Predicting aerobic granular sludge structural instability: An intelligent early-warning framework integrating convolutional neural network and fluorescence fingerprint features.

Aerobic granular sludge (AGS) was recognized as an innovative alternative superior to activated sludge processes, yet its development has been constrained by structural instability and the lack of early-warning methods for critical states. To address this limitation, an intelligent early-warning model (EPS-ResNet) based on multi-view convolutional neural networks was developed. This model achieved a 6±1-day advance prediction of AGS structural destabilization (accuracy:97.6%) by analyzing fluorescence characteristics in Excitation-Emission-Matrix Spectra (EEMs) of loosely/tightly bound extracellular polymeric substances (LB-EPS/TB-EPS). Through occlusion sensitivity analysis and fluorescence region segmentation, Region I (tyrosine-like proteins) of TB-EPS, Region IV (soluble microbial metabolites) of TB-EPS, and Region IV of LB-EPS were identified as the top three contributors to early-warning efficacy. Integrated microbiome analysis revealed that the superior early-warning performance of the model was primarily attributed to the capacity of key EEMs regions (I, IV, and V) of EPS to sensitively capture dynamic succession of dominant phyla and functional genera in AGS following shocks. Correlation analysis conducted through the Mantel test demonstrated that key dominant phyla responding to the model included Bacteroidota and Patescibacteria, while critical functional genera comprised Flavobacterium, Pseudazoarcus, Thauera, and Candidatus_Competibacter. An early-warning framework for abnormal states of AGS integrating scalability and mechanistic interpretation was developed in this study. This framework was demonstrated to be applicable not only for the early warning of AGS structural destabilization, but also extensible to the early detection of anomalies in biological treatment systems, thereby promoting the transformation of water treatment process operation and maintenance toward digitalization and intelligentization.