Simultaneous Detection of Defects in Electrical Connectors Based on Improved Convolutional Neural Network

Abstract

Electrical connector is a main channel for the transmission of signals and energy signals within electrical device. Even a tiny skew and a missing pin of the electrical connector have a vital impact on the reliability and stability of electrical devices. However, due to the high requirements for positioning, it is difficult for traditional machine vision and template matching technology to detect these tiny defects in multiple electrical connectors simultaneously in an accurate and high-throughput manner. To address this problem, we propose an electrical connector defect detection model based on an improved convolutional neural network (CNN). The model consists of three stages. First, image processing technology is used to quickly obtain a dataset of high-pixel images that highlight qualified pins, skew pins, and missing pins in a single electrical connector. Then, based on ResNet-152, three 3 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times3$ </tex-math></inline-formula> convolution kernels are used to replace the 7 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times7$ </tex-math></inline-formula> convolution kernels in the first layer of the network structure to expand the receptive field. A new smooth maximum unit (SMU) is used to replace Rectified Linear Unit (ReLU) to improve the network’s detection performance. Finally, a deformable convolutional network (DCN) is adopted to replace a part of the traditional convolution kernel, which improves the network’s adaptability to geometric changes and allows efficient extraction of the geometric and edge feature information of tiny pins. The experimental results show that the proposed model performs detection with an accuracy of 97.42% and 2.09% higher than ResNet-152 (95.33%), and with a duration of 0.988 s, 12.02% faster than the latter. This model can easily be applied in the industrial production inspection field to perform simultaneous detection of defects in multiple dense and tiny targets like pins.