Abstract
Downhole tool failures, including the breakage of polycrystalline diamond compact bits, the stalling of drilling motors, and the malfunction of measurement-while-drilling equipment, remain a primary cause of non-productive time, costly equipment damage, and heightened safety risks in both oil and gas drilling and water well operations.Traditional reactive maintenance strategies, which rely on post-failure analysis and simple threshold-based alarms on parameters such as weight on bit or vibration magnitude, have proven inadequate for anticipating sudden tool breakdowns, as they fail to capture complex precursor patterns and often produce an excessive number of false alarms.This paper proposes an edge artificial intelligence framework designed for real-time anomaly detection, continuously monitoring multivariate sensor streams that include triaxial vibration, torque, weight on bit, and rotational speed directly at the rig site.The methodology combines an unsupervised long short-term memory autoencoder, which learns the normal dynamics of drilling and flags deviations through reconstruction error, with a lightweight gradient boosting classifier that validates potential anomalies and reduces false positives.This hybrid model is quantized and deployed on an edge computing device (NVIDIA Jetson Orin Nano), enabling inference latencies below 30 milliseconds without relying on cloud connectivity.Experimental evaluation using field data from twelve wells demonstrates that the system detects pre-failure anomalies 15 to 30 seconds before critical tool failure, achieving 92% precision and reducing false alarms by 65% compared to conventional threshold methods.Consequently, edge-based predictive maintenance not only minimises downtime and fishing costs but also significantly enhances the operational resilience of drilling rigs as critical energy and water infrastructure.