Research On Detection Method Of End Face Defects Of Commutator Based On Machine Vision

As the core component of DC motor,the motor commutator has great influence on the performance of DC motor.With the development of intelligent manufacturing strategy,surface defect detection using machine vision has gradually become the mainstream method of motor commutator surface quality control.The process of motor commutator is complex,and its surface defects are widely distributed and varied.Due to the diversity of manifestations and the complexity of defect characteristics,it is difficult to establish a unified quantitative detection standard for many defects,and it is inevitable that the problems of missed detection and misdetection will occur when the traditional visual methods are used for detection.In view of the detection problems of the defects on the top and bottom surfaces of the commutator,this thesis uses deep learning technology to study the detection methods of the defects on the end surfaces,and develops detection algorithms and detection software.The main research contents include:Based on the analysis of testing requirements and difficulties,the overall scheme of the end face defect detection system of commutator is designed.The hardware of the system adopts the image acquisition equipment on the intelligent flexible manufacturing platform of the commutator.The top surface is illuminated by a high Angle ring light,and the digital camera is used for imaging.The bottom end face of the commutator is illuminated and imitated by the backlight,coaxial light and low-ring light respectively in the mode of single camera and multi-light source.The detection software is mainly composed of detection algorithm module and system monitoring module.Among them,the detection algorithm module realizes the detection of defects such as hook top crack,hook top burr and hook root crack on the top surface of the commutator,as well as brush damage on the bottom end surface.According to the image features of the defects near the top of the hook of the commutator,the semantics of the defects in the region were divided into four categories,and a defect detection algorithm based on semantic segmentation network was designed.In this algorithm,the improved channel attention mechanism is adopted to improve the feature screening ability of semantic segmentation network,and the model is pruned by using the channel attention weight distribution feature to improve the model’s operating efficiency.Using Focal Loss Function alleviates the problem of the decreased segmentation accuracy caused by uneven defect samples.Experimental results show that,compared with U-Net network,the improved network in this thesis has higher segmentation accuracy,reaching average Intersection over Union of 83.76%.Aiming at the problem that the defect characteristics of bakelite brush damage are complex and difficult to describe quantitatively,a bakelite brush loss detection algorithm based on residual network was designed.First of all,according to the analysis and statistics,the brush damage of bakelite is divided into four types according to the damage degree:brushless damage,light damage,moderate damage and heavy damage.In defect detection,the algorithm is used to classify the brush damage of each bottom face image.Then,in terms of network structure design,Inception module is introduced to improve the utilization ability of the network for multi-scale features,and weighted cross-entropy loss function is used to improve the classification effect of the network for different damage features and unbalanced samples.According to the experimental results,comparing with resnet18 network,the network used in this thesis has higher classification accuracy,with an average classification accuracy of 98.3%.Finally,this thesis tests the efficiency and accuracy of defect detection algorithm on the intelligent flexible production line server platform.The experimental results show that the deep learning detection algorithm designed in this thesis has a high accuracy on the intelligent flexible manufacturing platform of commutator.The detection accuracy of defects near the hook top reaches 99%,and the detection accuracy of defects with bakelite brush damage reaches 99.4%.The average detection time is about 0.25 s,which meets the requirements of intelligent flexible production line.