Ensemble pruning-based anomaly detection for improving data quality in ladle furnace temperature prediction

Molten steel temperature prediction is a crucial step in intelligent steelmaking. However, the accuracy of prediction models is often limited by low-quality industrial data, making it difficult to meet engineering requirements. To address this issue, this study investigates anomaly detection in ladle furnace production data to improve data quality and enhance the performance of prediction models. Existing anomaly-detection algorithms for ladle furnace data mostly rely on traditional single-model approaches, which lack robustness and stability under complex operating conditions and offer limited improvements to prediction models. This paper proposes an ensemble pruning-based anomaly-detection method. By designing a fitness function tailored for anomaly detection, the method better characterizes the performance of any sub-ensemble. In addition, a genetic algorithm incorporating a multi-population co-evolutionary strategy is introduced to search for the optimal sub-ensemble. Experiments on both public datasets and real-world ladle furnace data validate the effectiveness of the proposed method in terms of anomaly detection and prediction model enhancement. Results demonstrate that the proposed fitness function and optimization method successfully achieve optimal sub-ensemble selection, outperforming competing algorithms.