A machine learning approach for efficient and robust resistance spot welding monitoring

The estimation of the weld nugget diameter generated by the resistance spot welding process is a crucial element in the assessment of the overall quality of the weld and plays a major role in in-line process control. The process is crucial to produce end-products in many industries such as aviation, aerospace, automotive and other industrial areas. A modern car body contains typically several thousands of welds produced by resistance spot welding, setting an ideal scene for in-line process control. Current state-of-the-art monitoring methods are based on several features extracted from the dynamic resistance signal. However, the accuracy of those is generally not high. In this work, a method for predicting the nugget diameter based on the combination of unsupervised deep learning and Gaussian process regression is developed. Autoencoders are adopted to extract features from the dynamic resistance curve in a low-dimensional representation. These features embody underlying information on the process, possibly unobservable or not detectable by any other currently existing approach. Next, a Gaussian process regression model is trained to link those features to the target weld nugget diameter. Compared with the currently popular geometrical attributes approach, the results show that the model has a higher prediction accuracy in nugget diameter prediction, while remaining a low cost implementation in an industrial setting. These results are supported by several cases, derived directly from common industrial bottlenecks. Both cases indicate a strong potential with the new AE-GPR approach, with consistently improved results compared to the currently popular geometrical attributes approach.