Machine Learning for Flow Rate and AICD Status Prediction Using Distributed Acoustic Sensing Data

Abstract In this work, we present machine learning (ML) models for predicting bulk flow rate through a pipe and inflow from Autonomous Inflow Control Devices (AICDs) along the pipe using distributed acoustic sensing (DAS) data. The proposed methods consist of four steps: pre-processing the acoustic data, extracting the relevant acoustic features, conditioning the data for use in ML, and utilizing machine learning models for the prediction and classification of the flow rate and AICD status, respectively. The acoustic features used for training the machine learning models involve different time- and frequency-domain features of the measured DAS data. The performance and the relative importance of the proposed features are investigated across various frequency bands using statistical analysis to determine their correlation with the flow-induced signal. The most dominant acoustic signature generated by bulk flow is found to be at low frequency (∼1 – 100 Hz.), meanwhile the acoustic signature generated by inflow through an AICD is dominant at frequencies higher than 1 kHz. Regression and classification models using different techniques including multi-layer perceptron (MLP) and Light Gradient Boosting are supplied with the time- and frequency-domain features to predict flow rate and classify AICD status, respectively. We investigated the effectiveness of processing the acoustic data into features for machine learning through a series of various optional pre-processing steps (including channel averaging, normalizations, selection of specific frequency bands, filtering, etc.), calculation of temporal and frequency band features, and final post-processing for training and testing within a ML model. DAS data acquired from a comprehensive laboratory campaign, comprising both single-phase water and single-phase oil flowing conditions, was used to train and validate the machine learning models. Regression models estimate the bulk flow rate with average percentage error = 7-10 % and AICD flow rate with average percentage error = 4-5 %. The classification model predicts the AICD status with nearly 100% accuracy. This work demonstrates the viability of a ML-based DAS monitoring strategy for bulk flow rate estimation, AICD inflow rate (and allocation) estimation, as well as AICD status evaluation.