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

Area of Science:

• Environmental Science
• Water Resource Management
• Artificial Intelligence in Environmental Monitoring

Background:

• Illicit industrial discharges threaten urban water security.
• Tracking pollution sources is difficult due to data scarcity and opaque models.

Purpose of the Study:

• Develop an interpretable framework for pollutant profiling and industrial source tracking (InF-PaT).
• Address challenges of sparse, imbalanced pollution data.
• Enhance transparency and trust in AI-driven environmental monitoring.

Main Methods:

• Fused multi-source data and employed advanced feature engineering, including generative adversarial networks (GANs), Savitzky-Golay filtering, and principal component analysis (PCA).
• Utilized a staged modeling approach with soft sensing for pollutant prediction (R² > 0.91) followed by a multilayer perceptron for source tracking (accuracy > 0.96).
• Incorporated Shapley Additive Explanations (SHAP) for dual-level interpretability and transparent attribution.

Main Results:

• Achieved high-accuracy soft sensing of four key pollutants.
• Enabled precise industrial source tracking with over 96% accuracy.
• Identified key monitoring indicators (e.g., pH) and quantified data contributions (27.98%-38.68%).
• Provided transparent, evidence-based explanations for predictions.

Conclusions:

• The InF-PaT framework effectively bridges the gap between high-performance prediction and regulatory trust.
• Offers a pathway for intelligent and accountable governance of urban sewer networks.
• Demonstrates the utility of interpretable AI in addressing critical environmental challenges.