CN114925723B - Method for predicting residual service life of rolling bearing by adopting encoder and decoder - Google Patents

Method for predicting residual service life of rolling bearing by adopting encoder and decoder Download PDF

Info

Publication number
CN114925723B
CN114925723B CN202210487073.2A CN202210487073A CN114925723B CN 114925723 B CN114925723 B CN 114925723B CN 202210487073 A CN202210487073 A CN 202210487073A CN 114925723 B CN114925723 B CN 114925723B
Authority
CN
China
Prior art keywords
rolling bearing
output
time
encoder
decoder
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.)
Active
Application number
CN202210487073.2A
Other languages
Chinese (zh)
Other versions
CN114925723A (en
Inventor
张永平
卢瑾
徐森
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.)
Yancheng Institute of Technology
Yancheng Institute of Technology Technology Transfer Center Co Ltd
Original Assignee
Yancheng Institute of Technology
Yancheng Institute of Technology Technology Transfer Center 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 Yancheng Institute of Technology, Yancheng Institute of Technology Technology Transfer Center Co Ltd filed Critical Yancheng Institute of Technology
Priority to CN202210487073.2A priority Critical patent/CN114925723B/en
Publication of CN114925723A publication Critical patent/CN114925723A/en
Application granted granted Critical
Publication of CN114925723B publication Critical patent/CN114925723B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/04—Bearings
    • G01M13/045—Acoustic or vibration analysis
    • 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
    • G06N3/045—Combinations of 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
    • G06N3/047—Probabilistic or stochastic 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
    • G06N3/048—Activation functions
    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/80—Management or planning

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Acoustics & Sound (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention belongs to the technical field of bearing life prediction, and discloses a method for predicting the residual service life of a rolling bearing by adopting an encoder and a decoder, wherein a vibration signal of the rolling bearing is selected as a training set and a test set, a frequency domain amplitude signal is obtained through fast Fourier transform, and normalization processing is carried out; inputting the frequency domain amplitude signal of the training set after normalization processing as a convolutional neural network, extracting local information of a rolling bearing vibration signal and excavating deep features, taking the deep features as the input of an encoder and a decoder network, constructing a trend quantitative health index, and establishing a trend quantitative health index model; mining deep features of the frequency domain amplitude signals of the test set after normalization processing through a convolutional neural network, inputting a trend quantitative health index model, and obtaining trend quantitative health indexes of the test set; and performing linear regression on the trend quantization health index life percentage of the test set to obtain the residual life of the rolling bearing, thereby realizing residual life prediction.

