Research On Abnormal State Detection Of Blast Furnace Tuyere Image Based On Deep Learning

In the steelmaking industry,because the furnace is in a state of high temperature and high pressure during the blast furnace steelmaking process,how to ensure the safe and stable operation of the blast furnace is particularly important.The blast furnace operator observes the operation status of the blast furnace and the coal injection status through the blast furnace tuyere or the image transmitted from the blast furnace tuyere camera.Therefore,the working state of the blast furnace tuyere is an important basis for obtaining the internal information of the blast furnace.Due to the large number of blast furnace tuyere,the traditional detection method relying on manual experience judgment is timeconsuming and labor-intensive,and there are problems such as poor real-time detection and low accuracy of abnormal state detection,which in turn affects the safe operation of the blast furnace.In view of the above problems,this paper takes the blast furnace tuyere image as the research object,and constructs a blast furnace tuyere anomaly state detection model based on deep learning by improving the channel attention mechanism,and compares it with rep VGG and other models.The main work of the thesis is as follows:Aiming at the problems of low imaging of blast furnace tuyere images,lack of obvious differences in abnormal air ports,and difficulty in extracting features from existing deep learning models,this paper classifies image recognition of abnormal state of blast furnace tuyere vents as a fine-grained image classification task,improves the channel attention mechanism,and constructs an improved channel attention module SEB(Squeeze-and-Excitation Block),which fuses channel information to halve the number of channels and reduce redundant feature information.Then,the SEB module and the residual module are stacked on top of each other,the high and low feature information is integrated,an efficient channel attention residual network ESERNet(Efficient Squeezeand-Excitation Residual Network)is constructed,and the model is trained by using the preprocessed data set,and the optimal parameter combination of the model is obtained through the K-fold cross-validation experiment.Finally,the ESERNet model is compared with the SERNet,Res Ne Xt and rep VGG models.The results show that the average accuracy of ESERNet model is 97.10%,and the number of parameters is reduced by6.23%～56.71%,which greatly reduces the complexity of the model while ensuring high accuracy.Aiming at the problem that the channel attention mechanism is easy to ignore local feature information,this paper constructs the local channel attention module LSEB(Local Squeeze-and-Excitation Block).The LSEB module focuses on the local information of the feature image,obtains the weights of each region channel through convolution,and then performs global information integration.Finally,the local channel attention module and the residual module are combined to construct the LSERNet(Local Squeeze-andexcitation Residual Network).The model was trained using the blast furnace air outlet image dataset and compared with models such as Res Net50 and Res Ne Xt.The results show that the classification accuracy of the LSERNet model proposed in this paper is98.59%,which is 0.65%～1.17% higher than that of other models,and the feature information extraction ability of the LSERNet model is the strongest.