Deep Insight into Electrical Submersible Pump Maintenance: A Predictive Approach with Deep Learning

Abstract

Artificial lift systems play a crucial role in the oil and gas industry by maintaining or enhancing production rates through the conversion of kinetic energy into hydraulic pressure. However, identifying abnormal performance and preventing failures remains a major challenge. Failures occur when key parameters deviate from safe operating conditions, leading to downtime and loss of production volumes. To address this challenge, the objective is to establish repeatable frameworks for constructing predictive models through the utilization of deep learning algorithms. These models provide support for engineering and operational teams through a connected alarm system that sends prescriptive notifications, enabling prompt and informed decision-making. We approach the problem of predictive maintenance for electrical submersible pumps using a multi-task classification method. The different types of faults that can occur in these pumps are treated as individual tasks, and each task is represented by different severity grades to be predicted. To handle this multi-task classification problem, a unique model is trained end-to-end by defining a multi-task loss function. We have evaluated two different architectures for feature extraction, including a 1D CNN and an LSTM with an attention mechanism. The 1D CNN architecture consists of two convolutional and max pool layers, along with batch normalization to facilitate training. The LSTM architecture with attention was found to perform better than the vanilla LSTM in this multi-task classification problem. This study evaluated two architectures for their performance in the predictive maintenance of electrical submersible pumps: 1D CNN and LSTM with an attention mechanism. The results showed that the LSTM with attention using lookback architecture exhibited the best performance and was the easiest to train. The 1D CNN had a comparable performance but exhibited some overfitting in the current configuration. These findings highlight the potential of using 1D CNNs for this application and the importance of attention mechanisms in LSTM models. The maintenance of Artificial Lift Systems (ALs) requires significant resources and is traditionally performed through reactive process monitoring. An automated predictive maintenance solution using deep learning has been developed, including predictive models and a best practices guideline. This work introduces a novel automated system to reduce failures by analyzing real-time sensor data and statistical parameters to predict failures in ALs. The predictive tool supports work over plans including ESP replacement strategies, and reduces production losses.