, and the formula is as follows:
Figure FDA0002654315780000031
in the formula, ytRUL being the remaining life ratio of a certain bearing at time ttFor a bearing at time tThe remaining life, FLC, is the full life cycle of the bearing.
5. The method for predicting the residual life of a rolling bearing by combining automatic feature extraction and LSTM according to claim 1, wherein:
in the step 3, the feature vector F extracted by the automatic feature extraction module is expressed as F e RN：
F＝[f1,f2,…,fN] (6)
In the formula (f)1～fNRepresenting an N-dimensional characteristic of the rolling bearing at a certain moment;
the input of the LSTM residual life prediction model is specifically defined as xt∈RM×N：
Figure FDA0002654315780000032
In the formula, xtCharacteristic matrix for rolling bearing used for residual life prediction at time t, wherein
Figure FDA0002654315780000033
For the characteristic vector of the rolling bearing at time t, likewise, Ft-M+1And Ft-M+2Respectively representing the characteristic vectors of the rolling bearing at the moment t-M +1 and the moment t-M + 2;
the output of the LSTM residual life prediction model is the residual life ratio y of the rolling bearing at the moment tt。
6. The method for predicting the residual life of a rolling bearing by combining automatic feature extraction and LSTM according to claim 1, wherein:
in the step 5, the remaining life calculating step is:
(1) establishing a linear equation between the predicted residual life ratio and the operated time through linear regression;
t＝a·yt+b (8)
(2) calculating the full life cycle duration of the rolling bearing when ytWhen the life time is 0, the residual service life of the bearing is 0, and the running time, namely the full life cycle FLC is a.yt+b＝a·0+b＝b；
(3) The remaining service life of the bearing at any time t:
RULt＝FLC-t (9)
where t is the running time of the rolling bearing, ytFor the remaining life ratio of the rolling bearing after running to time t, FLC is the full life cycle, RUL, of the bearingtThe remaining life of the bearing after running to time t.
CN202010881816.5A 2020-08-28 2020-08-28 Rolling bearing residual life prediction method combining automatic feature extraction with LSTM Active CN112036547B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010881816.5A CN112036547B (en) 2020-08-28 2020-08-28 Rolling bearing residual life prediction method combining automatic feature extraction with LSTM

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010881816.5A CN112036547B (en) 2020-08-28 2020-08-28 Rolling bearing residual life prediction method combining automatic feature extraction with LSTM

Publications (2)

Publication Number Publication Date
CN112036547A true CN112036547A (en) 2020-12-04
CN112036547B CN112036547B (en) 2023-09-22

Family

ID=73586716

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010881816.5A Active CN112036547B (en) 2020-08-28 2020-08-28 Rolling bearing residual life prediction method combining automatic feature extraction with LSTM

Country Status (1)

Country Link
CN (1) CN112036547B (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112486096A (en) * 2020-12-09 2021-03-12 中国兵器装备集团自动化研究所 Machine tool operation state monitoring method
CN112949097A (en) * 2021-04-19 2021-06-11 合肥工业大学 Bearing residual life prediction model and method based on deep migration learning
CN113051689A (en) * 2021-04-25 2021-06-29 石家庄铁道大学 Bearing residual service life prediction method based on convolution gating circulation network
CN113075546A (en) * 2021-03-24 2021-07-06 河南中烟工业有限责任公司 Motor vibration signal feature extraction method and system
CN113255209A (en) * 2021-04-27 2021-08-13 燕山大学 Method for predicting residual life of bearing of gearbox
CN113326590A (en) * 2021-07-16 2021-08-31 北京博华信智科技股份有限公司 Rolling bearing service life prediction method and device based on deep learning model
CN113465923A (en) * 2021-06-08 2021-10-01 华能国际电力股份有限公司上海石洞口第一电厂 Residual life prediction method for grinding roller bearing of coal mill
CN113536982A (en) * 2021-06-28 2021-10-22 长江大学 Intelligent prediction method, system, equipment and medium for residual life of main reducer
CN113984833A (en) * 2021-10-29 2022-01-28 江苏徐工工程机械研究院有限公司 Environmental temperature equivalent and accelerated test method
CN114186500A (en) * 2022-02-16 2022-03-15 华中科技大学 Marine bearing residual life prediction method based on transfer learning and multiple time windows
CN115660198A (en) * 2022-11-07 2023-01-31 中国科学院空间应用工程与技术中心 Method for predicting residual service life of rolling bearing

