AWANDT: Assessing Welding Anomalies via Non-Destructive Tests

Detecting defects that may arise in weldings used in either critical or non-critical industrial applications is an extensive area of active research. As such, non-destructive tests are often required, mainly to preserve the integrity of normal samples while discarding defective ones. Amongst these types of tests, the most relevant are visual inspections which, however, require a considerable effort by a domain expert. Still, as these tasks require relevant efforts and could be biased by subjectivity and inexperience, the development of automated, objective tools that provide early warnings about the occurrence of one or more anomalies has become essential. To deal with these issues, this work proposes a framework for detecting anomalies in linear aluminum welding. Specifically, the framework starts by acquiring the three-dimensional representation of the welding using a 3D laser profiler. Afterward, a semi-supervised approach is followed by training a deep autoencoder on non-defective data samples, allowing the model to learn a mapping function that successfully reconstructs non-defective weldings with a small reconstruction error while providing a large error for abnormal data samples. Hence, by applying a proper threshold on the reconstruction error, which can be estimated via statistical analysis, the framework can provide real-time early warnings concerning surface defects on the welding. Different structures of autoencoder have been tested and found to reach an F1 score of over 90%.