The process industry is very important to the country, and it is the main tax revenue source of the state, what’s more, the development of it has a direct impact on the country’s economic base. But the production environment is quite complex for the process industry, especially for the industries of chemical, petroleum and metallurgy. The environment is usually in the high temperature and pressure or in the low temperature and vacuum environment, and even in the dangerous of explosion and gas leak. The results are very serious once the accident happens. Not only have the heavy losses of economic or human casualties, but also there maybe have irreparable environmental damage. Fault diagnosis is formed and developed to adapt the project. It analyzes the information which is obtained by the sensor systems and combined with the prior knowledge of the system. The skill analyzes the faults which have occurred or may occur in the future and determines the type, location, extent and the cause. The skill can increase the safety of equipment operation and reduce the costs which ensure the product quality firstly; what’s more, it can maximize avoid the serious in the process of subversion.The paper uses the data-driven approach for the fault detection after analyzing lots of previous methods which is used for the fault diagnosis in the process industry. The test results were compared and analyzed by applied to the Tennessee Eastman simulation platform. The main work is as follows:1) The paper uses the kernel independent component method for the process industry data which does not meet the Gaussian distribution and non-linear. It can get the corresponding nuclear independences though analyzing the normal historical data for the feature extraction. What’s more, it can establish the fault detection model though determining the control limits which is based on the run theory to achieve the fault detection. The independent components which are obtained by the method of the kernel independent component analysis will be as the input data to the least square support vector machines for the fault identification.2) The method is not suitable for the process industry which is as a large sample of data because the kernel independent component analysis uses the nuclear matrix which is the dimension of the nuclear matrix is equal to the square of the number of samples. So the paper uses the sparse kernel independent component analysis methods for the fault detection. |