Research On Data-driven Diagnosis Methods For Unanticipated Fault

Fault diagnosis for the unanticipated fault(UF) plays an important role in the condition monitoring/health management(CM/HM). In the age of big data, data-driven(DD)fault diagnosis methods become the state of the art. The main object systems for this dissertation are the satellite attitude control system(SACS) and the Tennessee-Eastman process(TEP), etc. The research on data-driven fault diagnosis for the UF is governed by the ideas of ‘framework to methods, stationary to non-stationary, multi-batch to singlebatch, whole-data to half-data, static to dynamic, theories to applications’. The key points of the research are as follows:(1) For UF, a general process model is constructed for the UF diagnosis. As the technical route of the whole dissertation, the model regularizes the process of the UF diagnosis and establishes the realization principles of the data-driven methods. For the stationary data, a UF diagnosis method based on the one-class multivariate statistical analysis(OCMSA) is introduced and analyzed, where the limitations as well as the compensation ideas are shown.(2) For the non-stationary data under the multi-batch condition, a detection statistics is constructed based on the smoothing preprocessing technique, whose effects on the detection are evaluated; the rules for the anticipated fault isolation(AFI) and the UF detection(UFD) are given based on the definition of the feature directions; a new contribution rate index is constructed. The research shows that the smoothing preprocessing can enhance the robustness of the computation and reduce the correlation of the non-stationary data; the feature direction can avoid the false determination caused by the changing magnitude of the fault; the new contribution rate index can avoid the phenomena of ”negative contribution” and ”fault smearing”.(3) For the non-stationary data under the single-batch condition, an improved detection statistics is constructed based on the time series modeling, whose effects on the detection are evaluated; The incremental and decremental formula for updating the inverse of the covariance matrix of the calibration residual is proposed and proved, based on which the incremental and decremental algorithm for the improved detection statistics is available. The research shows that the improved statistics can enhance the robustness of the computation as well as the adaptation of the detection statistics, and the proposed algorithm indeed can significantly reduce the computational complexity.(4) For the static model structure, A unified weight-framework is constructed for latent variable extraction, regression and detection; three theorems, namely, coefficient theorem, calibration theorem and detection theorem, are given and proved. The research shows that the theorems can successfully interpret the transform condition, calibration accuracy and detection performance, thus provide a theory for selecting among different methods.(5) For the dynamic model structure, the algorithm is given for identifying the stable kernel representation(SKR); the concept of feature directions for dynamic model is defined, based on which, the rules for AFI and UFD are provided; an optimal visualization algorithm is proposed for visualizing the fault information with high dimension. The research shows that the feature directions can be used for UFD in the dynamic model; the optimal visualization algorithm can offer plenty of space information to UFI.(6) For all the UF diagnosis methods above, a toolbox is designed and developed which contains all the novel methods in this dissertation as well as some traditional methods. The graphical user interface is friendly, full-functioned and easy-extended, which will benefit the academic communication and fault diagnosis applications.