Description

Method for predicting residual service life of rolling bearing by adopting encoder and decoder
Technical Field
The invention belongs to the technical field of bearing service life prediction, and particularly relates to a method for predicting the residual service life of a rolling bearing by adopting an encoder and a decoder.
Background
Since the concept of "industrial 4.0" is proposed in germany, countries in china, america, the sun, korea, etc. have developed their strategies for developing advanced manufacturing technologies, and their common feature is to combine the traditional industrial production with the modern information technology to raise the intelligence level of the manufacturing industry and further raise the core competitiveness of enterprises. In order to reduce production costs while maximizing production quantity and quality, enterprise operations must ensure that production equipment remains fully operational to maximize throughput of the production system. To keep production facilities fully efficient, enterprises often resort to "unyielding" maintenance strategies, with excessive maintenance also implying high costs. In certain industries, production equipment maintenance costs may account for 15% to 60% of the total production cost. The reasonable equipment maintenance is formulated, the overall efficiency of the production equipment is guaranteed to be maximized, the maintenance cost is reduced, and the method becomes one of the keys for improving the profitability of a manufacturing enterprise and enhancing the competitiveness of the manufacturing enterprise.
Currently, traditional manufacturing enterprises move towards the development of high quality and high efficiency, and more enterprises realize that the effective reduction of the maintenance cost of production equipment is crucial to the long-term development of the enterprises. In recent years, predictive maintenance techniques have been rapidly developed, and the goal thereof is to predict the remaining service life and potential failure of the equipment by using real-time monitoring data, thereby avoiding unexpected shutdown of the production equipment and reducing maintenance cost by maximizing the use of the remaining value of the production equipment. Therefore, a prerequisite and a primary task of predictive maintenance is the prediction of the remaining service life of the device. The mechanical manufacturing industry is the core of the industry and reflects the degree of industrialization, while the rolling bearing is called as an industrial joint as a key component in a transmission system of mechanical equipment, is widely applied to large-scale, high-speed, automatic and intelligent rotating mechanical equipment, is most prone to failure and performance degradation, and the operation state of the rolling bearing directly determines the energy of the mechanical equipment and has important influence on the health condition of the mechanical equipment. Therefore, the rolling bearing is used by monitoring the residual service life of the rolling bearing in time to fully utilize the rolling bearing.
With the rapid development of new technologies such as sensors, storage, network transmission and the like, a large amount of monitoring data are generated in the running process of the rolling bearing, the degradation information of the rolling bearing is mined by using the data, the accurate prediction of the residual service life is realized, and the method is a current research hotspot. The prediction of the residual service life of the rolling bearing is essentially a time series prediction problem, and a Recurrent Neural Network (RNN) has obvious superiority in the aspect of time series processing, so that the prediction method is widely applied to the field of prediction of the residual service life of the rolling bearing.
Rolling bearings are widely used as one of basic components of rotating machinery in a plurality of rotating machinery devices, and the operation state of the rolling bearing plays an important role in safe and reliable operation of the rotating machinery devices. In the industry, the rolling bearing is called as an industrial joint, and if the rolling bearing fails, equipment failure and economic loss are caused if the rolling bearing is light, and safety accidents and casualties are caused if the rolling bearing is heavy. Therefore, the method has great research significance for industrial production safety in accurately and timely predicting the residual service life of the rolling bearing.
Generally, existing failure prediction and health management methods can be divided into three major categories: physical model based methods, data driven methods, and hybrid methods of the two. The data driving method models degradation characteristics according to historical sensor data, and is wide in application range. Deep learning, one of the data-driven methods, has been used in various fields. In recent years, the deep learning method has certain application in the aspects of rolling bearing vibration signal feature extraction and residual service life prediction. Document [1] provides an integrated deep learning method for collaborative prediction of residual service life RUL of a multi-rolling bearing by combining time domain and frequency domain characteristics, and experimental results prove the effectiveness of the method. Document [2] proposes that a frequency domain, a time-frequency domain characteristic and an Auto Encoder (AE) compressed time domain characteristic are jointly input into a deep neural network for residual service life RUL prediction, and a better residual service life RUL prediction result is obtained. The document [3] inputs the peak value and the root mean square value from the wavelet coefficient into a Recurrent Neural Network (RNN) model to achieve the purpose of predicting the residual service life RUL of the rolling bearing. Document [4] proposes that the sum of energy entropies of intrinsic mode functions obtained by empirical mode decomposition is used as a state feature, and a Long Short Term Memory (LSTM) network is used for performing single-step prediction of a mechanical state, so that a good effect is obtained. The Convolutional Neural Network (CNN) is a deep learning method, has the characteristics of weight sharing, convolution operation, spatial pooling and the like, and can mine deep features in a large amount of data. In the literature [5], a convolutional neural network CNN fault diagnosis model is trained by constructing a feature matrix, and the classification effect is superior to that of methods such as AE. In the above studies, although the vibration signal feature extraction and the remaining service life RUL prediction are performed by the deep learning method, the feature extraction is performed manually and in advance by complicated signal processing, and the feature learning characteristic of the deep model is not exhibited.
[1]REN L，CUI J，SUN Y Q，et al .Multi-bearing remaining useful life collaborative prediction：A deep learning approach[J] .Journal of Manufacturing Systems，2017，43：248-256 .
[2]REN L，SUN Y Q，CUI J，et al .Bearing remaining useful life prediction based on deep autoencoder and deep neural networks [J] .Journal of Manufacturing Systems，2018，48：71-77 .
[3]AMAZ MALHI，RUQIANG YAN，et al .Prognosis of defect propagation based on recurrent neural networks[J] .IEEE Transactions on Instrumentation and Measurement，2011，60(03)：703-711 .
[4] Application of Chen Zai, liuyan, liu 21430and source long-short term memory neural network in mechanical state prediction [ J ]. University of great Netherage, 2018, 44 (01): 85-90.
[5]LU C，WANG Z Y，ZHOU B .Intelligent fault diagnosis of rolling bearing using hierarchical convolutional network based health state classification[J] .Advanced Engineering Informatics，2017，32：139-151 .
Through the above analysis, the problems and defects of the prior art are as follows:
(1) The existing rolling bearing vibration signal feature extraction method relies too much on expert experience.
(2) The capability of capturing the dependency relationship of the prediction of the residual service life of the rolling bearing under the condition of inputting a long sequence is rapidly attenuated, namely, the memory degradation phenomenon occurs, and the prediction precision is seriously restricted, namely, the problem of memory degradation caused by overlong sequence in the prediction of the residual service life of the rolling bearing is solved.
(3) Present data-driven remaining useful life RUL prediction methods still require a priori knowledge to extract features, construct health indicators (Hi), and set thresholds, which are inefficient in the big data era.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a method for predicting the residual service life of a rolling bearing by adopting an encoder and a decoder.
The invention is realized in such a way that a method for predicting the residual service life of a rolling bearing by adopting an encoder and a decoder comprises the following steps:
selecting rolling bearing vibration signals as a training set and a test set, obtaining frequency domain amplitude signals through fast Fourier transform, and carrying out normalization processing;
taking the frequency domain amplitude signal of the training set after normalization processing as the input of a convolutional neural network, extracting the local information of the vibration signal of the rolling bearing and excavating deep features, taking the deep features as the input of an encoder and a decoder network, constructing a trend quantization health index, and establishing a trend quantization health index model; mining deep features of the frequency domain amplitude signals of the normalized test set through a convolutional neural network, inputting a trend quantitative health index model, and obtaining trend quantitative health indexes of the test set;
