Anomaly Detection for Wind Tunnel Flow Field Based on Ensemble Learning

Anomalies in wind tunnel data can seriously affect data integrity, making efficient anomaly detection essential for improving data-driven methods. Research has shown that ensemble models for anomaly detection generally achieve better stability and performance than single models. However, current ensemble techniques often prioritize diversity excessively, leading to redundant base models that reduce overall effectiveness. To solve this issue, we introduce ensemble pruning into anomaly detection. First, we develop a hybrid ensemble generation method combining bagging and subspace techniques to better balance diversity and model accuracy. Next, we establish quantitative measurements for both diversity and accuracy, which form the joint optimization objectives for ensemble pruning. A multiobjective optimization algorithm is then applied to obtain the optimal pruned ensemble. We validate the method using real wind tunnel test data through three evaluation perspectives: anomaly detection performance, Mach number prediction accuracy, and process monitoring capability. Results demonstrated that our ensemble pruning–based approach significantly outperforms existing ensemble models and other pruning methods. The proposed method enhances detection reliability while maintaining computational efficiency, providing an improved solution for quality control in wind tunnel experiments.