基于改进LSTM神经网络的化工过程故障诊断

Fault diagnosis for chemical process based on an improved LSTM neural network

针对现代化工过程中数据非线性、高维度以及动态时序等特点,传统的故障诊断模型对化工过程的故障诊断精度较低.基于此,设计了一种基于改进的长短时记忆神经网络(LSTM)故障诊断方法.首先,将采集的故障数据输入卷积神经网络(CNN),对数据进行特征提取和降维;其次,将处理过的数据输入改进的LSTM网络,进行深层特征提取;最后,把提取的深层特征信息输入到注意力机制进行特征“聚焦”,实现特征融合后输入softmax分类器实现故障分类.由田纳西-伊斯曼(TE)过程诊断实验结果表明,基于改进的LSTM网络的故障诊断方法在故障分类精度、训练速度方面都更优于递归神经网络(RNN)、门控循环神经网络(GRU)、卷积神经网络(CNN)和深度自编码网络(DAEN),在实际化工过程的应用有一定的优势.

According to the characteristics of data nonlinearity,high dimension,and dynamic time sequence in modern chemical processes,traditional methods have low accuracy for fault diagnosis.Therefore,a fault diagnosis method based on improved long short-term memory(LSTM)neural network is designed in this paper.First,the collected fault data were input into a convolutional neural network(CNN)to extract the features and reduce the dimension of the data.Secondly,the processed data was input into the improved LSTM network for deep feature extraction.Finally,the extracted deep feature information was input into the attention mechanism for feature“focusing”to realize feature fusion,and then input into the softmax classifier to realize fault classification.Simulation results on the Tennessee-Eastman(TE)dataset demonstrate that the proposed method is better than recurrent neural network(RNN),gated recurrent unit(GRU),CNN,and deep auto-encoder network(DAEN)in fault classification accuracy and diagnose speed,which has certain advantages in the application of the practical chemical process.