Reaserch On Incipient Fault Diagnosis Based On PCA And Contribution Plots

With the development of science and technology, modern industrial processes have been much more complicated. If we cannot detect and diagnose the faults in the processes, then the whole sys-tem will be in abnormal situation, thus leading to system outage or industrial accidents. Hence, it is quite vital for us to guarantee the safety and reliability of the industrial processes. The technol-ogy of process monitoring and fault diagnosis, which is widely used in industrial area, can help us detect and diagnose the faults timely, and remind us to take immediate operations to avoid the occurrence of disasters.In this thesis, some issues on incipient sensor fault diagnosis and selection of the principal components (PCs) to establish PC model are focused. With the basic knowledge of principal component analysis (PCA) and contribution plots, some research has been done in the following aspects:1. The traditional fault detection methods, which are based on statistic indices, are not useful for incipient sensor faults. Meanwhile, reconstruction-based contribution plot (RBCP), which can guarantee the correct fault identification of single sensor fault with large magnitude, can sill lead to false fault identification when incipient sensor faults happen. A new type of contribution plot called average residual-difference reconstruction contribution plot (ARdR-CP) is proposed to overcome these mentioned problems. The ARdR-CP method can detect and identify incipient sensor faults simultaneously, and then the fault estimation equation derived from fault reconstruction theory is directly used to estimate the fault magnitude, thus forming the complete procedure of incipient sensor fault diagnosis.2. The relative contribution plot can achieve that variable contributions are statistically equal when there is no fault. When dealing with fault identification, however, the results are not satisfied. For the reason that the expectation expression of contribution plot can be ac-quired from relative contribution plot, another new type of contribution plot, namely average expectation-difference reconstruction contribution plot (AEdR-CP), is developed for incipi-ent sensor faults. The AEdR-CP method also contains the ability of detecting and identifying incipient sensor faults in the meanwhile.3. Almost all the conventional principal component selection methods do not consider the fault detection and identification performance, and the fault SNR method which takes the fault detection sensitivity into account can just deal with one type of sensor fault. In order to pick out the most suitable number of PCs, a new fault detection performance index indicating the fault detectable coverage is designed, and then a new fault identification performance index is proposed to judge whether false fault identification will take place or not. Moreover, detailed mathematical analysis about the influence of the number of PCs on the fault identification performance index is given. Finally, a new PC selection method based on performance optimization of fault detection and identification is provided.