The growing numbers, complexity, and data return of space missions are driving a need for automated fault detection and diagnosis. Traditional fault monitoring techniques possess limited ability to diagnose anomalies, requiring operators to investigate extensive amounts of telemetry to isolate root causes. An important hurdle to automating fault diagnosis for complex subsystems such as attitude determination and control is that interpreting anomalies to determine a root cause is highly contextual and inherently uncertain. This paper presents a method of using one-class support vector machines to train a long short-term memory neural network to diagnose time-varying faults in a simulated CubeSat attitude determination and control subsystem. Using elements of transfer learning, the one-class support vector machines are retrained on nominal flight telemetry, allowing the neural network to diagnose faults on the flight unit. The diagnoses are robust to missing data, false positives, modeling errors, contradicting information, and multiple faults. An application is demonstrated for the LightSail 2 solar sail satellite, where the system succeeds in diagnosing both known and previously unknown attitude determination and control faults.

太空任务的数量，复杂性和数据返回的不断增长，推动了对自动故障检测和诊断的需求。传统的故障监视技术诊断异常的能力有限，要求操作员调查大量遥测以找出根本原因。对于诸如姿态确定和控制之类的复杂子系统，自动进行故障诊断的一个重要障碍是，解释异常以确定根本原因在很大程度上取决于上下文并且具有内在的不确定性。本文提出了一种使用一类支持向量机训练长短期记忆神经网络，以在模拟的CubeSat姿态确定和控制子系统中诊断时变故障的方法。利用迁移学习的要素，对一类支持向量机进行了名义飞行遥测的再培训，从而使神经网络能够诊断飞行单元的故障。该诊断对于丢失数据，误报，建模错误，矛盾的信息以及多个故障具有鲁棒性。演示了LightSail 2太阳帆卫星的一个应用程序，该系统可成功诊断已知和以前未知的姿态确定和控制故障。