Research On Limited Data-Driven Life Prediction Methods And Applications Of Rolling Bearings Under Variable Operating Condition

One of the significant requirements for the intelligence of high-end machinery equipment is an efficient and stable prognostic and health management(PHM)system.The research of PHM not only includes relatively mature fault diagnosis technology but also contains machinery prognostics.The current data-driven based prognostics(DDP)researches have attracted the attention of many domestic and foreign scholars due to its flexible and convenient decision-making modeling process.However,its generalization performance restricts the application of DDP in actual application deployment.At the same time,DDP is highly dependent on the size and completeness of data samples.It is not difficult to speculate that when the scale of learnable data samples decreases,the generalization problem caused by the domain shifts from variable operating conditions may further deteriorate.This dissertation takes rolling bearing,an indispensable essential component in intelligent manufacturing application scenarios,as the research object,starting from the two aspects of degradation trend(DT)and remaining useful life(RUL)prediction,and fully considers potential variable working conditions in practical scenarios where monitoring samples are scarce.It takes a new approach and uses the idea of "few shot learning(FSL)" to realize the cross-domain predictions under limited samples(total data size).We focus on studying life prediction methods and corresponding derivative algorithms to overcome the dilemma of data-driven prediction methods in the case of limited monitoring samples.This dissertation’s implementation is significant to enrich and improve the theoretical methods of life prediction and can effectively improve the safety supervision level of rotating machinery in industrial sites.Its innovations and main contents are as follows:(1)In-depth study of the influence of the sample scale on data-driven algorithms and explain the dilemma of the current mainstream data-driven prediction research under limited data from the perspective of prediction error decomposition.Therefore,FSL is introduced,and the classification and regression processes in the original FSL are extended to prediction and extrapolation processes,which are the proposed limited-data-driven prognostics(LDDP)modes.And then,the two major themes of rolling bearing degradation trend and remaining life prediction,DT-LDDP and RUL-LDDP in short,are established,respectively.They define the form of input and output in the following DT and RUL predictions under limited samples and the general predictive modeling process.At the same time,the limited data samples set exist in the historical sample training and fine-tuning adaptation stages,which can effectively fit the actual industrial application scenarios and guide the construction of subsequent LDDP algorithms.(2)The two proposed limited-data-driven prognostics problems are both typical variablecondition prediction and decision-making tasks and are also affected by the domain shift problems.When the available training data is sharply reduced,the generalization error caused by this problem will increase significantly.In view of this,constructing a health indicator with general statistical properties helps improve the LDDP model’s generalization accuracy.Based on the classic unsupervised autoencoder network,a double adversarial learning based multi-domain distribution adaptation(DALMDA)algorithm is proposed.It first extracts the original signal’s multi-domain degradation features;then inputs them into the established discriminant module with an adversarial game function to realize the mapping of the common degradation features in the latent space between multi-source domains.At the same time,DALMDA quantifies the monotonicity,trend and robustness as the objective functions for the learnable parameter update and optimization of DALMDA.Finally,linear projection is used to generate a transferable health indicator that considers certain degradation and damage features,which is used in the initial monitoring and as the input of subsequent life prediction.(3)Aiming at the problem of weak generalization accuracy under variable working conditions caused by the limited amount of data in the prediction of rolling bearing degradation trend,a segmented degradation trend indicator with weak-stationary characteristics is constructed using the method of stationary subspace analysis.Moreover,a stationary subspaces-vector autoregressive with exogenous terms method(SSVARX)is proposed to conduct degradation trend prediction of rolling bearings.SSVARX first extracts the time and frequency domain degradation information with stationary characteristics from the multi-channel vibration signal and divides it into endogenous and exogenous terms required for autoregressive modeling according to the degradation law.Then a series of steps of order determination,stationarity test,and parameter estimation are carried out.On this basis,a vector autoregressive prediction model based on endogenous terms in time and frequency domains is established.This lightweight extrapolation prediction model is the first to describe the bearing degradation and damage from the perspective of the running stability of mechanical components and can complete the inference and predictions of the future moments only through the data before the prediction starting point under limited monitoring data samples.The proposed degradation trend prediction method can effectively predict the abnormal operation state in the future,and lay the foundation for the subsequent life prediction work.(4)Given the problem of weak generalization accuracy under variable working conditions caused by the limited amount of data in the residual life prediction of rolling bearings,a metalearning-based prototype LDDP algorithm is proposed: meta GRU,starting from the meta-learning theory and based on the deep recurrent unit(GRU).Meta GRU first divides the training data samples into subtasks and implements the life model training within the subtask level(inner loop)based on the gradient update strategy of back-propagation with time of GRU,and then aggregates all subtasks for across-subtask(outer loop)parameter update.Finally,a meta-learning prediction model that is easy to fine-tune with a small number of samples and can be generalized to unknown fields is established.On the basis of meta GRU and the excellent characteristics of domain alignment in domain adaptation,a prediction algorithm,statistical alignment based meta gated recurrent unit(SAMGRU),that integrates knowledge transfer and LDDP,is proposed.SAMGRU first constructs a high-order domain discrepancy measure to align the source domain and target domain data statistically,then performs the inner and outer loop parameter update of meta GRU,which endows the LDDP prediction model with the transferable ability across variable working conditions.At the same time,meta GRU further expands the compatibility of its basic prediction unit for variable input and output sample pair form.It proposes a novel LDDP algorithm(meta attention gated recurrent unit,MAGRU)that integrates the attention mechanism and encoding & decoding architecture.MAGRU introduces a GRU-based encoder-decoder network to realize the variable-length prediction function.At the same time,an attention layer is added to the constructed encoder-decoder network to further capture the contextual semantic information of the input features and enhance the generalization ability of the LDDP algorithm under variable working conditions.The proposed three algorithms can effectively accomplish rolling bearings’ prognostics tasks under variable working conditions.(5)Aiming at the problem of unlabeled historical samples and the problem of prediction reliability in the current limited data-driven forecasting,based on three proposed LDDP algorithms,the unsupervised modeling and prediction uncertainty research in LDDP are deeply explored.Firstly,the inner loop training of the meta GRU prototype algorithm is improved by clustering and assigning pseudo-labels to propose an unsupervised meta GRU algorithm(UMGRU)to deal with the challenges of massive unlabeled historical data in industrial applications.The posterior approximation idea in Bayesian theory improves the MAGRU algorithm,reconstructs its inner loop prediction unit through a variational inference approach,and finally proposes a probabilistic meta-learning LDDP prediction algorithm(Bayesian approximation enhanced probabilistic meta learning,BA-PML).BA-PML utilizes the Bayes by backprop method to learn the posterior distribution of the parameters of the inner loop prediction unit and converts the parameter uncertainty of the prediction unit into the prediction uncertainty.It extends the original point estimation prediction from LDDP to interval estimation predictions,enhancing LDDP’s reliability and quantifying the prediction uncertainty.(6)Based on the above LDDP solutions,IEEE PHM 2012 challenge(public data set),our selfbuilt ABLT-1A full-life fatigue test(laboratory data set),and a petrochemical pump performance degradation dataset(industrial field data)are individually cross-validated under different limited-data specifications.The analysis results show that the proposed series of LDDP algorithms can solve the weak generalization performance of DDP algorithms in small sample environments such as missing data.More importantly,UMGRU has achieved good cross-working-condition generalization results in the unlabeled environment of historical samples.BA-PML improves the reliability of LDDP methods,provides excellent prediction results,and quantifies the prediction uncertainty from the perspective of interval estimation.