A knowledge and data augmentation-based method for combustion state recognition in cogeneration systems

Accurate recognition of combustion states is crucial for the safe operation of cogeneration systems. However, in actual operation, the quantity of training samples across different combustion states is severely imbalanced, making it relatively difficult to train a deep model for accurate combustion state recognition. This paper proposes a knowledge and data augmentation-based combustion state recognition method of cogeneration systems to accurately identify imbalanced combustion states samples. Firstly, combustion states are labeled by combining the temperature characteristics of the system boiler with the environmental regulations on pollutant emissions. Simultaneously, input variables are selected based on the mechanism knowledge of pollutant formation. Then, a refined auxiliary classifier generative adversarial network (RACGAN), incorporating an independent classifier and self-attention module, is designed to obtain high-quality multi-class combustion states samples. Next, evaluation criteria are established to adaptively filter the generated samples, ensuring their accuracy and diversity. Finally, an improved Residual network (IResNet) model is trained using both generated and real samples to recognize combustion states. Experiments based on actual operational data from a heat and power company show that this method achieves high accuracy, stability and potential in combustion state recognition.