Deep feature learning for anomaly detection in gas well deliquification using plunger lift: A novel CNN-based approach

Abstract

Timely anomaly detection is critical for optimizing gas production in plunger lift systems, where equipment failures and operational issues can cause significant disruptions. This paper introduces a two-dimensional convolutional neural network (2D-CNN) model designed to diagnose abnormal operating conditions in gas wells utilizing plunger lift technology. The model was trained using an extensive dataset comprising casing and tubing pressure measurements gathered from multiple wells experiencing both normal and anomalous operations. Input data underwent a rigorous preprocessing pipeline involving cleaning, ratio calculation, window segmentation, and matrix transformation. Employing separate pre-training and transfer learning methods, the model's efficacy was validated through stringent testing on new, previously unseen field data. Results demonstrate the model's acceptable performance and strong diagnostic capabilities on this novel data from various wells within the operational block. This confirms its potential to fulfill practical field requirements by offering guidance for adjusting production systems in plunger lift-assisted wells. Ultimately, this data-driven, automated diagnostic approach provides valuable theoretical insights and technical support for sustaining gas well production rates.