Deep learning-driven detection of municipal sludge moisture content based on feature images of compressed slices.

Traditional sludge moisture content detection methods suffer from response lag and insufficient accuracy within the critical dewatering range of 70%-85%. To address these limitations, this study proposes a novel rapid detection strategy that integrates the rheological properties of non-Newtonian fluids with deep learning. By analyzing the dynamic evolution of surface textures under a controlled thermal field, we established a robust correlation between sludge surface morphology and moisture content. Systematic investigation identified the optimal experimental conditions as circular slices of 1 mm thickness heated at 140 °C. Among the Convolutional Neural Networks evaluated, the C3D model was selected for its superior capability in extracting spatiotemporal features from the drying process. The resulting 16-category classifier achieved 81.25% accuracy on an independent test set. Furthermore, field validations conducted at six wastewater treatment plants across diverse climatic regions (Beijing, Harbin, Shenzhen) demonstrated the method's robustness and engineering applicability. This research provides a non-contact, high-efficiency pathway for real-time monitoring and intelligent control of sludge deep dewatering.