Machine learning in intelligent robotic system
Mашинско учење интелигентног роботског система
AuthorDiryag, Ali Karkara A.
Committee membersBabić, Bojan R.
MetadataShow full item record
Nowadays, one of the most desirable features of every robotic system is the ability to adapt to the real world changing conditions. Similarly, failure prediction is equally important in different manufacturing environments in which repairs are often infeasible and failures can have disastrous consequences. In industrial robotics, failure prediction is helpful in reduction of a system down-time by identifying and repairing faulty components. Also, the reliability of a product manufacturing and increased human safety is ensured by implementing fault tolerance and failure prediction unit in the robotic system. It is known that the supervision and learning of robotic executions is not a trivial problem. In the 21st century, robots must be able to tolerate and predict internal failures in order to successfully continue performing their tasks. This doctoral dissertation presents a novel approach for prediction of robot execution failures based on machine learning technique - neural networks... (NNs). Real data consisting of robot forces and torques recorded immediately after the system failure are used for the NN training. Two types of neural networks are used: feedforward and recurrent (Elman) NNs. In total, 7 different learning algorithms and 24 NN architectures are implemented in order to find optimal solution for the problem of robot execution failures prediction. Each multilayer feedforward NN with different learning algorithm and architecture that consists of 1, 2, 3, or 4 hidden layers is evaluated several times, and the same NN architectures are trained using Elman recurrent NN. Experimental results indicate that Bayesian Regularization algorithm is the best choice for the prediction problem with prediction rate of 95.4545 percent, despite having the erroneous or otherwise incomplete sensor measurements invoked in the dataset. The experimental results show that the NN outperforms state-of-the-art algorithms, such as the Naïve Bayes, Decision Trees and Support Vector Machine based algorithms employed for the prediction of robot execution failures. Additionally, two independent failure prediction problems are treated in this dissertation. Several experiments in real time are conducted on an real nonholonomic mobile robot Khepera II in a laboratory model of manufacturing environment. First real world failure problem refers to the robot obstacle detection in indoor environment. Six infrared sensors mounted on the mobile robot are used to obtain information of the obstacle located left and right from the platform. Randomly generated failed sensor data is integrated into the training set so as to test the NN performance in this task. The result show that in over 96 percent of all tested cases NN recognized failed value, meaning that the obstacle location is successfully determined after the failed information is replaced with the expected one. Second real world problem refers to the failure prediction in a mobile robot trajectory tracking problem. Two independent trajectories are employed so as to objectively test the proposed intelligent approach. The tracking of the M-shaped and Labyrinth-type trajectories showed as a fairly easy task for the developed prediction method. In more than 99 percent of the cases, the neural network predicted the wheel command failure, which is next replaced with the desired value in order to successfully track chosen trajectory. The experiments show that a mobile robot can track desired trajectories with a minimal error in every control iteration, which evidence the robustness and the applicability of the proposed approach. Finally, all aforementioned experiments and obtained results indicate that the new method based on neural networks can successfully be applied for robot failure prediction, and also that novel neural network based control system of the mobile robot can be successfully used for solving obstacle detection and trajectory tracking problems in laboratory model of a manufacturing environment.