基于LSTM-DAE的化工故障诊断方法研究

现代化工过程愈加精密化、复杂化，使得化工过程数据呈现高度非线性、强耦合等特点，传统的故障诊断模型难以学习此类数据的有效特征表示，且无法挖掘隐含的时间序列信息。针对上述问题，提出了一种长短期记忆（long short-term memory，LSTM）网络与降噪自编码器（denoising auto-encoder，DAE）结合的LSTM-DAE化工故障诊断方法，用基于LSTM的特殊编码方式代替传统DAE模型的全连接网络编码方式，并结合高斯噪声处理和全连接网络解码层，搭建出基于LSTM-DAE的特征提取模型，最后由Softmax分类器输出故障诊断结果。该方法结合了DAE与LSTM的优点，确保了编码特征获取的有效性。使用田纳西-伊斯曼（Tennessee-Eastman，TE）过程数据设计所提方法与其他5类模型的对比实验，实验结果表明：在多故障诊断效果上，基于LSTM-DAE的化工故障诊断方法的训练集正确率达到了96.02%，测试集正确率达到了97.31%，平均误报率仅为0.65%，平均漏检率仅为3.19%，在6类模型中为最优；在单故障诊断效果上，基于LSTM-DAE的化工故障诊断方法能够提高对故障14的分辨能力，并缩短对故障18的检测延迟时间，有效降低了漏检率，表明所提方法能够在实际化工过程中进行有效的故障诊断。

Fault diagnosis of chemical process based on long short-term memory(LSTM)-denoising auto-encoder(DAE)

Modern chemical processes are becoming increasingly precise and complicated. It is difficult for traditional fault diagnosis (FD) models to learn the feature representation of raw data for high dimensional, nonlinear, and tightly coupled data. Furthermore, traditional FD models cannot extract the hidden time series information inside the raw data. Therefore, a novel FD model of chemical processes called LSTM-DAE based on long short-term memory (LSTM) and a denoising auto-encoder (DAE) has been proposed. By changing the full-connected encoding network of the traditional DAE model to a novel LSTM-DAE encoding network, and combining the Gaussian noise and fully-connected decoding network, a feature-extracting LSTM-DAE model has been established, with the final FD results given by a Softmax classifier. The proposed model combines the advantages of both DAE and LSTM, which ensures high efficacy in feature extraction. The experimental results for the Tennessee-Eastman (TE) process showed that in terms of multi-fault FD performance, the accuracy of the training set was 96.02%, the accuracy of the test set was 97.31%, the mean false alarm rate (FAR) was only 0.65%, and the mean miss detection rate (MDR) was only 3.19%, which is the best among all the six FD models. In terms of single-fault FD performance, the LSTM-DAE model can improve the resolution capability of fault 14 and reduce the delay time of fault 18, which reduces the MDR. The above analysis indicates that the proposed LSTM-DAE model can efficiently detect faults in actual chemical processes.