Research On Data-based Prognosis For Nonlinear Industrial Processes

With the rapid development of science and technology, both system structuresand equipment conditions of modern industrial processes are increasingly becomingmore complicated than before. Besides, the devices within industrial processes keepproducing huge amounts of input-output data during the long-term running. Thecontinuously growing amount of measured data, process nonlinearity as well asvarious uncertain disturbances play important roles in the safety and reliability ofindustrial processes. In respect of complicated nonlinear industrial processes withdifficulities in dynamic modelling and non-Gaussian distributed data, which resultin the infeasibility of traditional model-based approaches, the overall performanceis easy to be adversely affected by faults and improper conditions. The topic on howto effectively apply the measured big data to the key performance index prognosisof processes has become one of the hot issues in current academia and industry,aiming to improve the production efficiency and achieve the maximum benefits.Therefore, studying the prognosis issue for nonlinear industrial processes in thesense of big data is of great significance in both theoretical and practical fields.Based on the advantages as well as the theoretical overlap of partial leastsquares (PLS) and locally weighted projection regression (LWPR), this thesispresents a data-based algorithm called LWPR-MPLS for the prognosis of nonlinearprocesses by improving and integrating the two aforementioned methods. Theproposed algorithm just needs the dynamic non-Gaussian distributed input-outputmeasurements to provide on-line prognosis of key performance indexes fornonlinear processes, and four statistical indicators which can be extended to realtime monitoring as well as fault diagnosis. In order to determine the effectiveness ofthe proposed algorithm, several verifications have been implemented, which arerespectively based on the process data produced by MATLAB simulation and thereal datasets of proton exchange membrane fuel cell systems provided by FCLABResearch Federation, France. The results indicate that LWPR-MPLS algorithm isefficient in the on-line prognosis for nonlinear industrial processes on one hand; onthe other hand it succeeds in reflecting the abnormal phenomena of processes timelyand effectively via two modified statistical indicators, which contribute to efficientfault diagnosis by working with other indicators in process monitoring. Finally, thethesis extends the verified algorithm to the remaining useful life prediction of fuelcell systems. According to different failure thresholds, the corresponding values ofremaining useful life are estimated based on the variation of total voltage, which hasbeen chosen as the key performance index of the studied fuel cell system.