Sensor Fault Diagnosis Based On Structure Optimized PCA And Its Application

In industrial processes, accurate measurements are necessary in providing controllers and operators with a view of the process status. Sensor readings play a key role in assessments of the system state. The safety, reliability, and performance of complex systems with many sensors are largely dependent on the accuracy and reliability of the sensors. However, the sensors are not significantly more reliable than the systems being monitored, the indication of an abnormal state may be the result of a sensor failure rather than a system failure. Failure to identify the source of the indication of an "abnormal state" and take appropriate corrective action could result in expensive and unnecessary system shutdowns or, worse still, accidents that endanger both system and system personnel. Thus, it is very important for monitoring and diagnosis the sensor fault.To improve the correctness and speediness of sensor fault diagnosis based on principal component analysis (PCA), a set of structure optimization methods are proposed. Then, a sensor fault diagnosis system is designed and developed. The main contributions in this dissertation are listed as follows:1) A feature subspace based kernel PCA (FS_KPCA) method is presented to overcome the shortcoming of the standard KPCA method which is not appropriate to deal with large number of training data. FS_KPCA optimized the kernel matrix and reduces the computational cost of KPCA by constructing a lower-dimensional orthonormal basis on the feature subspace.2) The variable reconstruction algorithm based on FS_KPCA is derived. Then,sensor fault detection and diagnosis (SFDD) method based on FS_KPCA is proposed.Comparing to standard KPCA=based SFDD methods, the proposed method is more efficient in computation and needs less computer memory. Computer simulation on a non=isothermal CSTR process monitoring problem demonstrates the effectiveness and efficiency of the proposed method.3) Optimal design of the incidence matrix for Structured Residual Approach with Maximized Sensitivity (SRAMS) is discussed in this paper, in order to overcome the shortcoming of SRAMS that ignores the speed and sensitivity of fault diagnosis while considering the isolability of fault codes in the incidence matrix design. A set of optimal design indices including fault sensitivity and fault diagnosis sensitivity are defined, based on which two design algorithms are developed. The proposed approach is tested in the sensor fault detection and diagnosis of the reactor unit of the TE process, simulation results show that the diagnosis performance is better than original SRAMS. 4) A nonlinear sensor fault diagnosis method based on structured kernel principal component analysis (KPCA) is proposed. A set of KPCA submodels are constructed on reduced variables sets which are determined by the rows of incidence matrix. The scaled power values of the structured KPCA submodels constitute a structured residual vector. Faults isolating decisions are made according to the structured residual vector and the incidence matrix. Sensor faults sensitivity and critical sensitivity are defined, based on which an incidence matrix optimization algorithm is proposed to improve the performance of the structured KPCA. The proposed method is applied to a non-isothermal continuous stirred-tank reactor (CSTR), the results show that the diagnosis performance is better than the SRAMS method for nonlinear process.5) According to the characters of fluidized catalytic cracking units (FCCU), A sensor fault detection and diagnosis system based on structure optimized SRAMS structured for FCCU is designed and developed. The system has been implemented in a refinery and achieved the requirements of the project.Finally, the paper is concluded with a summary and prospect of future researches.