Multiscale cost-sensitive learning-based assembly quality prediction approach under imbalanced data

Assembly quality prediction of complex products is vital in modern smart manufacturing systems. In recent years, data-driven approaches have obtained various outstanding engineering achievements in quality prediction. However, the imbalanced quality label makes it difficult for conventional quality prediction methods to learn accurate decision boundaries, resulting in weak prediction capabilities. Moreover, the multiple working condition data information in the assembly system presents another challenge to quality prediction. To handle the above issues, a multiscale cost-sensitive learning-based assembly quality prediction approach is proposed in this paper. First, an improved Gaussian mixture model is developed to automatically partition the global multi-condition data into several diverse subspaces. Then, the local cost-sensitive learning models are employed to tackle imbalanced data in each subspace. Finally, by leveraging Bayesian inference, multiple local cost-sensitive learning models are integrated to obtain a global multiscale prediction model. To validate the effectiveness of the proposed method, the quality prediction comparative experiments are conducted on two real-world assembly systems. The favorable results demonstrate the superiority of the proposed method in assembly quality prediction.