Research on interpretable graph neural network agent model for siltation diagnosis in Urban-Scale sewer systems

Regularly diagnosing sewer siltation (RDS) is critical for informed dredging decisions and mitigating urban overflow pollution. However, traditional sewer siltation diagnosing frameworks face two major obstacles: high detection costs and low diagnostic efficiency. This study aims to develop a diagnostic method based on readily accessible and limited data to diagnose siltation in urban sewer networks, reducing the reliance on manual inspections. We have pioneered an innovative research paradigm that integrates urban spatiotemporal knowledge mapping, which includes: (1) identifying key external environmental factors and intrinsic pipeline characteristics influencing sewer siltation through mechanistic and literature analysis; (2) employing spatial analysis techniques to systematically collect and quantify indicators affecting siltation, constructing a spatial–temporal knowledge database of city scale to support quantitative analysis; (3) developing a graph-based diagnostic agent model that combines a long short-term memory network (LSTM) with a graph convolutional network (GCN) to achieve intelligent diagnosis of sewer siltation; and (4) employing the Integrated Gradients algorithm to enhance the interpretability of the machine learning model and quantitatively assess the contribution of different factors to sewer siltation. A case study conducted in Wuhan demonstrated that sewer siltation thickness is significantly correlated with wastewater flow characteristics, sewer connectivity, catering density, and the normalized vegetation index (NDVI). The accuracy of the proposed diagnostic method is 87.33%, indicating its effectiveness in diagnosing siltation conditions in sewer systems. Overall, this research enhances the accuracy of sewer siltation diagnostics, offers a novel perspective on multiscale analysis of urban sewer systems, and provides valuable guidance for urban-scale dredging planning. By preventing unnecessary siltation inspections, this method saves costs and resources, supporting urban wastewater management in achieving smarter and more efficient operations.