| description abstract | The process monitoring system for wind tunnel flow fields enables real-time monitoring of anomalies and faults within the flow field, allowing for the implementation of necessary adjustments to ensure the normal operation of the wind tunnel. Because of the performance limitations of single models, many ensemble monitoring models based on ensemble learning have been proposed. To ensure diversity within the ensemble, a large number of base learners are generated. However, it has been discovered that the redundancy resulting from a large number of base learners not only harms the performance of the ensemble model but also increases computational burden and storage overhead. To address this, this paper presents an ensemble monitoring model based on ensemble pruning. Specifically, given the unavailability of data labels in the wind tunnel flow-field data sets, one-class classifiers are employed as the base learners. After the ensemble generation stage, the performance of each base learner is estimated based on its correlation with the proxy model. Then, using the estimated performance, a dedicated base learner selection mechanism is proposed based on statistical testing to filter out redundant individuals. To validate the effectiveness of the proposed monitoring model, nine data sets from real wind tunnels were used for model training, and an additional nine data sets were used for testing. Experimental results demonstrated the significance of ensemble pruning in enhancing ensemble performance. | |
