[Objective] As municipal solid waste generation increases, waste incineration has become a mainstream treatment method due to its advantages in reduction, harmlessness, and resource recovery. To address the challenges posed by the diversity of flame combustion states an d the low efficiency of manual identification in grate-fired municipal solid waste incineration furnaces, this study proposes an improved flame combustion state recognition model, termed MSDRNeXt, based on the ConvNeXt architecture. The aim is to achieve i ntelligent identification of combustion conditions within the furnace, thereby reducing reliance on manual judgment and improving operational stability and safety in industrial applications.

[Methods] To accurately extract effective flame regions under comp lex conditions such as significant brightness fluctuations and dust occlusion, a Multi-Scale Feature Fusion (MSFF) module is first proposed. This module integrates multi -scale feature information and incorporates the Convolutional Block Attention Module (CBAM) to enhance fine-grained texture 2 features in shallow feature maps while suppressing irrelevant noise. The refined shallow features are further fused into deeper semantic feature maps, enabling a better balance between local texture representation and g lobal semantic understanding. In addition, data augmentation and normalization strategies are applied during training to further improve the model’s robustness to environmental variations. This design improves the robustness and generalization performance of the model in complex environments. Furthermore, considering the inherent asymmetry of air distribution on the left and right sides of the furnace, a Dual -Region Differential Pooling (DRDP) module is designed. This module partitions the feature space int o left and right regions and performs differential pooling to explicitly model spatial asymmetry. As a result, the model can effectively capture subtle inter-regional differences, which are critical for distinguishing combustion states with significant lat eral variations, such as biased combustion, raw material emergence, and material accumulation. The integration of MSFF and DRDP enhances the discriminative capability of the model while maintaining computational efficiency.

[Results] Experimental results on a dedicated waste incineration flame dataset demonstrate that the proposed MSDRNeXt model achieves an overall classification accuracy of 91.36%, outperforming the baseline ConvNeXt model and several comparative methods. In addition, the model achieves an inference speed of 277.85 frames per second (FPS), which satisfies the requirements for real -time monitoring in industrial scenarios. Meanwhile, the number of parameters is reduced by 53.43% compared with the baseline model, indicating a lightweight archit ecture with lower computational cost and memory consumption.

[Conclusion] The proposed MSDRNeXt model can efficiently and accurately identify flame combustion states in municipal solid waste incineration furnaces. It provides a reliable solution for real -time monitoring of the incineration process and supports the optimization of subsequent control strategies.