| With the rapid development and progress of manufacturing industry in my country,the intelligent reform of manufacturing technology cannot be delayed.As a non-standard product,high-frequency workpieces are widely used in some special occasions.During the processing of high-frequency workpieces,the relevant processing information of the workpiece is obtained by reading the label in the pallet carrying the workpiece,and the subsequent processing is further completed,which changes the traditional mode of manual input of information and realizes the intelligentization of workpiece processing.However,when the machining process is on the heat treatment stage,the workpiece will be separated from the pallet,resulting in the loss of labels containing machining information,affecting the subsequent intelligent machining process.Aiming at the problem of low recognition accuracy of high-frequency workpieces,this thesis uses deep learning technology to study the improvement method of network structure and loss function,realizing the re-association of workpieces and their labels,so as to improve the intelligence of workpiece processing.The main research contents of this thesis are as follows.(1)According to the characteristics of the high-frequency workpiece image dataset and the requirements of the network model in the recognition task,an improved Mobile Net V2 network model(Mobile Net V2 with Attention block and Feature fusion block,MAF)embedded in the attention and feature fusion module is proposed.Based on the Mobile Net V2 network,the calculation amount and parameter amount of the MAF model are reduced by further adjusting the width factor and resolution hyperparameters in Mobile Net V2;the expressive ability of the location features in the MAF model is improved by embedding the coordinative attention module;the width of the convolutional neural network is increased,the ability of the network to distinguish the feature information of the workpiece size is enhanced,and the robustness of the model is enhanced,by embedding the feature fusion module.Theoretical analysis and the experimental results show that,compared with the original Mobile Net V2 network,the improved MAF model can increase the image recognition speed by 10.45 ms,the recognition accuracy is increased by 1.86%,and the number of parameters is reduced by 17%,achieving a lightweight fast and accurate identification.(2)In order to further improve the recognition accuracy,according to the characteristics of large intra-class differences and small inter-class differences in the distribution of highfrequency workpiece image data,an improved joint loss function(Joint Loss,JL)is proposed that combines angle cosine loss and isolation loss.The angular cosine loss function is used to supervise the distance between the features of heterogeneous workpieces,which overcomes the misidentification caused by the distance of heterogeneous workpieces being too close,and realizes the good distinction of heterogeneous workpieces;On the basis of not changing the feature distance of different workpieces,the isolation loss is used to reduce the distance between the features of the same workpiece,the robust expression of the same workpiece is completed,and the recognition accuracy of the model for high frequency workpieces is improved.Theoretical analysis and the experimental results show that the MAF model using the JL loss function to supervise the training has a recognition accuracy of 98.40% on the highfrequency workpiece test set,which is 4.27% higher than that of the MAF model using the traditional Soft Max cross-entropy loss function. |
