Research On Fault Diagnosis Method Based On Deep Learning

The chemical process usually shows complex,high-dimensional,time-varying and non-Gaussian characteristics,which makes it more difficult to judge the state of chemical process.Those characteristics bring difficulties to the existing fault diagnosis methods.So far,deep learning technology has been widely used in fault diagnosis of complex chemical processes by the adavantages of its powerful feature extraction capabilities.Although the current deep learning methods used in chemical process fault diagnosis have improved in diagnosis accuracy and speed compared with traditional fault diagnosis techniques,this research is not mature and the performance of the network model is limited and needs to be improved.Based on deep learning technology,this paper aims at accurately extracting fault features,and conducts in-depth research from the perspective of feature fusion for the characteristics of multiple types of fault states,large similarities,large amounts of data,etc.,in complex chemical processes.(1)From the network structure level for merging the fault features,a complex chemical process fault diagnosis method based on deep learning multi-model fusion is proposed.This method constructs a parallel network model by using Long Short-Term Memory(LSTM)and Convolutional Neural Network(CNN)to extract fault features from both temporal and spatial characteristics respectively.Then the features extracted by the two methods are spliced and fused,and input into the Multilayer Perceptron(MLP)for further feature compression and feature extraction.This method makes the final extracted features of the network have both spatial and temporal characteristics,and the two aspects are integrated for diagnosis to improve the accuracy of fault diagnosis.(2)Feature fusion is carried out from the algorithm level,and a feature fusion method based on normalized convolutional network is proposed for fault diagnosis of complex chemical process.This method first inputs the simply processed raw data into the deep normalized convolutional network for fault state feature extraction.Then,the fault features extracted by the convolution module are input into the improved second-order pooling for feature fusion,and each state feature is further refined.Finally,the MLP performs the next feature extraction and compression.This method avoids the problems of gradient explosion and gradient disappearance by constructing a normalized convolutional network,and at the same time accelerates the network convergence speed.Combining the normalized convolutional network with the improved second-order pooling method further,the accuracy of the extracted fault features is improved and identification of the similar fault states can be better.