Skill Level Detection in ArcWelding towards an Assistance System forWorkers

Common post evaluation techniques (image based) in the welding industry imply several disadvantages interfering with further production pipeline optimization. The manufacturing processes are extended which increases time, costs and effort. The usage of an online assistance system supporting the welder through helpful information about the current welding quality could resolve these issues. Despite the simpler access to online generated data recorded with embedded (inside the arc welder) or attached inertial measurement units, common challenges like feature extraction and feature selection had to be mastered without losing sight of the decisive arc welding characteristics. Decoding and accentuating physical welding skills mathematically to an understandable data representation supporting the random forest classifier was the key task of this study. Smart arc welders recording 3D accelerometer, 3D gyroscope, voltage and current data streams in combination with supervised data recording sessions enabled a clean data acquisition. Additional features were determined building a 960-dimensional feature set. Concatenating a correlation matrix computation and a forward feature selection improved the balance of the model complexity and its generalization and led to an optimized feature set including 15 features. The final optimization provided an extensive grid search. The resulting model reflected the test welders' expertise even more accurately than the expert and novice labels in the data validated from a welding expert. A quantitative evaluation reached a 72% F1 score and a 76% accuracy.