Research On Detecting Bearing-cover Defects Based On Improved YOLOv3

Detecting defect is a branch of target-detection in the field of computer vision.It is widely used in production,daily maintaining of equipment and other occasions.Bearing,as an essential part of various mechanical equipment,will cause faults or even casualties once damaged.Bearing-cover plays a key role in preventing foreign bodies from invading the raceway.It ensures that grease only acts on the rolling body and raceway,meanwhile prevents it overflow,so it is particularly important to detect the defects of bearing-covers.Nowadays,most of the detection of defects on bearingcovers used artificial inspection,it was time-consuming and easy to leak.The traditional algorithms of detecting defect extract features by characteristic operators.The methods obtain the characteristics of covers which are the shallow information and fewer information.The detecting results of using these methods are poor and in a low efficiency.Traditional algorithms cannot meet the requirements of real-time detection and the accuracy of detection.In this work,in order to ensure that the trained model has good generalization,a large-scale dataset of bearing-cover defects is established.The dataset covers four different types of defects to meet the normal requirements of detection.In this thesis,the improved YOLOv3 algorithm is used to detect the bearing-cover defects.Firstly,a new feature extraction network,Bottle Neck Attention Network(BNA-Net),is proposed,the network is composed of the bottleneck structure of residual,the max-pooling layer and the SE attention module.The bottleneck structure can extract more features and deepen the network,the max-pooling layer plays a role of reducing the dimensionality of feature maps.Attention mechanism can focus on learning one’s target characteristics.The attention can forecast the location and classification characteristics by obtaining more details and locational information.Secondly,the use of the unifying feature module makes the output features unify and integrate features which are output by the last four attention residual modules in BNA-Net.This processing will form the attention prediction subnet which will be combined by global basic features.Thirdly,the defect localization subnet is composed of a multi-scale building block and other convolution layer.The idea of multi-scale feature fusion is introduced into the multi-scale building block.Among them,the stacking operation between dilated convolution kernel and ordinary convolution kernel not only reduces the number of parameters,but also expands the receptive field.It preserves the important information of contexts and reduces the complexity of the models.Finally,according to the feature that the size of slight defects on bearing-cover are many times smaller than other defects,a new outputting branch is created.On above basis,we make the operation of regression and classification.The use of the new branch greatly improves the detection results of slight defects.Experiment results show that the proposed method in this thesis can detect the defects accurately which are on the surface of bearing-covers and maintain the efficiency of detection efficiently.The m AP reaches 69.74%,which is 16.31%,13.4%,13%,10.9%,and 7.2% more than that of YOLOv3,Efficient Det-D2,YOLOv5,YOLOv4,and PP-YOLO,respectively.