Research On Data-driven Design Methods Of Fault Diagnosis And Fault-tolerant Control Systems

The increasing demands on safety and reliability of dynamic systems have receivedmuch attention from people. These requirements have extended from those nuclear re-actor, chemical industry process, aerospace vehicle systems, etc. to new systems such asautonomous vehicles or intelligent robots. An efective fault diagnosis and fault-tolerantcontrol is of prime importance for the strong support on safety and reliability. Hence,the study on fault diagnosis and fault-tolerant control technology has both theoretical andpractical importance.Although the model-based fault diagnosis and fault-tolerant control theory has beenwell-established, it is still difcult to establish mathematical models by means of the frstprinciple for complicated plants. On the other hand, a large amount of historical datafrom regular sensor measurements, event-logs and records are often available. Motivatedby this observation, it is of great interest to design fault diagnosis and fault-tolerant con-trol schemes based on efcient data-driven design methods of fault diagnosis and fault-tolerant schemes instead of model-based methods.This dissertation focuses on the data-driven design of fault diagnosis and fault-tolerant control systems for discrete linear time-invariant (LTI) systems. The main contri-butions of this dissertation can be summarized as follows:For discrete LTI systems, a data-driven design method of robust fault diagnosissystems is proposed, in which the well-established parity space and parity vectors areidentifed directly by of-line data. Based on the H2performance index, residual genera-tions for robust fault detection, robust reduced order fault detection, robust fault isolationand robust fault identifcation are constructed for sensor and actuator faults, respectively.Compared with the existing data-driven design method of fault diagnosis, the proposedmethod in this dissertation generates residuals keeping sensitivity to faults and robustnessto disturbances, and thus improves the robustness of fault diagnosis systems.For discrete LTI systems, this dissertation proposes a data-driven design methodof parametrization controllers. The proposed method studies the data-driven design ofthe extended Internal Model Controllers (EIMC) based on the identifed parity space andthe residual generation, the data-driven design of observer-based Youla parametrization controllers for SISO and MIMO systems based on the reduced order parity space and fullstate observers, respectively.Based on Youla parametrization, a residual data-based tuning method for Youla pa-rameters, denoted as residual-based tuning (RbT), is proposed. This method exploitsresidual signals for estimating the gradient of the cost function and based on an iterativedescent algorithm to tune Youla parameters for disturbance rejection. In the theoreticalview, the convergence issue, the global optimum condition and the asymptotic accuracyare given. In the view of engineering, an extended RbT (ERbT) algorithm is proposed tooptimize Youla parameters online.Based on a two degrees-of-freedom (2-DOF) Youla parametrization structure, thisdissertation proposes a novel residual-based iterative feedback tuning method, denotedas residual iterative feedback tuning (RIFT). This method feeds back residuals into theclosed loop, estimates the gradient of the cost function via set of experiments and thenachieves a fault-tolerant control based on the iterative descent algorithm. For single-input-single-output (SISO) and multiple-input-multiple-output (MIMO) plants, two RIFTalgorithms are developed, respectively. The number of experiments of MIMO plants isfurther discussed.