Fault Detection And Identification For Complex Industrial Processes Based On Data-Driven Methods

Under the promotion of network and information,the scale of modern process industry system is becoming larger and more complex.Process monitoring and fault diagnosis are significant for ensuring the safety operation and product quality of industrial processes.Due the the complex mechanism of system,various factors that affectinig process operation,data-driven multivariate statistical process monitoring methods have received widespread attention and developed rapidly.Traditional multivariate statistical process monitoring usually assumes that system is operating under stable conditions,the monitor variables are independent and the variables are linearly correlated.While there are equipment degradation,measurement errors and process noise in practical industrial process,the process monitoring data has the characteristics of non-stationary and so on,it is necessary to carry out specific analysis and processing when using the monitoring data for fault diagnosis.Based on the in-depth analysis of the high-demensional characteristics,nonlinear correlation and strong dynamics of complex process variables,this dissertation focuses on feature extraction,fault detection and classification algorithms.The main research contents are as follows:Considering the distribution structure of industrial process monitoring variables will be changed when fault occures,which makes it difficult to separate faults.A fault classification algorithm based on sparse local preservation projection is proposed.Firstly,considering various fault types will causing different effects on distribution of the monitoring variables,local preserving projection is introduced to extract feature of faults.Then,considering that reconstruction based on sparse representation can effectively reduce the redundant information among monitoring data and improve the the interpretability of projection transformation.The sparsity constraint on the projection matrix is added during the construction of the local preservation projection objective function.And the solving the projection matrix is transformed into an iterative quadratically constrained quadratic program through relaxation processing.The proposed method effectively improving fault separability of reduced dimension.Considering the problem of incomplete extraction of high-dimensional spatial distribution features in process monitoring,this dissertation proposes global and local preserving projection based on Fisher discriminant analysis algorithm for process monitoring.Firstly,considering traditional statistical monitoring algorithms such as PCA,PLS and LPP only preserving just global or local feature of original spatial,this dissertation proposes global and local preserving projection method based on manifold learning.The GLPP algorithm can preserve both local neighbor feature and global structure in dimension reduction at the same time.Since manifold learning focuse on sample representation but on pattern classification,this dissertation considering combining Fisher discriminant analysis to construct global and local projection objective functions with discriminative performanc.Then,the transformation matrix is obtaind by solving the generalized sinigular value to ensure the separability of fault modes in lowerdimensional manifold.Finally,in order to reduce the influence of process noise and measurement errors,Kernel density estimation is proposed to solving the statistical indicator control limit,which further improves process monitoring performance.Considering the difficult extraction of nonlinear features and weak interpretability of fault classification in complex industrial processes,this dissertation proposes Fisher sparse representation algorithm based on deep belief network for fault classification.Firstly,deep belief network is applied to fit the nonlinear process monitoring data and eliminate redundant information to acqure the features of normal and fault modes.And the feature extracted is used to initialize dictionary of corresponding fault,this greatly reduces the computational complexity of dictionary learning.Secondly,the sparse representation objective function is constructed based on Fisher discriminant analysis,in which sparsity constraint,discrimination constrain,and minimal reconstruction error are introduced.This promise the discrimination of both sparse coefficients and dictionary solved by iterative solution with quadratic programming.On the one hand,the proposed method ensures the performance of sparse coefficients and reconstruction errors in distinguishing fault modes.On the other hand,the proposed method afford better fault classification performance with more explanatory result.Considering the problem of dynamic characteristics of variables in complex industrial processes lead to large fault detection delay and low classification accuracy,canonical global and local preserving projection analysis method is proposed in this dissertation.Firstly,time series autocorrelation of monitoring variables is considers in model construction,on the other hand,the distribution of variables changes after fault occurring.Canonical variante analysis is applied to construct the observation matrix with time lags.Then the Laplacian matrix of past observations and future obeservations are aqured on the perspective of GLPP algorithm.The objective function of canonical GLPP analysis algorithm is constructed based on the purpose of preserving both time series information and spatial structure of variables.Transformation matrix solved can be used to project testing data into reduced dimensional,combining statistical indexes and corresponding control limits,the proposed method greatly reduces fault detection delay and false alarm rate,and effectively improves the accuracy of fault classification.