From pollutant profiling to source attribution: An interpretable staged machine learning framework for sewer...

Illicit industrial discharges into urban sewer networks pose a severe threat to water security, yet source tracking is hindered by data scarcity within the network and model opacity. This study developed an interpretable framework for pollutant profiling and industrial source tracking (InF-PaT). To address the challenge of sparse and imbalanced pollution data, InF-PaT fused multi-source data and introduced a comprehensive feature engineering workflow that leveraged a generative adversarial network to augment the training dataset. This workflow was complemented by Savitzky-Golay filtering and principal component analysis to extract robust features from high-dimensional data. To dismantle the black-box barrier, a staged modeling strategy was proposed. InF-PaT first achieved high-accuracy soft sensing of four key pollutants (R² > 0.91), translating complex raw data into physically meaningful indicators. These predictions then served as interpretable intermediate variables for a multilayer perceptron to perform precise source tracking (accuracy > 0.96). The framework's transparency was further fortified by Shapley Additive Explanations, thereby providing dual-level interpretability. This approach globally identified key monitoring indicators (e.g., pH) for soft sensing and source identification and quantified multi-source data contributions (27.98%-38.68%), while establishing transparent, evidence-based attribution explanations for each individual prediction. InF-PaT thus bridges the critical gap between high-performance prediction and regulatory trust, offering a robust pathway toward the intelligent and accountable governance of urban sewer networks.