Equipment Reliability DNA: Decoding Failure Patterns with eDNA Analytics

Abstract

Abstract Reliable failure data is the starting point for practical reliability engineering in oil and gas operations. In many assets, however, the most useful failure information is recorded in free-text maintenance work-order narratives rather than in consistently populated structured fields. As a result, historical records are difficult to use directly for Reliability-Centered Maintenance (RCM), Root Cause Analysis (RCA), Reliability, Availability, and Maintainability (RAM) studies, bad-actor analysis, and predictive-maintenance development. Before the Equipment Data and Analysis (eDNA) platform was introduced, it was estimated that approximately 80% of maintenance failure records were not immediately suitable for reliability analysis because the failure descriptions and classification fields were incomplete, inconsistent, or dependent on local terminology. Reliability engineers therefore had to read work orders manually, interpret the maintenance narrative, and map each event to an ISO 14224-aligned failure taxonomy. This work consumed engineering time and was difficult to scale across the volume of corrective-maintenance transactions generated by multiple offshore and onshore assets. PTTEP developed eDNA to convert this unstructured maintenance history into reviewable, standardized reliability data. The platform combines SAP maintenance work orders, equipment master data, and operating-context information with a Generative Artificial Intelligence (GenAI) workflow based on Large Language Models (LLMs). The workflow interprets narrative text, proposes failure mode, failure mechanism, failed part, action, and severity classifications aligned with ISO 14224, and routes the output through reliability-engineer validation. SAP, and Plant Data Management System are integrated through a centralized cloud data lake, with results are visualized through Power BI dashboards and Asset Performance Management (APM) tools. Across nine operating assets, eDNA has processed multi-year historical failure records and has materially reduced manual data-cleansing effort. The model currently reports more than 65% classification accuracy and continues to improve through engineering feedback and prompt refinement. More importantly, eDNA has created a common MTBF/MTTR baseline for cross-asset benchmarking, bad-actor identification, and future predictive-maintenance development. This paper describes the business problem, data workflow, AI classification approach, validation process, deployment results, and lessons learned from the eDNA implementation.