CN114861979A - Rolling bearing residual life prediction method based on LSTM and TDNN - Google Patents

Rolling bearing residual life prediction method based on LSTM and TDNN Download PDF

Info

Publication number
CN114861979A
CN114861979A CN202210349547.7A CN202210349547A CN114861979A CN 114861979 A CN114861979 A CN 114861979A CN 202210349547 A CN202210349547 A CN 202210349547A CN 114861979 A CN114861979 A CN 114861979A
Authority
CN
China
Prior art keywords
bearing
training
lstm
residual life
tdnn
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.)
Withdrawn
Application number
CN202210349547.7A
Other languages
Chinese (zh)
Inventor
席霄鹏
张汇鑫
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.)
Shandong University of Science and Technology
Original Assignee
Shandong University of Science and Technology
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 Shandong University of Science and Technology filed Critical Shandong University of Science and Technology
Priority to CN202210349547.7A priority Critical patent/CN114861979A/en
Publication of CN114861979A publication Critical patent/CN114861979A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • 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
    • 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
    • G06N3/08—Learning methods
    • 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
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Business, Economics & Management (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a rolling bearing residual life prediction method based on LSTM and TDNN, which comprises the following steps: (1) extracting time domain characteristics of an original vibration signal of a training bearing to form a time domain characteristic set, and performing minimum-maximum normalized processing on each characteristic; (2) constructing a health factor by using at least two nonlinear functions, then constructing a mapping relation between training characteristics and the health factor and between the health factor and the residual life percentage by sequentially using LSTM and TDNN, and constructing an LSTM-TDNN residual life prediction model; (3) inputting vibration data of a tested bearing, obtaining a residual life estimated value according to an LSTM-TDNN prediction model after feature extraction and normalization processing, and giving comprehensive evaluation of prediction accuracy. The rolling bearing provided by the invention has more accurate residual life prediction results, realizes high-precision and strong-conservative unification, can be effectively applied to rolling prediction maintenance tasks, and reduces the maintenance cost.

