Machine Learning of the Ultrasound Signal Travel Path Effect in Estimating the Residual Life of the US Army Vehicles

Abstract

In this work, neural networks based machine learning (ML) is proposed for continuously monitoring the residual life of structural parts in the US Army vehicles using ultrasound signals. From a carefully selected set of experiments, a 65 GB data set comprising of ultrasound signals were obtained. Several scales of wavelet decompositions were examined for these signals and the statistical content in a 11-scale wavelet decomposition were selected as the input features to the ML algorithms. Using the knowledge of the statistical outliers and an autoencoder algorithm, the spurious signals coming from the far away placed sensors were removed. Using a five-layer neural network based linear regression ML algorithm, the damage was estimated up to a 95.25% aggregated correlation over the entire range of the damage in various stages of progression even in the presence of variations in the incoming data samples with regards to the ultrasound excitation frequency and in the relative orientation of the ultrasound travel path with respect to the damage orientation. When the variation in the incoming data samples is reduced by controlling for either the ultrasound excitation frequency or the ultrasound travel path, the algorithm estimated the damage up to an aggregated correlation in the range of 99.71-99.86 Accounting for such variations is important because the ultrasound actuator-sensor systems cannot be expected to be deployed at precise locations relative to the damage location and orientation because such information may not be available while the damage is still nascent The ability to differentiate damage signals coming from different structural parts and learning the evolving damage from those signals will lead to increased sustainment of vehicles in the US Army operations.