Fault detection and monitoring system using enhanced principal component analysis for the application in wastewater treatment plant

Abstract

Fault detection and monitoring is essentially important in wastewater treatment to ensure that safety, environmental regulations compliance, maintenance and operation of the Wastewater Treatment Plant (WWTP) are under control. Many researchers have developed methods in fault detection and monitoring such as fuzzy logic, parameter estimation, neural network and Principal Component Analysis (PCA). In studies involving data and signal model approach, PCA is the most appropriate method used in this work. Besides when using PCA, the dimensionality of the data, noise and redundancy can be reduce. However, PCA is only suitable for data with mean constant or steady state data. The use of PCA can also increase false alarm and produce false fault in a plant such as WWTP. Modifications of PCA need to be done to overcome the problems and hence, enhanced methods of PCA are proposed in this work. The enhanced methods are Multiscale PCA (MSPCA) and Recursive PCA (RPCA), which are appropriate for offline monitoring test and online monitoring test, respectively. To see the effectiveness of the methods, they were applied into the european Co-operation in the field of Scientific and Technical Research (COST) simulation benchmark WWTP. The results from the simulation plant were then applied in a real WWTP, IWK Bunus Regional Sewage Treatment Plant (RSTP). The data of WWTP involved are Dissolved Oxygen (DO), Biochemical Oxygen Demand (BOD), Chemical Oxygen Demand (COD) and Nitrate (SNO). In analysis for both plants, faults were detected when the confidence limit is over 95% and confidence limits in the range of 90-95% were considered for alarm region in the data, using Hotelling's T2 and residual. Finally, simulation results of the proposed methods were compared and it was found that the enhanced methods of PCA (MSPCA and RPCA) were able to reduce false alarm and false fault in the analysis of fault detection by 70% for steady state influence and dynamic influence and hence provides more accurate results in detecting faults in the process data.