Classification of acoustic emission data from buckling test of carbon fibre panel using unsupervised clustering techniques

Acoustic emission is widely used for mechanical diagnostics and to characterise damage in composite materials. Distinction between different damage mechanisms is still one of the major challenges and remains an unresolved issue. The objective of cluster analysis is to separate an acoustic emission data set into multiple classes that reflect different acoustic emission sources. This article is concerned with the implementation of unsupervised clustering techniques to classify acoustic emission transients from a carbon fibre laminate buckling test. A new approach to signal feature extraction was utilised, whereby principal components provide signal features that represent the greatest data variance while remaining linearly uncorrelated with each other; feature selection was undertaken using a hierarchical clustering method and finally a cluster analysis was performed using k-means and Fuzzy C-means techniques. The aim of the work is to reduce the data required in the classification process, thereby reducing the processing time and computational power required, without significantly affecting the classification result. Thus, an approach which is more suited to online processing, allowing fast and efficient processing and storage of data is provided. The proposed unsupervised clustering analysis was able to separate acoustic emission signals into two different clusters that were correlated to the damage mechanisms observed. The results show that the clustering groups have a good fit with ultrasonic C-scan and digital image correlation strain data. The application of a clustering process that uses the most effective acoustic emission features as input data is an objective method, and this investigation shows that it may be a useful complement in the field of non-destructive evaluation.