Research On Abnormal Detection Of Relay Outer Surface Based On Deep Learning

Intelligent surface inspection technology based on deep learning is in great demand in many fields such as positioning inspection and surface anomaly determination in manufacturing processes,and is gradually replacing the inefficient manual inspection methods.However,in practical applications,the complex industrial scenarios,multi-task crossover working environment and the specificity of the inspection task itself hinder the promotion of the technical approach.In this study,a complete set of detection solutions including system construction,hardware selection,software development and deep learning algorithm design are designed to address the appearance abnormality of automotive electronic relays.The work of this study is briefly described as follows.(1)In response to the problems of unbalanced and ineffective enhancement of online enhancement methods,a data enhancement method with parallel online and offline methods is proposed.The offline enhancement methods such as mosaic and mixup are used in combination with the traditional online enhancement to analyze the statistical characteristics of the data and build an efficient database management model.(2)For the detection difficulties of relay outer surface defect targets with background viscosity,intra-class differences and inter-class correlation,we propose the Tri-DFPN module with feature-dense cross-layer connectivity and self-attentive feature fusion,establish bottom-up enhancement branches to supplement the localization information in the feature map,and expand the feature transfer from local layers to richer contextual information.(3)To address the low-quality problem of the suggestion frame,a three-stage cascade module correction module is designed to build a Cascade Tri-DFPN defect detection network.By resampling the suggestion frame and training the single-range detector step by step with the three-stage IOU,we effectively solve the mismatch problem of the suggestion frame in the training and inference sessions,and combine with the dense feature pyramid to improve the detection performance of the model for relays at multiple scales.(4)A complete relay appearance abnormality detection process is designed for the light and other production parameters of the field implementation environment,etc.It includes hardware selection,acquisition system construction,UI software design and deep learning algorithm design implementation.The field construction and online testing of the detection process was completed,achieving a 94.58% recall rate of defective samples and a single inference test time of 0.077 s,meeting the speed and accuracy requirements.The study not only provides a complete solution for relay appearance inspection,but also has reference significance for surface inspection of other products.