Description

Rolling bearing residual life prediction method based on LSTM and TDNN
Technical Field
The invention relates to the technical field of residual life prediction of rolling bearings, in particular to a residual life prediction method of a rolling bearing based on LSTM and TDNN.
Background
The rotary machine is taken as a core part of modern industrial equipment, and plays an important role in the fields of aerospace, water conservancy and hydropower, chemical metallurgy and the like. The rolling bearing is used as a key part of the rotary machine, affects the running state of the whole mechanical equipment, and the occurrence of a fault may cause that the mechanical equipment cannot realize the set function, so that the industrial production is affected and economic loss is brought, and the life safety of practitioners is endangered. Therefore, the method has important guiding significance for predicting the service life of the rolling bearing.
At present, methods for predicting the residual life of a rolling bearing are mainly divided into two categories, namely an analytical model-based method and a data-driven method, and the data-driven method also comprises a random process-based method and an artificial intelligence-based method. Due to the complexity of the bearing operating environment and the service working condition, an accurate failure mechanism model is difficult to establish; the method based on the random process needs to select the degradation process in advance, and the type of the degradation process directly influences the prediction precision; the artificial intelligence based method has strong data processing capacity, does not need physical mechanism and expert experience, can continuously acquire information training of the existing data and update the network model, and shows huge development prospect in the field of residual life prediction.
The existing rolling bearing residual life intelligent prediction method mostly uses a convolution neural network or a circulation neural network as a main body to complete data preprocessing and residual life estimation, and has the following problems: firstly, neglecting the existence of unequal sequence and time delay, due to the difference of working conditions, installation positions, environments and the like, the actual service time of a bearing is different, the problem of unequal sequence exists when a prediction model is trained, and in addition, the time delay is difficult to quantify when the residual service life at the current moment is estimated by means of historical operation information; secondly, the performance index reflecting the health state of the bearing, namely the health factor acquisition process is complex and the diversity is neglected, and the transportability of the method is poor; thirdly, the evaluation index only evaluates the prediction accuracy, but the specific conservative prediction or aggressive prediction lacks evaluation, and the unification of high accuracy and strong conservation is realized as far as possible in practice.
Disclosure of Invention
The invention aims to provide a rolling bearing residual life prediction method based on LSTM and TDNN, and improve the rolling bearing residual life prediction accuracy.
In order to achieve the purpose, the invention adopts the technical scheme that:
a method for predicting the residual life of a rolling bearing based on LSTM and TDNN comprises the following steps:
(1) collecting a full-life original vibration signal of a training group rolling bearing under a certain working condition as a training set, and collecting a full-life original vibration signal of a testing group rolling bearing under the same working condition as a testing set;
(2) extracting time domain characteristics of a training set bearing, carrying out minimum-maximum normalization processing on each characteristic, and constructing a training characteristic set X l ：
Figure BDA0003579009220000021
Wherein, l represents the number of the bearing of the training set (l is 1,2.. q), m represents the total number of the features, m is 36, n represents the total number of the samples of the bearing of the l training set,
Figure BDA0003579009220000022
an mth normalized feature representing an nth sample of an lth training bearing;
(3) during the training process, p health factors are constructed
Figure BDA0003579009220000023
And combines it with the sampling duration, i.e.:
Figure BDA0003579009220000024
Figure BDA0003579009220000025
normalization of health factors to [0,1]Obtaining a normalized health factor model
Figure BDA0003579009220000026
Figure BDA0003579009220000027
Where t is time, p represents the function type, p ≧ 2, t k The kth sample sampling duration corresponding to the l training set bearing, T being the total sampling duration, θ l (t)∈(0,1]，
Figure BDA0003579009220000028
Is theta of the l-th bearing l (t) and
Figure BDA0003579009220000029
the mapping relation between the two;
