Research On Intelligent Instrument Reading System Based On Inspection Robot

With the increasing level of industrial automation,inspection robots are playing an increasingly important role in industrial production.Accurately reading instrument data is crucial in the process of instrument inspection by these robots.However,traditional manual inspection methods are inefficient and prone to reading errors.Therefore,in order to improve the efficiency and quality of industrial production,a research study was conducted on an intelligent instrument reading system for inspection robots,aiming to achieve real-time and accurate instrument reading functionality.The main research work and achievements of this paper are as follows:(1)A trackless navigation intelligent instrument inspection system,composed of a mobile inspection robot and a host base station,was proposed to address the challenges posed by the diverse types and wide distribution of instruments in factory environments.The inspection robot utilizes its onboard depth camera for autonomous navigation to access multiple inspection points.Instrument images are captured using a visible light camera and transmitted to the base station.Corresponding reading algorithms were designed to achieve instrument reading functionality.Additionally,a visual monitoring interface was developed for real-time monitoring of the inspection process.Experimental results demonstrated that the system achieved full scene coverage,with an instrument recognition accuracy of 95%.(2)To overcome the impact of navigation and localization errors on image acquisition by inspection robots,resulting in incomplete or small target images,an intelligent acquisition solution based on robot visual servoing technology was proposed.After the robot navigates to the inspection target point,the system collects real-time instrument images and uses visual servo technology to adjust the robot’s posture until the system collects instrument images with appropriate positions and sizes.Experimental tests conducted in a controlled environment validated the effectiveness of this solution,showing that it enabled accurate,fast,and automated image acquisition,thereby improving the efficiency and quality of the acquisition process.(3)Addressing the reading challenges posed by different types of instruments,such as dialtype and digital display-type instruments,a solution based on deep learning was proposed and implemented for instrument detection and reading recognition.Firstly,the YOLOv4-tiny object detection network was employed to obtain instrument regions and type labels.Experimental results on the Baidu_Dataset,Pointer-10K dataset,and self-built dataset demonstrated mean average precision(mAP)of 100%,73.65%,and 99.89%,respectively,with a speed of 50 frames per second.Subsequently,corresponding reading algorithms were designed based on the detection results.For dial-type instruments,image preprocessing was applied to improve image quality,and a U-net network was utilized to segment the dial and pointer.Readings were calculated based on geometric relationships and label prior information.Experimental results revealed that the overall relative error of dial-type instrument readings could be as low as 0.89%.For digital display-type instruments,similar image preprocessing was performed,followed by projection-based segmentation of the digital display area,and character recognition was achieved using the CRNN network.Experimental results showed that the recognition accuracy on the NUM_DATA dataset and POINT_NUM dataset reached 97.07% and 91.10% respectively.