Welding defect detection is still often performed by the human visual inspection or with non destructive tests. These quality inspections methods can be time consuming and can have an important error rate. In this paper, we propose an approach for the detection of welding faults through the detection of abnormal subsequences of the welding voltage signal. The approach is based on the One-Class SVM with distance substitution kernels. The One Class-SVM has been used in many works for the detection of abnormal subsequences. However, often after transforming the subsequences into a set of features. Nevertheless, finding the relevant features for anomaly detection may be challenging. Dealing with the raw subsequences in distance-based approaches, on the other hand, are known to be effective and can generalize well to different problems but they often suffer from high computational cost, which may restrict their application, especially with the need of real time predictions in an industrial context. We show in this work that the One-Class SVM can be successfully used directly with the raw subsequences. This is achieved by employing distance substitution kernels. These class of kernels has not yet gained widespread adoption for time series anomaly detection. The results show that the approach is both accurate and fast, which makes it more suitable for real-time welding monitoring. We further propose an approach for automatic diagnosis of welding defects.
In this paper, an approach is proposed for the detection of abnormal time series subsequences using
the One-Class SVM with an application to defect detection of two automatic welding processes. The
One Class-SVM has been used in many works for the same purpose but only after transforming the
subsequences into a set of feature vectors. However, finding the relevant features that allow anomaly
detection may be challenging. Methods dealing with the raw subsequences, on the other hand, can be
easily generalized to other problems and are known to be efficient. Nevertheless, they often suffer from
high computational cost since they involve computing the distances to all the reference subsequences,
which may restrict their application, especially with the need of real time predictions in most of the
applications. We show in this work that, contrary to what is indicated in the literature, the One-
Class SVM can be used directly with the raw subsequences. This is achieved by employing distance
substitution kernels, which are a class of kernels that allow to use dissimilarity measures in the
One-class SVM formulation. The results show that the approach has a similar efficiency of those
suggested in the literature that operate with raw subsequences in terms of accuracy. Furthermore,
because it just requires the distances to the support vectors, the approach has a low time complexity
in the prediction phase, making it more suitable for real-time monitoring. We further show that the
One-Class SVM with non-metric dissimilarity remains efficient for aberrant subsequence identification.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.