Machine Learning Fault Detection and Diagnosis
The increasing numbers and complexity of small spacecraft and missions is a driving a growing need for automated fault detection, isolation, and recovery. Traditional methods possess limited ability to perform on-board fault diagnosis, leaving ground operators to interpret telemetry and isolate root causes. This could prove overwhelming to operators of large constellations of small satellites. Machine learning methods offer an alternative approach that could benefit greatly from the data generated by these constellations. Unfortunately, previous studies have found these methods susceptible to high rates of false positives when detecting anomalies. Fault isolation also remains challenging due to the difficulty of obtaining examples of faulty data needed to train supervised classifiers. We present a new data-driven fault detection and isolation architecture that uses one class support vector machines to provide a running time series of fault signals to a long short-term memory neural network for isolation. The support vector machines are trained to detect anomalies entirely from nominal data while a fault simulator can simulate the activations of the various detectors under different scenarios. Since the long short-term memory network does not receive raw telemetry the requirement to provide perfectly realistic examples of faulty data is relaxed. The ability for long short- term memory to appreciate time dependent context also ensures the system is robust to spurious false positives from the detectors.