<p>To address stage recognition under complex damage evolution and signal superposition in steel fiber-reinforced concrete (SFRC) subjected to dry-heat curing, this study develops an acoustic emission (AE)-based classification framework using a Random Forest model that incorporates the <i>b</i>-value feature and Optuna-based hyperparameter optimization. SFRC specimens with fiber contents of 1-3% were cured at 50&#xa0;°C, and a hybrid AE dataset was constructed. The feature set combines conventional time-domain parameters with the evolutionary trend information captured by the <i>b</i>-value. Considering the imbalanced stage-label distribution, in which the post-peak stage forms a long tail because steel-fiber bridging prolongs residual deformation and generates abundant AE activity, Optuna was employed to improve model configuration for this practical data condition. Incorporating the <i>b</i>-value notably improved early-stage recognition compared with a baseline using only time-domain parameters: compaction-stage precision increased from 6.1 to 64.4%, and Cohen’s Kappa improved from 0.079 to 0.366. After optimization, the model achieved an overall test accuracy of 87.36% and a post-peak recall of 95.7%, along with improved precision for the first three stages. Feature-importance analysis indicates that the <i>b</i>-value contributes the largest share (≈ 50%), while AF and Duration provide additional calibration information. From an engineering-monitoring perspective, the high post-peak sensitivity may be useful for early warning; however, the trade-off between false positives and sensitivity should be considered when applying the model in practice.</p>