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101855440A (en) * 2007-11-07 2010-10-06 飞思卡尔半导体公司 Knock signal detection in automotive systems
CN103728663A (en) * 2012-10-16 2014-04-16 中国石油化工股份有限公司 Time-frequency analysis method
US20170300605A1 (en) * 2016-04-19 2017-10-19 General Electric Company Creating predictive damage models by transductive transfer learning
CN109710636A (en) * 2018-11-13 2019-05-03 广东工业大学 A kind of unsupervised industrial system method for detecting abnormality based on depth migration study
CN109726524A (en) * 2019-03-01 2019-05-07 哈尔滨理工大学 A kind of rolling bearing remaining life prediction technique based on CNN and LSTM
CN110941928A (en) * 2019-11-26 2020-03-31 哈尔滨理工大学 Rolling bearing residual life prediction method based on dropout-SAE and Bi-LSTM
CN111579243A (en) * 2020-06-17 2020-08-25 大连理工大学 Rolling bearing intelligent diagnosis system based on deep migration learning

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101855440A (en) * 2007-11-07 2010-10-06 飞思卡尔半导体公司 Knock signal detection in automotive systems
CN103728663A (en) * 2012-10-16 2014-04-16 中国石油化工股份有限公司 Time-frequency analysis method
US20170300605A1 (en) * 2016-04-19 2017-10-19 General Electric Company Creating predictive damage models by transductive transfer learning
CN109710636A (en) * 2018-11-13 2019-05-03 广东工业大学 A kind of unsupervised industrial system method for detecting abnormality based on depth migration study
CN109726524A (en) * 2019-03-01 2019-05-07 哈尔滨理工大学 A kind of rolling bearing remaining life prediction technique based on CNN and LSTM
CN110941928A (en) * 2019-11-26 2020-03-31 哈尔滨理工大学 Rolling bearing residual life prediction method based on dropout-SAE and Bi-LSTM
CN111579243A (en) * 2020-06-17 2020-08-25 大连理工大学 Rolling bearing intelligent diagnosis system based on deep migration learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
刘聪;李言俊;张科;: "基于伪维格纳分布的图像融合", 光学学报, no. 10, pages 2716 - 2720 *

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112486096A (en) * 2020-12-09 2021-03-12 中国兵器装备集团自动化研究所 Machine tool operation state monitoring method
CN113075546A (en) * 2021-03-24 2021-07-06 河南中烟工业有限责任公司 Motor vibration signal feature extraction method and system
CN112949097B (en) * 2021-04-19 2022-09-16 合肥工业大学 Bearing residual life prediction model and method based on deep migration learning
CN112949097A (en) * 2021-04-19 2021-06-11 合肥工业大学 Bearing residual life prediction model and method based on deep migration learning
CN113051689A (en) * 2021-04-25 2021-06-29 石家庄铁道大学 Bearing residual service life prediction method based on convolution gating circulation network
CN113255209A (en) * 2021-04-27 2021-08-13 燕山大学 Method for predicting residual life of bearing of gearbox
CN113255209B (en) * 2021-04-27 2022-06-07 燕山大学 Method for predicting residual life of bearing of gearbox
CN113465923A (en) * 2021-06-08 2021-10-01 华能国际电力股份有限公司上海石洞口第一电厂 Residual life prediction method for grinding roller bearing of coal mill
CN113536982A (en) * 2021-06-28 2021-10-22 长江大学 Intelligent prediction method, system, equipment and medium for residual life of main reducer
CN113536982B (en) * 2021-06-28 2023-02-24 长江大学 Intelligent prediction method, system, equipment and medium for residual life of main reducer
CN113326590B (en) * 2021-07-16 2021-10-29 北京博华信智科技股份有限公司 Rolling bearing service life prediction method and device based on deep learning model
CN113326590A (en) * 2021-07-16 2021-08-31 北京博华信智科技股份有限公司 Rolling bearing service life prediction method and device based on deep learning model
CN113984833A (en) * 2021-10-29 2022-01-28 江苏徐工工程机械研究院有限公司 Environmental temperature equivalent and accelerated test method
CN113984833B (en) * 2021-10-29 2024-03-01 江苏徐工工程机械研究院有限公司 Environment temperature equivalent and acceleration test method
CN114186500A (en) * 2022-02-16 2022-03-15 华中科技大学 Marine bearing residual life prediction method based on transfer learning and multiple time windows
CN115660198A (en) * 2022-11-07 2023-01-31 中国科学院空间应用工程与技术中心 Method for predicting residual service life of rolling bearing

