ZHU Yonghong, DAI Chenyu, LI Manhua
(School of Mechanical and Electronic Engineering, Jingdezhen Ceramic University, Jingdezhen 333403, Jiangxi, China)

Abstract: At present, the firing temperature of ceramic roller kiln is mainly detected by thermocouples. Due to the easy aging of thermocouples, the temperature detection accuracy of thermocouples gradually becomes low so as to affect the firing quality of ceramic products. For this problem, an intelligent temperature detection method of fusing flame image feature recognition based on deep learning with thermocouple point detection data instead of thermocouple. This method is that a multi-scale feature extraction network based on the shift window visual self-attention mechanism is adopted for the flame image of ceramic roller kiln, convolutional neural network and local and remote features of transformer branch is used for retaining more image information to obtain more accurate flame image features which are fused with thermocouple point detection data, thus, the temperature of ceramic roller kiln can be accurately detected. In the multi-scale feature extraction network model, firstly, an auto-encoder network based on multi-layer transformer is used for extracting shallow and multi-scale deep features, and then multiple features are fused into transformer and convolutional neural network to make it be able to capture feature information, finally, the data obtained from the thermocouple point detection is input into the front network to achieve the fusion of the flame image features and the key point detection temperature data by the feature level information fusion. Experimental results show that the fusion network model proposed in this paper is 1.75% higher in average feature recognition accuracy and 2.67% lower in average error generation than the convolutional neural network fusion method, which is superior to the convolutional neural network branch or transformer branch image fusion in most indicators. Hence, the method proposed in the paper is effective and feasible.
Key words: ceramic roller kiln; deep learning; information fusion; intelligent temperature detection