An effective approach for early fuel leakage detection with enhanced explainability

Leakage detection at service stations with underground storage tanks containing hazardous products, such as fuel, is a critical task. Early detection is important to halt the spread of leaks, which can pose significant economic and ecological impacts on the surrounding community. Existing fuel leakage detection methods typically rely on statistical analysis of low-granularity inventory data, leading to delayed detection. Moreover, explainability, a crucial factor for practitioners to validate detection outcomes, remains unexplored in this domain. To address these limitations, we propose an EX plainable F uel L eakage D etection approach called EXFLD, which performs online fuel leakage detection and provides intuitive explanations for detection validation. EXFLD incorporates a high-performance deep learning model for accurate online fuel leakage detection and an inherently interpretable model to generate intuitive textual explanations to assist practitioners in result validation. Unlike existing explainable artificial intelligence methods that often use deep learning models which can be hard to interpret, EXFLD is a human-centric system designed to provide clear and understandable insights to support decision-making. Through case studies, we demonstrate that EXFLD can provide intuitive and meaningful textual explanations that humans can easily understand. Additionally, we show that incorporating semantic constraints during training in the ANFIS model enhances the semantic interpretability of these explanations by improving the coverage and distinguishability of membership functions. Experimental evaluations using a dataset collected from real-world sites in Australia, encompassing 167 tank instances, demonstrate that EXFLD achieves competitive performance compared to baseline methods, with an F2-score of 0.7969. This dual focus on accuracy and human-centric explainability marks a significant advancement in fuel leakage detection, potentially facilitating broader adoption. • Address challenge of explainable fuel leakage detection at underground storage tanks. • Propose an explainable ML-based method for early fuel leakage with validation support. • Temporal fusion transformer applied for real-time fuel leakage detection. • ANFIS employed to provide intuitive textual explanations. • Proposed method offers intuitive explanations while ensuring detection accuracy.