Automated Detection of Rig Events from Real-Time Surface Data Using Spectral Analysis and Machine Learning

Abstract

The authors present a method for automated, high-fidelity detection of rig events characterized by complex temporal signals, such as downlinking, or wave-induced heave affecting floating rigs. These can adversely impact other systems utilizing relevant data streams, for example downlinking via mud pulse telemetry can interfere with detection of pressure changes that might indicate hole cleaning problems. Identifying these events using classification techniques applied to time-domain data is difficult, hence spectral (frequency domain) techniques, combined with Machine Learning (ML), were applied to solving this problem. Surface measurements from a variety of wells, fields, regions, service companies and operators were used to develop and validate the detection methods. Data was preprocessed using time-frequency analysis, and then input to discriminative classifiers to identify rig events of interest. For downlinking state detection, high recall and precision scores (both >93%) were achieved on independent holdout well data, and thus false positive rates were low. Successful detection was demonstrated on wells separate from the training data, hence the method is expected to generalize to new well operations. The detection method enhances situational awareness, and can actively support other software in improved automated decision-making by providing operational context in real-time, such as suppression of false warnings from monitoring pressure or modelled ECD for detecting signs of poor hole cleaning. These techniques are not limited to downlinking or heave detection, and can be applied more generally to scenarios with complex periodic signals.