(4) establishing an LSTM prediction model, taking q groups of training bearing characteristics as the input of the LSTM model under the p health factors established in the step (3),
Figure BDA00035790092200000210
as output, adjusting the number M of hidden layer network layers 1 And number of nodes Q 1 Training an LSTM model to obtain a mapping relation F of training bearing characteristics and each health factor p l And further obtaining a bearing characteristic-health factor model based on the LSTM, namely:
Figure BDA00035790092200000211
wherein, F p l Is the first training bearing characteristic X l Mapping to p-th health factor, F (X) ═ F (X) 1 )∪F(X 2 )...∪F(X q )；
(5) Establishing a TDNN prediction model, and respectively outputting p training set health factors output by the LSTM model in the step (4)
Figure BDA00035790092200000212
Inputting TDNN model, training set bearing residual life percentage
Figure BDA00035790092200000213
As output, adjust the number of hidden layers M 2 And number of nodes Q 2 Verifying parameters such as the number P of checks, the number d of time delay steps and the like, training a TDNN prediction model to obtain a mapping relation between each health factor and the percentage of the residual life, and further obtaining a TDNN-based health factor-residual life prediction model, namely:
Figure BDA0003579009220000031
wherein,
Figure BDA0003579009220000032
T c is the current elapsed life, T 100 Is the total life, G p Is a mapping relation between the p-th health factor and the percentage of the remaining life;
(6) extracting time domain characteristics of the bearing of the test set, carrying out minimum-maximum normalized processing on each characteristic, and constructing a test characteristic set:
Figure BDA0003579009220000033
where T represents the test set bearing number (T ═ 1), m represents the total number of features, m ═ 36, i represents the total number of samples for the T-th test set bearing,
Figure BDA0003579009220000034
an mth normalized feature representing an ith sample of the Tth set of training bearings;
(7) inputting the bearing characteristics of the test set into the LSTM-based bearing characteristic-health factor model constructed in the step (4), namely outputting p health factor clusters of the bearing of the test set, and respectively calculating the average values of the p health factor clusters
Figure BDA0003579009220000035
And calculating the average value of each health factor
Figure BDA0003579009220000036
Namely:
Figure BDA0003579009220000037
(8) averaging the test set health factors obtained in step (7)
Figure BDA0003579009220000038
Inputting the residual life percentages of the bearings to be tested of p test sets into a TDNN model, evaluating the accuracy of a prediction result by using at least one of an RMSE function and a Score function as an evaluation index, and finding an optimal health factor-residual life prediction model, namely a health factor-residual life prediction model corresponding to the minimum value of the RMSE function or/and the Score function;
(9) and (5) taking the optimal health factor-residual life prediction model determined in the step (8) as a residual life prediction model of the rolling bearing, and calculating the residual life of the bearing to be measured.
Further, the total number n of samples of the ith training set bearing in the step (1) refers to a set of vibration signals collected every 10s to 60s during the accelerated life test of the ith training set bearing.
Further, the time domain features extracted in the step (2) are 18-dimensional time domain features of the full-life-cycle vibration signals of the training set bearing in the horizontal direction and the vertical direction; and (4) extracting time domain characteristics in the step (6) which are 18-dimensional time domain characteristics of the full-life-cycle vibration signals of the bearing of the test set in the horizontal direction and the vertical direction.
Further, the health factors constructed in the step (3) are at least two, preferably 4, health factors constructed based on exponential, power, logarithmic and polynomial nonlinear functions, and the calculation formulas of the health factors constructed based on exponential, power, logarithmic and polynomial nonlinear functions are respectively:
Figure BDA0003579009220000041
HI 2 (t)＝1-θ r
Figure BDA0003579009220000042
Figure BDA0003579009220000043
wherein r is the order of theta, r ∈ N * 。
Further, the calculation formulas of the RMSE function and the Score function in step (8) are respectively:
Figure BDA0003579009220000044
Figure BDA0003579009220000045
where N is the total number of sample points,
Figure BDA0003579009220000046
i.e. the difference between the predicted value of the remaining lifetime and the true value at the z-th sample point.
The invention has the beneficial effects that:
