Maximum Likelihood-Based Neural Network for Monitoring Wastewater Treatment Processes

Wastewater treatment processes (WWTPs) contribute significantly to water recycling and energy conservation, because their failures can lead to the potential risk of adverse environmental impacts. WWTPs are acknowledged to be largely nonlinear systems with a series of biological reactors. Additionally, the presence of process uncertainties and measurement noises makes it difficult for the purpose of fault detection. This paper focuses on the data-driven realization of fault detection for sludge bulking and toxic impact in WWTPs under measurement noises. The fault detection models employ nonlinear auto-regressive with exogenous input (NARX) neural networks to construct the residual generator through the maximum likelihood approach. The maximum likelihood-based neural network can optimize both the model parameters and covariance matrix related to stochastic noises, based on which the model delivers more robustness performance against measurement noises than the existing work. There fore, the nonlinearity and measurement noises of WWTPs can be processed simultaneously. The superiority and effectiveness of the maximum likelihood-based neural network including NARX neural networks, long short-term memory neural networks, and gate recurrent unit neural networks are validated through the benchmark simulation model no.1, where the successful detection rates for de tecting sludge bulking faults and toxic impact faults are proved to be significantly higher than the least squares-based methods.