Design and performance improvements for fault detection in tightly-coupled multi-robot team tasks

This paper presents our current work to improve the design and performance of our previous work: SAFDetection, a sensor analysis based fault detection approach that is used to monitor tightly-coupled multi-robot team tasks. We improve this prior approach in three aspects. First, we show how Principal Components Analysis (PCA) can be used to automatically generate a small number of sensor features that should be used during the learning of the model of normal operation. Second, we implement three different algorithms for clustering sensor data in SAFDetection and compare their fault detection rates on physical robot team tasks, to determine the best technique for clustering sensor data while learning the model of normal team task operation. A third improvement we present is to modify the state transition probability from constant to a time-variant variable to describe the operation of the robot system more accurately. Our results show that a PCA feature selection approach, combined with a soft classification technique and time-varying transition probabilities, yields the best fault detection results.