Predicting Just in Time: Prognostic Health Management for Oilfield Equipment.

Abstract

Delays in the hydraulic fracturing stimulation process due to equipment issues may result in Non-Productive Time (NPT) and often associated with high costs. Both operators and service companies are interested in implementing a system to predict these events in advance so that they can proactively react and prevent failures before they occur. Our objective was to develop, implement, and demonstrate the capabilities of such a system. Based on historical data, we identified the most NPT-prone subcomponents of stimulation equipment. We combined failure reports and maintenance records with sensors data to train Machine Learning (ML) models. The final solution was deployed on the cloud to predict remaining useful time (RUL) in real-time and to start specific notifications sending whenever RUL dropped below the preset value. We analyzed 20,000+ failure records and identified the most NPT-prone equipment, subcomponents, failure modes, and root causes. We found out that root causes are very scattered. We evaluated the applicability of physics-based and data analytics approaches for Top-10 root causes by NPT duration and NPT count. When NPT count per root cause is low (<10), the trained ML models had insufficient accuracy. In this case, only subject matter experts and understanding of physics behind performance deterioration allowed us to boost data-driven models’ accuracy. We started with testing on historical data. Then we continued testing on real-time data without field intervention. This stage was required to check and to build trust in the predictive ability of the models. After initial testing, we eliminated some of the models and switched to monitoring with field interventions. For six months of operating, the model helped to decrease the NPT related to the selected subcomponent by 20 times. This work presents the process of enabling Intelligent Diagnostics for hydraulic fracturing equipment by implementing failure identification and prediction models. It combines both domain knowledge (failure modes of the equipment, sensors data) and recent advances in ML (time series analysis, data processing, and feature engineering). The developed system to be further extended to different equipment, as the workflow described in the present study is not specific to particular equipment.