Abstract: According to the characteristics of process industries,the issue of low recall in identifying defective products caused by imbalanced class was addressed.To extract effective features from high-dimensional data,the advantages of one class F-score and mRMR in feature extraction were combined to effectively reduce the feature dimension and extract valuable features.Then,the Wasserstein Generative Adversarial Network (WGAN) algorithm was employed to augment the quantity of defective product.Subsequently,the focal loss function was optimized with class weights to enhance the recognition rate of hard case.Furthermore,leveraging the LightGBM algorithm in conjunction with a threshold movement strategy,a quality prediction model was constructed based on WGAN and hard case mining techniques.Finally,the proposed model was applied to the open-source SECOM dataset,and the result indicated that the presented approach effectively enhanced the recall rate of defective products while maintaining overall accuracy,which provided a scientific and practical method for in-depth exploration of the intricate mapping relationship between critical production factors and product quality,as well as facilitating intelligent quality prediction efforts.

Key words: high-dimensional data, Wasserstein generative adversarial network, Focal Loss function, hard case mining, LightGBM algorithm, threshold moving, product quality prediction

摘要： 针对连续性工业生产特点,重点关注类别不平衡造成的不合格样本召回率低问题。为了从高维数据提取有效特征,结合one class F-score和最小冗余最大相关性在特征提取方面的优势,有效降低特征维度并提取有价值特征;利用Wasserstein生成对抗网络(WGAN)方法扩增不合格样本数量;通过类别权重优化Focal Loss函数以提高困难样本识别率;通过轻量级梯度提升机算法结合阈值移动策略,构建基于WGAN数据增强和难例挖掘技术的质量预测模型(WGAN_Focal Loss_LGB(TM))。将所提模型应用于开源SECOM数据集,验证了所提方法的有效性。

关键词: 高维数据, Wasserstein生成式对抗网络, Focal Loss函数, 难例挖掘, 轻量级梯度提升机算法, 阈值移动, 产品质量预测