Prediction of Activated Sludge Sedimentation Performance Using Deep Transfer Learning

Abstract

Activated sludge (AS) bulking is a significant challenge in AS processes, and therefore, predicting the settling performance of AS is essential to maintaining the long-term stable operation of wastewater treatment plants (WWTPs). In this study, AS samples taken from 42 WWTPs and three laboratory reactors were imaged and labeled with sludge volume indexes to predict the sedimentation performance of AS based on deep learning models. A tagged AS image database was established with 105,695 images. Comparing five different deep learning algorithms suggested that the ImageNet-trained lightweight MobileNetV3-Large model obtained optimal performance. This model achieved an accuracy of 98.06%, with the F1-Score values for AS nonbulking, limitedbulking, and bulking categories of 98.8, 95.4, and 98.4%, respectively. These findings demonstrate that this model can precisely predict the process of AS transitioning from nonbulking to bulking and provide early warning during the limited-bulking stage to facilitate timely regulation by WWTPs.