Evaluation of machine learning techniques for structural health monitoring using ultrasonic guided waves under varying temperature conditions

Abstract

The implementation of machine learning methods for structural health monitoring applications has proven to be very effective, especially in detecting damage and compensating for environmental and operational conditions. The use of guided waves in this area has also shown to be a powerful tool due to their sensitivity to structural changes in the propagation medium. In this work, two strategies for detecting damage and distinguishing their positions and for dealing with temperature variations without an additional classical temperature compensation technique are investigated. For this purpose, four unsupervised dimensionality reduction learning methods were used and compared: Principal Component Analysis, Kernel Principal Component Analysis, t-distributed stochastic neighbour embedding and Autoencoder. The first strategy (score plot) consists of using the latent dimensions directly to distinguish the data points of different states of the structure, and the second (DI plot) proposes a method to use Q- and T2-statistics, which have been proposed in previous work for PCA, computed using the compressed representation of the monitoring data. To this end, monitoring data from intact and damaged states of a 500x500x2 mm plate of carbon-fibre–reinforced polymer recorded by 12 piezoelectric transducers at different temperatures are examined. As a reversible damage model, a 10 mm thick aluminium disc is placed at four different locations on the plate. The results primarily show the success of the methods used with DI plot in detecting damage regardless of varying temperature. The autoencoder in the first strategy also demonstrates promising performance in detecting and distinguishing the position of the damage, even in the presence of varying temperature conditions.