Key points are not available for this paper at this time.
Activated sludge bulking critically undermines the operational stability and effluent quality of wastewater treatment plants (WWTPs), posing significant environmental and economic sustainability challenges. While microscopic image analysis offers the potential for early detection, conventional techniques suffer from low adaptability and delayed predictive capabilities. This study presents an artificial intelligence (AI)-driven framework leveraging convolutional neural networks (CNNs) for early detection of sludge bulking from microscopic images. A data set of 1200 Gram-stained images was generated by inducing filamentous bulking and recovery in a lab-scale sequencing batch reactor (SBR), covering six distinct classes: Normal, Early stage, Mild, Severe, Recovering, and Recovered. Among the evaluated CNN architectures, InceptionV3 demonstrated optimal performance, achieving a validation accuracy of 92.88% ± 2.81% and a macro-F1 score of 92.85% ± 2.82%. Crucially, progressive unfreezing during transfer learning enhanced model generalization while optimizing computational efficiency (overall accuracy: 87.38% ± 0.94%; AUC: 0.98). Class activation mapping (CAM) revealed sensitivity to critical morphological precursors such as filamentous branching, which were detectable 2–3 days before traditional sludge volume index (SVI)-based thresholds were breached. This CNN-based approach enables real-time automated bulking diagnosis, improving accuracy while reducing reliance on manual expertise and delayed SVI measurements, thereby supporting sustainable and smart WWTP operations.
Gao et al. (Mon,) studied this question.