Design Of Flatness Recognition Model Based On Improved Neural Network

Strip steel is a basic industrial production material.With the rapid development of social economy,various industries have higher and higher quality requirements for strip steel.Flatness is an important performance index used to measure the quality of strip steel products,and strip shape recognition is an important part of cold rolling process control,which has a direct impact on product quality.Based on the artificial intelligence algorithm,this paper conducts in-depth research on the flatness recognition model.The main research work is divided into two parts,one is the research on the image-based flatness recognition method,and the other is the research on the flatness recognition model based on the contact flatness detection method.The main work is as follows:Aiming at the apparent flatness defects,the image-based flatness recognition method has the advantages of non-contact detection,intuitiveness,and intelligence.In this paper,Gaussian kernel function is applied to image feature extraction,and Gaussian convolution is proposed.Compared with traditional convolution,Gaussian convolution introduces nonlinearity into the convolution operation at the cost of adding a small number of parameters and computational complexity.A HN-CNN composed of standard convolution,Gaussian convolution and other nonlinear convolutions is proposed.It can extract linear and nonlinear features of input information at the same time and realize effective feature fusion,which improves the capacity of the model.And expressive ability,enhance the generalization ability of the model.And establish a flatness recognition model based on HNCNN to effectively reveal the relationship between optical characteristics and flatness defects,and realize accurate non-contact flatness defect recognition.The simulation results show that the convergence effect and recognition accuracy of the HN-CNN proposed in this paper are better than standard convolution.In addition,the flatness recognition model based on HN-CNN can better identify flatness defects,and compared with existing recognition methods,the average accuracy is increased by 5.6%.For potential flatness defects,in order to solve the problems of insufficient feature extraction and poor adaptive ability when using Euclidean distance,PCA and other methods to extract characteristics of flatness stress value signals,automatic encoder(AE)is used instead Euclidean distance,PCA and other methods.That is,an unsupervised learning method is used to extract the features of the flatness signal,and combined with the RBF neural network,the AE-RBF flatness recognition model is designed.And in view of the shortcomings of genetic algorithm,an improved adaptive genetic algorithm(IAGA)with dynamic changes in crossover rate and mutation rate is designed as the evolution progresses,which effectively improves the training effect of the model.Finally,the simulation experiment shows that the designed IAGA optimized AE-RBF flatness recognition model can better complete the flatness defect recognition.And its average recognition accuracy of test samples is higher than existing methods.