Abstract: The welding gun is a key device in engineering manufacturing, and its stability and reliability are crucial for the continuity of the production line and the quality of products. Faced with the challenge of data imbalance in welding gun fault prediction, a welding gun fault prediction method based on the Jaya-RUSBoost improved model is proposed. By combining undersampling, ensemble learning, and parameter setting optimization, this method achieves data balance and improves the accuracy of fault prediction. First, a RUSBoost fault prediction model is constructed, and experiments are designed to evaluate the impact of hyperparameter values on model performance, thereby determining the optimal range of model parameters. Subsequently, the Jaya metaheuristic algorithm is designed to iteratively optimize the parameters of the RUSBoost model, in order to seek the optimal fault prediction parameter model. The results of the case study show that compared with the traditional RUSBoost algorithm, the proposed algorithm has increased the average fault prediction accuracy and F1 value by 9.43% and 8.41% respectively on five welding guns. Moreover, compared with various machine learning models, the accuracy and other indicators have also been significantly improved. The welding gun fault prediction method proposed in this paper has high practical application value and broad prospects for promotion, and can provide strong support for the intelligent maintenance of welding equipment.