A multimodal fusion framework for predicting the life cycle of aerobic granular sludge: Integration of fluorescence...

Aerobic granular sludge (AGS) undergoes four stages throughout its life cycle, namely Adaptation, Growth, Maturation, and Disintegration. The dynamic evolution of these stages alters granule morphology and structural integrity, thereby affecting wastewater treatment performance and system stability. Therefore, early identification of potential stage transitions in AGS is important for operational management. Recent advances in artificial intelligence (AI) have provided new opportunities for predicting the states of complex biological treatment systems. In this study, a multimodal machine learning (MML) model for AGS full life cycle evolution analysis was constructed by integrating three-dimensional excitation-emission matrix (3D-EEM) spectral images with operational physicochemical parameters, achieving life cycle stage identification and short-term prediction. The results showed that the MML model achieved an accuracy of 0.95, outperforming the spectral image-based model (0.72) and the numerical-based model (0.90). Multi-time-span prediction further demonstrated stable short-term performance (Δt ≤ 7 days) with F1-scores above 0.85. In addition, improved SHapley Additive exPlanations (SHAP) and guided backpropagation (GBP) were used to analyze feature contributions. Overall, the proposed MML modeling strategy was further extended into a framework for AGS life cycle monitoring, providing early indications of phase evolution and supporting process assessment.