Faults in robot operations are risky, particularly when robots are operating in the same environment as humans. Early detection of such faults is necessary to prevent further escalation and endangering human life. However, due to sensor noise and unforeseen faults in robots, creating a model for fault prediction is difficult. Existing supervised data-driven approaches rely on large amounts of labelled data for detecting anomalies, which is impractical in real applications. In this paper, we present an unsupervised machine learning approach for this purpose, which requires only data corresponding to the normal operation of the robot. We demonstrate how to fuse multi-modal information from robot motion sensors and evaluate the proposed framework in multiple scenarios collected from a real mobile robot.