Heterogeneous Data Analytics for Fault Diagnosis and Accommodation of a Mobile Robot with Dislocated Actuator Faults

With the increased popularity and complexity of present autonomous vehicles, system fault diagnosis has become a key field of research to ensure the longevity and safety in the industry. These highly dynamic and nonlinear systems can become unstable with slight perturbations which is a major problem to be addressed as they become more widely available in the civilian sector. This thesis aims to study this problem by developing a novel fault-tolerant control system via a heterogeneous learning approach. An small-scale Unmanned Ground Vehicle (UGV) is used as a test bed which is subjected to dislocated suspension faults. Fault data based on the perturbed trajectory of the faulty UGV is recorded and stored in a database. Spectral clustering is used to detect fault patterns in the unlabeled database in order to distribute the data into fault clusters. The clustered data can then be used to develop PID control laws to accommodate the loss in performance due to each fault. This is done by implementing a genetic algorithm to optimize a set of PID gain vectors for each fault cluster in the database. Long Short-Term Memory (LSTM) neural networks are developed for real-time fault diagnosis based on UGV trajectory data. Finally, a deep ensemble learning model composed of n LSTM networks is distributed to a cloud-based compute cluster. The results of the ensemble are then aggregated to produce a robust unbiased prediction. This prediction will then be used to select the corresponding PID gain vector best suited to track a given trajectory.