CubeSats suffer from low reliability and have little to no Fault Detection, Isolation and Recovery mechanisms onboard. Advanced CubeSat missions such as the Lunar Meteoroid Impact Observer (LUMIO), will use more complex attitude determination and control systems, increasing the need for advanced fault detection. Traditionally this requires model-based fault detection methods which are complicated, computation heavy, and highly sensitive to disturbances. Machine learning has proven proficient at fault detection in several non-space related applications, but training data including spacecraft faults is not readily available. In this research an especially lightweight unsupervised learning method for fault detection is designed for the LUMIO attitude determination and control components. The result is a system capable of detecting artificially induced bias, calibration error, and drifting measurement faults on the scale of 0.1 milliradians per second in the IMU, with no false alarms being raised. The method was tested on simulated LUMIO telemetry from the IMU and reaction wheels as well as on real spacecraft telemetry from the OPS-SAT sun sensor, star tracker, reaction wheels, and IMU. In both cases, excellent fault detection and false alarm performance was observed, highlighting the potential of this method for application in CubeSat AOCS fault detection and isolation.