Quality assurance of clinched joints using explainable machine learning

Quality assurance (QA) of clinched joints is predominantly performed by destructive testing. Recently, non-destructive evaluation (NDE) methods received increasing attention as a potential alternative. However, the inherently indirect measurement of underlying effects poses a significant challenge to its broader application. To tackle this, two experimental data sets, containing a total of 43 potential process deviations and defects are established using transient dynamic analysis (TDA). On these, several machine learning (ML) models are trained to detect the underlying deviations. The best-in-class model is used to identify a frequency band at which a classification accuracy of 88.58% across all 43 classes is achieved. Further analysis of the most discriminative model features reveals the importance of measuring both excitation as well as specimen response. This lays the foundation for further research towards defect specific in-line measurements of mechanical joints, further improving joint reliability. • Experimental data set of potential process deviations and defects in clinched joints. • Representative engineered features as scalars from time series. • ML model for clear differentiation among 43 deviations with 88% accuracy. • Explainability analysis of ML model reveals importance of excitation measurement.