Integrating fluid–solid coupling domain knowledge with deep learning models: An automatic and interpretable diagnostic system for the silting disease of drainage pipelines

Abstract

An accurate and robust diagnostic model for the silting disease of drainage pipelines is significant for the numerical simulation of urban waterlogging and assessment of risk areas. Various long short-term memory (LSTM) neural networks have been adopted for intelligently diagnosing pipeline siltation owing to the strong correlation between pipeline siltation and flow characteristics over time. However, LSTM often has strong randomness and uncertainty when diagnosing the collocation points of non-training set features and rules and violates the basic fluid–solid coupling physical mechanism. To fully integrate fluid–solid coupling domain knowledge with data-driven neural networks to improve the accuracy of the diagnostic model, this study proposes a knowledge-data collaboratively driven model based on hard constraint projection (HCP) and theory-guided loss function correction (TLFC) for the intelligent diagnosis of drainage pipeline siltation (IDPS). This diagnostic model converts physical fluid–solid coupling constraints, such as the governing equations and boundary/initial conditions, into a mathematical simplification that is easy to handle through discretisation. It adjusts input data sequence and optimises diagnostic results from intelligent algorithms through HCP on the hyperplane and theory-guided loss function correction. The performance of the self-diagnosing system integrating fluid–solid coupling domain knowledge with deep learning models was verified through experiments based on a full-scale prototype. The experiment results indicate that the proposed model had a good ability of robustness to resist the noisy input observations when no less than 600 boundary points were adopted. Compared with typical LSTM, TLFC-based LSTM and HCP-based LSTM, the proposed algorithm achieved highest performance regarding diagnostic accuracy. The mean absolute percentage error of the proposed system was kept below 2%. Furthermore, owing to the knowledge-data collaboratively driven mechanism, the proposed system can extrapolate and accurately diagnose situations outside the training dataset range.