Satellite failures account for heavy, irreparable damages, especially when associated with the Power System which is the heart of a satellite. Anomalies in Satellite Power System (SPS) can lead to complete failure of the mission. This demands the need to understand the causes of power system related failures. Huge number of sensors installed in a satellite system conveys information regarding the health of the system. The conventional manual level checking of sensors can be augmented with data driven fault diagnosis approach to reduce the false alarm and burden on operating personnel. The latter has the advantage of exploiting the interrelationship between sensor measurements for fault diagnosis. In this work, Convolutional Neural Network (CNN) is trained on satellite telemetry data for sensor fault detection in SPS. Various processing schemes in time and frequency domains were explored to process the input data to CNN. Promising results were obtained with combination of Stockwell transform (S-transform) and CNN for data processing and classification, respectively. Advanced Diagnostics and Prognostics Testbed (ADAPT), a publicly-available dataset was analysed and used for validating the proposed algorithm, yielding an accuracy as high as 96.7%, precisison of 0.9, F1 score of 0.95 and AUC equal to 0.976.

卫星故障造成了巨大的，无法弥补的损失，尤其是与作为卫星心脏的电力系统相关联时。卫星电源系统（SPS）异常会导致任务完全失败。这就需要了解与电力系统相关的故障的原因。卫星系统中安装的大量传感器传达有关系统运行状况的信息。传感器的常规手动液位检查可以通过数据驱动的故障诊断方法来增强，以减少误报和操作人员的负担。后者的优势是可以利用传感器测量之间的相互关系进行故障诊断。在这项工作中，对卷积神经网络（CNN）进行了卫星遥测数据培训，以用于SPS中的传感器故障检测。探索了时域和频域中的各种处理方案，以处理输入到CNN的数据。结合斯托克韦尔变换（S-transform）和CNN分别进行数据处理和分类，获得了可喜的结果。分析了公共诊断数据集Advanced Diagnostics and Prognostics Testbed（ADAPT），并用于验证该算法，该算法的准确率高达96.7％，精确度为0.9，F1得分为0.95，AUC等于0.976。