Thomas Swarbrick
ChampionX
Electrical submersible pump (ESP) systems are a common form of artificial lift for oil production wells. The ability to detect broken or twisted shafts is critical for optimizing production efficiency of these ESP systems. This problem is particularly relevant in the Permian Basin due to the harsh operating conditions, including sand production, high gas content and frequent pump restarts. This paper presents a deep learning-based approach utilizing Long Short-Term Memory (LSTM) networks to detect shaft failures in ESP systems in real-time.
The proposed model leverages time-series data collected from both downhole and surface sensors installed in typical ESP systems. By analyzing these temporal patterns, the LSTM model is trained to recognize the dynamic signatures indicative of shaft breakage, even under noisy and variable operating conditions. Experimental results demonstrate that the LSTM-based model can effectively predict shaft failures with high accuracy, outperforming alternative methods such as rule-based and statistical models. This work highlights the potential of deep learning techniques in enhancing the reliability and operational efficiency of ESP systems in challenging field environments.