Recognition of Combustion Conditions in MSWI Process based on Hierarchical-Split ResNet

Municipal solid waste incineration (MSWI) has significant advantages in the treatment of municipal solid waste, however, the instability of the solid waste combustion state can lead to a reduction in the operational efficiency of the whole system. In addition, the judgement of the MSWI state mainly relies on the operators' own experience to control the combustion, leading to the identification results with substantially subjective and arbitrary. In order to address these challenging problems, a Hierarchical-Split residual neural network is proposed for MSWI process combustion state recognition, which is modeled by collecting combustion images, combining the experience of field experts and the incinerator grate structure, and classifying the combustion state into four categories: normal burning, partial burning, channelling burning, and smothering. Aiming at the problem of incomplete image information extracted by the traditional single-scale convolution, the Hierarchical-Split feature extraction module is utilized not only to enhance the ability of the model to extract feature information, but also to improve the recognition accuracy of the model. Finally, the proposed method is validated by taking the flame data of a municipal waste treatment plant in Beijing as an example. The experimental results show that the method effectively improves the generalization and robustness, and has good recognition accuracy.