Research On Parts Recognition Of Simulated Workshop Based On Convolutional Neural Network

In modern assembly workshops,studying the visual recognition of bridge cranes for workshop parts is of great significance in the process of unmanned operation.Due to the continuous increase of labor costs in the modern manufacturing industry,especially in the field of heavy industry,the work intensity of labor workers is high,and they often work in dangerous environments,and the working workshop environment is often harsh.This paper is simulated by the platform environment of the unmanned assembly workshop,builds a three-dimensional model of the bridge crane,completes the research on the identification of the gear and bearing parts of the workshop within the frame of the bridge crane,and selects a binocular camera with cmos image sensor in the artificial light source With the assistance of the simulating workshop environment,the image acquisition of gears and bearing parts was completed,and the calibration experiment of the binocular camera was completed to obtain the specific position and coordinates of the object to be measured.Using machine vision instead of manually identifying workshop artifacts,performing preprocessing operations such as image enhancement processing on the obtained images,improving the neural network structure of the convolutional neural network structure YOLOv3,and constructing a feature-enhanced X-YOLOv3 network structure.Thereby improving the recognition accuracy rate of the collected data set,and obtaining a more ideal recognition effect.The research methods of this article are as follows:First,the simulation environment of the unmanned workshop and the experimental platform for identifying parts under the bridge crane environment are built.The preparation work for the identification of the bearing and gear parts in the workshop and the design scheme of the bridge crane structure are completed to realize the unmanned bridge crane.The goal of the run.Aiming at the operation of the bridge crane’s carts and trolleys,a positioning and control scheme for the bridge crane is proposed,which combines the simulation of the workshop platform with the bridge crane as the main body and the recognition of the binocular camera.Comprehensive consideration of various factors affecting data collection,choose CAM-AR0135-3T16 high frame rate variable baseline USB3.0 binocular camera and microscope LED ring light source.After completing the calibration experiment,collect photos of the data set of the workshop gear and bearing parts.After obtaining the photos of the data set of the workshop parts,the data set is image-enhanced to solve the problem of insufficient data set and prepare for the subsequent application of convolutional neural network for recognition.Then perform the image annotation work on the data set of the workshop artifacts,use the Labeling tool to generate the Annotations file in the xml format required for the training model,and apply the normalization process to reduce the time for data calculation during the model training process,thereby improving The speed of training the model.Using Python programming software as the language for writing image recognition applications,the convolutional neural network structure of the target detection network YOLOv3 was improved,and the feature-enhanced X-YOLOv3 network structure was constructed,which realized that a smaller data set can be obtained.Better recognition effect of workshop workpieces.After the feature enhancement,the accuracy and loss value of workshop workpiece recognition have been significantly improved,and a relatively ideal recognition effect has been achieved.This paper is used a convolutional neural network structure framework for visual recognition.According to the location,size,structure and other characteristics of the parts,the convolutional neural network recognition model is constructed,the collected picture information is preprocessed,and the target area of the image is extracted.For real-time detection and recognition,the running results of the software prove that the accuracy of detection and recognition of the convolutional neural network model X-YOLOv3 has been significantly improved compared to before the improvement.