(1) according to the method for predicting the residual life of the rolling bearing, the LSTM neural network and the TDNN neural network are combined, an LSTM model is built for each group of training bearings under a certain working condition, the problem of inconsistent sequence is solved, the TDNN neural network is combined to obtain the mapping relation between the health factor and the residual life percentage, in the process, the TDNN neural network is used for further utilizing historical information, time delay existing in the prediction process is quantized, and the TDNN only needs past data, so that the advance prediction is realized;
(2) the method considers the influence of factors such as the operation condition, the environment, the abrasion degree and the like of the bearing on the diversity of the health factors, constructs various health factors which are in a nonlinear relation with time, establishes an LSTM-TDNN model based on different types of health factors, evaluates the prediction accuracy by adopting RMSE and Score evaluation indexes, ensures the prediction precision, realizes conservative prediction, is more beneficial to taking maintenance measures in advance for the rolling bearing, and avoids sudden failure of the rolling bearing.
Drawings
FIG. 1 is a flowchart of a method for predicting the remaining life of a rolling bearing according to the present invention;
FIG. 2 is four time domain characteristic diagrams of a bearing1-1 horizontal direction full life cycle vibration signal;
FIG. 3 shows the condition of the bearings 1-3 in the health factor of
Figure BDA0003579009220000047
An image of the remaining life percentage prediction result;
FIG. 4 shows the condition of the bearings 1-3 in the health factor of
Figure BDA0003579009220000051
An image of the remaining life percentage prediction result;
FIG. 5 shows the condition of the bearings 1-3 in the health factor of
Figure BDA0003579009220000052
An image of the remaining life percentage prediction result;
FIG. 6 shows the condition of the bearings 1-3 in the health factor of
Figure BDA0003579009220000053
An image of the remaining life percentage prediction result;
FIG. 7 shows the bearings 1-3 in
Figure BDA0003579009220000054
And comparing the obtained residual life percentage prediction accuracy with the obtained graph.
Detailed Description
The invention provides a method for predicting the residual life of a rolling bearing based on LSTM and TDNN, which is further described in detail below in order to make the purpose, technical scheme and effect of the invention clearer and clearer. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Referring to fig. 1, the invention provides a rolling bearing residual life prediction method based on LSTM and TDNN, comprising the following steps:
(1) under a certain working condition, selecting original vibration signals of a plurality of rolling bearings in a full life cycle as a training set, and using original vibration signals of the rest group of bearings in the full life cycle as a test set; in the step, in order to obtain the original vibration signal of the whole life cycle of the rolling bearing, a bearing accelerated life test needs to be carried out on the rolling bearing of a training set and a testing set, a unidirectional acceleration sensor is respectively fixed in the horizontal direction and the vertical direction of the testing bearing through a magnetic seat, the sampling frequency, the sampling interval and the sampling duration are set, and a dynamic signal collector is used for collecting the vibration signals in the horizontal direction and the vertical direction and storing the vibration signals;
(2) extracting 18-dimensional time domain characteristics of the full-life periodic vibration signals of the bearings in the horizontal direction and the vertical direction of each training set, carrying out minimum-maximum normalized processing on each characteristic, and constructing a training characteristic set X l ：
Figure BDA0003579009220000055
Wherein l represents a training set bearing number (l is 1,2.. q), m represents a total number of features, m is 36, n represents a total number of samples of the l training set bearing, namely a set of vibration signals collected by the l training set bearing every 10-60 seconds in an accelerated life test,
Figure BDA0003579009220000056
an mth normalized feature representing an nth sample of an lth training bearing;
(3) during the training process, p kinds of nonlinear health are constructedFactor(s)
Figure BDA0003579009220000057
And combines it with the sampling duration, i.e.:
Figure BDA0003579009220000058
Figure BDA0003579009220000059
the constructed nonlinear health factors are at least two, preferably 4, of the health factors constructed based on the exponential, power, logarithmic and polynomial nonlinear functions, and the calculation formulas of the health factors constructed based on the exponential, power, logarithmic and polynomial nonlinear functions are respectively as follows:
Figure BDA0003579009220000061
HI 2 (t)＝1-θ r (4)
Figure BDA0003579009220000062
Figure BDA0003579009220000063
normalizing the health factor to [0,1 ] according to equation (7)]Obtaining normalized health factor
Figure BDA0003579009220000064
Figure BDA0003579009220000065
