Improvement Of Industrial Process Monitoring Method Based On Independent Component Analysis

With the development of computer and information science, computer integrated manufacturing systems are widely used in modern industrial. The collection and saving of massive industrial production data create favorable conditions for industrial process monitoring. Industrial process monitoring technology is significant to modern industry: to improve the management of production equipment; to ensure product quality and stability of the system; to avoid major accidents; to produce potential economic and social benefits. Multivariate statistical process monitoring is an important branch of industrial process monitoring. This paper studies the process monitoring method with multivariate statistics based on Independent Component Analysis(ICA).The process monitoring method with multivariate statistics based on Independent Component Analysis(ICA)takes the process monitoring with indexes of PCA. It does not take full use of the features of ICA. According to the characteristic of the mutual independence of the components separated with ICA, this paper propose a process monitoring index called the probability density index. The index estimates the probability densities of each component with the kernel density estimation in order to obtain the joint probability density. Then, use the joint probability density to judge whether the process state is abnormal. By TE process simulation, it can make a conclusion that the new probability density index detects more fault points and is more effective comparing with traditional indexes.The process monitoring method with multivariate statistics based on independent component analysis(ICA) is mainly used for fault detection, but it is not effective for fault classification. Fast and accurate fault classification has important significance in the industrial process monitoring. For this reason, combining with extreme learning machines, a method called ICA-ELM for fault classification is proposed. Compared with the traditional neural network algorithm, ELM only requires to set the number of hidden neurons, and it avoids multiple iterations and local minimum problem. Firstly, ICA-ELM extracts the fault features with ICA, and then trains the networks with ELM, so as to realize fault classification. ICA-ELM is tested with TE process data and compared with other fault classification methods. The experimental result shows that the accuracy of ICA-ELM is higher, training speed of ICA-ELM is faster, and ICA-ELM is a simple rapid method for fault classification.