Research Of Appearance Inspection Of Resistance Spot Welding Based On Semantic Segmentation

As an efficient and concise welding technology,Resistance Spot Welding(RSW)has been widely used in automated production pipelines.Its welding quality largely affects the quality of the whole product.RSW is in a high temperature and impact process environment,which is prone to defects such as fake welding,desoldering,and spattering.Non-destructive testing is an important quality inspection method.Nondestructive testing research for resistance spot welding includes intrinsic signal testing,radiographic testing,ultrasonic testing,eddy current leakage testing and appearance inspection.However,there are still shortcomings in terms of testing cost and stability,which cannot fully adapt to the online testing of industrial pipelines.This paper studies the RSW inspection method based on machine vision,and focuses on the image segmentation in visual inspection.The proposed method achieves good results as a result of applying deep learning-based semantic segmentation to industrial image processing.The main research works include:1.An image acquisition system was built based on machine vision,including hardware and software parts,which enables image acquisition and processing automatically.An image pre-processing process was designed to obtain a stable region of interest(ROI).2.The image segmentation methods were compared and analyzed.The applicability of various methods for industrial online detection is analyzed according to the two indicators of segmentation accuracy and running time.Based on the classical methods,segmentation of RSW images achieved a MIOU(mean intersection over union)of 61.91%over the five foreground categories in the image.3.The semantic segmentation method based on deep learning is studied in detail.The effects of the encoder and decoder were analyzed.The hyperparameters were determined based on experiments.Aiming at the problem of class imbalance,data augmentation methods such as scaling,gamma adjustment,and hue adjustment were compared,and a loss function with balanced weights was proposed.Finally,the MIOU increased by 1.3%,and relieved the class imbalance problem.4.A classifier based on decision tree is studied,where the tree structure and the threshold of each node are designed based on expert knowledge.The classifier takes the features extracted from the segmentation results as input.Compared with the classical methods,the semantic segmentation method improves the classification accuracy by 5.85%on average.On the dataset composed of edge-level images,the semantic segmentation method improves the accuracy by 26.10%.Finally,the image acquisition system based on semantic segmentation achieved an average quality inspection accuracy of 99.46%.Among the randomly sampled images,the absolute error of the distance feature does not exceed 10 pixels.The average image segmentation running time is 234.54 milliseconds,and the average time of the classifier is 43.24 milliseconds,which satisfies the requirements of automatic RSW online detection.