Abstract: In modern chemical processes, timely and accurate fault diagnosis is important for enhancing the safety and reliability. Data-driven fault diagnosis methods have been regarded as a promising approach in the last decades of research for increasingly complex chemical processes. Data-driven fault diagnosis methods can greatly reduce the dependence on human experience, and realize end-to-end fault diagnosis by automatically extracting features. However, most existing research assumes training and testing data come from the same distribution, while a chemical process may have multiple working conditions. On the one hand, the fault diagnosis performance of the model will deteriorate when the process is run under new working conditions. On the other hand, due to the low probability of failure, some operating conditions may have few fault data in history. To address these issues, in this work, a novel fault diagnosis method, DA-CycleGAN, is proposed for multimode chemical processes. This study is the first to overcome the degradation of model diagnosis performance when only normal data are available under new working conditions. It notes that the normal data is available under any working condition. A two-dimensional CycleGAN is used to capture the temporal and spatial features of fault data. And fault data is generated by combining fault features and normal data under new operating conditions, thus filling a blank in new working conditions for fault data. Furthermore, the domain adaptation method is used to minimize the distribution differences between historical fault data and generated data and to improve the fault diagnostic performance under new operating conditions. To test the performance of this method, four working conditions of the Tennessee-Easthman (TE) process are used in the experiment. The results on twelve condition-changed fault diagnosis tasks show that this method can improve the average fault diagnosis rate by more than 3% compared to the model trained using only historical fault data.

Key words: fault diagnosis, transfer learning, CycleGAN, domain adaptation

摘要： 化工生产过程中可能会存在着多种运行工况。一方面，当生产工况变化时，故障诊断模型的性能会恶化。另一方面，由于故障是小概率事件，因此在新工况下只能获取到正常运行数据。为解决这些问题，提出一种基于二维循环一致性生成对抗网络(Cycle-Consistent Generative Adversarial Networks, CycleGAN)的多工况故障诊断方法。该方法首先利用二维CycleGAN捕获历史工况下各类故障的时空域特征，并将这些特征添加到新工况的正常数据中，用以构建新工况下的伪故障数据集。其次，使用领域自适应方法减少真实故障数据与生成数据间的分布差异，使故障知识更好地迁移到新工况中。利用田纳西-伊斯曼(TE)过程的十二个变工况故障诊断任务对所提算法进行性能测试，实验结果表明，该算法相比于直接利用历史工况下故障数据建模能提高平均故障诊断率3%以上，能有效提高模型在新工况下的故障诊断性能。

关键词: 故障诊断, 迁移学习, 循环一致性生成对抗网络, 领域自适应