Diagnosing activated sludge dysfunction: A full-scale wastewater treatment study.

Activated sludge systems treating municipal wastewater frequently experience performance dysfunction, yet conventional diagnostics rely on isolated parameter monitoring that is difficult to integrate for prioritising interventions. Here, we report an integrated diagnostic framework combining machine learning interpretation, microbial community profiling, and multi-parameter assessment to characterise and prioritise the likely drivers of critical operational failures in a full-scale plant (50,000 m3/day). SHAP (SHapley Additive exPlanations) analysis of stacking ensemble models revealed chemical oxygen demand (COD) removal as the dominant predictive feature (importance 0.261, 2.3-fold greater than total nitrogen), providing an objective basis for prioritising candidate interventions; we emphasize that SHAP quantifies predictive contribution and does not by itself establish causation. Converging evidence pointed to extreme substrate limitation under an over-aged biomass regime as the most likely primary driver of dysfunction: excessive mixed liquor suspended solids (MLSS, 7010-9313 mg/L), severely depressed specific oxygen uptake rate (SOUR, 0.029-0.172 mg O2/(g MLSS·h), only 1-2% of reference values), prolonged solids retention time (SRT, > 45 days), and critical food-to-microorganism (F/M) ratio (0.017 kg BOD5/(kg MLSS·day)). Pronounced methylotroph enrichment (17.5-25.7% vs. < 2% typical) and metazoan proliferation (78-94% occurrence, Sludge Biotic Index = 8.2) were consistent with aged, low-loaded sludge conditions. The convergence of multiple independent lines of evidence supports a set of candidate optimization strategies, proposed here as testable interventions rather than validated solutions: biomass reduction to 3500-4500 mg/L, stepwise methanol supplementation (20-25 mg/L; ≈ US$0.008 per m3 treated, ≈ US$130,000 per year, justified by a denitrification carbon balance), and advanced respirometry-based monitoring. This explainable-AI-assisted framework provides a reproducible, transferable approach for prioritising wastewater-treatment optimization from routine monitoring data; the machine-learning analysis identifies predictive associations rather than causal mechanisms.