Unsupervised Deep Learning for Instrumented Infrastructure: A Case Study

Abstract

Deep learning methods have recently shown great success in numerous fields, including finance, healthcare, linguistics, robotics, and even cybersports. Modern computer hardware makes it possible to train very large deep neural networks that can handle datasets of high dimensionality and rich structure. The area of unsupervised deep learning addresses modelling underlying data distributions when no ground truth labels are present. Probabilistic deep learning studies models that provide a measure of uncertainty for a prediction. We present a case study where a state-of-the-art family of deep probabilistic unsupervised models called Variational Auto-Encoder (VAE) is applied to data accrued from a network of fibre-optic sensors installed within a composite steel-concrete half-through railway bridge. Our goals were to: 1) automatically characterise the response of the bridge under stimuli based on sensor measurements, and 2) based on this characterisation, determine when a train passes across a bridge. Based on the VAE model, we present an algorithm to automatically identify the "train event" points. Our case study illustrates how state-of-the-art deep learning methods can be applied to a civil infrastructure engineering problem without directly modelling the physics of the objects or using any labels.