Research On Defog And Combustion State Recognition Of Rotary Kiln Coal-fired Image Based On Deep Learning

Rotary kilns have been used in many industrial fields,such as smelting metals,producing cement,producing iron and steel,etc.Under the high-temperature environment in the kiln,the material mixture is in the process of movement,and more complicated physical and chemical reactions usually occur.The combustion state and stability have an important influence on the sintering quality of materials,energy consumption and pollutant emissions.Due to the complexity of the sintering process of the rotary kiln,mathematical modeling of the sintering process usually has certain difficulties.Real-time monitoring of the combustion status in the kiln is the key and difficult point of stability control during the coal-fired sintering process.The accuracy of the detection directly affects the level of automatic control of the rotary kiln,which is very important for the application of the rotary kiln in the industrial industry.The sintering state of the rotary kiln during combustion can be well expressed through the image of coal-fired flame.For example,the radiant heat and combustion effect of the rotary kiln during sintering can be reflected from the brightness information.The traditional rotary kiln detection system directly takes temperature as the measurement target,such as thermocouples,high-temperature colorimetric measuring instruments and other detection methods.Because of the limitations of the on-site environment and the structure of the kiln body,these methods cannot accurately detect the sintering condition of the sintering belt,and cannot make the rotary kiln intelligent detection system obtain better results.In recent years,the method of using flame images to identify combustion status has been applied to intelligent detection systems by many scholars.The method of using coal-fired images to identify combustion status has attracted the attention of academic circles at home and abroad.The current processing methods for coal-fired flame images mainly include segmenting the images and extracting artificial features for pattern recognition.These image processing methods usually have the problems of high algorithm complexity and dependence on human experience.And the image-based detection method usually uses clear images as the recognition object,and the harsh on-site environment in front of the kiln will seriously affect the quality of the flame image.The flame images taken are often blurred and low in definition because the kiln is often filled with smoke and dust.In this paper,for the status of combustion image recognition of rotary kiln flame images affected by smoke and dust,a method of defogging and combustion state recognition method for rotary kiln flame images is proposed.The research workof this paper mainly includes the following contents:1.Propose a method of defogging flame image of rotary kiln based on convolutional neural network.This method is designed on the basis of re-arranging the atmospheric scattering model.It does not estimate the transmission matrix and atmospheric light values separately as most existing models,but directly generates clear images through lightweight convolutional nerves.The design of this end-to-end defogging method is easily embedded in other deep learning models(for example,faster R-CNN),which improves the performance of advanced tasks for blurred images.Experimental results on synthetic fog images,natural blur images and rotary kiln flame image data sets prove that the algorithm is superior to existing technologies in terms of PSNR,SSIM and subjective visual quality.In addition,when the convolutional neural network in this paper is connected to faster R-CNN and the joint pipeline is trained from beginning to end,experiments show that the detection performance of blurred images has been greatly improved.2.Propose a method based on transfer learning to identify the combustion state of rotary kiln.Convolutional neural networks have great advantages in feature extraction,and have certain invariance to operations.The current neural network model has the characteristics of large amount of calculation and high calculation resources,and the deep convolutional neural network model is prone to local optimization problems,so transfer learning has become an ideal choice.In order to improve the overall recognition performance,this paper applies the transfer learning method to classify images,compares the parameter initialization model with the transfer learning model,and compares it with other traditional image classification methods.The experimental results show that this method can obviously improve the accuracy of the combustion state of the rotary kiln,has good robustness and generalization ability,and can overcome the traditional image recognition algorithm's reliance on manual extraction of features.Studies have shown that the use of convolutional neural networks has good performance on the rotary kiln flame defogging in terms of measurement indicators and visual effects,which lays the foundation for advanced image processing;and recognizes the combustion of rotary kiln based on deep learning State,this method has high recognition accuracy and fast processing speed,which is superior to traditional image recognition methods.