UAV Anomaly Detection Using Active Learning and Improved S3VM Model
Dawei Pan, Longqiang Nie, Kang Weixin, Zhe Song · 2020
Unmanned aerial vehicles can complete various specific tasks and play an increasingly important role in various fields. However, UAVs often have anomalies during flight, which is likely to cause huge losses. Therefore, the use of data-driven anomaly detection methods has attracted attention. Because UAV sensors have less labeled data and more unlabeled data, semi-supervised support vector machine (S3VM) classification method is introduced. Considering that unlabeled instances may not have the information content and the distribution assumptions used may not reflect the real time series distribution well, the pure S3VM classification method may not necessarily achieve the ideal classification effect. For existing time series data, active sampling is used to mine the most valuable data of the classifier model through margin sampling (MS). In addition, due to the overlapping of different types of data, the classification boundary of the S3VM classifier cannot be located in a low-density region. By adjusting the distance of most samples to the classification boundary, the classification boundary is in a low-density region, which meets the theoretical design requirements of the classifier. Experiments show that combining MS active learning and improved S3VM classification method will achieve the ideal classification effect. Using the UAV time series data, according to the classification prediction results of the classifier model during the operation stage of the training data training, an estimate of the prediction uncertainty is given. Anomaly detection is realized by comparing the predicted value with the uncertainty interval for classification. This paper uses three sets of UAV channel telemetry data for experimental verification. Using MS active learning and the improved S3VM algorithm compared with SVM and S3VM algorithms, it verifies the effectiveness of the algorithm in different UAV data sets for anomaly detection.