A multi-parameter fire detection method for aircraft cargo compartments based on CNN-Swin Transformer

This study proposes a multi-parameter fire detection method for aircraft cargo compartments based on a convolutional neural network (CNN)-Swin Transformer fusion architecture, aiming to address the high false alarm rates and low accuracy of traditional single-sensor detection systems. A multi-sensor data acquisition system was built to generate a multimodal time-series dataset comprising carbon monoxide mass fraction, dual-wavelength smoke characteristics (blue light power, red light power, and Sauter mean diameter), temperature, and humidity. An innovative fusion architecture was designed, integrating a multi-scale 1D dilated convolutional module with an improved Swin Transformer module. The multi-scale convolutional module effectively captures local transient features, while the transformer module’s window-based multi-head self-attention mechanism models global cross-sensor dependencies with low computational complexity. Furthermore, a dynamic position encoding mechanism was introduced to embed fire-related physical features into positional representations, improving the model's ability to perceive fire evolution patterns. An adaptive feature fusion gating mechanism was used to dynamically integrate local and global features. The loss function combines weighted cross-entropy and focal loss to handle class imbalance effectively. The experimental results show that the proposed model achieves a macro-average F1-score of 0.98 on a simulated cargo compartment fire dataset, with only a 4.2% decrease in F1-score under 5% noise disturbance, outperforming mainstream baseline models. Ablation studies confirm the synergistic effects of each module, and decision boundary visualization indicates that the model’s decision-making mechanism aligns well with the physical principles of fire dynamics. This study proposes a high-precision, low-latency, and well-interpretable solution for fire detection in confined spaces, meeting the stringent requirements of civil aviation cargo compartment safety monitoring.