Research On Fault Detection Of Industrial Process Based On Multivariate Statistical Analysis

With the continuous improvement of the automation level of unit,the scale of the thermal control system is constantly expanding,and the complexity of the system is increasing.The following problems of it are the number of fault points has increased and their concealment becomes stronger.The existing fault detection and diagnosis methods of thermal control system can not meet the requirements of the production site,so it is necessary to study the advanced fault detection and diagnosis technologies applicable to thermal control system.Considering the characteristics of the thermal control system,the multiple statistical analysis method is used to study it.As an emerging multiple statistical analysis technology,non-negative matrix factorization can only perform pure additive operations in the process of data dimensionality reduction.The decomposition results can be interpreted as a partial description of the process data.To some extent,the non-negative matrix factorization method captures the essence of intelligent data description.Therefore,the non-negative matrix factorization method is better than the traditional multiple statistical analysis method in the data description ability.At present,the application of non-negative matrix factorization theory in the field of fault detection and diagnosis is still in its infancy.In this paper,the generalized projection non-negative matrix factorization algorithm is used as the core algorithm.Then,the fault detection and diagnosis model based on generalized projection non-negative matrix factorization is extended around several problems in the thermal control system.In summary,the research work of this paper mainly includes the following aspects:(1)By drawing on the idea of embedding linear projection in the projective non-negative matrix factorization algorithm,a new non-negative matrix factorization improvement algorithm,the generalized projection non-negative matrix factorization algorithm,is proposed.The new method not only improves the derivation of the iterative rules,but also relaxes the non-negative constraints to the original data matrix.The coefficient matrix obtained by the decomposition of the generalized projection non-negative matrix factorization algorithm has better orthogonality and sparsity.In addition,the convergence of the generalized projection non-negative matrix factorization algorithm is also proved in theory.(2)Based on generalized projection non-negative matrix factorization algorithm,a new fault detection and diagnosis model suitable for thermal control system is constructed.Tow monitoring statistics TG2 and SPEG are designed for detecting the process fault.Their upper control limits are determined by the kernel density estimation method.In order to successfully separate the process variable causing the fault,the contribution plots method based on TG2 and SPEG are also designed.(3)In view of the data missing that often occurs in modern industrial processes,the robustness of the generalized projection non-negative matrix factorization monitoring model is studied.Considering the poor robustness of this monitoring model when the test set is incomplete,a method combining the second-order Markov chain model with the generalized projection non-negative matrix factorization monitoring model is proposed to deal with the fault detection problem when the test set is incomplete.The test shows that the fault detection accuracy of the method is still more than 90%when the loss rate of the test set is up to 30%.(4)In view of the shortcoming of the generalized projection non-negative matrix factorization algorithm in data classification,this work attempts to introduce the optimal classification of Fisher discriminant analysis in the process of solving the generalized projection non-negative matrix factorization algorithm,and proposes a new supervised learning method,called Fisher generalized projection non-negative matrix factorization algorithm.Then the problem of fault detection and diagnosis is transformed into a multi-class data classification problem.The experimental results show that the Fisher generalized projection non-negative matrix factorization algorithm can deal with the multiple fault problems very well.