Machine Vision On-line Detection Algorithm And Adaptive Research And Development Of Commutator Hook Defects

Commutator is one of the key parts of the motor,which has high processing quality requirements and large market demand.However,manual quality inspection has many problems,such as slow speed,high cost,and the subjective influence of the staff on the inspection standard is difficult to quantify.Automatic inspection based on machine vision can make up for its disadvantages.However,the fluctuation of the actual production conditions has an impact on the imaging characteristics of the commutator,resulting in a large discrete distribution of image gray,edge and other features.In the early stage,the detection algorithm based on the independent experimental platform has a large fluctuation of false alarm rate and false alarm rate in the actual production,with poor adaptability.In this paper,a set of defection detection system for commutator hook is designed,which can adapt to the fluctuation of actual working conditionsAccording to the requirements of commutator hook defect detection,the overall software and hardware design of the detection system is completed.In view of the influence of the fluctuation of the actual production conditions on the imaging characteristics of the workpiece,this paper collected the image samples of different batches of workpiece in the actual production,combined with the typical image samples,analyzed the influence of the factors such as the quality of the workpiece,the wear of the brush wheel on the imaging,and proposed the adaptive design idea of the hook defect detection algorithm under the actual conditions.In order to solve the problem that the discrete distribution of image edge,gray level and other features in actual production makes it difficult to segment the defect area of straight hook surface completely,this paper designs a set of cascade segmentation structure,and proposes a strategy of segmentation from coarse to fine,with global to local layer by layer,which has good adaptability to the actual working conditions.There are three layers in the structure,the first layer is based on the template matching of edge features to complete the coarse positioning of the straight hook region in the global region;the second layer is based on the BP neural network knot model,which transforms the region positioning problem into the pixel point classification problem,and completes the fine positioning of the straight hook region in the local region;the third layer is based on the method of double threshold segmentation to complete the segmentation of the surface defect region.Aiming at the problem that the hook area can not be accurately segmented and the hook defects can not be distinguished by simple quantitative indicators,this paper proposes a hook shape defect detection algorithm based on feature coding and machine learning.In the feature extraction stage,considering the real-time requirements,the SIFT feature descriptors are improved and the computation is reduced;in the feature coding stage,the number of hook defect samples is small and the samples are unbalanced,so the random sampling of the original features has the problem of uneven features.Therefore,a classification sampling strategy is proposed to improve the description ability of the visual vocabulary library;finally,the random forest is constructed The classifier completes the defect recognition and classification.Finally,the adaptability,accuracy and detection efficiency of the algorithm are verified by experiments.Samples of different blank batches,different wear states of brush wheels and different wear states of dies are collected as the test library.Finally,the experiments show that the algorithm can meet the actual production needs.