The existing flame radiation image temperature measurement technology has measurement errors due to the coking problem of the detector lens. There is an urgent need for an online monitoring method that can intelligently eliminate the coking interference. An online monitoring method for the temperature field of power station boilers that integrates flame radiation images and convolutional neural network (CNN) is proposed. Firstly, the detector is calibrated via a blackbody furnace, and the relationship between the monochromatic radiation intensity of the detector and the image intensity is established. Secondly, a CNN model suitable for flame image processing is designed, and the training set is constructed by using the non-coking flame radiation intensity images of the boiler collected on-site to establish the flame radiation intensity image restoration model. Finally, the measurement accuracy of this method is verified by using the simulated coking flame images. The results show that the temperature measurement accuracy decreases with the reduction of the number of training sets. When the number of flame images in the learning set is 3 000, the relative error of temperature measurement is 1.4%. The temperature measurement accuracy decreases as the coking area increases. When the coking area is 30%, the maximum relative error of temperature measurement is 0.7%. Furthermore, studies show that when the model of the detector trained by the learning set calculates the coked images of other detectors, the temperature measurement error will increase, with the maximum relative error reaching 34.6%. This indicates that when applying this method, the detectors of each burner need to be trained separately. The proposed method can intelligently eliminate the interference of coking on the flame radiation image, achieve high-precision online monitoring of the temperature field, and provide reliable technical support for the safe operation and combustion optimization of power station boilers.

radiation thermometry  /  temperature field  /  pulverized coal flame images  /  power plant boiler  /  convolutional neural network