and performing linear regression on the trend quantitative health index of the test set and the life percentage of the test set to obtain the residual life of the rolling bearing.
Optionally, the method specifically includes:
selecting partial data of a rolling bearing vibration signal under different working conditions as a training set, and performing fast Fourier transform on an original vibration signal of the training set to obtain a frequency domain amplitude signal;
secondly, normalizing the frequency domain amplitude signal to be used as characteristic input, and using the service life percentage as an output training model;
thirdly, the frequency domain amplitude signal after the normalization processing is used as the input of a convolutional neural network, and the convolutional layer in the convolutional neural network is used for traversing the whole input data sequence so as to extract the local information of the vibration signal and mine deep features;
inputting the deep features extracted from the convolutional layer into an encoder and decoder network of a recurrent neural network based on an attention mechanism, wherein the encoder is a layer of bidirectional long-short term memory network, and the decoder is a layer of long-short term memory network and an attention mechanism; constructing a trend quantitative health index and establishing a trend quantitative health index model by utilizing the advantages of a long-time memory network and a short-time memory network of a recurrent neural network on time series data and the attention of an encoder network and a decoder network based on an attention mechanism on important characteristics;
fifthly, fast Fourier Transform (FFT) is carried out on non-whole-life time domain vibration signals under different working conditions in the test set to obtain frequency domain amplitude signals, normalization processing is carried out, deep features mined in the third step are combined with the trend quantitative health index model in the fourth step, and trend quantitative health indexes of the test set are obtained;
and sixthly, performing linear regression by using a least square method according to the trend quantization health index of the test set and the service life percentage of the test set, thereby obtaining the residual service life of the rolling bearing.
Optionally, the second step specifically includes:
Figure 391894DEST_PATH_IMAGE001
wherein
Figure 649831DEST_PATH_IMAGE002
Representing N-dimensional characteristic input of a rolling bearing at time t under a certain working condition, N =2048 t ∈[0,1]Indicating the percentage of life degradation output of the rolling bearing at time t; d tra Representing vibration signal data of a rolling bearing under a certain working condition in a training set, wherein R is a frequency domain amplitude characteristic matrix; and T is the service life running time of the rolling bearing.
Optionally, in the third step, local abstract information of the data is automatically extracted by using convolution operation, local link and weight sharing characteristics of the convolutional neural network CNN to mine deep features.
Optionally, the third step specifically includes:
inputting the frequency domain amplitude signal after the rolling bearing normalization processing into a convolution layer, wherein the specific convolution layer operation is as shown in formula (1):
Figure 987271DEST_PATH_IMAGE003
(1)
in the formula:
Figure 427480DEST_PATH_IMAGE004
is the ith convolution kernel of the l layer
Figure 44537DEST_PATH_IMAGE005
The weight value of each of the plurality of the weight values,
Figure 570196DEST_PATH_IMAGE006
for the j th convolved local region r in the l th layer, representing the convolution operation, W is the convolution kernel width, and the modified linear element activation function is used to output for each convolution
Figure 445749DEST_PATH_IMAGE007
Performing nonlinear transformation, specifically expressed as formula (2):
Figure 225617DEST_PATH_IMAGE008
(2)
in the formula:
Figure 528422DEST_PATH_IMAGE007
for convolutional layer output values, f (-) is the activation function ReLU,
Figure 275798DEST_PATH_IMAGE009
is composed of
Figure 705774DEST_PATH_IMAGE007
And (4) an activation value obtained by modifying the linear unit activation function.
Optionally, the specific implementation process of the encoder and decoder network in the fourth step is as follows:
the encoder is responsible for extracting the context information c of the input sequence, and the decoder calculates the output sequence one by one according to the context information c and the data of the previous sequence; output y for a certain sequence of positions t t Expressed by equation (3):
Figure 854995DEST_PATH_IMAGE010
(3)
wherein h is t Information representing the hidden layer at time t, g (\8729;) is an activation function; for the output, the output y at time t t Is the hidden layer information h at that moment t Last moment output y t-1 And a function of the input context information c.
Optionally, the attention mechanism in the fourth step is specifically implemented as follows: the attention-drawing mechanism is actually equivalent to adding a gate a before the global information c participates in the output calculation, and the gate a will input data x according to the time t t Determining the output y at the time of calculating t t When it comes to which part of the global information c is more important, it tells the neural network that the output y is being calculated t When the global information is the most concerned, the global information is the most concerned; global information c is used more effectively when calculating output; the specific formula of the attention mechanism is as follows:
Figure 63123DEST_PATH_IMAGE011
in the formula, h t Is the output variable of the decoder at time t,
Figure 48528DEST_PATH_IMAGE012
it is the output variable of the encoder at time s,
Figure 303DEST_PATH_IMAGE013
the importance of the input information at the time s to the output at the time t is calculated in some way, and finally normalized by the softmax function as the attention gate a at the time t t This attention gate a t And encoder output
Figure 4031DEST_PATH_IMAGE014
Performing dot product operation, allowing the information with the most important correlation with the output calculation at the time t to flow in, and blocking the flow of unimportant information; in which the calculation is carried out
Figure 383060DEST_PATH_IMAGE013
There are three ways, W a And v a Are trainable parameters in the network.
Optionally, the specific implementation process of the long-short time memory network and the bidirectional long-short time memory network in the fourth step is as follows: forgetting door f in long-and-short-term memory network t The information that determines what proportion is retained in the network is calculated as:
Figure 590181DEST_PATH_IMAGE015
（7）
in the formula: x is the number of t Is an input sequence; h is t-1 The state memory quantity at the last moment;
Figure 345648DEST_PATH_IMAGE016
activating a function for sigmoid; w f To forgetA weight matrix of the gate; b t A bias for a forgetting gate; f. of t The state of forgetting the door;
and an input gate i t Selectively memorizing new information in the cell state, and the calculation formula is as follows:
Figure 203882DEST_PATH_IMAGE017
in the formula:
Figure 284971DEST_PATH_IMAGE018
candidate values for cell status; c t A new cell state; tan h (·) is a hyperbolic tangent function; w i Is the weight matrix of the input gate; w c A weight matrix that is a state of the cell; b i Is the bias of the input gate; b c A bias that is a cellular state; i all right angle t Is the status of the input gate;
output gate o t The currently output information is determined, and the calculation formula is as follows:
Figure 979389DEST_PATH_IMAGE019
in the formula: w is a group of o Is a weight matrix of the output gate; b o Is the offset of the output gate; o. o t Is the state of the output gate;
the mathematical transformation formula of the bidirectional long-time and short-time memory network comprises the following steps:
Figure 538546DEST_PATH_IMAGE020
wherein: h is t 、
Figure 251287DEST_PATH_IMAGE021
Respectively outputting a front propagation layer, a back propagation layer and a back propagation layer at the time t; w 1 、W 3 Weight matrices, W, for the input layer and the backward propagation layer, respectively 3 、W 5 From the forward and backward propagation layers to the own propagation layer, respectivelyA weight matrix; w 4 、W 6 The weight matrixes from the forward propagation layer to the output layer and the backward propagation layer to the output layer are respectively set; o is t Is the output value of the final output gate; the function f (-) represents the cell calculation process; the function g (-) represents the function for stitching the forward and backward propagation results.
Optionally, the trend quantized health indicator Hi in the fifth step is a customized indicator, real-time estimation of the health state of the rolling bearing is realized, a trend quantized health indicator Hi is proposed for predicting the remaining service life, and the trend quantized health indicators Hi at time t can use respective RUL t Divided by the initial time RUL o ；
Figure 988430DEST_PATH_IMAGE022
（15）
Obtain the data located in the interval [0, 1]]The Hi value sequence in (1) has the failure point that when Hi reaches 0 t And reaches 0.
Optionally, the first step specifically includes:
(1) Reconfiguring Fourier transform points by configuring a port configuration channel according to the points of the acquired vibration signals of the rolling bearing;
(2) After receiving a signal for starting the fast Fourier transform, taking out interference data from the data storage, converting the interference data into a frequency domain amplitude signal and inputting the frequency domain amplitude signal into the programmable logic device;
