Abstract: As a complex nonlinear large-scale system, wastewater treatment plant (WWTP) faces problems such as bad working environment, complex control objectives. These problems often lead to fault of the system, so it is urgent to develop the efficient monitoring technology. This study proposed a new fault detection technology, namely, the forecast component-relevance vector machine, which combines the relevance vector machine with the forecastable component analysis. Firstly, the forecastable component algorithm is used to extract features information from the collected data of WWTPs, in order to remove duplicate features and redundant information. Then the relevance vector machine model is trained by offline data. In order to verify the superiority of the proposed method, the forecast component-relevance vector machine and another three methods (RVM, PCA-RVM, ICA-RVM) are used to monitor the wastewater treatment Benchmark Simulation Model 1 (BSM1) platform provided by the International Water Association. Experiments show that the fault detection accuracy of the forecast component-relevance vector machine is higher than the other three methods.

Key words: wastewater treatment, fault detection, relevance vector machine, forecastable component analysis, feature extraction