为了解决催化裂化系统中复合故障识别的挑战，并在数据不平衡和小样本情况下提高故障诊断的准确率，以满足工业实时过程的需求，本研究提出了一种基于深度类别监督堆栈自编码器的故障诊断方法。该方法通过在自编码器中引入类别信息，增强了对类别信息的关注，同时利用混合损失函数协调特征保真度与分类精度，在保持了模型结构的简洁性的同时，有效提升了故障诊断准确率。对催化裂化装置反应-再生系统故障数据集的验证结果表明，与堆栈自编码器、多层感知器、深度信念网络、t分布随机邻域嵌入、主成分分析等方法相比，该方法显著提高了诊断准确率，并缩短了模型的训练时间。

The reaction-regeneration system in catalytic cracking unit is an important core equipment for the secondary treatment of crude oil. The increased operational complexity of catalytic cracking units, which is often operated under high pressure, leads to a significant increase in the probability of faults. Therefore, efficient and stable fault diagnosis and condition monitoring are of great significance to ensure the safe operation of catalytic cracking unit. The deep learning methods represented by stacked autoencoder (SAE) is an effective feature extraction method, which is widely used to extract features from collected data quickly and can preserve the original structure of the data thus it is particularly suitable for dealing with chemical fault data. However, it has limited ability as an unsupervised model to deal with classification problems and cannot fully utilize the category information. In order to improve the performance of fault diagnosis in catalytic cracking systems against data imbalance and small sample problems, this work proposes a fault diagnosis method based on deep category-supervised stacked autoencoders (DCSAE). The proposed model extracts deep features layer by layer by stacking multiple category-supervised self-encoders, and introduces category-supervised information to improve the recognition of different fault types by effectively learning useful features with high dimensional unbalanced data and few samples. Then, Bayesian classifiers are utilized for fault diagnosis. Finally, the proposed method is validated on a reaction-regeneration system dataset, and compared with the SAE, the multilayer perceptron (MLP), deep belief network (DBN), the t-distributed stochastic neighbor embedding (TSNE), and the principal component analysis (PCA), the proposed method achieves an accuracy of 96.0%. In addition, the proposed method performs well in dealing with fault data and achieves the highest diagnostic accuracy. In summary, the proposed method provides an innovative idea for the fault diagnosis of catalytic cracking unit, and also provides a valuable reference for the practice in related engineering fields.