Abstract:

The data of chemical processes often contains dynamic timing characteristics, and traditional fault detection has low usage of dynamic information, which limits the fault diagnosis performance. To address this problem,a new method of chemical process fault diagnosis based on an attention?enhanced encoder?decoder network model (AEN) was proposed. The coding part uses the LSTM to extract the feature information of the process data and combine it with the attention mechanism to utilize the dynamic information among the process data more effectively； the decoding part uses the LSTM and combines the context vector provided by the attention mechanism to provide more accurate state information for the softmax regression, and finally, the softmax regression is used to obtain the probability value of the fault category for each sample data. The introduction of the attention mechanism improves the efficiency of the model in using process dynamic information in the time domain. The proposed method is experimented with using Tennessee Eastman process data and compared with the results of standard PCA?SVM, DBN and ResNet, and the results show that the proposed method is more effective in diagnosing faults.

Key words: Fault diagnosis, LSTM networks, Attention mechanism, Softmax regression, Encoder?decoder network

摘要：

化工过程的数据往往含有动态时序特性，传统故障检测对动态信息的使用率较低，限制了故障诊断性能。针对这个问题，提出了一种基于注意力增强的编解码网络模型的化工过程故障诊断新方法。编码部分利用LSTM提取过程数据的特征信息，结合注意力机制，更加有效地利用过程数据间的动态信息；解码部分利用LSTM并结合注意力机制提供的上下文向量，为归一化指数的回归提供更加精准的状态信息，最后利用归一化指数回归得到各个样本数据的故障类别概率值。结果表明，注意力机制的引入，提高了模型在时域下对过程动态信息的使用效率。针对本文提出的方法，利用田纳西伊士曼过程数据进行了实验，并与标准的PCA?SVM、DBN和ResNet的结果进行了对比。结果表明，该方法诊断故障的效果更加理想。

关键词: 故障诊断, 长短期记忆网络, 注意力机制, 归一化指数回归, 编解码网络