Research On Methods For Remaining Useful Life Prediction Of Products Driven By Condition Monitoring Big Data

The rapid advancements of sensor,information and communication technology,and Internet-of-Thing(IoT)technique,etc,the condition monitoring(CM)technique has been adopted on a large scale in various products.In order to implement real-time online monitoring toward the health condition of the products,different types of sensors have been deployed on some specific parts of the products to acquire data by high sampling frequency.Consequently,the growing amount of sensors result in the continuous collection of high volume data,which inevitably creates an industrial big data environment.Under such a circumstance,the available massive CM data can be leveraged to continuously track the product's health status and make an accurate and reliable estimation about its remaining useful life(RUL).With these actionable information in hand,catastrophic failures can be avoided by the means of advance warning,maintenance schedule can be further optimized to reduce the cost of unnecessary and expensive maintenance actions,and meanwhile,the reliability,availability,and safety of the products can also be improved.Thus,the research of this dissertation will focus on prognostic approaches for RUL prediction of products.In the era of industrial big data,the amount of products group monitored and the number of the sensors for each product are large,massive CM data are acquired by the high sampling frequency,products are working under dynamic operating conditions and multiple failure modes coexist,and the failure physical models for some complex products are unavailable.Aiming at aforementioned practical challenges,this dissertation firstly conducts the research on RUL prediction method based on degradation trajectory similarity based prognostic method,then develops the deep learning(DL)-based prognostic framework to tackle the issue of RUL prediction of products under dynamic operating conditions and multiple failure modes coexisting.In addition,particle filter(PF)technique and DL-based model are strategically integrated to develop the hybrid prognostic approach.Finally,in the bootstrap implementation framework,the deep convolutional neural network(DCNN)-based prognostic approach is proposed to quantify prediction interval of the RUL without utilization of the physical model of the products.The main contents and innovations of this dissertation are summarized as follows:(1)Development of an improved degradation trajectory similarity-based prognostic(TSBP)method toward RUL prediction of the products.Considering the scenario where there are plenty of similar products with massive historical CM data,traditional TSBP method can accurately provide the point estimation of the RUL about the operating products.However,this might impose restrictions on the application of TSBP method to some safety-critical products,such as aircraft engine,etc.Thus,it is neccesary to conduct in-depth study about the TSBP method to break through the limitation of point estimation of the RUL.This dissertation proposes an improved TSBP method,which integrates the kernel density estimation(KDE)technique and ?-criterion.(2)Development of a deep learning(DL)-based prognostic method.Modern engineering systems generally work under dynamic operating conditions and multiple failure modes coexist.Most of the traditional data-driven prognostic methods still lack an effective model to handle this prognostic issue.Due to the fact that bi-directional long short-term memory(BLSTM,one kind of DL-based model)networks can effectively extract long-term temporal dependences hidden in the time series data and has powerful nonlinear modelling capacity.Thus,this dissertation proposes a BLSTM-based approach that can provide an end-to-end prognostic solution.And the proposed DL-based method also can obtain more accurate RUL prediction results compared with other state-of-theart data-driven prognostic methods.(3)Development of a hybrid prognostic method that integrates particle filter(PF)technique with DL-based model.For some specific products that the critical failure mode is known and the failure physical model has been established,traditional hybrid prognostic method is based on the physical model of the products and leverage the CM data to estimate the probability density function(PDF)of the RUL.However,a limitation associated with traditional hybrid approaches is that numerous cumbersome steps including but not limited to feature extration,selection,reduction and regression analysis about the degradation indicator of the products.And the aforementioned processing steps take advantage of in-depth domain knowledge.In addtion,the prognosis performance greatly relies on these specific steps,which might make the traditional hybrid approaches inefficient and not that robust.Thus,this dissertation proposes an enhanced DL-based hybrid prognostic method in the PF framework,which fully utilizes BLSTM networks and feed-forward neural networks(FFNN)to automatically extract and select features and implement regression analysis to obtain the degradation indicator of the products.(4)Development of a novel DCNN-Bootstrap prognostic method.For some complex products or engineering systems,the corresponding failure modes and failure physical models are unavailable,thus the aforementioned PF-DL fusion prognostic method cannot effectively quantify the RUL prediction interval.Accordingly,this dissertation proposes a generalized prognostic approach based on DCNN and is implemented in the Bootstrap framework.The proposed DCNN-Bootstrap prognostic method do not require any prior physical information about the products,such as its failure physical model and the corresponding parameters statistical distribution,and this novel characteristic will enhance the applicability of the DL-based prognostic method toward the complex products and engineering systems.