An Evaluation of Fault-tolerant Robotic Controller Using Evolutionary Computation

The aim of the research reported in this thesis was to design and construct several Fault-Tolerant Controllers for a simulated multi-joint planar robotic arm system. Analysis and comparisons of the convergence rates of the controllers developed will assist in determining and understanding the effectiveness of the controller to adapt itself as the number of faults in the system increases. This research designed four Fault-Tolerant Controllers by combing two well-known Optimisation Algorithms: Genetic Algorithm (GA) and Particle-Swarm Optimisation (PSO) with two well understood robotic controllers: Artificial Neural Network (ANN) and Lookup Table (LUT). The effectiveness of the controller is measured by how fast it can recover when different numbers of faults are applied to it. A Fault-Tolerant Controller can be constructed using either active or passive Fault-Tolerant Control. Passive Fault-Tolerant Control is typically achieved via redundancy, such as having backup components integrated into the system. When faults occur, the system can remain operational by quickly switching to the backup component. Active Fault-Tolerant Control actively updates its parameter and/or architecture to adapt itself to compensate the effects of faults. Results have shown that the Fault-Tolerant Controller PSO-LUT has the fastest convergence rate over all faults combinations applied to it, followed by PSO-ANN, GA-LUT and GA-ANN. Results from all the controllers have shown that the overall performance of the controllers is acceptable. However, in some instances the controllers can become stuck near a local optimal for a significant period before converging to a solution