A Study On Dynamic Control And Fault Diagnosis For Reconfigurable And Modular Robots

Robot abilities increase greatly as the development of robot technology, and robot applications spread into more and more fields, so people hope they can complete more complex tasks. As we know, robot can do many kinds of different tasks through programming, however, how many tasks it can realize after all is constrained to its mechanical structure. Sometimes, maybe robot can choose optimal configuration or structure with different parameters to adapt to different jobs when tasks are designated in advance. This is very possible. But robot cannot form optimal structure to finish some unpredictable, nonstructural environment or variant tasks, such as nucleus scrap heap recycle bin, cosmic space station and lunar base etc. Thus, it is very difficult or impossible to design a robot which can achievement wider task demand. Reconfigurable and modular robot emerges, it can change its kinematical and dynamical parameters due to the different environment to accommodate diverse environment and complete some more complex work. Scholars in the world have deeply studied the reconfigurable and modular robot, but focus on its automatic generation in kinematics and dynamics, kinematics calibration and distributed control. There are many challenging, just underway and marginal aspects, for example, inverse kinematics solution, uncertainty control, fault detection & diagnosis and tolerant control. These problems still need scholars pay much attention and effort. Hence, It is very necessary and significant to study further on reconfigurable and modular robot.This paper systemically studies on inverse kinematics solution on configurable robot with simulated annealing based genetic algorithm, dynamical uncertainty control based on RBF neural networks and neurofuzzy methods and fault diagnosis by nonlinear observer and neurofuzzy observer. Main content and innovation are as follows.Because inverse kinematics of reconfigurable and modular robot has not an exclusive solution, so simulation results for improved numerical iteration algorithm in reference [140] are given firstly, and then simulated annealing based genetic algorithm is deeply researched to guarantee the optimal solution. Also simulation compared to numerical iteration algorithm verifies the proposed genetic algorithm is very more effective and advantageous.Radial Basis Functions in neural networks use Gaussian function. How to determine its centers and widths is studied based on desired trajectory data clustering method. But it ignores the uncertainty influence, so it is just an approximate method, however it doesn't need uncertainty bound, and very effective through simulations.Study on compensation control based on RBF neural networks due to model and unmodelled uncertainty. There are two compensation schemes. One is under assumption that acceleration be measured; a RBF neural network is proposed with rich input information. And adaptive weights and robustifying term are designed with bounded network reconstruction error and estimated reconstruction error. The other is velocity observer based RBF neural network compensator. In this scheme, acceleration input is cut down in order to reduce system cost and complexity. Through Lyapunov stability estimated error and tracking error are analyzed and proved. Simulation results in two schemes show that RBF neural network based control scheme for reconfigurable and modular robot is very effective and simple.BP based self- adaptive neurofuzzy controllers are designed. A novel BP based neurofuzzy method considered as main controller to approximate inverse dynamic model, auxiliary CTC is used to compensate modeling and unmodelled uncertainty. In training process, sub-goal method is used to avoid output oscillation and saturation. Not only structure, but also inference mechanism is simpler than conventional BP base neurofuzzy scheme. But it also has disadvantages: it is not convenient for stability analysis, in bad real time ability and strict with the uncertainty.RBF base neurofuzzy compensation control structure is proposed. Self-adaptation laws on center and reciprocal width parameters in membership function are designed through Taylor decomposition. After that, velocity observer based neurofuzzy algorithm is derived. Anyway, neurofuzzy controller has a bit more complex structure than neural networks, but this scheme improves self-adaptation and control quality through more systemic updates and self-adaptive inference mechanism. Finally, simulations are studied and prove that neurofuzzy controller make system arrive at good dynamic information and realize uncertainty compensation well.Actuator fault diagnosis problem for reconfigurable and modular robot is studied. Improvement and simulation results are given for nonlinear fault observer based on [136]. Sufficient condition for stable fault observer is obtained by choosing appropriate Lyapunov function. Compared with reference [136], this nonlinear observer doesn't depend on fault's second derivative, so conservatives are reduced and assumptions are simple. As a result, we find that output oscillation problem exists because of sign function in observer. So base on this, its assumptions are throwing off and RBF based neurofuzzy fault observer is proposed, which can realize accurate estimation for actuator fault of reconfigurable and modular robot. More importantly, compared with reference [135] etc. with neural networks for estimating fault, neurofuzzy possesses stronger advantage and self-adaptation. Online ability of proposed neurofuzzy observer overcomes the inefficiency and difficulty to obtain optimal parameters combination resulting from the parameters tuning manually.The conclusion and the perspective of future research are given at the end of the paper.