A generative model-based coevolutionary training framework for noise-tolerant softsensors in wastewater treatment processes

Abstract Data-driven softsensors have gained widespread application in process monitoring and quality prediction, offering advantages over traditional measurement techniques by mitigating their limitations and costs. However, the effectiveness of softsensor models is often hindered by noise in data acquisition, posing significant challenges for model training. To tackle this issue, this study introduces a coevolutionary training framework based on generative models to mitigate the impact of noise corruption. The framework employs a denoising variational autoencoder to extract global and local features from auxiliary data, enhancing population distribution and constructing a deep nonlinear representation to counter noise effects. Additionally, a dual population coding method inspired by evolutionary computation is proposed, enabling the coevolution of network parameters and structure. The proposed multiobjective evolutionary network optimization with denoising strategy (MENO-D) demonstrated exceptional performance in various experiments. On a water quality prediction dataset, the MENO-D-trained softsensor model achieved the lowest prediction error under 10% and 20% noise interference. Further, on the WWTP benchmark dataset across three weather conditions, MENO-D-trained softsensor model exhibited competitive accuracy and robustness.