An Evaluation Method of Acceptable and Failed Spot Welding Products Based on Image Classification with Transfer Learning Technique

The1 spot welding technique is widely used in the industrial production line, but it suffers inconsistent quality. Therefore, the evaluation of the spot-welding product is of great importance for the industrial production. Many destructive and nondestructive methods have been used in the product evaluation, but they are inefficient and hard to be applied in the mass production. In recent year, machine vision method has been used to differentiate the acceptable and failed spot welding products according to their solder joint images. This opened new opportunities for the spot welding product quality evaluation using digital image technique. However, this method cannot achieve general performance on different spot-welding products as well as ideal classification accuracy. In this work, a novel method which based on the transfer learning technique was proposed to classify the spot-welding products according to their solder joint images. The GoogLeNet was used to extract the features of the solder joint image, which is pretrained on the ImageNet. Then a multilayer perceptron (MLP) was used to classify these images. Our method achieved a final classification accuracy of 96.99% on a testing set included 334 images.