Improving Identification of Wind Tunnel Systems Using Clustering-Based Ensemble Outlier Detection Model

Abstract

This paper proposes an outlier ensemble based on the clustering technique to improve the identification accuracy of wind tunnel systems, which are very critical pieces of test equipment in the field of aerospace. The use of clustering has two objectives in our outlier ensemble, that is, generating diverse training subsets and improving the robustness of base detectors. By analyzing the data characteristics of wind tunnel systems, we propose a hybrid criterion to determine the most appropriate clustering algorithm and the corresponding clustering number. This criterion is constituted by a qualitative and a quantitative criterion. The qualitative criterion is implemented first to eliminate several candidates from alternative clustering algorithms. Moreover, the quantitative criterion is used to determine the ultimate algorithm. In addition, a robust base detector is also developed with the assistance of the selected clustering algorithm. Finally, we verify the proposed detection model in two ways. In an offline application, the model is verified through two system identification models. In an online application, it is verified only through the performance of outlier detection.