Where t is time, p represents the function type, p ≧ 2, t k The kth sample sampling duration corresponding to the l training set bearing, T being the total sampling duration, θ l (t)∈(0,1]R is the order of θ, r ∈ N * ，
Figure BDA0003579009220000066
Is theta of the l-th bearing l (t) and
Figure BDA0003579009220000067
the theta in the formulas (3) to (6) is theta l (t)；
(4) Establishing an LSTM prediction model, taking q groups of training bearing characteristics as the input of the LSTM model under the p health factors established in the step (3),
Figure BDA0003579009220000068
as output, adjusting the number M of hidden layer network layers 1 And number of nodes Q 1 Training an LSTM model to obtain a mapping relation F of training bearing characteristics and each health factor p l Constructing an LSTM-based bearing feature-health factor model, namely:
Figure BDA0003579009220000069
wherein, F p l Is the first training bearing characteristic X l Mapping to the p-th health factor, F (X) ═ F (X) 1 )∪F(X 2 )...∪F(X q )；
(5) Establishing a TDNN prediction model, and respectively outputting p training set health factors output by the LSTM model in the step (4)
Figure BDA00035790092200000610
Inputting TDNN model, training set bearing residual life percentage
Figure BDA00035790092200000611
Adjusting the number of hidden layers as outputM 2 And number of nodes Q 2 Verifying parameters such as the number P of checks, the number d of time delay steps and the like, training a TDNN prediction model to obtain a mapping relation between each health factor and the percentage of the residual life, and constructing a TDNN-based health factor-residual life prediction model, namely:
Figure BDA00035790092200000612
wherein,
Figure BDA00035790092200000613
T c is the current elapsed life, T 100 Is the total life, G p Is a mapping relationship between the health factor and the percentage of remaining life;
(6) extracting 18-dimensional time domain characteristics of the bearing full-life periodic vibration signals in the horizontal direction and the vertical direction of the test set, carrying out minimum-maximum normalized processing on each characteristic, and constructing a test characteristic set:
Figure BDA0003579009220000071
where T represents the test set bearing number (T ═ 1), m represents the total number of features, m ═ 36, i represents the total number of samples for the T-th test set bearing,
Figure BDA0003579009220000072
an mth normalized feature representing an ith sample of the Tth set of training bearings;
(7) inputting the bearing characteristics of the test set into the LSTM-based bearing characteristic-health factor model constructed in the step (4), namely outputting p health factor clusters of the bearing of the test set, and respectively calculating the average values of the p health factor clusters
Figure BDA0003579009220000073
Namely:
Figure BDA0003579009220000074
(8) averaging each health factor obtained in step (7)
Figure BDA0003579009220000075
Inputting the residual life percentages of the bearings to be tested of p test sets into a TDNN model, evaluating the accuracy of a prediction result by using at least one of an RMSE function and a Score function as an evaluation index, and finding an optimal health factor-residual life prediction model, namely a health factor-residual life prediction model corresponding to the minimum value of the RMSE function or/and the Score function;
(9) and (5) taking the optimal health factor-residual life prediction model determined in the step (8) as a residual life prediction model of the rolling bearing, and calculating the residual life of the bearing to be measured.
The above calculation formulas of the RMSE function and the Score function are respectively:
Figure BDA0003579009220000076
Figure BDA0003579009220000077
where N is the total number of sample points,
Figure BDA0003579009220000078
i.e. the difference between the predicted value of the remaining lifetime and the true value at the z-th sample point.
Example 1
The method for predicting the residual life of the rolling bearing based on the LSTM and the TDNN is further described below by combining specific embodiment data.
The simulation environment and parameters of this embodiment are selected as follows:
simulation environment
Model: intel (R) core (TM) i3-9100 CPU @3.60GHz 3.60 GHz;
operating the system: windows 10 professional edition;
software: matlab R2020 a.
Parameter setting
The original data of this embodiment is obtained from XJTU-SY rolling bearing accelerated life test conducted by the Lei Asia teaching team of mechanical engineering college of Western Ann university and Changjiang Yang science and technology Co., Ltd, in Wang et al, IEEE Transactions on Reliability,2018,69(1): 401-.
Taking the working condition 1 in the above document as an example, the detailed description of the experimental steps is performed, and the prediction result of the remaining life percentage is given:
