Anomaly Detection in Robotic Welds - Investigation of LSTM Autoencoder Model Performance

Gas metal arc welding (GMAW) is commonly used for joining metals. Despite the widespread adoption of robotic GMAW, welding errors still occur frequently [1]. They can be costly and time consuming to discover and fix after welding has been completed. This paper presents a method for detecting welding anomalies using unsupervised machine learning on sound data. Earlier attempts have yielded unsatisfactory results, we propose using a long short-term memory (LSTM) autoencoder model to detect welding anomalies in sound from the flux cored arc welding (FCAW) method. The main findings are that the model is well suited, and that it outperforms previous methods.