Research on structured FCN algorithm for combustion chamber temperature field prediction in industrial boilers

To control and perform nondestructive testing (NDT) on the internal temperature field of a boiler, it is necessary to detect the temperature field within the boiler’s temperature field. Obtaining a real-time, high-precision temperature field reconstruction method has always been a crucial requirement for boiler power plants. Therefore, this study proposes a Lightweight Hybrid Data Preprocessing (LHDP) framework and a physics-guided FCN-StructNet-TF deep learning architecture for the real-time reconstruction of the temperature field in a coal-fired boiler combustion chamber. Compared to FCN-32s, the FCN-StructNet-TF model adopts a 4-level structured multi-connectivity and multi-channel attention mechanism, which can improve spatial resolution to 0.1 m. Validated on a 310 t/h industrial boiler, this framework scored 0.892 using the entropy weight TOPSIS method, indicating an exceedingly high level. Compared to other novel reconstruction frameworks, this framework achieves a 57.8% improvement in processing time and requires only 0.78 s to complete one reconstruction plus plotting on low-computing-power platforms. This method exhibits high accuracy, efficiency, and generalization capability under limited labeled data scenarios, providing a novel approach for the intelligent monitoring of coal-fired boilers.