Damage quantification of mining explosion-proof equipment using thermal flame dynamics and improved LSTM.

To address the lack of quantitative characterization methods for flame propagation during internal non-ignition detonation tests of explosion-proof equipment, this study proposes a damage severity quantification method using an improved LSTM model with thermal imaging feature fusion. A detonation flame thermal imaging dataset is constructed, where flame contours are segmented and features extracted via a pre-trained Faster R-CNN. Temporal sampling generates flame image sequences, from which key spatiotemporal features-area, centroid velocity, aspect ratio, circularity, and area change rate-are derived to describe dynamic flame propagation. Based on this, the improved LSTM captures time-varying behavior: a first-layer bidirectional LSTM fuses forward and backward temporal data to detect propagation direction changes, multi-layer stacking extracts high-level features for adapting to damage scale variation, and Dropout between layers mitigates overfitting. Experimental results show prediction errors remain within 2% across 5-20 cm damage levels. The near-zero shift in error distribution confirms the effectiveness of bidirectional fusion, deep feature modeling, and regularization in capturing flame temporal dynamics. This research provides a data-driven approach for accurate damage diagnosis and supports safety certification and maintenance of mining explosion-proof equipment, with implications for coal mine safety.