Cascade Fault Detection and Diagnosis for the Aluminium Smelting Process using Multivariate Statistical Techniques

Abstract

Real-time fault detection and diagnosis for the aluminium electrolysis process is difficult to perform because the process measurements are dynamic, multivariate and limited. This problem motivates the use of multivariate statistical techniques, particularly Principal Component Analysis (PCA) and Partial Least Square (PLS), in this research. The objective of this research is to design and develop a new PCA/PLS based system for aluminium smelting process that can detect and diagnose faults effectively. As a result of the development of the new system, the main research question is: Does a system based on PCA and PLS, effectively detect and diagnose faults in aluminium smelting process? In order to address the above question, the research involved several steps. A taxonomy of aluminium process fault detection and diagnosis systems was first identified with four key elements: techniques, knowledge, usage frequency and mode of results. Pilot studies were then run to address selection of variables and dynamic behaviour. Finally, the new 'Cascade' fault detection and diagnosis system was developed in four stages: (1) detecting faults using Multiway-PCA (MPCA), (2) discovering abnormal patterns using MPCA, (3) diagnosing faults using MPCA and Multiway-PLS (MPLS), and (4) integrating the functions of detection and diagnosis to develop a new system. The evaluation of the new system using aluminium smelting data shows that this system is effective to detect and diagnose faults. This research has contributed to the development of fault detection and diagnosis systems of the aluminium smelting process by investigating the application of multivariate statistical techniques. Firstly, a new design for a MPCA/MPLS based system, in which the alumina feeding cycle was treated as a batch operation, has created a new way in which to consider the dynamics of the process during alumina feeding. Secondly, the occurrence of cascade-like patterns during anode changing has been solved by using multiple models. Thirdly, abnormal patterns based on alumina concentration versus resistance curves have been discovered. Finally, the developed fault detection and diagnosis taxonomy has enabled researchers to communicate the key elements of the system clearly. The application of this system is expected to assist operators to detect faults and diagnose anode faults, effectively.