Signal Processing and Machine Learning for Effective Integration of Distributed Fiber Optic Sensing Data in Production Petrophysics

Abstract

Distributed fibre optic sensing (DFOS) is progressively being considered in the mix of customary surveillance tools for oil and gas producing assets. Its applications are beyond monitoring of wells for production and reservoir optimization, including detection of well integrity risks and other well completion failures. However, while DFOS can uniquely yield time-dependent spatially distributed measurements, these are yet to be routinely used in formation evaluation and production logging workflows. The large volumes and complexity of time- and depth-dependent data produced by DFOS often require the usage of Digital Signal Processing (DSP) to reduce the amount of stored data and data-driven techniques such as machine learning (ML) for analysis. Distributed sensing data is sampled at high rates; up to 10,000 samples per second and depth for Distributed Acoustic Sensing (DAS), and one sample per minute and depth for distributed temperature sensing (DTS). The high sampling rate in time, across hundreds or thousands of meters, creates a big data problem. Consequently, managing and transferring data acquired in the field to an expert analyst is extremely challenging. Even when these data management challenges are overcome, the amount of data itself is still not suitable for manual analysis. Starting from edge computing for feature extraction, we illustrate the principles of using DSP and ML to effectively handle the challenges of analyzing time-dependent distributed data from DFOS. Results enable integration of DFOS with customary formation evaluation and production surveillance workflows. Feature extraction, a crucial DSP step used to generate inputs to ML, reduces data size by orders of magnitude while ML models analyse continuous data streams from the field. We derive thermal features from DTS data effectively characterizing Joule Thomson effects. Moreover, we combine DTS thermal features with acoustic features from DAS in supervised ML for multiphase downhole inflow predictions. In so doing, we have successfully applied ML on DFOS for real-time detection of sand production, production and injection profiling, and well integrity surveillance. With use cases in a range of well completion types and well operating conditions, we demonstrate an endto- end system of DFOS that effectively integrates DAS and DTS into routine analysis techniques for Formation Evaluation Specialists and Production Petrophysicists.