(1) in the embodiment, five groups of original vibration signals of the bearings with the full life are selected through multiple tests, and original acceleration vibration signals of Bearing1-1, Bearing1-2, Bearing1-4 and Bearing1-5 are selected as training sets, that is, original acceleration vibration signals of 4 groups of training bearings, q is 4 and Bearing1-3 are selected as test sets;
(2) extracting time domain characteristics of each training Bearing and carrying out minimum-maximum normalization processing to construct a characteristic matrix set [ X ] of Bearing1-1, Bearing1-2, Bearing1-4 and Bearing1-5 training bearings 1 ] 36×123 、[X 2 ] 36×161 、[X 4 ] 36×122 And [ X ] 5 ] 36×52 Wherein four time domain features of the maximum, minimum, peak and peak-to-peak values of the bearing1-1 are shown in FIG. 2;
(3) four health factors, namely the formulas (3) to (6) are constructed based on exponents, logarithms, exponents and polynomial nonlinear functions, and multiple experiments show that when r is 3, the health factors are obviously biphasic, so that r is set to be 3, and each health factor is obtained according to the formulas (3) to (6)
Figure BDA0003579009220000081
Then obtaining the normalization according to the formula (7)
Figure BDA0003579009220000082
(4) Establishing an LSTM prediction model, taking 4 groups of training bearing characteristics as the input of the LSTM model respectively under the 4 health factors established in the step (3),
Figure BDA0003579009220000083
as output, adjusting the number M of hidden layer network layers 1 And number of nodes Q 1 Parameters such as initial learning rate L, learning rate attenuation D, packet loss rate B, maximum round number R and the like are set as shown in table 1, an LSTM model is continuously trained and updated according to a loss function MSE, and a mapping relation F of training bearing characteristics and each health factor is obtained p l Constructing a bearing characteristic-health factor model based on the LSTM;
(5) establishing a TDNN prediction model, and respectively outputting 4 training set health factors output by the LSTM model in the step (4)
Figure BDA0003579009220000084
Inputting TDNN model, training set bearing residual life percentage
Figure BDA0003579009220000085
As output, adjust the number of hidden layers M 2 And number of nodes Q 2 Verifying parameters such as the number P of checks, the number d of time delay steps and the like, wherein the parameters are set as shown in table 1, continuously training and iteratively learning the TDNN model for multiple times according to a Levenberg-Marquardt algorithm, continuously correcting the TDNN prediction model to obtain a mapping relation between each health factor and the percentage of the residual life, and constructing a TDNN-based health factor-residual life prediction model;
(6) in this embodiment, the bearings 1-3 are used as the bearings to be tested, the time domain characteristics of the bearings to be tested in the test set are extracted and subjected to min-max normalization processing to obtain the characteristic set [ X ] of the bearings to be tested 3 ] 36×158 Inputting the parameters into the LSTM-based bearing characteristic-health factor model constructed in the step (4), namely outputting 4 health factor clusters of the bearing in the test set, and respectively calculating the average value of the 4 health factor clusters
Figure BDA0003579009220000091
Averaging the obtained average values of each health factor
Figure BDA0003579009220000092
Inputting the health factor-residual life prediction model based on the TDNN constructed in the step (5), and outputting residual life percentages of the bearings to be tested in 4 test sets under different health factors, as shown in fig. 3 to 6;
(7) and (3) evaluating the accuracy of the percentage of the residual life of the bearing to be tested by adopting two evaluation indexes of an RMSE function and a Score function, as shown in figure 7, under four health factors, the two evaluation index values of the RMSE function and the Score function are smaller, which shows that the prediction accuracy is higher, and HI 4 Type health factor, both of which are minimal, therefore, HI was selected 4 The residual life prediction model under the type health factor is used as an optimal prediction model, the residual life of the bearing to be tested is calculated according to the residual life percentage of the bearing to be tested output by the prediction model, the conservative prediction can be realized on the basis of ensuring the precision, and the reasonable allocation of resources can be realized by taking maintenance measures on the bearing to be tested.
TABLE 1
Figure BDA0003579009220000093
The parts which are not described in the invention can be realized by adopting or referring to the prior art.
It is to be understood that the above description is not intended to limit the present invention, and the present invention is not limited to the above examples, and those skilled in the art may make modifications, alterations, additions or substitutions within the spirit and scope of the present invention.

