Airbag Leak Point Detection Based On Deep Learning

Airbags are an important component of many devices,and its air tightness is a key index of the equipment.The method of smearing can be used to detect the leak point of the airbag,and the location of the leak point can also be seen intuitively,but the time and labor required is relatively high,also there may be some missed leak points due to long-term manual observation.In recent years,deep learning has developed by leaps and bounds thanks to the advancement of convolutional neural networks and the substantial improvement of computing capacity.Also it has gradually highlighted its dominant position in the field of object detection.This paper is based on object detection method using deep learning.It is applied to the analysis of the leak point detection in an airbag.The network design theory is discussed from multiple perspectives,and the network can realize rapid and precise leak point detection.First of all,this paper starts with data.Data collection,data filtering and data labeling has been made to build an airbag leak point detection dataset,which is divided into training set and validation set.In order to meet the three requirements of m AP,recall and detection speed at the same time,this paper analyzes the basic modules,concatenation methods,and overall architecture of the backbone network used for object detection.At the same time,a same object detection algorithm but with different backbone networks is used for training and validation.Also the characteristics of each backbone network were analyzed in order to design a backbone network suitable for the detection of leak points of an airbag.In addition,this paper analyzes object detection algorithms based on different design concepts.It mainly selects one one-stage detection algorithm,two two-stage detection algorithms,and one keypoints-based detection algorithm for experiment.After the experiment,It is recognized that the structure of a two-stage detection network is more suitable in the leak point detection scenario.At the same time,this topic also considers the limitation of the object detection network due to the small size of the dataset made,so a variety of data augmentation methods is exploited to expand the training set to improve the detection m AP,recall and generalization ability.Finally,based on the previous experimental results and characteristics of the leak point detection scenario an object detection network is designed with rapid detection capability,which also meets the requirements of m AP,recall and speed.