Quality assessment of resistance spot welding process based on dynamic resistance signal and random forest based

A scheme for online quality monitoring of resistance spot welding (RSW) process is proposed to effectively determine the rate of spot weld quality. In this work, the random forest (RF) classification featuring with dynamic resistance (DR) signals which were collected and processed in the production environment was carried out. The obtained results demonstrated that the constructed RF model based on DR profile features adequately distinguished high-quality welds from the other unacceptable welds such as inadequate sized welds and expulsions. Variable importance evaluation of RF was implemented against the input features. It showed that two DR slopes for nugget nucleation and growth (v₂, v₃) and dynamic resistance (R_γ) in the final half cycle play the most significant roles in achieving more accurate results of classification, while absolute gradient ∇_max is useful in detecting minor expulsion from pull-out failure. In addition, shunting effect in consecutive welds was tentatively investigated via the DR curves, accounting for noticeable declines in the stage I of DR. The results revealed that shunted welds beyond minimum weld spacing do not significantly undermine the accuracy of classification. The implementation of RF based on the combination of welding parameters and DR features improves the accuracy of classification (98.8%) with ntree = 1000 and mtry = 4, as weld current significantly distinguished situations where DR features solely achieve accuracy (93.6%). The incorporation of the RF technique into online monitoring system attains a satisfying RSW quality classification accuracy and reduces the workload on destructive tests.