Deep learning approaches for force feedback based void defect detection in friction stir welding

• Published in: Journal of advanced joining processes 2022-06, Vol.5, p.100087, Article 100087
• Main Authors: Rabe, P., Schiebahn, A., Reisgen, U.
• Format: Article
• Language: English
• Subjects: Deep learning; Force feedback; Friction stir welding; Non destructive quality monitoring; Void detection
• ISSN: 2666-3309; 2666-3309
• DOI: 10.1016/j.jajp.2021.100087

The Friction Stir Welding (FSW) process is known as a solid state welding process comparably stable against external disturbances in its steady state. Therefore, the process is commonly applied with fixed welding parameters utilizing axial force control or position control strategies. External and internal process disturbances, introduced by the workpieces, gap tolerances, tool wear or machine/tool inadequacies are rarely monitored and no conclusion towards the weld seam quality is drawn. In most applications the sole process monitoring feature is the axial force and quality control measures are often only applied to external weld seam features. Based on prior analytic process analysis and quality models based on analytic algorithms, machine learning approaches are employed to find suitable quality criteria to detect internal weld seam defects (void defects) over a wide range of welding parameters. The recorded welding force data is examined using Long Short-Term Memory (LSTM) Networks, Bidirectional Long Short-Term Memory (BiLSTM) Networks and different versions of Convolutional Neural Networks (CNN), differing in Layer architecture and Node count (and thus resulting complexity). In an effort to reduce data and speed up development and training, manual feature extraction was performed. Features were selected based on the cyclical nature of the FSW-process and prior research into in-plane process force data quality relevancy. The features as well as raw recorded force data were used to train previously mentioned networks. Overall the viability of the approaches was shown by a high rate of correct void defect classification, with CNNs exceeding 93% test set classification rate. Network training had to be closely monitored as a small dataset and large number of input features increased the risk of overfitting.