Machine learning-based algorithm for the classification and prediction of multi-type spot weld quality

Due to the complexity of body resistance spot welding, there are still problems existing in predicting the quality of resistance spot welding, such as less samples between special welded plates and low welded quality joints point, and the distribution of the dataset changing with the welding conditions. To solve this problem, this paper proposes an algorithm for spot welding quality classification and prediction that distinguishes between multiple types of welded joints. Firstly, dynamic resistance signals of spot welding points with different sheet materials and parameter combinations are collected, and denoising and analysis are conducted. The spot welding feature dataset is obtained from the dynamic resistance signals, and the distribution of feature datasets for different types of welding points and different quality intervals of the same type of welding points is analyzed. Then, principal component analysis (PCA) is applied to reduce the dimensionality of the feature dataset and obtain the principal component dataset. Finally, a backpropagation neural network (BPNN) is used, where the principal component dataset is trained and classified by a Naive Bayesian classifier, with the different quality interval datasets as input and the quality scores as output. Through training and testing of the neural network model, accurate quality prediction is achieved for different types and different quality welding points.