Also Published As

Publication number Publication date
CN112036547B (en) 2023-09-22

Similar Documents

Publication Publication Date Title
CN112036547A (en) Rolling bearing residual life prediction method combining automatic feature extraction with LSTM
CN109460618B (en) Rolling bearing residual life online prediction method and system
CN110286333B (en) Fault diagnosis method for lithium power battery system
CN106197999B (en) A kind of planetary gear method for diagnosing faults
CN113158814B (en) Bearing health state monitoring method based on convolution self-encoder
CN106596116A (en) Vibration fault diagnosis method of wind generating set
CN111046945A (en) Fault type and damage degree diagnosis method based on combined convolutional neural network
CN110907066A (en) Wind turbine generator gearbox bearing temperature state monitoring method based on deep learning model
CN112257528B (en) Wind power gear box fault diagnosis method based on wavelet transformation and dense connection expansion convolutional neural network
CN104748962A (en) Planetary gear box intelligent diagnosis method based on stacking automatic encoding machine
CN106650122B (en) A kind of equipment variable parameter operation methods of risk assessment
CN108444696A (en) A kind of gearbox fault analysis method
CN109214356A (en) A kind of fan transmission system intelligent fault diagnosis method based on DCNN model
CN114266197A (en) Method for diagnosing equipment fault of hydropower station
CN111812507A (en) Motor fault diagnosis method based on graph convolution
WO2023065580A1 (en) Fault diagnosis method and apparatus for gearbox of wind turbine generator set
CN111878322B (en) Wind power generator device
CN114048688A (en) Method for predicting service life of bearing of wind power generator
CN115060491A (en) Fan bearing operation health degree assessment method and system based on multi-source data
CN113743674A (en) Energy storage output prediction method, system, equipment and medium based on deep learning
CN112163474B (en) Intelligent gearbox diagnosis method based on model fusion
Cao et al. Remaining useful life prediction of wind turbine generator bearing based on EMD with an indicator
CN115165364A (en) Wind turbine generator bearing fault diagnosis model construction method based on transfer learning
CN110554667A (en) convolutional Neural Network (CNN) based intermittent industrial process fault diagnosis
CN112213103A (en) Fault diagnosis method, device, system and medium for rail transit rolling stock bearing

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
CB02 Change of applicant information
CB02 Change of applicant information

Address after: 221000 No. 6, Keji Road, Xuzhou Economic and Technological Development Zone, Xuzhou City, Jiangsu Province

Applicant after: XCMG Hanyun Technology Co.,Ltd.

Address before: No.6, Keji Road, Gulou District, Xuzhou City, Jiangsu Province

Applicant before: JIANGSU XUGONG INFORMATION TECHNOLOGY Co.,Ltd.

GR01 Patent grant
GR01 Patent grant