Scholarlink

<p>Resistance spot welding (RSW) is a critical process in thin-walled structure manufacturing, where online quality diagnosis based on process signal features is a promising way to unleash the power of artificial intelligence (AI) to ensure product reliability. However, assembly deviations in mass production introduce random fluctuations in fit-up conditions, compromising weld quality and challenging the generalization of AI diagnosis models from standard conditions to abnormal conditions, resulting in a significant decrease in model prediction accuracy. To address this challenge, we propose a novel feature extraction method combining autoencoder-based compression and domain adversarial training, thereby improving the generalization performance of the diagnosis model. The autoencoder compresses high-dimensional process signals into low-dimensional features, while a Transformer-based attention mechanism enhances feature expressive capability and model interpretability. Domain adversarial training further minimizes feature distribution discrepancies across varying fit-up conditions. A multi-sensor signal dataset under standard and abnormal fit-up conditions, consisting of 1652 welds with measured nugget diameters, was established. The method’s performance was validated using standard-condition training data and abnormal-condition test data, comparing it against handcrafted feature extraction and three established algorithms (PCA, Isomap, and LLE). Results demonstrate superior generalization: the proposed method exhibits only a 12.53% performance degradation on the test set, significantly lower than the 64.87% degradation observed with the handcrafted method. Additionally, the proposed method improves the weld quality diagnosis accuracy from 81.04% to 92.18% and reduces <i>RMSE</i> of nugget diameter prediction by nearly 34.4%. These findings confirm that the proposed method can effectively enhance the generalization ability of the quality diagnosis model for out-of-distribution data, providing support for the realization of an intelligent RSW process.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing the generalization performance of resistance spot weld quality diagnosis through autoencoder and domain adversarial training

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • Qiang Song,
  • YuJun Xia,
  • JianNan Zhang,
  • WenLong Xu,
  • YongQuan Jia,
  • MingQing Wan,
  • YuChen Yuan,
  • YongBing Li

摘要

Resistance spot welding (RSW) is a critical process in thin-walled structure manufacturing, where online quality diagnosis based on process signal features is a promising way to unleash the power of artificial intelligence (AI) to ensure product reliability. However, assembly deviations in mass production introduce random fluctuations in fit-up conditions, compromising weld quality and challenging the generalization of AI diagnosis models from standard conditions to abnormal conditions, resulting in a significant decrease in model prediction accuracy. To address this challenge, we propose a novel feature extraction method combining autoencoder-based compression and domain adversarial training, thereby improving the generalization performance of the diagnosis model. The autoencoder compresses high-dimensional process signals into low-dimensional features, while a Transformer-based attention mechanism enhances feature expressive capability and model interpretability. Domain adversarial training further minimizes feature distribution discrepancies across varying fit-up conditions. A multi-sensor signal dataset under standard and abnormal fit-up conditions, consisting of 1652 welds with measured nugget diameters, was established. The method’s performance was validated using standard-condition training data and abnormal-condition test data, comparing it against handcrafted feature extraction and three established algorithms (PCA, Isomap, and LLE). Results demonstrate superior generalization: the proposed method exhibits only a 12.53% performance degradation on the test set, significantly lower than the 64.87% degradation observed with the handcrafted method. Additionally, the proposed method improves the weld quality diagnosis accuracy from 81.04% to 92.18% and reduces RMSE of nugget diameter prediction by nearly 34.4%. These findings confirm that the proposed method can effectively enhance the generalization ability of the quality diagnosis model for out-of-distribution data, providing support for the realization of an intelligent RSW process.