Stacked Autoencoder Network Based Fault Diagnosis

Due to the increasing developments of information technologies,data-driven fault diagnosis has become one of the most popular researches in multivariable process control.However,the massive high-dimensional measurements that accumulated by distributed control systems have brought great computational and modeling complexity to traditional fault diagnosis algorithms,which fail to take advantage of the higher-order information for online estimation.With the problems of low-order,dynamics,nonlinearity,multi-mode and incipient faults,this thesis proposes several novels fault diagnosis methods based on stacked autoencoder network in view of deep learning and statistical analysis,which are summarized as follows,(?)A stacked auto-encoder based fault diagnosis method is put forward after introducing deep learning into multivariable industrial process.The underlying high-order correlations are learned by enjoying the great expressive power of stacked auto-encoder network,and detection rate is improved,especially for incipient fault.(?)A geometry framework is expounded on stack sparse auto-encoder,with capabilities of multiple nonlinear mapping and complex function approximation,from which the detection ability of stack auto-encoder network based fault diagnosis is described.With the consideration of time-correlation,a weighted time series fault diagnosis method is presented.(?)An interpretation of auto-encoder net is presented based on Taylor expansion,which motivates the representation ability of incipient faults by Taylor's high-order term o(xn).Due to the fact that the nearest neighbors over time are not necessarily the nearest spatial neighbors in dynamic process,a dynamic reconstruction based multilayer network technology is proposed for fault diagnosis without increasing the network structure.The reconstruction can increase the distinguishable distance between categories while maintain its separability.(?)As traditional statistical analysis technologies fail to take advantage of the higher-order information,a higher-order correlation based multivariate statistical approach is proposed by employing a stacked auto-encoder network.The more layers stacked,the higher order correlations are learned.Then,three detection statistics are presented to monitor whether the process is remaining in-control hierarchically.Moreover,only normal historical data is used in training phase can avoid the problem of data-unbalance(?)A stacked autoencoder network based adaptive thresholding method is proposed for dynamic processes by taking the importance of online data into consideration.Multimode identification and fault diagnosis are integrated under a global representation modeling,which reduces the costs of model switching.Moreover,an interpretation based on the reformulated sigmoid function is presented to explore the expressive powers of auto-encoder net.The stacked auto-encoder based fault diagnosis methods can reveal the process state and evolution trajectory in a more detailed way.Successful applications and simulation experiments demonstrate the effectiveness of the present algorithms,which enrich the achievement of data-driven fault diagnosis,and also suggest the necessity and potential of further research.