Monitoring large-scale industrial systems for wastewater treatment processes with process noise using data-driven NARX approach

Abstract

Wastewater treatment processes (WWTPs) are large-scale systems comprising multiple biological reactors, which are essential for preventing water pollution and promoting water reuse. Safety assessment and accurate process monitoring are crucial for maintaining the effluent quality of WWTPs. However, the presence of uncertainties and process noise degrades the performance of fault detection models, posing significant challenges to reliable monitoring. This paper proposes a data-driven fault detection framework for monitoring failures in wastewater treatment processes affected by impulsive noise. The fault detection model employs nonlinear autoregressive with exogenous input (NARX) neural networks to construct the residual generator with the aid of robust continuous mixed $p$-norm optimization. Robust continuous mixed $p$-norm combines multiple error $p$-norms to enhance the cost function with diverse error information, minimizing it to produce adaptive gains that adjust the training gain based on data quality at each step. When impulsive noise occurs, the correction term for parameter estimation approaches zero, enabling the model to achieve greater robustness against impulsive noise compared to existing methods. Additionally, the fault detection model incorporates an adaptive moment estimation-based variable-step algorithm to enhance convergence by adaptively adjusting the learning rate. The proposed method is applied to the benchmark simulation model no. 1, and experimental results demonstrate that it achieves accurate detection rates for monitoring WWTPs.