(3) Storing the frequency domain amplitude signal output by the programmable logic device into data storage; and after the conversion is finished, sending a conversion finished signal for storage.
By combining all the technical schemes, the invention has the advantages and positive effects that: according to the method, a frequency domain amplitude signal obtained by Fast Fourier Transform (FFT) is subjected to normalization processing and then used as the input of a convolutional layer of a convolutional neural network, deep features are mined, and the problem that a traditional feature extraction method depends on expert experience too much is avoided. The invention adopts the encoder and decoder network of the recurrent neural network based on the attention mechanism to construct the trend quantitative health index, thereby further predicting the residual service life of the rolling bearing and being suitable for predicting the residual service life of the rolling bearing.
Aiming at the problem that the rolling bearing whole life signal sequence is too long to be beneficial to training of the convolutional neural network, the invention provides the method for compressing the information sequence in the time dimension by using the convolutional layer in the convolutional neural network, improves the capacity of the convolutional neural network for processing the long sequence by using an attention mechanism, and finally realizes the prediction of the residual life of the rolling bearing.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present application, and it is obvious for those skilled in the art that other drawings can be obtained from the drawings without creative efforts.
Fig. 1 is a flowchart of a method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to an embodiment of the present invention.
Fig. 2 is a schematic diagram of a method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to an embodiment of the present invention.
Fig. 3 is a diagram of a model of an encoder and a decoder according to an embodiment of the present invention.
Fig. 4 is a schematic diagram of a verification data experiment platform provided in the embodiment of the present invention.
Fig. 5 is a schematic diagram of a predicted value of the residual life of a bearing according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Example 1:
as shown in fig. 1, a method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to an embodiment of the present invention includes:
s101: selecting rolling bearing vibration signals as a training set and a test set, obtaining frequency domain amplitude signals through fast Fourier transform, and carrying out normalization processing;
s102: taking the frequency domain amplitude signal of the training set after normalization processing as the input of a convolutional neural network, extracting the local information of the vibration signal of the rolling bearing and excavating deep features, taking the deep features as the input of an encoder and a decoder network, constructing a trend quantization health index, and establishing a trend quantization health index model; mining deep features of the frequency domain amplitude signals of the normalized test set through a convolutional neural network, inputting a trend quantitative health index model, and obtaining a trend quantitative health index;
s103: and performing linear regression on the trend quantitative health index of the test set and the life percentage of the test set to obtain the residual life of the rolling bearing.
According to the method, a frequency domain amplitude signal obtained by Fast Fourier Transform (FFT) is subjected to normalization processing and then used as the input of a convolutional layer of a convolutional neural network, deep features are mined, and the problem that a traditional feature extraction method depends on expert experience too much is avoided. The invention adopts the encoder and decoder network of the recurrent neural network based on the attention mechanism to construct the trend quantitative health index, thereby further predicting the residual service life of the rolling bearing and being suitable for predicting the residual service life of the rolling bearing.
Aiming at the problem that the rolling bearing whole life signal sequence is too long to be beneficial to training of the convolutional neural network, the invention provides the method for compressing the information sequence in the time dimension by using the convolutional layer in the convolutional neural network, improves the capacity of the convolutional neural network for processing the long sequence by using an attention mechanism, and finally realizes the prediction of the residual life of the rolling bearing.
Example 2:
as shown in fig. 2, on the basis of embodiment 1, the method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to an embodiment of the present invention specifically includes:
the method comprises the steps of firstly, selecting partial data of a rolling bearing vibration signal under different working conditions as a training set, and performing Fast Fourier Transform (FFT) on an original vibration signal of the training set to obtain a frequency domain amplitude signal;
second, normalize the frequency domain amplitude signal and input it as a feature, and output the life percentage as a training model, for example
Figure 419411DEST_PATH_IMAGE001
In which
Figure 516680DEST_PATH_IMAGE002
Representing N-dimensional characteristic input of a rolling bearing at time t under a certain working condition, N =2048 t ∈[0,1]Indicating the percentage of life degradation output of the rolling bearing at time t; d tra Representing vibration signal data of a rolling bearing under a certain working condition in a training set, wherein R is a frequency domain amplitude characteristic matrix; t is the service life running time of the rolling bearing;
thirdly, the frequency domain amplitude signal after the normalization processing is used as the input of a convolutional neural network CNN, and the convolutional layer in the convolutional neural network CNN is used for traversing the whole input data sequence so as to extract the local information of the vibration signal and mine deep features;
fourthly, inputting the deep features extracted from the convolutional layer into an encoder and decoder network of a recurrent neural network based on an attention mechanism, wherein the encoder is a layer of bidirectional long-short-term memory network BiLSTM, and the decoder is a layer of long-short-term memory network LSTM and the attention mechanism; constructing a trend quantitative health index and establishing a trend quantitative health index model by utilizing the advantages of a long-term and short-term memory network LSTM of a recurrent neural network on time series data and the attention of an encoder and decoder network based on an attention mechanism on important characteristics;
fifthly, fast Fourier Transform (FFT) is carried out on non-whole-life time domain vibration signals under different working conditions in the test set to obtain frequency domain amplitude signals, normalization processing is carried out, deep features mined in the third step are combined with the trend quantitative health index model in the fourth step, and trend quantitative health indexes of the test set are obtained;
and sixthly, performing linear regression by using a least square method according to the trend quantization health index of the training set and the life percentage of the test set, thereby obtaining the residual life of the rolling bearing.
The method comprises the steps of carrying out frequency domain pretreatment on all rolling bearing original time domain vibration signals of 17 groups in total of a training set and a test set by using Fast Fourier Transform (FFT), converting the signals into frequency domain amplitude signals by using the Fast Fourier Transform (FFT), and selecting the first half of the frequency domain amplitude signals as input. At the moment, the frequency domain amplitude signal obtained after the frequency domain amplitude transformation can more visually represent the fault characteristics, can better reflect fault information and is more beneficial to the extraction and identification of the characteristics.
According to the method, the convolutional layer is used for automatically extracting data local abstract information to mine deep features, and the problem that the traditional feature extraction method is too dependent on expert experience is solved. Aiming at the problem that the rolling bearing whole life signal sequence is too long to be beneficial to training of the neural network, the invention provides the method for compressing the information sequence on the time dimension by using the convolution layer in the neural network, and the method for predicting the residual life of the rolling bearing by using the attention mechanism to improve the capacity of the circulating neural network for processing the long sequence. The invention relates to a pure data-driven rolling bearing residual service life prediction method based on a small amount of priori knowledge.
According to the method for predicting the RUL of the rolling bearing by the trend quantization health index based on the encoder and the decoder of the attention system, the frequency domain amplitude signal obtained by Fast Fourier Transform (FFT) is subjected to normalization processing and then used as the input of the convolution layer, deep features are mined, and the problem that the traditional feature extraction method excessively depends on expert experience is solved; and then, an encoder and decoder network based on an attention mechanism is adopted to construct a trend quantization health index, so that the residual service life of the rolling bearing is further predicted, and the method is suitable for predicting the residual service life of the rolling bearing.
Example 3:
