Edge Computing Based Pump Component Defect Visual Inspection System

With the rapid development and widespread application of machine vision,more and more enterprises have started to focus on and apply AI technology to improve efficiency and reduce costs.As one of the core components of the automotive transmission system,the quality of the oil pump directly affects the performance of the car in terms of power,economy,reliability,and safety.In the automotive industry,the quality inspection of oil pumps are extremely important links,which are of vital importance for ensuring the quality of cars and user experience.This paper focuses on the research of pump defect inspection.Through a deep investigation and analysis of the defect inspection scenarios and algorithms of domestic and foreign,combined with the actual industrial environment requirements of the pump manufacturing enterprises,a pump defect visual Inspection system based on machine vision,object detection algorithm and edge-embedded development platform is proposed.The application of deep learning technology solves many difficult problems in pump defect inspection,improves the accuracy and real-time performance of defect inspection,and provides an efficient and reliable means of quality control for pump production.The main research contents and innovative points are as follows:(1)To address the misalignment issue in pump component defect inspection,a novel algorithm called YOLOX-FA based on the YOLOX object detection network model is proposed.This algorithm proposed six-parameter representation method firstly,uses a decoupling head structure and a Tanh activation function,achieves 0°-360° full-angle object heading detection,and simultaneously detects object position,size,category,and orientation information.The experimental results show that the YOLOX-FA algorithm achieved a m AP@0.5angle45° performance index of 77.64%.In practical applications,the detection accuracy of the algorithm was 94.51%,indicating that YOLOX-FA algorithm has excellent performance and practical value.(2)To address the problem of decreased detection accuracy of orientation-free targets and overlapping detection boxes of similar targets in the pump quality inspection system,optimization solutions were proposed for the six-parameter representation method and NMS algorithm,respectively.The orientation-free targets were accurately described using the angle-free six-parameter representation method.The NMS algorithm was improved by grouping and secondary filtering strategies,which effectively eliminated the problem of overlapping detection boxes of similar targets.The ablation experiment results showed that the detection accuracy of YOLOX-FA algorithm in detecting all missing parts and misaligned orientations on the pump surface increased from 41.74% to 79.83%,demonstrating the effectiveness of the proposed optimization methods.Moreover,using a single network model alone,the YOLOX-FA algorithm completed all detection tasks in the pump quality inspection system,which is conducive to the edge-side transplantation of the algorithm.(3)In order to develop a visual defect inspection system for pump components based on the YOLOX-FA algorithm on Jetson edge computing devices,a compact network design scheme is proposed.Depthwise separable convolution is used to replace some ordinary convolutions,significantly reducing the model parameter quantity and shrinking the model size by about 50%.The model optimizer in the Tensor RT framework is used for model serialization and algorithm deployment,further reducing the model size to 57 MB and improved the performance of the algorithm on Jetson edge computing devices.In conclusion,the defect inspection system for pump components developed based on the YOLOX-FA algorithm performs stably and achieves high detection accuracy on Jetson edge computing devices.It can also be applied to quality inspection of other industrial products and has significant industrial application value.