Fault Detection And Simultaneous Fault Identification Based On Representation Learning With Autoencoders

In industrial production,the reliability and safety of the system is especially im-portant.Once the system fails,it will bring a lot of immeasurable losses,consequently,an effective fault detection and identification system should be established.In recent years,with the development of computer technology and data storage technology,a large amount of process data has been recorded,which provides favorable support for fault detection and recognition methods based on representation learning.The fault detec-tion and recognition method based on representation learning has wide generality and is suitable for large-scale process industry.In the actual process industry,the actual variables obey a variety of different distribu-tions,and there is a highly nonlinear relationship between the variables.The importance of different variables is also different.Traditional fault detection methods usually have restrictive hypotheses,such as linear hypothesis,Gaussian distribution hypothesis,etc.These hypotheses ignore the characteristics of the process industry,making the various methods unable to achieve the expected fault detection effect.At the same time,a variety of faults sometimes occur simultaneously in actual industrial production.However,si-multaneous fault data are more difficult to acquire and present complex intrinsic patterns than independent faults.Based on the research of fault detection and simultaneous fault identification,this paper proposes a fault detection and simultaneous fault identifica-tion strategy based on automatic coding representation learning on the basis of in-depth understanding of process characteristics,data characteristics and fault detection and iden-tification methods based on representation learning.The specific research contents are as follows:(1)The traditional principal component analysis(PCA)method has Gaussian hy-pothesis and linear hypothesis for variables,that is,the system variables are assumed to obey Gauss distribution,and the relationship between variables is linear.These hypothe-ses are contrary to the characteristics of the actual process industry and it is difficult to accurately describe the system.Even the kernel principal component analysis(KPCA)method,which can deal with nonlinear problems,can not solve the problem that vari-ables obey multiple distributions.The estimation of the T~2statistic control limit in fault detection requires that the variable obey the Gaussian distribution.In view of the above situation,this paper proposes a nonlinear fault detection strategy for Gaussian representa-tion learning.The original variables of the system are mapped to the representation space obeying the Gaussian distribution by using the variational autoencoder(VAE),and the monitoring statistic is established in the representation space.The VAE method is a type of deep neural network with a multi-layered nonlinear structure that enables it to handle complex nonlinear relationships between variables.At the same time,the representations learned by VAE are Gaussian-distributed,so it is easy to establish monitoring statistic and estimate the corresponding control limit.The proposed algorithm makes full use of the characteristics that VAE can extract Gaussian representation and deal with nonlinear relations,and improves the performance of traditional methods for fault detection.(2)The distance-based fault detection method is very sensitive to the variables of the system,and the detection effect is very good when the variables are important.There-fore,distance-based fault detection methods require critical representation information.Aiming at this situation,this thesis proposes a fault detection method based on stacked denoising autoencoder(SDAE)robust representation learning.The SDAE applies a local denoising criteria to perform a layerwise pre-training process,stacking successive layers together,and obtaining the desired deep neural network model through global fine-tuning.Through denoising training and stacked initialization,SDAE can gain a deeper network structure and learn robust representations.The SDAE architecture is learned from the data,which ensures that the learning model can truly reflect the data characteristics,so that the robust representations learned by SDAE play an important role in the subse-quent distance-based fault detection algorithm.After extracting robust representations,the k nearest neighbor(k NN)method is applied to the representation space to establish monitoring statistic and perform fault detection.The proposed method makes full use of the advantages of SDAE to extract robust representation information,and reduces the influence of invalid variables on distance-based fault detection methods.(3)The first two methods are aimed at whether the system has a fault,and put forward the corresponding detection strategies,only know whether the system has a fault,can not know the exact type of fault.In practice,it is sometimes necessary to know the exact type of fault and propose corresponding system recovery strategies for different types of faults.Simultaneous faults are often more difficult to identify than independent faults.Simultaneous faults are composed of multiple independent faults,and the feature patterns are entangled.In view of the above situation,this thesis proposes a simultaneous fault identification algorithm based on the sparse representation learning of stacked sparse autoencoder(SSAE).There are a large amount of unlabeled data in actual industrial systems,and these unlabeled data also contain a lot of useful information of the system.Therefore,the SSAE is used to perform pre-training of the deep neural network layer by layer in a large amount of unlabeled data,and the pre-trained weights are used as the initialization weights of the deep network.On this basis,the deep neural network for classification is finely adjusted using independent fault data and a small amount of simultaneous fault data.Pre-training in a large amount of unlabeled data using the SSAE method can obtain a part of the information of the system,and these sparse representations are very useful for subsequent classification tasks.Due to the above merits,the proposed method achieves good simultaneous fault identification performance.(4)The above-mentioned simultaneous fault identification algorithm takes into ac-count the fact that the simultaneous fault features are entangled together.Because the network structure and training process are simple,the computational cost is relatively low.However,in practice,simultaneous fault data are sometimes difficult to obtain,which results in a small amount of data for faults or even the missing data of some simultaneous faults.To solve this problem,this thesis proposes a simultaneous fault identification algo-rithm based on SSAE residual transfer representation learning.The proposed method first uses SSAE for pre-training,and retains part of the system information to obtain sparse representations.On the basis of layer-by-layer pre-training,a deep residual network is constructed.Since independent fault data are much more than the simultaneous fault data,and the simultaneous faults are combinations of many independent faults,the inde-pendent fault data can provide some information to the simultaneous faults.Therefore,the deep residual network is pre-trained in the independent fault data set,and then the pre-trained weights are transferred to the deep residual network used to predict all faults,including independent and simultaneous faults.Finally,this deep residual network is globally fine-tuned across all fault data sets.Due to the use of deeper residual network and the complex transfer process,the algorithm requires higher computational costs.The proposed algorithm utilizes the advantages of transfer learning and residual network to improve the performance of simultaneous fault identification.The fault detection methods proposed in this thesis are verified with the nonlinear numerical simulation system and Tennessee Eastman(TE)process,and the excellent fault detection results are obtained.The simultaneous fault identification algorithms are applied into Solid Oxide Fuel Cell(SOFC)power generation system,and good simultaneous fault identification performances are also obtained.