A Deep Fuzzy Feature Extraction Method for Non-stationary Wastewater Treatment Processes

To address the challenge where static models struggle to capture state transitions in non-stationary wastewater treatment processes caused by influent load fluctuations and complex biochemical reactions, a deep fuzzy feature extraction method with a dynamic update mechanism is proposed. First, a supervised stacked autoencoder is designed to learn latent features; by embedding target label information into the loss function, the model prioritizes the extraction of features highly correlated with effluent quality. Second, a fuzzy learning inference layer is embedded within the feature space to map abstract features into membership functions and logical rules, thereby enhancing model interpretability. Finally, a state-evolution-driven dynamic updating mechanism is constructed, and a comprehensive nonstationarity index is developed to quantify operational condition variations. Based on the quantified changes, model parameters are adaptively adjusted to improve tracking capability under nonstationary conditions. Experimental results demonstrate that the prediction accuracy is above 0.90, and the proposed model outperforms basic recurrent neural networks and attention-based time-series prediction models in terms of feature representation.