Fault Diagnosis Of Complex Industrial Processes Based On K-nearest Neighbors

The efficient operation of the modern complex industrial processes is significantly important due to the increasing demands on product quality. Fault diagnosis methods are essential to improve the reliability, safety of the automatic control systems and guar-antee efficient operation of the industrial processes. As the increasing developments of information technologies, the costs of data acquirement and storage reduce. Massive data obtain from industrial processes, which promote the significant development of fault diagnosis methods. As a kind of representative methods, multivariate statistical process monitoring becomes one of the popular research subjects and produces many impressive academic and industrial results.The traditional multivariate statistical process monitoring methods assume that the data are all acquired from single operation mode, the process variables are linear cor-related and Gaussian distributed. However, as the industrial processes become more and more complex, the acquired data usually contradict above assumptions. To over-come these problems, this thesis proposes several novel fault detection and isolation methods based on the classical k-nearest neighbor rule (kNN), which are summarized as follows:(1) To deal with the fault detection of industrial processes with high dimensional variables, nonlinear, multimode, and non-Gaussian distributed data, a new fault detec-tion method based on random projection and kNN rule is proposed. Compared to the PCA method for dimensionality reduction, the proposed method combines the advan-tages of random projection in distance preservation (in the expectation) and kNN rule in dealing with the problems of multimodality, nonlinearity and non-Gaussian that often coexist in complex indutrial processes.(2) To deal with the fault isolation of multiple sensors faults, a novel isolation in-dex has been provided based on k-nearest neighbor rule. In the frame of multivariate statistical process monitoring, the commonly used fault isolation methods are contri-bution analysis methods, such as contribution plot, partial decomposition contribution and Reconstruction-based contribution. However, these contribution analysis methods suffer from smearing effect, thus, prone to misdiagnosis, i.e.,. the non-faulty sensor is easy to be erroneously identified as faulty sensor. Moreover, theoretical analysis shows that these contribution analysis cannot isolate multiple sensors faults. The proposed kNN-based isolation index is defined in the original measurement space, so it doesn’t affect by the fault smearing and can isolate multiple sensors faults.(3) For the problem of fault reconstruction, a kind of representative fault recon-struction method, the PCA-based reconstruction fault identification method, has been analyzed. The theoretical analysis and experimental illustrations show that the PCA-based reconstruction method cannot unique identify the true faulty sensors in some special cases, especially when a small fault affect one or several sensors and cause the variables correlation changed, in this case the PCA-based reconstruction method usu-ally cannot uniquely isolate the true sensor fault. In addition, the estimation of the fault magnitude by the PCA-based reconstruction method is inaccuracy. We try to solve the problem of fault reconstruction using the k-nearest neighbor regression. An improved fault reconstruction method using k-nearest neighbor regression is presented. The pro-posed method can deal with the fault reconstruction problem where multiple variables are affected by the faults. Moreover, the proposed method do not impose restriction on the data distribution and can be used to deal with nonlinear and non-Gaussian data.Finally, the conclusions and the future research works are given.