Search

CN-115879241-B - Cross-domain bearing life prediction method and system based on single-source domain generalization

CN115879241BCN 115879241 BCN115879241 BCN 115879241BCN-115879241-B

Abstract

The invention relates to the field of bearing life prediction, in particular to a method and a system for predicting the life of a cross-domain bearing based on single-source domain generalization. In order to solve the defect of low prediction precision of the service life of the bearing in the prior art, the predictor adopted in the invention consists of a data preprocessing module and a service life prediction module, wherein the data preprocessing module is used for constructing a data pair sequence consisting of a time domain characteristic peak-to-peak value time sequence of a vibration data sample and a time sequence normalization value as the input of the service life prediction module so as to predict the residual service life of the bearing in the vibration data sample acquisition time. The method is beneficial to avoiding the data distribution difference of different bearings and ensures the accuracy of a predictor obtained by learning the source bearing for predicting the target bearing. Therefore, the data distribution diversity is increased through the source bearing data, the generation of the anti-fact data on the basis of increasing the data diversity is avoided, and the generalization capability of the prediction model is improved.

Inventors

  • XU JUAN
  • MA BIN
  • Lv Zengwei
  • FAN YUQI
  • Ding Xiu

Assignees

  • 合肥工业大学

Dates

Publication Date
20260508
Application Date
20221215

Claims (9)

  1. 1. A single-source-domain-generalization-based cross-domain bearing life prediction method is characterized by comprising the steps of firstly obtaining a predictor, wherein the predictor is obtained through machine learning, the predictor comprises a data preprocessing module and a life prediction module, the data preprocessing module is used for processing vibration data of a bearing, the life prediction module takes the data processed by the data preprocessing module as input data, and the life prediction module predicts the residual life of the bearing according to the input data; the processing of the vibration data of the bearing by the data preprocessing module comprises the following steps: SA1, obtaining vibration data of a bearing , Representing the first of the bearings The vibration data samples are data composed of a plurality of data points acquired in unit time according to the set acquisition frequency, Is composed of vibration data samples collected in continuous time, N is ordinal number, and N is vibration data The number of the vibration data samples is 1-N; SA2, acquiring input time sequence , ={ , ,..., ,... I is ordinal number, 1 +.i, For step size, i=n-m+1; representing an ith time series sample; Representation of J-th data of (i.e. vibration data sample) Time domain characteristic peak-to-peak values of (a); 1.ltoreq.j.ltoreq.m; SA3, pair input time series Normalization processing is carried out to make Normalized value is ; SA4, outputting data by the data preprocessing module as follows , ; The vibration data D of the bearing is obtained by screening from the original data H of the bearing, and the screening of the vibration data D comprises the following steps: SB1, recording the original data H of the bearing as , The method comprises the steps of representing the (R) th vibration data sample collected by a bearing in bearing monitoring, wherein 1R R represents the total number of the collected vibration data samples, extracting a time domain characteristic peak-to-peak value set from original data H, and extracting a time domain characteristic peak-to-peak value set from the original data H For a pair of Obtaining degradation data with monotonic trend using an isotonic regression algorithm ; Representing degradation data The r-th value of (a); SB2, constructing sliding window pair degradation data with length z Sliding to obtain degradation gradient of each window data, and obtaining the kth window value The corresponding degradation gradient is ,1≦k≦R-z+1; SB3, construction of degenerate continuous Domain , And Representing degenerate continuous domains, respectively Is set to the start point value and the end point value of (c), The initial value of (1) is , The initial value of (1) is ; SB4, calculation Degradation gradient of all window values The average value of (1) is denoted as Av, and is determined Whether the following condition 1-2 is satisfied, executing step SB5 if satisfied, and executing step SB6 if not satisfied; Condition 1: On which there is a continuous fluctuation domain , Respectively represent continuous fluctuation domains Is set to the start point value and the end point value of (c), Degradation gradient corresponding to all window values 、 、......、 Are all greater than 0, and The number of window values C-z+1-b+1 is larger than or equal to a set value C0, and C0 is larger than 0; Condition 2: The number of window values of the degradation gradient which are corresponding to the degradation gradient is Q, wherein Q is greater than or equal to a set value Q0, and Q0 is greater than or equal to C0; SB5, update degenerate continuous domain , = , = Then return to SB4; SB6 to Final degradation point, obtaining original data As vibration data I.e. = 。
  2. 2. The method for predicting the service life of a cross-domain bearing based on single-source domain generalization as claimed in claim 1, wherein the step SB4 specifically comprises the following sub-steps: SB41, calculate The mean value Av of the degradation gradient of all window values is obtained; SB42, ream = T is s as initial value; SB43, calculation Upper window value Is of the degradation gradient of (2) ; SB44, judgement If so, updating t=t+1 and returning to step SB43, otherwise, executing step SB45; SB45, judging whether t-b is greater than or equal to C0, if not, updating t=t+1, returning to step SB42, if yes, executing the following step SB46; SB46, order data Corresponding degradation gradient 、 、......、 The data quantity greater than the average value Av is recorded as Q, whether the Q is greater than or equal to a set value Q0 is judged, if not, t=t+1 is updated, the step SB42 is returned, and if yes, the data is made As a continuous wave domain I.e. = 。
  3. 3. The method for predicting life of cross-domain bearing based on single source domain generalization of claim 1, wherein said life prediction module comprises a preliminary prediction network, an inverse normalization unit and an output unit, said preliminary prediction network is used for obtaining Corresponding first predicted value And Corresponding second predicted value The inverse normalization unit is used for performing inverse normalization on the second predicted value Performing inverse normalization processing to obtain inverse normalized predicted values The output of the output unit is the output of the predictor, and the output result of the output unit is the residual life sequence , For bearing in vibration data sample A residual life prediction value in acquisition time; Is a learnable affine parameter.
  4. 4. The single-source domain generalization-based cross-domain bearing life prediction method according to claim 3, wherein the method comprises the following steps: Representation of Is used for the average value of (a), Representation of Gamma and beta are both learnable affine parameters; in order to set the constant of the device, Take value over interval 10 -6 ,10 -5 .
  5. 5. The single-source domain generalization-based cross-domain bearing life prediction method according to claim 4, wherein the method comprises the following steps: 。
  6. 6. The single source domain generalization based cross-domain bearing life prediction method of claim 3, wherein the predictor obtaining comprises the steps of: S1, constructing and initializing a predictor consisting of a data preprocessing module and a life prediction module; S2, obtaining vibration data D S of the source bearing, Construction of multiple learning samples { | I=n-m+1, the vibration data sample is data constituted by a plurality of data points acquired in a unit time according to a set acquisition frequency, Is composed of vibration data samples acquired over a continuous time period; Representation of The nth vibration data sample of (a); for source bearing in vibration data sample Is used to determine the remaining life in terms of acquisition time, Making the predictor machine learn the learning sample to iterate the parameters until the loss of the predictor obtains a set loss threshold value, and fixing the predictor; The predictor's loss function is: source bearing in vibration data sample for predictor output Is provided for the time remaining life prediction.
  7. 7. The single-source-domain-generalization-based cross-domain bearing life prediction method of claim 3, wherein the preliminary prediction network is constructed based on a single-layer gating cycle unit GRU.
  8. 8. A single-source-domain-generalization-based cross-domain bearing life prediction system comprising a memory in which a computer program is stored, which when executed is adapted to implement the single-source-generalization-based cross-domain bearing life prediction method of any of the preceding claims 1-7.
  9. 9. The single-source domain generalization based cross-domain bearing life prediction system of claim 8, further comprising a processor coupled to the memory, the processor for executing a computer program to implement the single-source domain generalization based cross-domain bearing life prediction method of any of claims 1-7.

