A study on spot welding quality judgment based on improved generative adversarial network and auto-encoder. (15th February 2022)

• Record Type: Journal Article
• Authors: Wang, Bing
• Is Part Of: Mechanical systems and signal processing. Volume 165(2022)
• Publication Date: 2022-02-15
• Subjects: Generative adversarial network -- Auto-encoder -- HMM -- Spot welding -- Quality judgment; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.108318
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A GAN was employed in this paper to generate unqualified spot welding joint samples and expand the sample dataset until the number of each kind of unqualified spot welding joint samples turned almost the same with the number of qualified ones. An AE was adopted to learn the features of both qualified and unqualified spot welding joint samples, so that with the assistance of personal experience, the feature vector set of each kind of spot welding joint samples could be obtained. A hidden Markov model (HMM) was utilized at the end to judge the spot welding quality, it is a classifier for AE. Abstract: To address the problems that only a few unqualified spot welding joint samples are obtained during spot welding, and traditional feature extraction methods are unable to obtain hidden deep-seated quality features of spot welding, an improved method was proposed in this paper to judge the spot welding quality. The proposed method integrated the generative adversarial network (GAN) and the auto-encoder (AE) to construct a network structure having the functions of sample data generation, feature extraction and pattern recognition. It also presented appropriate improvements to address the insufficient diversification of generated samples in the standard GAN. An improved generative adversarial network (IGAN) was first employed to expand the samples dataset of unqualified spot welding joints, which was followed by the selection of feature vector of the samples through AE combined with personal experience. A hidden Markov model (HMM) was utilized at the end to judge the spot welding quality. The spot welding joint samples data of stainless steel railway carriage roof were employed in this study to prove that the proposed improved method is feasible. It ensured that the generated data can fit the waveform curve of spot welding joint samples with unqualified quality as far as possible, at the same time, spent shorter time for training and testing data sets and provided a higher classification accuracy than other classification methods.