Machine Learning Provides Reliable ESP-Condition Monitoring

Abstract

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 229219, “Translating Physics Into Intelligence: Machine Learning For Reliable ESP-Condition Monitoring,” by Temirlan Zharkynbek, Sergio A. Caicedo, SPE, and Cristina Hernandez Labrador, AIQ, et al. The paper has not been peer-reviewed. This research focuses on combining physics-based expert rules with machine learning (ML) to improve the detection of failure-related events in electrical submersible pumps (ESPs). The goal is to create a generalized framework for abnormal-event detection that addresses the challenge of limited labeled data and sensor data, providing a practical solution for ESP management and maintenance planning. By using ML as a denoising layer over the physics rules, the method bridges rule-based and data-driven approaches. ESP reliability remains a decisive lever for production uptime, yet unplanned shutdowns still account for a significant share of lifting costs. Years of field digitalization have turned each pump into a high-frequency data source, but limited ground-truth failure labels and heterogeneous sensor layouts still hinder the move from reactive maintenance to predictive decision-making. This study tackles that constraint by treating physics knowledge as a source of weak supervision. A library of expert rules—encapsulating known degradation signatures—runs continuously on sliding windows of normalized sensor trends, producing numeric scores that act as soft labels. Rather than stopping at these rule outputs, the authors trained a model-agnostic supervised learner to denoise and generalize the aggregated scores across wells with varying sensor availability and operating envelopes. The resulting model serves as a single, field-deployable anomaly index that requires no manual labeling while retaining the interpretability of its physics roots. Previous work introduced a two-stage hybrid pipeline for ESP-failure-related abnormal-event handling. The two stages were as follows: 1. Semisupervised detection: A long short-term memory (LSTM) autoencoder trained on expert-labeled “normal” periods flags any reconstruction-error spikes as abnormal events. 2. Physics-rule-based classification: A library of physics-inspired rules diagnoses each flagged event. That workflow proved accurate but labor-intensive because subject-matter experts had to annotate large “normal” spans initially. The current contribution is to invert the sequence such that the following takes place: - Physics rules are run continuously (on every sliding window, not only alarms). - Numeric scores are treated as weak labels. - A supervised model is trained to denoise and generalize these scores, eliminating the need for manual “normal” tags. The result is a label-efficient, physics-guided ML framework that delivers earlier, cleaner abnormal-failure-related event detection. Motivation for Physics‑Based Diagnostic Rules. Early alarm systems successfully highlighted operational changes, such as pump-speed adjustments or choke-setting shifts that affect wellhead pressure, and revealed issues beyond the pump, enabling comprehensive monitoring of the entire well system rather than focusing solely on the ESP. For each predefined alarm category, the rule engine calculates a similarity score that quantifies how closely the current sensor-trend pattern aligns with that category’s signature. The category with the highest score is flagged as the primary diagnosis for the window. Ranking alarms in this way surfaces the most-probable root cause while still preserving secondary scores for additional context, thereby reducing ambiguity when different failure mechanisms share overlapping symptoms.