Description

Cross-domain bearing life prediction method and system based on single-source domain generalization Technical Field The invention relates to the field of bearing life prediction, in particular to a method and a system for predicting the life of a cross-domain bearing based on single-source domain generalization. Background Bearings are important components of rotary machines and are critical to the operation of industrial machines. Once the bearings fail, they directly affect the safety of the whole equipment and produce a certain economic loss. Thus, the prediction of the remaining life of the bearing has an important role in the prognosis and health management of the device. Existing deep learning models have met with great success in predicting the remaining life of a bearing, but they assume that training data and test data follow the same data distribution. However, in the actual life remaining prediction scenario of different bearings, vibration data collected on unknown bearings cannot be guaranteed to be distributed identically to training vibration data of the bearings due to different working conditions or materials of the different bearings. It is not practical to further collect full life cycle vibration data for each type of bearing. If the existing learning model is trained on one bearing and applied directly to an unknown bearing, the performance of the model will be significantly degraded. In the prior art, the single-source domain generalization method has wide application prospect for predicting the residual life of an unknown bearing, and can model from a domain with high source domain generalization capability, increase the diversity of data distribution and enable better results to be obtained on other unknown target domains relatively. However, the single-source domain generalization method is not applied to the aspect of predicting the residual life of the bearing at present, and cannot be directly applied to the aspect of predicting the residual life of the bearing. The main reasons for this are two points, one is that the vibration data set of the bearing is time series data of the whole life cycle, and a large amount of tiny fluctuation data exists in the early stage. These early data have little or no negative impact on the description of the bearing degradation tendency. Thus, for more accurate life predictions, the vibration dataset of the entire life cycle should be more reasonably partitioned to reduce interference with early data. Secondly, the existing Shan Yuanyu generalization method increases the diversity of data distribution through a data enhancement technology. However, such image-oriented enhancement techniques often suffer from uncertainty that creates "anti-facts data", i.e., a complete inconsistency with the actual vibration data distribution of the bearing, thereby deteriorating the performance of the model. Disclosure of Invention In order to solve the defect of low accuracy of bearing life prediction in the prior art, the invention provides a single-source-domain-generalization-based cross-domain bearing life prediction method and system, which can realize high-accuracy cross-domain prediction. The invention adopts the following technical scheme: The method comprises the steps of firstly obtaining a predictor, wherein the predictor is obtained through mechanical learning, the predictor comprises a data preprocessing module and a life predicting module, the data preprocessing module is used for processing vibration data of a bearing, the life predicting module takes the data processed by the data preprocessing module as input data, and the life predicting module predicts the residual life of the bearing according to the input data; the processing of the vibration data of the bearing by the data preprocessing module comprises the following steps: SA1, obtaining vibration data D= { D 1,d2,…,dn,…,dN},dn of a bearing to represent an nth vibration data sample of the bearing, wherein the vibration data sample is data formed by a plurality of data points acquired in unit time according to a set acquisition frequency, D is formed by vibration data samples acquired in continuous time, N is ordinal number, and N is the number of the vibration data samples in the vibration data D, and N is equal to or greater than 1 and N is equal to or less than N; SA2, the acquisition input time sequence x= { X 1,x2,…,xi,…,xI }, I is ordinal number, 1+.i+.i, m is step size, i=n-m+1, x i represents the ith time series sample; Representing the j-th data in x i, namely the time domain characteristic peak value of the vibration data sample d i+j-1, wherein j is equal to or less than 1 and m is equal to or less than m; SA3, normalizing the input time sequence X to make the normalized value of X i be SA4, outputting data by the data preprocessing module as follows Preferably, the vibration data D of the bearing is obtained by screening from the raw data H of the bearing, and the screening of t