Process Monitoring And Fault Diagnosis Based On Multivariate Statistical Methods

Nowadays, the automation control of industrial process tends to be large-scale and complicated constantly. Consequently, once the control system or equipment breaks down, it would not only take hours or days to fix the trouble, but also lead to a huge economic losses, and even injury. In order to maintain the system working efficiently and reliably, it is necessary to implement a good control for the industrial process. Yet, for all of this, it is important to stuy a process monitoring and fault diagnosis system, which can detect the fault in real-time and judge the source and type of fault, and this has important theoretical significance and broad application value.Multivariate statistical process control is an important branch in the research of process monitoring and fault diagnosis area. By analyzing and interpretating the on-line process measurement data, we could detect and identify the abnormal condition appeared in the process. In this paper, Tennessee-Eastman (TE) process is used as the background. Wavelet denosing is combined with the tranditional process monitoring method-principal component analysis (PCA) and independent component analysis (ICA). And a new fault diagnosis method is improved. Details are as follows:(1) The wavelet threshold denoising method is introduced which includes offline and online applications based on the analysis of wavelet.(2) Process monitoring based on principal component analysis (PCA) and wavelet denosing combined with PCA are introduced.(3) Process monitoring based on independent component analysis (ICA) is introduced. By using the wavelt for denosing and PCA for reducing dimension, wavelet denosing is integrated with PCA and ICA for Process monitoring.(4) A new fault diagnosis method base on mixed similarity is introduced. In this method, by making full use of the gaussian and non-gaussian information of the process data, the distance similarity based on PCA subspace is combined with the cosine similarity based on ICA subspace for fault diagnosis. The TE simulation shows that wavelet denosing combined with PCA has a good performance than PCA on fault detection radio, but is not always a valid method on detecting partial special faults. What’s more, the method that integrates wavelet denosing and PCA-IC A has a better performance on fault detecting, especially similar to fault3. In addition, the improved fault diagnosis method based on the mixed similarity could identify some faults which are difficult to identify by variable contribution plots.