Deep transfer network (DTN) has been widely used for classification tasks, which introduces maximum mean discrepancy (MMD) based loss function to extract similar latent features and reduce the discrepancy of distributions across the source and target data. However, it is a little challenging to apply deep transfer learning for fault classification tasks in industrial chemical processes, since process variables have physicochemical properties and occupy different impacts in reflecting the process status. Hence, features extracted from these process variables also have different contributions for domain adaptation in DTN. In this paper, a linear discriminant analysis (LDA)–based DTN is proposed for fault classification ofchemical processes, in which a weighted MMD is designed for domain adaptation. First, the LDA algorithm is introduced to determine how much influence each variable can distinguish samples from source and target domains. Then, a corresponding weight is assigned to each feature variable to design the weighted MMD for network transferring. A Tennessee Eastman (TE) process and a real hydrocracking process are used to validate the fault classification performance of the LDA-based DTN. The results indicate LDA-based DTN has better generalization performance and classification accuracy than traditional DTN.

深度传输网络（DTN）已广泛用于分类任务，它引入了基于最大平均差异（MMD）的损失函数，以提取相似的潜在特征并减少源数据和目标数据之间分布的差异。然而，将深层转移学习应用于工业化学过程中的故障分类任务是一个挑战，因为过程变量具有物理化学性质，并且在反映过程状态方面具有不同的影响。因此，从这些过程变量中提取的特征对于DTN中的域适应也具有不同的作用。在本文中，基于线性判别分析（LDA）的DTN被提出用于化学过程的故障分类，其中加权MMD被设计用于域自适应。第一，引入了LDA算法来确定每个变量可以将样本与源域和目标域区分开的程度。然后，将相应的权重分配给每个特征变量，以设计用于网络传输的加权MMD。田纳西州伊斯曼（TE）过程和实际的加氢裂化过程用于验证基于LDA的DTN的故障分类性能。结果表明，基于LDA的DTN具有比传统DTN更好的泛化性能和分类精度。田纳西州伊斯曼（TE）过程和实际的加氢裂化过程用于验证基于LDA的DTN的故障分类性能。结果表明，基于LDA的DTN具有比传统DTN更好的泛化性能和分类精度。田纳西州伊斯曼（TE）过程和实际的加氢裂化过程用于验证基于LDA的DTN的故障分类性能。结果表明，基于LDA的DTN具有比传统DTN更好的泛化性能和分类精度。