Research On Recognition Method Of Coal-fired Flame Condition Based On Deep Transfer Learning

Coal-fired kilns are widely used in metallurgy,power generation and other industrial fields.They are a kind of complex industrial objects with large hysteresis and strong nonlinearity,which make mathematical modeling difficult and difficult to effectively control.The current monitoring of coal-fired conditions mainly relies on “manual fire watching”,which directly affects product quality and energy consumption.Because the coal-fired flame image contains a wealth of coal-fired condition information,the recognition of coal-fired condition based on the coal-fired flame image is the key to realize the stability control of coal-fired kilns.In recent years,deep learning has shown excellent performance in the field of image recognition.Since a well-performed deep learning model requires a large number of labeled samples,and in the actual production of coal-fired kilns to prevent abnormal operating conditions,it is difficult to obtain large-scale labeled coal-fired flame images.In response to this problem,transfer learning technology has been widely used in image recognition to improve the recognition effect.Therefore,in view of the low recognition accuracy of coal-fired flame conditions and the need for a large number of labeled coal-fired flame images for supervised deep learning,this paper studies the recognition method of coal-fired flame conditions based on deep transfer learning to improve the recognition accuracy of coal-fired conditions,which is of great significance for industrial energy saving and consumption reduction,and improvement of production quality.The main research content is divided into the following points:(1)Study the basic structure,optimization algorithm and training process of convolutional neural network,and analyze its advantages and disadvantages in the field of image recognition.At the same time,the basic principles and classification of transfer learning methods and the advantages of transfer learning in deep learning are explained and analyzed.(2)In order to improve the recognition accuracy of coal-fired flame condition and solve the problem of insufficient data of labeled coal-fired flame images,this paper proposes a coal-fired flame condition recognition method based on deep CNN model transfer.On the coal-fired flame image data set,fine-tune the four models pre-training models includes Alex Net,VGGNet,Goog Le Net and Res Net to realize the recognition of coal-fired flame conditions,and evaluate the transfer performance of the four models in the task of coal-fired flame recognition.The results show that the method proposed in this paper has a better recognition effect of coal-fired conditions,and the transfer learning model based on Res Net has the best performance recognition effect.(3)In order to solve the problem that traditional transfer learning methods cannot capture the domain-specific characteristics of coal-fired flame images,this paper proposes a coal-fired flame condition recognition method based on unsupervised deep transfer features.Using the unsupervised learning method DBN combined with the pre-training model Res Net,the domain-specific features of the coal-fired flame image are further extracted on the basis of the general transfer feature,and then the coal-fired flame condition recognition is realized.The experimental results show that the fusion of transfer learning and unsupervised learning methods has higher recognition accuracy of working conditions than traditional transfer learning methods and advanced supervised fine-tuning methods,and further reduces the dependence on labeled coal-fired flame image data.(4)In order to further optimize the performance of the coal-fired flame condition recognition model,enhance the robustness of the model and avoid over-fitting,a coal-fired condition recognition method based on ensemble deep transfer learning is proposed.On the basis of unsupervised deep transfer features,the base classifiers are constructed by using the better-performing pre-training models VGGNet,Goog Le Net and Res Net,and the base classifier is trained using the data set sampled by the Bagging algorithm.Then,the output of each base classifier is integrated through the relative majority voting method to improve the model performance and the accuracy of working condition recognition.The experiment verifies the effectiveness of the method.