Claims (6)

1. A method for predicting the residual life of a rolling bearing based on LSTM and TDNN is characterized by comprising the following steps:
(1) collecting a full-life original vibration signal of a training group rolling bearing under a certain working condition as a training set, and collecting a full-life original vibration signal of a testing group rolling bearing under the same working condition as a testing set;
(2) extracting time domain characteristics of the bearing of the training set and minimizing each characteristicMaximum normalization processing, constructing a training feature set X l ：
Figure FDA0003579009210000011
Wherein, l represents the number of the bearing of the training set (l is 1,2.. q), m represents the total number of the features, m is 36, n represents the total number of the samples of the bearing of the l training set,
Figure FDA0003579009210000012
an mth normalized feature representing an nth sample of an lth training bearing;
(3) during the training process, p health factors are constructed
Figure FDA0003579009210000013
And combines it with the sampling duration, i.e.:
Figure FDA0003579009210000014
Figure FDA0003579009210000015
normalization of health factors to [0,1]Obtaining normalized health factor
Figure FDA0003579009210000016
Figure FDA0003579009210000017
Where t is time, p represents the function type, p ≧ 2, t k The kth sample sampling duration corresponding to the l training set bearing, T being the total sampling duration, θ l (t)∈(0,1]，
Figure FDA0003579009210000018
Is theta of the l-th bearing l (t) and
Figure FDA0003579009210000019
the mapping relationship between the two;
(4) establishing an LSTM prediction model, taking q groups of training bearing characteristics as the input of the LSTM prediction model under the p health factors established in the step (3),
Figure FDA00035790092100000110
as output, adjusting the number M of hidden layer network layers 1 And number of nodes Q 1 Training an LSTM model to obtain a mapping relation F of training bearing characteristics and each health factor p l Constructing an LSTM-based bearing feature-health factor model, namely:
Figure FDA00035790092100000111
wherein, F p l Is the first training bearing characteristic X l Mapping to the p-th health factor, F (X) ═ F (X) 1 )∪F(X 2 )...∪F(X q )；
(5) Establishing a TDNN prediction model, and respectively outputting p training set health factors output by the LSTM prediction model in the step (4)
Figure FDA0003579009210000021
Inputting TDNN model, training set bearing residual life percentage
Figure FDA0003579009210000022
As output, adjust the number of hidden layers M 2 And number of nodes Q 2 Verifying the checking number P and the time delay step number d parameters, training a TDNN prediction model, and obtaining a mapping relation between each health factor and the residual life percentageThe method constructs a health factor-residual life prediction model based on TDNN, namely:
Figure FDA0003579009210000023
wherein,
Figure FDA0003579009210000024
T c is the current elapsed life, T 100 Is the total life, G p Is a mapping relation between the p-th health factor and the percentage of the remaining life;
(6) extracting time domain characteristics of the bearing of the test set, carrying out minimum-maximum normalized processing on each characteristic, and constructing a test characteristic set:
Figure FDA0003579009210000025
where T represents the test set bearing number (T ═ 1), m represents the total number of features, m ═ 36, i represents the total number of samples for the T-th test set bearing,
Figure FDA0003579009210000026
an mth normalized feature representing an ith sample of the Tth set of training bearings;
(7) inputting the bearing characteristics of the test set into the LSTM-based bearing characteristic-health factor model constructed in the step (4), namely outputting p health factor clusters of the bearing of the test set, and respectively calculating the average values of the p health factor clusters
Figure FDA0003579009210000027
Namely:
Figure FDA0003579009210000028
(8) averaging the test set health factors obtained in step (7)
Figure FDA0003579009210000029
Inputting the health factor-residual life prediction model based on TDNN, outputting the residual life percentage of the bearing to be tested of p test sets, and utilizing at least one of an RMSE function and a Score function as an evaluation index to evaluate the accuracy of a prediction result and find an optimal health factor-residual life prediction model, namely the health factor-residual life prediction model corresponding to the minimum value of the RMSE function or/and the Score function;
(9) and (5) taking the optimal health factor-residual life prediction model determined in the step (8) as a residual life prediction model of the rolling bearing, and calculating the residual life of the bearing to be measured.
2. The method for predicting the residual life of a rolling bearing based on LSTM and TDNN as claimed in claim 1, wherein the total number n of samples of the training set bearing in step (1) is the set of vibration signals collected every 10-60 s of the training set bearing in the accelerated life test.
3. The method for predicting the residual life of a rolling bearing based on LSTM and TDNN as claimed in claim 1, wherein the time domain features extracted in step (2) are 18-dimensional time domain features of a full life cycle vibration signal of a training set bearing in horizontal and vertical directions; and (4) extracting time domain characteristics in the step (6) which are 18-dimensional time domain characteristics of the full-life-cycle vibration signals of the bearing of the test set in the horizontal direction and the vertical direction.
4. The method for predicting the remaining life of a rolling bearing based on LSTM and TDNN according to claim 1, wherein the health factors constructed in the step (3) are at least two of health factors constructed based on exponential, power, logarithmic and polynomial nonlinear functions, and the calculation formulas of the health factors constructed based on exponential, power, logarithmic and polynomial nonlinear functions are respectively:
Figure FDA0003579009210000031
HI 2 (t)＝1-θ r
Figure FDA0003579009210000032
Figure FDA0003579009210000033
wherein r is the order of theta, r ∈ N * 。
5. The LSTM and TDNN-based rolling bearing residual life prediction method according to claim 4, wherein the health factors configured in the step (3) are 4, i.e. p is 4.
6. The method for predicting the residual life of a rolling bearing based on LSTM and TDNN according to claim 1, wherein the RMSE function and the Score function in step (8) are respectively calculated by the following formula:
Figure FDA0003579009210000034
Figure FDA0003579009210000035
where N is the total number of sample points,
Figure FDA0003579009210000036
i.e. the difference between the predicted value of the remaining lifetime and the true value at the z-th sample point.
CN202210349547.7A 2022-04-02 2022-04-02 Rolling bearing residual life prediction method based on LSTM and TDNN Withdrawn CN114861979A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210349547.7A CN114861979A (en) 2022-04-02 2022-04-02 Rolling bearing residual life prediction method based on LSTM and TDNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210349547.7A CN114861979A (en) 2022-04-02 2022-04-02 Rolling bearing residual life prediction method based on LSTM and TDNN

