Fiber-Optic Acoustic-Based Disturbance Prediction in Pipelines Using Deep Learning

Abstract

The problem of detecting nonstationary disturbances in a pipeline is demonstrated using a predictive framework based on high spatial resolution fiber-optic acoustic sensors. We show that the root-mean-square (RMS) acoustic power is related to flow and density changes in the fluid. However, in practice, fluid parameters are not known at the resolution of the acoustics. In an experimental study, we trained long-short-term memory (LSTM) networks to exploit hidden patterns in an acoustic time series to predict the RMS acoustic power. We found LSTM perform efficiently and shows improvement over baseline neural network predictor, and its strength lies in discriminating sequential order from spatial input data. The system is verified on 25 m resolution fiber-optic acoustic data. Results show promise in predicting anomalous disturbances despite unknown pipe and fluid parameters..