Unsupervised Autoencoder-Based Anomaly Detection Under Limited Failure Data for Oil Industry

This paper presents a deep autoencoder-based methodology for unsupervised anomaly detection in centrifugal pumps under limited failure data conditions, focusing on real-world applications in the oil and gas industry. The autoencoders are designed to reconstruct input sequences from SCADA data, including pressure, temperature, vibration, and electric current, with anomaly detection derived from the reconstruction error. Anomaly detection is based on a moving-minimum anomaly score (MMAS), designed to suppress transient fluctuations and highlight consistent deviations. Each autoencoder was trained on pre-fault operational windows and evaluated on sequences encompassing registered failures. The autoencoder detected anomalies up to 74 days before the actual fault detection by the maintenance team, with consistent warnings of two to three days in additional cases. The architecture, training parameters, and data preprocessing steps are proposed, including pressure differential computation and idle-state filtering. The findings demonstrate the viability of unsupervised deep learning in industrial fault detection scenarios characterized by limited failure annotations in an industrial environment.