Deep Learning Based Contribution Index For Nonlinear Process Fault Diagnosis

Process monitoring plays a key role in maintaining production safety and ensuring product quality in the industrial process.Fault diagnosis is an important part of process monitoring,which is aimed to extract fault information in the process and help engineers locate and recover faults.With the increasing expansion of industrial production scale and the increasing complexity of industrial process,it is difficult to deal with the diagnosis task in the pratical process using the traditional linear method.Therefore,designing fault diagnosis method for nonlinear process has become an important research topic in the field of process monitoring.Data mining technology,such as deep learning,has developed rapidly in recently years,which makes it possible to extract complex process features and fault information from process data.However,due to the poor interpretability of deep learning model,deep learning based fault diagnosis has been difficult to achieve interpretable fault identification.Taking deep neural network model as the starting point,this paper explores and studies the fault diagnosis methods based on deep learning.The main research area of this paper includes the following parts:(1)In order to process the stored off-line fault data effectively,an industrial offline fault pattern clustering method based on Sequence Discriminative Feature Extraction Network(SDFEN)is proposed.The proposed SDFEN includes two parts: a prediction network and a local information extraction network.Firstly,a set of contribution rate sequences are calculated as supervision for the training of prediction network.The local information extraction network further extracts the local information in the data by reconstruction and neighbor samples prediction.DTW is used to select neighbor samples,and a parallel strategy with Euclidean distance based pre-selection is designed to reduce the time consumption of the neighbor searching process.Using the extracted discriminantive features to train Gaussian mixture model can get a more accurate clustering result.(2)In order to deal with monitoring task in multi-modal and nonlinear process,Mixture Sequential Network is proposed for feature extraction,and monitoring indexs are further constructed.The proposed network consists of encoder,modal identification network and decoder part.The encoder is composed of a recurrent neural network,which can extract dynamic features from the industrial sequence.Modal identification network is used to identify data modes and give modal information.The decoder part is composed of multiple sub-decoders,which are used to purified the features in a single mode and prevent the confusion of the process information in multi-modal process.In addition,based on the proposed network,a weighted reconstruction error and a weighted contribution plot are designed for on-line fault detection and diagnosis,which can give better monitoring results in the multimodal nonlinear process.(3)In order to deal with the fault diagnosis task in nonlinear process,a deep learning based method named Locally Linear Back-propagation Based Contribution is proposed.This method constructs a local linear model on the trained auto encoder,and combines the idea of reconstruction based contribution to suppress the smearing effect,which is caused by the propagation of fault information.The contribution index for fault diagnosis is optimized to give a more accurate fault diagnosis.(4)In order to make better use of the nonlinear features extracted by the auto encoder to deal with the fault diagnosis task in nonlinear process,an Adaptive SemiDefinite Matrix Based Contribution is proposed.This method optimizes the construction process of local linear model by estimating the output of hidden layer.Moreover,in order to reduce the redundancy of variables,an adaptive correlation weight matrix based on the maximum information coefficient is added in the calculation of contribution index,which further improves the diagnosis performance.(5)In the nonlinear process with strong dynamics,the fault information will propagate dynamically among various variables over time,which seriously affects the accuracy of traditional fault diagnosis methods.Therefore,a fault diagnosis method based on Deep Dynamic Residual Space is proposed to solve this problem.This method calculates the prediction residual through a GRU based autoregressive network,and maps the original data to the deep dynamic residual space.Then,a dynamic principal component analysis is further constructed to extract the remaining process information,and the fault diagnosis is completed by constructing reconstruction based contribution.The proposed method can effectively suppress the influence of process dynamics on fault diagnosis and give better diagnosis results.