Performance analysis and comparison of machine learning algorithms for predicting nugget width of resistance spot welding joints

Recently, machine learning algorithms have been extensively utilized in resistance spot welding (RSW) applications to develop non-destructive weldability assessment systems to predict nugget width of RSW joints. However, different predictive models have different prediction performance that can be highly inconsistent. It is critical to compare predictive models and determine the efficient model(s). To the best of our knowledge, a comprehensive analysis and systematic prediction performance comparison of RSW nugget width prediction models have not been performed. This paper presents a statistical performance comparison methodology based on bootstrapping and hypothesis testing techniques to systematically compare the prediction performance of predictive models and determine the efficient model(s). Also, a deep neural net (DNN) nugget width prediction model is developed, analyzed, and compared with prior models. Bootstrapping is applied to generate sampling distributions for each predictive model, and statistical comparison tests are employed to analyze and compare the performance of each predictive model and identify statistically significant performance differences. Results of this analysis indicate that DNN, developed for RSW nugget width prediction in this paper, outperforms previous models.