The distribution of actual industrial process data is complex, and variations in data distribution caused by equipment wear and changing operating conditions can easily lead to model mismatch, presenting a severe challenge to fault diagnosis methods that assume data follows a Gaussian distribution. In this context, we propose a novel fault detection method based on generative models in this paper. Firstly, historical data are used to train a denoising diffusion probabilistic model (DDPM) to generate data. Secondly, both the training set and the generated data are input to an autoencoder, and a data evaluation metric is constructed to filter high-quality out of distribution features. Subsequently, positive and negative sample pairs are constructed based on these features, and an improved supervised contrastive learning detection model is designed to extract unique features of normal data under the supervision of virtual fault samples. Finally, the effectiveness and superiority of the proposed method are validated through the Tennessee Eastman simulation process.

中文翻译：

---

一种改进的去噪扩散概率模型监督对比学习在工业过程中的故障检测


实际工业过程数据的分布很复杂，设备磨损和运行条件变化导致数据分布变化，很容易导致模型不匹配，对假设数据遵循高斯分布的故障诊断方法提出了严峻的挑战。在此背景下，我们在本文中提出了一种基于生成模型的新型故障检测方法。首先，使用历史数据训练去噪扩散概率模型 （DDPM） 以生成数据。其次，将训练集和生成的数据都输入到自动编码器中，并构建数据评估指标以过滤出高质量的分布特征。随后，基于这些特征构建正负样本对，并设计了改进的监督式对比学习检测模型，以在虚拟故障样本的监督下提取正常数据的独特特征。最后，通过田纳西州伊士曼仿真过程验证了所提方法的有效性和优越性。