on the basis of the embodiment 2, in the third step provided by the embodiment of the invention, the data local abstract information is automatically extracted by using the convolution operation, the local link and the weight sharing characteristics of the convolutional neural network CNN to mine deep features.
According to the method, the convolutional layer is used for automatically extracting data local abstract information to mine deep features, and the problem that the traditional feature extraction method is too dependent on expert experience is solved.
Example 4:
as shown in fig. 3, on the basis of embodiment 2, the frequency domain amplitude signal after normalization processing is used as an input of a convolutional neural network CNN, and a convolutional layer in the convolutional neural network CNN is used to traverse the whole input data sequence to extract local information of a vibration signal and mine deep features, which specifically includes:
inputting the frequency domain amplitude signal after the rolling bearing normalization processing into a convolution layer, wherein the specific convolution layer operation is as shown in formula (1):
Figure 349507DEST_PATH_IMAGE003
(1)
in the formula:
Figure 257551DEST_PATH_IMAGE004
is the ith convolution kernel of the l layer
Figure 441408DEST_PATH_IMAGE005
The weight value of each of the plurality of the weight values,
Figure 76789DEST_PATH_IMAGE006
for the jth convolved local region r in the ith layer, representing the convolution operation, and W being the convolution kernel width, the activation function of the modified linear unit (ReLU) is used to output for each convolution
Figure 780434DEST_PATH_IMAGE007
Performing nonlinear transformation, specifically expressed as formula (2):
Figure 374226DEST_PATH_IMAGE008
(2)
in the formula:
Figure 514220DEST_PATH_IMAGE007
f (-) is the activation function modified linear unit for convolutional layer output values,
Figure 512481DEST_PATH_IMAGE009
is composed of
Figure 585479DEST_PATH_IMAGE007
The activation value obtained by the modified linear unit ReLU activation function.
The method for mining deep features adopts convolution operation in an artificial neural network, which is an operation for effectively extracting local information and compressing input data; convolution operation is carried out on the time dimension of the characteristic, through training, on one hand, the turning point of the rolling bearing entering the damage stage from the health stage can be effectively found out from the local continuous characteristic parameters, and on the other hand, the length of the characteristic parameters on the time dimension can be greatly compressed.
Example 5:
on the basis of embodiment 2, a concrete implementation process of the encoder and decoder network in the fourth step provided by the embodiment of the present invention is as follows:
the encoder is responsible for extracting context information c of the input sequence, and the decoder calculates the output sequence one by one according to the context information c and the data of the previous sequence. Output y for a certain sequence of positions t t Can be expressed by equation (3):
Figure 835326DEST_PATH_IMAGE010
(3)
wherein h is t Information representing the hidden layer at time t, g (\8729;) is an activation function; for the output, the output y at time t t Is the hidden layer information h at that time t Last moment output y t-1 And a function of the input context information c; wherein the output y from the last moment t-1 Ensures the output of the current momenty t And y t-1 And the context information c of the input data guarantees y t The accuracy of (2).
The method comprises the steps of constructing an encoder by using a bidirectional long-time and short-time memory network, inputting the features extracted by a convolutional neural network, and outputting the features as a hidden layer; a long-time memory network and a short-time attention mechanism are used as a decoder, a hidden layer is input, and the remaining service life of the rolling bearing of the test set is output. For the characteristic parameters of the whole life cycle of the rolling bearing, the analysis of various characteristics shows that whether the data driving algorithm can learn the global variation trend of the characteristics is the key for accurately predicting the residual life of the rolling bearing, and the local variation trend of the characteristics has a large amount of noise instead. On the other hand, after the convolution layer is compressed, the length is reduced but still longer. Therefore, a bidirectional LSTM network is proposed to be used as an encoder, so that the output of the encoder is ensured to contain data information transmitted from the positive direction and the negative direction, and the global communication of the information in the time dimension is further improved. And finally, a single-layer LSTM network with an attention gate is used as a decoder, so that effective information is purposefully extracted while global information is reviewed, and the health state index of the rolling bearing is accurately predicted.
The invention utilizes the residual service life prediction method of the rolling bearing of the encoder and the decoder based on the attention mechanism, and uses the bidirectional long-time memory network BilSTM as the encoder, thereby ensuring that the output of the encoder comprises data information transmitted from the positive direction and the negative direction, and further improving the global communication of the information on the time dimension; and finally, a single-layer LSTM network with an attention gate is used as a decoder, so that effective information is purposefully extracted while global information is reviewed, and the health state index of the rolling bearing is accurately predicted.
Example 6:
on the basis of embodiment 2, a specific implementation process of the fourth attention mechanism provided by the embodiment of the present invention is as follows: the attention-drawing mechanism is actually equivalent to adding a gate a before the global information c participates in the output calculation, and the gate a will input data x according to the time t t Determining the output y at the time of calculating t t When it comes to which part of the global information c is more important, it tells the neural network that the output y is being calculated t Which part of the global information is most interesting. Thereby making more efficient use of the global information c in computing the output. The specific formula of the attention mechanism is as follows:
Figure 462616DEST_PATH_IMAGE011
in the formula, h t Is the output variable of the decoder at time t,
Figure 705379DEST_PATH_IMAGE012
it is the output variable of the encoder at time s,
Figure 367304DEST_PATH_IMAGE013
the importance of the input information at the time s to the output at the time t is calculated in some way, and finally normalized by the softmax function as the attention gate a at the time t t This attention gate a t And encoder output
Figure 788052DEST_PATH_IMAGE014
Performing dot product operation, allowing the information with the most important correlation with the output calculation at the time t to flow in, and blocking the flow of unimportant information; wherein the calculation is carried out
Figure 168218DEST_PATH_IMAGE013
There are three ways, W a And v a Are trainable parameters in the network.
The invention adopts the encoder and decoder network of the recurrent neural network based on the attention mechanism to construct the trend quantitative health index, further predicts the residual service life of the rolling bearing, and is suitable for predicting the residual service life of the rolling bearing. The encoder and decoder based on the attention mechanism can also use BiGRU as the encoder and GRU and attention mechanism as the decoder to predict the residual life of the rolling bearing.
Because the final purpose of the residual life of the rolling bearing is to realize accurate residual life measurement in the early stage of the degradation stage of the rolling bearing, in order to realize the point of the neural network, the training data is intercepted for a time length, and a section of data including a normal stage and a degradation stage in the whole life cycle of the rolling bearing is randomly selected as input. And then, taking the energy value characteristic sequence of the sub-frequency spectrum band extracted from the length of the rolling bearing vibration signal as the input of the neural network. Therefore, the method is a Data enhancement (Data enhancement) measure to some extent, can reduce the possibility of the over-fitting phenomenon of the neural network in the training process, and improves the prediction effect of the neural network on the test Data. Considering the degradation law of rolling bearings, i.e. the health phase takes the most part of the life cycle of a rolling bearing, while the degradation and damage phases only take a small part of the final time. Therefore, when the vibration signal with random length is selected, the random range selected at the starting point is set to be within the range of 0% to 50% of the whole life cycle of the rolling bearing, and the random range selected at the finishing point is set to be within the range of 75% to 100% of the whole life cycle of the rolling bearing.
Example 7:
on the basis of the embodiment 5, the specific implementation process of the long-short time memory network LSTM and the bidirectional long-short time memory network BiLSTM provided by the embodiment of the present invention is as follows: compared with the RNN, the long-term memory network LSTM has newly added cell states and 3-gate structures. Forgetting door f in long-time memory network LSTM t The information that determines what proportion is retained in the network is calculated as:
Figure 949092DEST_PATH_IMAGE015
（7）
