Machine learning data analytics based on distributed fiber sensors for pipeline feature detection

Distributed acoustic fiber optic sensors (DAS) enable spatially distributed monitoring of perturbations and contain rich multidimensional information that can be used in structural health monitoring. Machine learning based on physics-based simulations can make a breakthrough in traditional data analysis methods to improve their efficiency and performance, solving a series of problems such as huge data volume, low data processing speed, data signal-to-noise ratio, etc. Here, the relationship of DAS response and corrosion type are studied. First, we present a systematic theoretical study of the potential of direct coupling of quasi-distributed acoustic sensing (q-DAS) with guided ultrasound typically used for real-time pipeline health monitoring. To investigate properties of scattered acoustic waves and the performance of DAS and q-DAS in identifying defects, we use finite element analysis to simulate the response in a variety of pipeline structures including welds, clamps, defect types, and sensor installations representing various corrosion patterns expected in practice. A specific emphasis will be placed upon simulating and modeling pitting corrosion defects and contrasting with other types of corrosion observed in practice. We also aim to compare and analyze signal characteristics due to different kinds of corrosion types and structures, and to enhance machine learning algorithms for detection and size prediction of major pipeline structural changes and corrosion types. Ultimately, results of simulated DAS and q-DAS sensor networks are analyzed by a neural network-based machine learning algorithm for defect identification through supervised learning. To evaluate and improve effectiveness, we estimate model uncertainty and identify features of simulated results that contribute most to the model performance and efficacy.