Anomaly detection on a hybrid kinematic machine

Abstract: Detecting anomalies is a promising and current research subject that can have useful applications, for example in the field of robotics. In this thesis, anomaly detection is investigated using a hybrid kinematic machine, which is a pick-and-place robot that excels at moving objects at high speed and with great reach. Three different types of anomalies have been chosen to be studied in this thesis; collision, the robot dropping an object, and weight offset between the expected weight that the robot is carrying and the actual weight being carried. The data were gathered from the robot control system, as well as sensors on the robot when the robot was moving in a specified trajectory cycle. Then, the previously mentioned anomalies were introduced to the robot. The anomaly detection was conducted by using different anomaly detection models which have certain characteristics. The models were first trained using normal data. Then the trained models and a devised threshold value were used to evaluate how well the models were able to detect the anomalies. The results are promising, especially for collision detection and weight offset detection. Detecting an object being dropped, however, seems more challenging. The experiments also indicate that a good model of the robot dynamics is of great importance when detecting anomalies. The results also indicate that the most important features for detecting anomalies are the torque data from the control system and data from an accelerometer at the endpoint of the robot. The most promising models for anomaly detection are the local outlier factor model and the autoencoder, which is a type of artificial neural network. Further work could investigate more varied anomalies that are harder to detect with more advanced models, while also focusing on the models being computationally fast enough to be applicable in a real-time system.