Establishing an effective early warning mechanism for the rocket final assembly process (RFAP) is crucial for the timely delivery of rockets and the reduction of additional production costs. To solve the unsystematic design of warning indicators and warning levels in RFAP and address the problem of low warning accuracy caused by imbalanced data distribution, this paper redesigns the warning indicators and warning levels in a systematic way, and develops a novel multiclass imbalanced learning method based on dynamic sampling algorithm (DyS) and improved ensemble neural network (IENN). The DyS algorithm dynamically determines the training set after oversampling the minority class, while the IENN can effectively suppress the oscillation in the iterative process of the DyS algorithm and improve the overall classification accuracy by removing the redundant and ineffective networks from the ensemble neural network. The experiment results indicate that the proposed method outperforms other methods in terms of accuracy and stability for early warning of tardiness in RFAP.

为火箭的最终组装过程（RFAP）建立有效的预警机制，对于及时交付火箭和降低额外的生产成本至关重要。为解决RFAP中警告指标和警告等级的非系统性设计，解决数据分布不均引起的警告精度低的问题，本文系统地重新设计了警告指标和警告等级，并提出了一种新颖的基于多类不平衡学习的方法。动态采样算法（DyS）和改进的集成神经网络（IENN）的研究。DyS算法在对少数类进行过采样后，动态确定训练集，IENN可以有效地抑制DyS算法的迭代过程中的振荡，并通过从集成神经网络中删除多余和无效的网络来提高整体分类的准确性。实验结果表明，所提出的方法在准确性和稳定性方面优于其他方法。