Research On Fault Detection Method Based On Manifold Learning For Nonlinear Process

With the development and progress of science and technology,the scale of modern industry is becoming more and more ambitious,and the structure is becoming more and more complex.It is of great significance to study how to further improve the reliability and stability of industrial system.In this thesis,the nonlinear industrial process is taken as the object and manifold learning theory is combined to study the fault detection problems of nonlinear,multivariable and multimodal industrial processes respectively.The specific work is as follows:1)Aiming at the problems of high dimensionality and nonlinearity of monitoring data in industrial process and its vulnerability to noise,a fault detection method based on improved isometric mapping and support vector machine is proposed.Firstly,aiming at the problem that the manifold learning algorithm is highly susceptible to noise,an standardized residual isometric mapping algorithm is proposed.Based on the construction statistics,the data manifold is reconstructed,and the statistical residuals are used to measure the sample data.The residual analysis is performed to isolate noise in the confidence interval,so that the low-dimensional nonlinear principal elements in the high-dimensional manifold under the noise environment are extracted,and the robustness of the manifold learning algorithm is improved.Then combined with the characteristics of structure risk minimization of support vector machine,the fault detection model is constructed,and the radial basis kernel function suitable for the process monitoring signal is selected to train the data to realize the fault detection under the nonlinear monitoring data with noise.2)The thesis studies the fault detection method in the process of multimodal industry,and proposes a fault detection method based on improved tangent space alignment algorithm.In the method,the structural variables are firstly used to describe the relationship between the multivariate variables in the monitoring data and to construct the global coordinates,and the nonlinear main features of the mutual relations are extracted by minimizing the cost.Then establish an incremental learning mechanism in algorithm,and the matrix similar statistics is built to effectively maintain the data scale and improve the rapidity of fault detection.Finally,theT~2and SPE monitoring quantities are established to monitor the industrial system online and to accurately identify system faults.3)The TE process was used as the object for simulation experiments.First,Gaussian noise is added to TE process to simulate the actual industrial process,and the applicability of the fault detection model is verified by changing the proportion of noise and the type of fault.Secondly,under the different operating modes of the process,the fault detection method proposed in this paper is used for simulation test,and the test results are compared with other fault detection methods,the test results verify the effectiveness of the fault detection method proposed in this paper.