Research on FT-transformer-based intelligent discrimination model for roof stability classification of retreating coal roadways

• A hybrid FT-Transformer–SVM model is proposed for intelligent roof stability classification under small-sample conditions. • Self-attention-based feature weighting combined with GMM clustering achieves 100% consistency with manual grading. Cross-validation, ablation, and hard-sample tests confirm strong robustness and generalization stability. • An engineering software system enables real-time prediction, zoning, and visualization of roof stability in coal mines. Roof stability classification in retreating coal mine roadways is essential for safe production and cost-effective support design. Conventional approaches such as RMR and Q-systems often oversimplify geological conditions, while machine learning models require large datasets and lack interpretability. To address these limitations, this study proposes a hybrid framework that integrates FT-Transformer feature extraction, Gaussian Mixture Model (GMM) clustering, Principal Component Analysis (PCA), and Support Vector Machine (SVM) classification. Seven quantitative indicators representing lithological structure, stress environment, and mining disturbance are used as inputs. The FT-Transformer employs self-attention to adaptively weight features and capture nonlinear interactions. GMM provides unsupervised clustering with sign-corrected attention weights, enabling automated and physically interpretable grading. PCA reduces dimensionality, and SVM achieves robust classification under small-sample conditions. Experiments on borehole and mechanical parameter data demonstrate that the proposed method achieves 100% consistency with manual expert grading, with cross-validation accuracy ranging from 98.3% to 100%, and recall and F1-scores consistently above 0.98. Standard deviations across folds remain near zero, confirming model stability. Ablation tests further verify the contribution of each module, while hard-sample evaluations confirm strong generalization. An engineering case study illustrates the end-to-end workflow, highlighting the system’s practical value for data-driven roadway support design and roof stability risk identification.