Working Condition Recognition Algorithm Based On Res3D CNN And Imbalance Classification

Rotary kiln has extensive research and application in many industrial fields such as metallurgy,iron and steel,cement,etc.Identifying the working conditions of the rotary kiln is the focus of the production process.It not only affects the quality of the output,but also relates to the energy consumption and pollution emissions of the rotary kiln.With the rise of computer vision technology,many researchers have begun to study image-based recognition methods and have achieved good results.This article uses deep learning methods to study the issues of multi-channel feature extraction of flame images,network information transmission and imbalance classification,and proposes a model that can efficiently represent the flame motion state.This article has carried out the following research work:Aiming at the problem that it is difficult for a single frame image to reflect the dynamic change information of the flame in the time dimension and the flame image is disturbed resulting in the poor robustness of the features extracted from it,a multichannel spatiotemporal feature extraction network is proposed,which is based on three-dimensional convolution.It can extract the spatiotemporal features of the image sequence to make up for the traditional method’s inability to obtain the flame dynamic change information.At the same time,it fully considers the characteristics of the flame image with multiple details and different feature scales.Four parallel threedimensional convolutions of different sizes are used to enrich the spatiotemporal feature expression of the flame and enhance the robustness of the flame image features obtained.Experimental results show that this method has achieved better performance than traditional two-dimensional and three-dimensional convolution methods.Aiming at the problem that the learning ability is difficult to maintain after the network level is deepened,and the flame image feature may lose information in the network,a three-dimensional residual network is proposed,which is based on threedimensional convolution and has the characteristics of shortcut connection and data dimension conversion.Shortcut connection avoids information loss to a certain extent,guarantees the learning efficiency of the network,and data dimension conversion greatly reduces the amount of network parameters and reduces the computational cost of the model.Experimental results show that this method can effectively retain the information in the network,improve the efficiency of flame image recognition,and at the same time increase the training speed of the model.Aiming at the problem of the imbalanced number of samples in various states of the rotary kiln,a multi-channel residual network is proposed.This network combines the advantages of multi-channel and residual network.It enriches feature expression while preventing information loss,and can efficiently characterize flame images.This article choose to use the Adaboost classifier based on SVM to classify the combustion state samples of the rotary kiln.Adaboost is a common algorithm to solve the problem of data imbalance.It is easy to use,efficient,and can be detected in real time on the industrial site;SVM is a mature classification algorithm that can process nonlinear,high-dimensional,and small sample data,and is suitable for processing highdimensional spatiotemporal characteristics of rotary kilns.Using the framework of the combination of multi-channel residual network and Adaboost,using real data from the industrial site,a large state-owned enterprise industrial site rotary kiln operating conditions were identified and compared.The results show that the framework has better recognition accuracy for each state of the rotary kiln.At the same time,the recognition accuracy of small sample data is better than other methods.