A Novel Dynamic Fault Detection and Diagnosis Algorithm based on PCA and RF – An Application on a Water Distribution System in Northern Colombia

• Published in: Procedia computer science Vol. 251; pp. 777 - 782
• Main Authors: De-la-Cruz, Aldair, Portnoy, Ivan, Lombana, Antony
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 2024
• Subjects: Fault Detection; Fault Diagnosis; PCA; Fault Detection; RF; Fault Diagnosis; PCA
• ISSN: 1877-0509; 1877-0509
• DOI: 10.1016/j.procs.2024.11.184

The proper operation of water distribution systems (WDS) is essential to the correct functioning of cities and the well-being of their inhabitants. Hence, faults in these systems threaten the continuity and quality of the vital water supply while representing a significant economic and environmental issue, as water losses due to pipe breaks and leakages are substantial worldwide. Data-driven and machine-learning approaches for fault detection and diagnosis (FDD) have proven to be helpful tools to help manage WDS, providing reliable monitoring capabilities. The academic literature shows the emerging relevance of such methods. As for the Colombian context, a literature review shows a need for such applications in real distribution systems. This research proposed a novel FDD method based on PCA and RF, which can capture high-order process dynamics and provide a reliable monitoring capacity. The method feeds on the process sensor signals and augments the input data matrix by including the variables’ time differences. The proposed method was validated on a real WDS in Barranquilla, Northern Colombia. The method was assessed using four major historical datasets. Validation results showed that the method can accurately detect mild and severe faults while controlling for the false alarm rate (about 3%) when monitoring non-faulty operation conditions. Furthermore, the method achieved a robust diagnostic accuracy (99.57%) with the validation data. The proposed method's performance for monitoring tasks demonstrates its effectiveness, constituting a valuable tool to aid operators in real-time early FDD. Future research will focus on including multi-source data (other than the currently available sensor signals) and the effects of time-drifting and process aging.