Abnormal time series detection under unsupervised and multi-working conditions for power system of unmanned equipment

Abstract

Due to the importance of the power subsystem in unmanned equipment (UE), it needs accurate anomaly detection methods to ensure its safe and stable operation. Under the condition of unsupervised and extremely lack of abnormal samples, deep learning methods and machine learning methods are difficult to apply. In addition, the operating conditions of UE are variable and abnormal modes are diverse, which bring greater challenges to the anomaly detection task. To solve the above problems, this paper proposes a new anomaly detection method for time series of unmanned equipment power system (UEPS). Firstly, a similarity measurement method based on series decomposition and fast dynamic time warping is introduced to improve the accuracy of series similarity measurement. Then, taking the time series as samples, the density clustering algorithm is applied to separate the categories with small sample number as anomaly categories. Finally, based on the normal class labels obtained by clustering, the support vector data description algorithm is used to realize new abnormal time series detection in the absence of abnormal samples. In the anomaly detection experiments on power system data in Unmanned Aircraft System Mission, the method proposed in this paper shows excellent performance, and effectively solves the problem of unsupervised, multi-working conditions and lack of abnormal samples.