in the formula: x is a radical of a fluorine atom t Is an input sequence; h is t-1 The state memory quantity at the last moment;
Figure 481836DEST_PATH_IMAGE016
activating a function for sigmoid; w f Weight for forgetting gateA matrix; b t A bias for a forgetting gate; f. of t The state of forgetting the door;
and an input gate i t Selectively memorizing new information in the cell state, and the calculation formula is as follows:
Figure 588332DEST_PATH_IMAGE017
in the formula:
Figure 190215DEST_PATH_IMAGE018
candidate values for cell status; c t A new cell state; tan h (·) is a hyperbolic tangent function; w i Is the weight matrix of the input gate; w c A weight matrix that is a state of the cell; b i Is the bias of the input gate; b c A bias that is a cellular state; i.e. i t The status of the gate is entered.
Output gate o t The currently output information is determined, and the calculation formula is as follows:
Figure 791092DEST_PATH_IMAGE019
in the formula: w o Is a weight matrix of the output gate; b o Is the offset of the output gate; o t The state of the output gate.
The mathematical transformation formula of the bidirectional long-short time memory network BilSTM comprises the following steps:
Figure 162030DEST_PATH_IMAGE020
wherein: h is t 、
Figure 173849DEST_PATH_IMAGE021
Respectively outputting a front propagation layer and a rear propagation layer at the time t; w is a group of 1 、W 3 Weight matrices, W, for the input layer and the backward propagation layer, respectively 3 、W 5 The weights from the forward propagation layer and the backward propagation layer to the self propagation layer respectivelyA matrix; w 4 、W 6 The weight matrixes from the forward propagation layer to the output layer and the backward propagation layer to the output layer are respectively set; o is t Is the output value of the final output gate; the function f (-) represents the cell calculation process; the function g (-) represents the function for stitching the forward and backward propagation results.
Example 8:
on the basis of the embodiment 2, the trend quantitative health index Hi in the fifth step provided by the embodiment of the invention is a customized index, and the real-time estimation of the health state of the rolling bearing is realized. Trending quantification of the health indicator Hi is an important tool in residual useful life RUL prediction. To reduce this manual and empirical process, or to reduce the need for expensive a priori expert knowledge, the remaining useful life RUL prediction presents a simple trending quantitative health indicator Hi. Formally, for the existing rolling bearing degradation signal, the trending quantitative health indicator Hi at time t can be used with the respective RUL t Divided by the initial time RUL o ：
Figure 13760DEST_PATH_IMAGE022
（15）
Obtain the data located in the interval [0, 1]]The Hi value sequence in (1) has the failure point that when Hi reaches 0 t To 0). Thus, hi can more easily predict, regress, and calculate the remaining useful life RUL.
The invention provides an encoder and decoder model of a recurrent neural network with attention mechanism in order to mine the relationship between the characteristics and the Hi value and utilize the relationship. After the training of the neural network model is completed, a series of characteristics are input into the trained neural network model, the neural network outputs a series of rolling bearing health state values Hi, and finally linear regression is carried out according to the series of rolling bearing health state values, so that the predicted value of the residual life of the rolling bearing is calculated.
Example 9:
on the basis of embodiment 1, obtaining a vibration signal of a rolling bearing as a training set and a test set according to an embodiment of the present invention, and obtaining a frequency domain amplitude signal through fast fourier transform specifically includes:
(1) Reconfiguring Fourier transform points by configuring a port configuration channel according to the points of the collected vibration signals of the rolling bearing;
(2) After receiving a signal for starting the fast Fourier transform, taking out interference data from the data storage, converting the interference data into a frequency domain amplitude signal and inputting the frequency domain amplitude signal into the programmable logic device;
(3) Storing the frequency domain amplitude signal output by the programmable logic device into data storage; and sending a conversion completion signal for storage after the conversion is completed.
According to the invention, the dynamic bearing vibration signal is converted into the frequency domain amplitude signal by adopting the fast Fourier transform, the programmable logic device can reduce the development cost and period, reduce the occupation of internal logic resources and improve the reliability of the fast Fourier transform; the frequency domain amplitude signal obtained after the frequency domain amplitude transformation can more visually represent the fault characteristics, can reflect fault information and is more favorable for characteristic extraction and identification.
Example 10:
based on example 1, the present invention performs validation analysis on a model using a data set from a data challenge of PHM 2012. According to the division of the PHM2012 data set, the training set is the full life data of the bearings 1, 2, 3, 1, 3, and the test set is the non-full life data of the rest of the bearings. The detailed partitioning results are shown in table 1:
TABLE 1 Experimental data (PHM 2012 bearing data set)
Figure 402016DEST_PATH_IMAGE023
And performing linear regression by using a least square method according to the trend quantization health index of the test set and the life percentage of the test set, thereby obtaining the residual life of the rolling bearing.
The method specifically comprises the following steps: the predicted value of the residual life of the bearing is calculated by performing linear regression according to the series of trend quantitative health index values of the bearing, and the prediction result of the bearing 2_7 is shown in fig. 5 by taking the bearing 2_7 as an example.
As can be seen from table 1, the total life data of the bearing 2_7 was 230 groups, the non-total life data was 172 groups, and the intersection point of the predicted fitted straight line and the X axis (i.e., P =0 bearing complete failure) was 228.3 groups. Since the data sampling interval is 10 s per set, the actual remaining lifetime is (230-172) × 10=580 s, and the predicted value of the remaining lifetime is (228.3-172) × 10=563 s.
The trend quantized health index of the test set and the life percentage of the test set are adopted to carry out linear regression, so that the residual life of the rolling bearing is obtained, the obtained residual life of the rolling bearing is more accurate and effective, the linear regression is simple to realize, the operation efficiency is high, the trend quantized health index is constructed, the residual service life of the rolling bearing is further predicted, the method is suitable for predicting the residual life of the rolling bearing, and important data is provided for production of enterprises.
Example 11:
on the basis of the embodiments 1 to 10, the principle of the verification data experiment platform provided by the embodiment of the invention is as follows: selecting rolling bearing vibration signals as a training set and a test set, obtaining frequency domain amplitude signals through fast Fourier transform, and carrying out normalization processing; taking the frequency domain amplitude signal of the training set after the normalization processing as the input of a convolutional neural network, extracting the local information of the vibration signal of the rolling bearing and excavating deep features, taking the deep features as the input of an encoder and a decoder network, constructing a trend quantitative health index, and establishing a trend quantitative health index model; and (3) excavating deep features of the frequency domain amplitude signals of the normalized test set through a convolutional neural network, inputting a trend quantitative health index model, acquiring a trend quantitative health index, acquiring the trend quantitative health index of the test set, and performing linear regression to obtain the residual life of the rolling bearing.
According to the method, a frequency domain amplitude signal obtained by Fast Fourier Transform (FFT) is subjected to normalization processing and then used as the input of a convolutional layer of a convolutional neural network, deep features are mined, and the problem that a traditional feature extraction method depends on expert experience too much is avoided. The invention adopts the encoder and decoder network of the recurrent neural network based on the attention mechanism to construct the trend quantitative health index, thereby further predicting the residual service life of the rolling bearing and being suitable for predicting the residual service life of the rolling bearing.
It should be noted that the embodiments of the present invention can be realized by hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. Those skilled in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such code being provided on a carrier medium such as a disk, CD-or DVD-ROM, programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier, for example. The apparatus and its modules of the present invention may be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of hardware circuits and software, e.g., firmware.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention, and the scope of the present invention should not be limited thereto, and any modifications, equivalents and improvements made by those skilled in the art within the technical scope of the present invention as disclosed in the present invention should be covered thereby.

