Big Data Analytics for Advanced Fault Detection in Wastewater Treatment Plants

Fault detection in wastewater treatment plants (WWTPs) presents difficult challenges, highlighted by the nonlinear, nonstationary nature of operations and the varying fault intensities that often are neglected. While big data analytics promise transformative results, they come with their own challenges in process monitoring such as handling vast datasets, ensuring real-time responsiveness, and coping with imbalanced data distributions. With our 609-days simulation of the Benchmark Simulation Model 2 (BSM2), yielding datasets as expansive as 876,960 samples for each of the 31 measurements considered here, the inherent issues become more obvious. To address this challenging process monitoring problem, our research introduces a novel fault detection framework, handling both imbalanced data distribution and big data complications. The core of this framework includes two critical components. The first one is a wavelet-based feature analyzer which utilizes the wavelet energy and entropy information for each measurement to extract the most valuable and critical features. The other element is the enhanced neural network classifier which deals with imbalanced data distribution. This classifier partitions the data into multiple segments, subsequently determining the BSM2 operational condition (i.e. normal or fault) for each distinct segment. The proposed detection framework has demonstrated the capability to accurately identify the operational condition of the large-scale BSM2 dataset, achieving a False Alarm Rate (FAR) of less than 10%. The promising results obtained from this framework can facilitate future research on developing digital twins for WWTPs.