Fault detection for highly coupled signals in the piercing process based on multivariate variational mode decomposition

The tremendous success in fault diagnosis of engineering applications with nonstationary signals has been shared for many decades. Most of the signal pre-processing approaches have been studied by many scholars with admirable results. However, for determining a faulty event from multiple signal sources, the research still leaves some gaps that can be filled, especially for the progressive piercing process. In this paper, we have conducted a comprehensive analysis of multi-punch signal processing to accomplish punch condition prediction. We also defined multiple factors and compared the impact of each factor on model performance, including signal decomposition method, feature selection, and machine learning model. Finally, we propose a comprehensive working flow that applies Multivariate Variational Mode Decomposition (MVMD) to decompose highly coupled signals, selects Full-band frequency energy features, and employs Random Forest as the primary regressor, achieving an accuracy of up to 80%. Moreover, we also consider the computation time for data processing and model training to define the working flow. The model that achieves the shortest computation time in our workflow while achieving the best prediction model performance.