UAV sensor data anomaly detection using predictor with uncertainty estimation and its acceleration on FPGA

Anomaly detection is a common requirement in automatic data analysis, condition monitoring, diagnostics and prognostics. Prediction is a type of data-driven anomaly detection method. However, to quantify the uncertainty of the predictor is a challenging issue for determining the anomalous state of current observed sample or sequence. To address this issue, this work presents a prediction based anomaly detection method with support vector machine (SVM) predicting uncertainty optimal estimation. As a result, both point anomaly and fragment anomaly can be detected with high detection performance. What is more, considering the high real-time demands of actual industrial applications, FPGA based vector processor acceleration is implemented. Thus, this work can meet the embedded system based data anomaly detection requirements, i.e., unmanned aerial vehicle. Experimental results illustrate that anomaly detection FPR and FNR are 5.88% and 2.20%, respectively, and the speedup is 2.6 compared with PC, which indicates good application prospects.