Due to the hidden nature and complexity of resistance spot welding weld nugget formation, how to avoid the time-consuming and money-consuming problem of traditional defect diagnosis methods and accurately grasp the weld nugget status is still an urgent problem. In this paper, an improved GAN model is proposed to solve the corresponding problem by combining the weld nugget defects with the dynamic resistance curve. Aiming at the problem that traditional GAN algorithms are prone to pattern collapse, this paper utilizes a variational autoencoder integrated with a channel attention mechanism as the generator part of the generative adversarial network, which helps the model pay better attention to the high-weight part of the defective sample data and combines the encoding and decoding processes to highlight defective features, thus reconstructing the defective samples with higher quality. Convolutional neural networks are then utilized to identify the features of the generated samples and diagnose the type of weldment defects. The test results show that the proposed scheme is highly reliable and the model outperforms other schemes in diagnosing welded nugget defects under the same conditions, avoiding undesirable effects such as underfitting. The validation of the actual dataset shows that, compared with other diagnostic methods that generally have an accuracy rate of less than 75%, the accuracy of the weld nugget defects diagnosis of this paper's method reaches more than 94%, which is a positive impetus to the development of auto body welding diagnosis. •Provides guidance for diagnosis in the presence of small samples of weld nugget defects•Proposes a method for diagnosing weld defect types that incorporates the characteristics of the spot-welding process.•An improved generative adversarial network is used to reconstruct weld nugget samples with defect characteristics.•Performance comparison demonstrates the effectiveness of the proposed method in diagnosing weld nugget defect types.