IRMAE-AKDE: A Novel Solution to Deep Imbalanced Regression for Performance Prediction of Rolled Steel Plate

Yield strength (YS) is a critical mechanical property that directly impacts steel plate performance. Accurate YS prediction is essential to ensure consistent product quality in industrial applications. However, current data-driven approaches often struggle with imbalanced data distributions, resulting in suboptimal performance, particularly in regions with limited sample availability. To this end, we propose a novel approach, named IRMAE-AKDE, which integrates the implicit rank-minimizing autoencoder (IRMAE) with adaptive bandwidth kernel density estimation (AKDE) to improve YS prediction accuracy under imbalanced conditions. Specifically, IRMAE constructs compact latent spaces to effectively represent sample features, while AKDE dynamically adjusts label distributions based on sample density, allowing for more robust predictions across regions with limited data. We design adaptive cost-sensitive weights for each sample to train and optimize the model end-to-end. The proposed approach is evaluated on benchmark datasets and the STEEL-DIR dataset of approximately 3500 hot-rolled steel plate samples, each with 35 attributions and the corresponding YS values. Compared with the state-of-the-art methods, IRMAE-AKDE outperforms in the med-shot and few-shot regions of the AILRONS, COMPATIV, TEMPERING, and STEEL-DIR datasets. In practical applications, IRMAE-AKDE reduces the mean absolute error (MAE) by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4.7\%, 14.1\%$ </tex-math></inline-formula>, and 4.4% and the error geometry mean by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$5.2\%, 10.7\%$ </tex-math></inline-formula>, and 3.3% across the overall, med-shot, and few-shot regions, respectively. These results underscore the effectiveness of our approach in tackling the challenges of YS prediction in steel plate production and advancing its applicability in deep imbalanced regression.