Publications (1)

Publication Number Publication Date
CN114861979A true CN114861979A (en) 2022-08-05

Family

ID=82629110

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210349547.7A Withdrawn CN114861979A (en) 2022-04-02 2022-04-02 Rolling bearing residual life prediction method based on LSTM and TDNN

Country Status (1)

Country Link
CN (1) CN114861979A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116011109A (en) * 2023-01-13 2023-04-25 北京控制工程研究所 Spacecraft service life prediction method and device, electronic equipment and storage medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116011109A (en) * 2023-01-13 2023-04-25 北京控制工程研究所 Spacecraft service life prediction method and device, electronic equipment and storage medium
CN116011109B (en) * 2023-01-13 2023-09-08 北京控制工程研究所 Spacecraft service life prediction method and device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
CN110361176B (en) Intelligent fault diagnosis method based on multitask feature sharing neural network
Fan et al. A novel machine learning method based approach for Li-ion battery prognostic and health management
Yan et al. A dynamic multi-scale Markov model based methodology for remaining life prediction
CN112785091A (en) Method for performing fault prediction and health management on oil field electric submersible pump
CN110633792A (en) End-to-end bearing health index construction method based on convolution cyclic neural network
CN109829136B (en) Method and system for predicting residual life of degradation equipment with random jump
Di et al. Ensemble deep transfer learning driven by multisensor signals for the fault diagnosis of bevel-gear cross-operation conditions
CN111289250A (en) Method for predicting residual service life of rolling bearing of servo motor
CN111459144A (en) Airplane flight control system fault prediction method based on deep cycle neural network
CN111783362A (en) Method and system for determining residual service life of electric gate valve
CN114861349B (en) Rolling bearing RUL prediction method based on model migration and wiener process
CN111709577B (en) RUL prediction method based on long-range correlation GAN-LSTM
CN114048688A (en) Method for predicting service life of bearing of wind power generator
CN117290771A (en) Rotary machine fault diagnosis method for generating countermeasure network based on improved auxiliary classification
CN114692507A (en) Counting data soft measurement modeling method based on stacking Poisson self-encoder network
CN114925723B (en) Method for predicting residual service life of rolling bearing by adopting encoder and decoder
CN114861979A (en) Rolling bearing residual life prediction method based on LSTM and TDNN
Tang et al. Prediction of bearing performance degradation with bottleneck feature based on LSTM network
CN117521512A (en) Bearing residual service life prediction method based on multi-scale Bayesian convolution transducer model
CN116595319A (en) Prediction method and system applied to rail transit motor health state evaluation
CN118114118A (en) CNDT-based fault diagnosis method for typical weapon equipment
CN117407665A (en) Retired battery time sequence data missing value filling method based on generation countermeasure network
Luo et al. A novel method for remaining useful life prediction of roller bearings involving the discrepancy and similarity of degradation trajectories
CN115825732A (en) Intelligent diagnosis method for open-circuit fault of permanent magnet synchronous motor driving system with associated characteristics
CN115292820A (en) Method for predicting residual service life of urban rail train 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
WW01 Invention patent application withdrawn after publication
WW01 Invention patent application withdrawn after publication

Application publication date: 20220805