Exploring Modern Feature Extraction Techniques for Improved Offshore Fault Detection in Oil and Gas Operations

Abstract Fault detection in offshore oil and gas operations is vital due to the high risks and costs associated with unforeseen failures. This study benchmarks a variety of feature extraction techniques for this critical task, with a particular focus on RandOm Convolutional KErnel Transform (ROCKET), aiming to improve the detection and prevention of failures. ROCKET is evaluated alongside wavelet-based features, statistical features, and those derived from self-supervised learning (SSL) approaches, all using a consistent linear evaluation protocol. This standard method is commonly employed to assess the quality of features in representation learning. Experiments were conducted using a Stratified Group K-Fold (SGKF) cross-validation framework on a multiclass classification problem. Results show that ROCKET consistently outperformed all baseline models across multiple metrics, including macro F1 score, accuracy, balanced accuracy, and Matthews Correlation Coefficient (MCC). Notably, ROCKET achieved a balanced accuracy significantly higher than that of a recent transformer-based model while maintaining a comparable F1 score. These findings underscore the effectiveness of ROCKET's training-free feature extraction and its ability to generate highly discriminative representations suitable for downstream classification. Moreover, the use of a simple linear classifier highlights the strength of the extracted features, suggesting further gains may be possible with non-linear classifiers. This study reaffirms the value of lightweight and scalable methods for fault detection in offshore oil and gas operation.