Research And Implementation Of Edge Intelligence Defect Detection Based On Model Compression

The rapid development of intelligent manufacturing field brings about the connection of massive heterogeneous terminal devices,accompanied by the rapid growth of network edge device quantity and data generation.Most traditional industrial manufacturing systems use servers as computing centers to collect data generated by terminal devices and make decisions.Terminal devices only have data sending and storage functions.With the wide application of artificial intelligence technology represented by deep learning in manufacturing industry,edge intelligence technology emerges.The emergence of edge intelligence means that many functions in industrial manufacturing can be directly supported by edge devices,and model training,deployment and inference can also be completed on edge devices.For the demand of defect classification detection for varistor,based on deep learning and model compression technology,thesis realizes the edge intelligent detection scheme for pressure-sensitive resistor defects,deploys deep neural network models on resourcelimited edge devices,and realizes surface defect classification detection for pressuresensitive resistors at the edge end.Aiming at the problem that convolutional neural network has many parameters,large inference computation amount and difficult to deploy on resource-limited edge end,a model compression technology combining model pruning and quantization is proposed.While ensuring the accuracy of model detection,the parameter amount of model is reduced to 1/5～1/22 of original one,inference speed is increased by 30%～80%,real-time performance of convolutional neural network detection at edge end is ensured,and efficiency of pressure-sensitive resistor defect classification is improved.A defect detection system for varistor based on Jetson Nano is built.Using Tensor RT inference engine,a compressed lightweight model for defect classification and detection was accelerated and deployed on an edge device,with a detection accuracy of 95.4% and a frame rate of 31 FPS.Thesis adopts an edge intelligent detection solution based on deep learning and model compression,which realizes the classification detection of surface defects of pressure-sensitive resistors at the edge.Through the combination of model pruning and quantization,the model compression method improves the model inference efficiency and reduces the required computational resources,achieving the effective deployment of convolutional neural network models at the edge.It has certain research significance and application value,and can enable edge intelligent devices to be applied in manufacturing,thereby improving the real-time and intelligence of production.