Explainable ESP Diagnostics Enabled by Physics-Guided Ai

Abstract Electrical Submersible Pump (ESP) failures can cause prolonged downtime and significant replacement costs. AI models detect abnormal patterns in real-time ESP sensor data very well, but with insufficient labels and inconsistent data, they struggle to distinguish between regular operational changes and failure-related events. A physics-guided rule layer, built from years of operator and SME knowledge, adds two benefits: it classifies each anomaly as a specific failure or normal operation, and it generates labeled feedback for targeted retraining when operating regimes shift. Together, these steps reduce false alarms and improve diagnosis in data-limited operations. Sixteen real-time ESP signals were modeled across 35 operational and failure scenarios, including gas interference, partial intake restriction, and broken shaft, using a transient ESP simulator to produce physically consistent responses without large, labeled datasets. Dynamic behavior-mapping functions normalize and weight signals by physical relevance, producing ranked diagnostic scores. The rules were implemented as an event-classification module that integrates with AI surveillance systems and complements existing anomaly detectors, including tags for regular operational changes that form a feedback loop for selective AI retraining. Validation across 35 simulated scenarios showed complete differentiation of event classes when all 16 signals were available. When seven signals were treated as frozen or removed to mimic field data gaps, the module continued to separate ESP-related failures from unrelated anomalies with high reliability, thereby reducing false positives relative to anomaly detection alone. This prioritization supports faster event assessment and reduces unnecessary interventions. Based on typical ESP replacement timelines, avoiding one misdiagnosed failure can prevent about 10 to 14 days of deferred production. The approach was first tested on 20 wells and deployed on +79 wells, with classification outputs reviewed by ESP subject matter experts and field engineers to confirm validity and operational fit. Labels corresponding to regular operational changes were routed to the AI layer to trigger targeted retraining, thereby enhancing model stability and adaptability as operating regimes evolved. Unlike conventional time-series AI classification methods that require extensive labeled datasets and focus on a limited set of failure modes, this solution integrates simulator-consistent signal mapping with a domain-rule layer that captures years of SME knowledge to complement AI anomaly detection. The rules classify failures and regular operational changes and generate feedback labels to retrain the AI anomaly model, enabling reliable, explainable decisions despite sparse sensors and limited labels.