Claims (10)

1. A method for predicting the remaining service life of a rolling bearing using an encoder and a decoder, comprising:
selecting rolling bearing vibration signals as a training set and a test set, obtaining frequency domain amplitude signals through fast Fourier transform, and carrying out normalization processing;
taking the frequency domain amplitude signal of the training set after normalization processing as the input of a convolutional neural network, extracting the local information of a rolling bearing vibration signal and excavating deep features, taking the deep features as the input of an encoder and a decoder network, and inputting the deep features extracted by the convolutional layer into the encoder and decoder network of a cyclic neural network based on an attention mechanism, wherein the encoder is a layer of bidirectional long-short time memory network, and the decoder is a layer of long-short time memory network and an attention mechanism; constructing a trend quantitative health index and establishing a trend quantitative health index model by utilizing the advantages of a long-time memory network and a short-time memory network of a recurrent neural network on time series data and the attention of an encoder network and a decoder network based on an attention mechanism on important characteristics; mining deep features of the frequency domain amplitude signals of the normalized test set through a convolutional neural network, inputting a trend quantitative health index model, and obtaining trend quantitative health indexes of the test set;
and performing linear regression on the trend quantitative health index of the test set and the life percentage of the test set to obtain the residual life of the rolling bearing.
2. The method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to claim 1, comprising:
the method comprises the steps of firstly, selecting partial data of a rolling bearing vibration signal under different working conditions as a training set, and performing fast Fourier transform on an original vibration signal of the training set to obtain a frequency domain amplitude signal;
secondly, normalizing the frequency domain amplitude signal to be used as characteristic input, and using the service life percentage as an output training model;
thirdly, the frequency domain amplitude signal after the normalization processing is used as the input of a convolutional neural network, and the convolutional layer in the convolutional neural network is used for traversing the whole input data sequence so as to extract the local information of the vibration signal and mine deep features;
inputting the deep features extracted from the convolutional layer into an encoder and decoder network of a recurrent neural network based on an attention mechanism, wherein the encoder is a layer of bidirectional long-short term memory network, and the decoder is a layer of long-short term memory network and an attention mechanism; constructing a trend quantitative health index and establishing a trend quantitative health index model by utilizing the advantages of a long-time memory network and a short-time memory network of a recurrent neural network on time series data and the attention of an encoder network and a decoder network based on an attention mechanism on important characteristics;
fifthly, fast Fourier Transform (FFT) is carried out on non-whole-life time domain vibration signals under different working conditions in the test set to obtain frequency domain amplitude signals, normalization processing is carried out, deep features mined in the third step are combined with the trend quantitative health index model in the fourth step, and trend quantitative health indexes of the test set are obtained;
and sixthly, performing linear regression by using a least square method according to the trend quantitative health index of the test set and the service life percentage of the test set, thereby obtaining the residual service life of the rolling bearing.
3. The method for predicting the remaining useful life of a rolling bearing using an encoder and a decoder according to claim 2, wherein the second step specifically comprises:
Figure FDA0004011430920000021
wherein x is t ∈R N×1 Representing N-dimensional characteristic input of a rolling bearing at time t under a certain working condition, N =2048 t ∈[0,1]Indicating the percentage of life degradation output of the rolling bearing at time t; d tra Representing vibration signal data of a rolling bearing under a certain working condition in a training set, wherein R is a frequency domain amplitude characteristic matrix; and T is the service life running time of the rolling bearing.
4. The method for predicting the remaining useful life of a rolling bearing using an encoder and a decoder according to claim 2, wherein the third step automatically extracts data local abstract information to mine deep features using convolution operations, local links and weight sharing characteristics of a convolutional neural network.
5. The method for predicting the remaining useful life of a rolling bearing using an encoder and a decoder according to claim 2, wherein the third step specifically comprises:
inputting the frequency domain amplitude signal after the rolling bearing normalization processing into a convolution layer, wherein the specific convolution layer operation is as shown in formula (1):
Figure FDA0004011430920000022
in the formula:
Figure FDA0004011430920000023
is the jth 'weight of the ith convolution kernel at level l, <' >>
Figure FDA0004011430920000024
For the j th convolved local region r in the l th layer, representing convolution operation, W is the width of convolution kernel, and y is output to each convolution by using modified linear unit activation function l(i,j) Performing nonlinear transformation, specifically expressed as formula (2):
a l(i,j) ＝f(y l(i,j) )＝max{0,y l(i,j) } (2)
in the formula: y is l(i,j) For convolution layer output values, f (-) is an activation function modified linear unit, a l(i,j) Is y l(i,j) And (4) an activation value obtained by modifying the linear unit activation function.
6. The method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to claim 2, wherein the network of the encoder and the decoder in the fourth step is implemented by:
the encoder is responsible for extracting the context information c of the input sequence, and the decoder calculates the output sequences one by one according to the context information c and the data of the previous sequence; to pairOutput y at a sequence of positions t t Expressed by equation (3):
p(y t |y t-1 ,y t-2 ,...,y 1 ,c)＝g(h t ,y t-1 ,c) (3)
wherein h is t Information representing the hidden layer at time t, g (-) being an activation function; for the output, the output y at time t t Is the hidden layer information h at that moment t Last moment output y t-1 And a function of the input context information c.
7. The method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to claim 2, wherein the attention mechanism of the fourth step is implemented by: the attention mechanism is introduced to add a gate a before the global information c participates in the output calculation, and the gate a will input data x according to the time t t Determining the output y at the time of calculating t t When it comes to which part of the global information c is more important, it tells the neural network that the output y is being calculated t When the global information is the most concerned, the global information is the most concerned; global information c is used more effectively when calculating output; the specific formula of the attention mechanism is as follows:
Figure FDA0004011430920000031
Figure FDA0004011430920000032
Figure FDA0004011430920000033
in the formula, h t Is the output variable of the decoder at time t,
Figure FDA0004011430920000034
it is the output variable of the encoder at time s,
Figure FDA0004011430920000035
the importance of the input information at the time s to the output at the time t is calculated in some way, and finally normalized by the softmax function as the attention gate a at the time t t This attention gate a t And the encoder output->
Figure FDA0004011430920000041
Performing dot product operation, allowing the information with the most important correlation with the output calculation at the time t to flow in, and blocking the flow of unimportant information; wherein it is calculated->
Figure FDA0004011430920000042
There are three ways, W a And v a Are trainable parameters in the network.
8. The method for predicting the remaining service life of a rolling bearing using an encoder and a decoder according to claim 2, wherein the specific implementation process of the long-time and short-time memory network and the bidirectional long-time and short-time memory network in the fourth step is as follows: forgetting door f in long-and-short-term memory network t The information that determines what proportion is retained in the network is calculated as:
f t ＝σ(W f ·[h t-1 ,x t ]+b t ) (7)
in the formula: x is the number of t Is an input sequence; h is t-1 The state memory quantity at the last moment; σ (-) is a sigmoid activation function; w is a group of f A weight matrix for a forgetting gate; b is a mixture of t A bias for a forgetting gate; f. of t The state of forgetting the door;
and an input gate i t Selectively memorizing new information in the cell state, and the calculation formula is as follows:
i t ＝σ(W i ·[h t-1 ,x t ]+b i ) (8)
Figure FDA0004011430920000043
in the formula:
Figure FDA0004011430920000044
candidate values for cell status; c t A new cell state; tan h (·) is a hyperbolic tangent function; w is a group of i Is the weight matrix of the input gate; w is a group of c A weight matrix that is a state of the cell; b is a mixture of i Is the bias of the input gate; b c A bias that is a cellular state; i all right angle t Is the status of the input gate;
output gate o t The currently output information is determined, and the calculation formula is as follows:
o t ＝σ(W o ·[h t-1 ,x t ]+b o ) (10)
h t ＝o t *tanh(C t ) (11)
in the formula: w o Is a weight matrix of the output gate; b o Is the offset of the output gate; o. o t Is the state of the output gate;
the mathematical transformation formula of the bidirectional long-time and short-time memory network comprises the following steps:
h t ＝f(w 1 x t +w 2 h t-1 ) (12)
h′ t ＝f(w 3 x t +w 5 h′ t+1 ) (13)
O t ＝g(w 4 h t +w 6 h′ t ) (14)
wherein: h is t 、h′ t Respectively outputting a front propagation layer, a back propagation layer and a back propagation layer at the time t; w is a 1 、w 3 Weight matrices, w, for the input layer and the backward propagation layer, respectively 3 、w 5 Respectively are weight matrixes from a forward propagation layer and a backward propagation layer to a self propagation layer; w is a 4 、w 6 The weight matrixes from the forward propagation layer to the output layer and from the backward propagation layer to the output layer are respectively; o is t Is the most importantThe output value of the final output gate; the function f (-) represents the cell calculation process; the function g (-) represents the function for splicing the forward and backward propagation results.
9. The method of claim 2, wherein the trend-based quantized health indicator Hi of the fifth step is a customized indicator for estimating the health status of the rolling bearing in real time, and the remaining life prediction provides a trend-based quantized health indicator Hi, and the trend-based quantized health indicators Hi at time t are each implemented using a respective RUL t Divided by the initial time RUL 0 ：
Figure FDA0004011430920000051
Obtain the data located in the interval [0, 1]]The Hi value sequence in (1), the failure point is when Hi reaches 0,LUL t And reaches 0.
10. The method for predicting the remaining useful life of a rolling bearing using an encoder and a decoder according to claim 2, wherein the first step specifically comprises:
(1) Reconfiguring Fourier transform points by configuring a port configuration channel according to the points of the collected vibration signals of the rolling bearing;
(2) After receiving a signal for starting the fast Fourier transform, taking out interference data from the data storage and converting the interference data into a frequency domain amplitude signal to input into a programmable logic device;
(3) Storing the frequency domain amplitude signal output by the programmable logic device into data storage; and after the conversion is finished, sending a conversion finished signal for storage.
CN202210487073.2A 2022-05-06 2022-05-06 Method for predicting residual service life of rolling bearing by adopting encoder and decoder Active CN114925723B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210487073.2A CN114925723B (en) 2022-05-06 2022-05-06 Method for predicting residual service life of rolling bearing by adopting encoder and decoder

