Fault Detection Of Catenary Droppers Based On Faster R-CNN And EfficientNet Network

The catenary dropper is an important component for maintaining the safe and stable operation of the catenary in electrified railways.Currently,the fault detection of the catenary dropper mainly relies on manual inspection,which is inefficient.It has become a trend to use deep learning technology for intelligent inspection of catenary droppers.The Faster R-CNN network has high detection accuracy and can identify and locate targets in complex environments;At the same time,the EfficientNet network can achieve a good balance between detection accuracy and efficiency,and is suitable for the classification of dropper status.Therefore,a catenary dropper fault detection method based on Faster R-CNN and EfficientNet networks is proposed to locate and classify catenary droppers to determine their fault types.Aiming at the problem that there are multiple catenary droppers overlapping in the catenary dropper images of railway crossings and other sections,which leads to the traditional network detection of catenary droppers prone to missed detection and false detection,a catenary dropper location algorithm based on an improved Faster R-CNN network is proposed.Firstly,a CBAM hybrid attention mechanism is added after the main feature extraction network to improve the feature extraction ability of the main feature extraction network for catenary droppers and enhance the attention to key information in the image;Secondly,in order to improve the regression accuracy of the prediction frame,the intersection to union ratio(IOU)algorithm in the traditional network is replaced by the EIOU algorithm,thereby optimizing the target detection evaluation indicators;Then,in order to reduce the missed detection caused by overlapping multiple droppers,the traditional NMS algorithm in the network is replaced with the Soft-NMS algorithm to avoid erroneous deletion of candidate boxes;Finally,a comparative experiment is conducted with traditional networks,SSD,and YOLO-v3 networks to verify the positioning effect of the improved network.The average accuracy of the improved Faster R-CNN network reached 98.29%,which is2.94% higher than that of the traditional network,and can accurately detect catenary droppers in the image.Aiming at the small difference between the characteristics of the slack state and the normal state of the dropper,which leads to the inaccurate fault detection of the traditional EfficientNet network,a dropper fault detection algorithm based on the improved EfficientNet network is proposed.Firstly,in order to improve the multi-scale feature extraction capability of the network,a parallel convolution layer of 5×5 is added to the first module,which is combined with the original convolution layer of 3×3 to generate a multi-scale feature fusion feature map,thereby further obtaining image feature information;Secondly,in order to improve the network’s ability to perceive subtle differences in suspenders in different states,the ECA channel attention mechanism is introduced,and two adaptive convolutions of 1×1 are used to replace the fully connected layer in the SE module to obtain the weight of each channel,improving the network classification accuracy;Finally,the effectiveness of the improved algorithm is verified through ablation experiments,and compared with ResNet-50 and MobileNet-V2 classification networks.The improved EfficientNet network has the highest accuracy of 95.19%,which is 2.81% higher than the traditional network.It can effectively detect loose and broken droppers.