Torabi, Niloofar (2018-01). Real-Time Machine Learning Based Open Switch Fault Detection and Isolation for Multilevel Multiphase Drives. Doctoral Dissertation.
Thesis
Due to the rapid proliferation interest of the multiphase machines and their combination with multilevel inverters technology, the demand for high reliability and resilient in the multiphase multilevel drives is increased. High reliability can be achieved by deploying systematic preventive real-time monitoring, robust control, and efficient fault diagnosis strategies. Fault diagnosis, as an indispensable methodology to preserve the seamless post-fault operation, is carried out in consecutive steps; monitoring the observable signals to generate the residuals, evaluating the observations to make a binary decision if any abnormality has occurred, and identifying the characteristics of the abnormalities to locate and isolate the failed components. It is followed by applying an appropriate reconfiguration strategy to ensure that the system can tolerate the failure. The primary focus of presented dissertation was to address employing computational and machine learning techniques to construct a proficient fault diagnosis scheme in multilevel multiphase drives. First, the data-driven nonlinear model identification/prediction methods are used to form a hybrid fault detection framework, which combines module-level and system-level methods in power converters, to enhance the performance and obtain a rapid real-time detection. Applying suggested nonlinear model predictors along with different systems (conventional two-level inverter and three-level neutral point clamped inverter) result in reducing the detection time to 1% of stator current fundamental period without deploying component-level monitoring equipment. Further, two methods using semi-supervised learning and analytical data mining concepts are presented to isolate the failed component. The semi-supervised fuzzy algorithm is engaged in building the clustering model because the deficient labeled datasets (prior knowledge of the system) leads to degraded performance in supervised clustering. Also, an analytical data mining procedure is presented based on data interpretability that yields two criteria to isolate the failure. A key part of this work also dealt with the discrimination between the post-fault characteristics, which are supposed to carry the data reflecting the fault influence, and the output responses, which are compensated by controllers under closed-loop control strategy. The performance of all designed schemes is evaluated through experiments.