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210487073.2A CN114925723B (en) 2022-05-06 2022-05-06 Method for predicting residual service life of rolling bearing by adopting encoder and decoder

Publications (2)

Publication Number Publication Date
CN114925723A CN114925723A (en) 2022-08-19
CN114925723B true CN114925723B (en) 2023-04-07

Family

ID=82807465

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210487073.2A Active CN114925723B (en) 2022-05-06 2022-05-06 Method for predicting residual service life of rolling bearing by adopting encoder and decoder

Country Status (1)

Country Link
CN (1) CN114925723B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115660198B (en) * 2022-11-07 2023-08-01 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 (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112101220A (en) * 2020-09-15 2020-12-18 哈尔滨理工大学 Rolling bearing service life prediction method based on unsupervised model parameter migration

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109726524B (en) * 2019-03-01 2022-11-01 哈尔滨理工大学 CNN and LSTM-based rolling bearing residual service life prediction method
US11276410B2 (en) * 2019-09-13 2022-03-15 Microsoft Technology Licensing, Llc Convolutional neural network with phonetic attention for speaker verification
CN110633792B (en) * 2019-10-22 2022-03-22 西安交通大学 End-to-end bearing health index construction method based on convolution cyclic neural network
CN110912598B (en) * 2019-11-22 2020-08-28 中原工学院 Large-scale MIMO system CSI feedback method based on long-time attention mechanism
CN114420105A (en) * 2020-10-13 2022-04-29 腾讯科技（深圳）有限公司 Training method and device of voice recognition model, server and storage medium
CN113139278A (en) * 2021-04-08 2021-07-20 上海工程技术大学 Deep learning-based bearing residual life prediction method and system
CN113465923B (en) * 2021-06-08 2023-04-11 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
CN113642414A (en) * 2021-07-19 2021-11-12 北京航空航天大学 Method for predicting residual service life of rolling bearing based on Transformer model
CN113947017A (en) * 2021-10-18 2022-01-18 北京化工大学 Method for predicting residual service life of rolling bearing

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112101220A (en) * 2020-09-15 2020-12-18 哈尔滨理工大学 Rolling bearing service life prediction method based on unsupervised model parameter migration

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于MSCNN-LSTM的滚动轴承剩余寿命预测方法;胡城豪等;《中国测试》;20200930(第09期);106-113 *

Also Published As

Publication number Publication date
CN114925723A (en) 2022-08-19

Similar Documents

Publication Publication Date Title
CN109726524B (en) CNN and LSTM-based rolling bearing residual service life prediction method
Zhao et al. A recurrent neural network approach for remaining useful life prediction utilizing a novel trend features construction method
Wu et al. Data-driven remaining useful life prediction via multiple sensor signals and deep long short-term memory neural network
Su et al. An end-to-end framework for remaining useful life prediction of rolling bearing based on feature pre-extraction mechanism and deep adaptive transformer model
Rai et al. An integrated approach to bearing prognostics based on EEMD-multi feature extraction, Gaussian mixture models and Jensen-Rényi divergence
Chen et al. Mechanical state prediction based on LSTM neural netwok
Chang et al. Remaining useful life prediction for rolling bearings using multi-layer grid search and LSTM
CN112986827A (en) Fuel cell residual life prediction method based on deep learning
Zhao et al. Probabilistic remaining useful life prediction based on deep convolutional neural network
CN114925723B (en) Method for predicting residual service life of rolling bearing by adopting encoder and decoder
CN113723007A (en) Mechanical equipment residual life prediction method based on DRSN and sparrow search optimization BilSTM
CN111914897A (en) Fault diagnosis method based on twin long-short time memory network
Wang et al. A remaining useful life prediction model based on hybrid long-short sequences for engines
Kargah-Ostadi Comparison of machine learning techniques for developing performance prediction models
CN114266201B (en) Self-attention elevator trapping prediction method based on deep learning
Tang et al. Prediction of bearing performance degradation with bottleneck feature based on LSTM network
Deng et al. Remaining useful life prediction of machinery: A new multiscale temporal convolutional network framework
Chang et al. Temporal convolution-based sorting feature repeat-explore network combining with multi-band information for remaining useful life estimation of equipment
Shi et al. A dual attention LSTM lightweight model based on exponential smoothing for remaining useful life prediction
Zhao et al. A two-stage machine-learning-based prognostic approach for bearing remaining useful prediction problem
CN114486262B (en) CNN-AT-LSTM-based bearing residual service life prediction method
CN114861349A (en) Rolling bearing RUL prediction method based on model migration and wiener process
Zhang et al. Remaining useful life prediction of lithium-ion batteries based on TCN-DCN fusion model combined with IRRS filtering
CN113052060B (en) Bearing residual life prediction method and device based on data enhancement and electronic equipment
CN112633550B (en) RNN-based contact network fault trend prediction method, equipment and storage medium

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
GR01 Patent grant
GR01 Patent grant