焊接作为现代工业制造的核心工艺，其质量直接决定了焊接结构的力学性能与服役安全性。然而，由于材料质量、工艺参数和操作技术等因素的影响，可能导致气孔、裂纹、未焊透和夹渣等缺陷的产生。超声检测技术凭借其非破坏性、高灵敏度、强穿透力和实时动态检测等优势，已成为焊接质量评价的主要检测方法之一。但现有超声检测缺陷识别方法存在诸如信号特征提取可靠性不高、缺陷样本数量少且不平衡以及传统机器学习模型泛化能力差等不足。因此，本文将深度学习方法引入焊接缺陷智能检测领域，重点开展以下几个方面的研究：

（1）设计并制备含有夹渣、裂纹、气孔和未焊透四种典型缺陷的焊接试样5件，利用数字式超声探伤仪CTS-4020采集缺陷回波信号。基于连续小波变换、马尔科夫变迁场和格拉姆角场等方法，将一维超声回波信号转换为二维图像，制备了焊接缺陷图像数据集。采用AlexNet、VGGNet和GoogleNet网络对数据集进行评价，并通过加噪方式对数据集进行扩容。

（2）针对焊接缺陷样本数量少且数据不平衡以及识别准确率低等问题，提出一种改进DCGAN-ResNet的焊接缺陷识别方法。首先，基于Wasserstein距离与谱归一化改进DCGAN，并利用增强后的数据集验证其有效性。然后，在ResNet的基础上，利用组卷积和通道注意力机制改进残差块，从而提升重要特征的关注度以及网络的非线性表达能力。最后，通过超参数优化和交叉验证优化超参数以减少过拟合风险。实验结果表明，改进DCGAN-ResNet方法在焊接缺陷图像特征提取、识别准确率和效率等方面均具有更好的效果。

（3）针对焊接缺陷超声检测中多类不平衡数据分布和模型轻量化需求，提出一种改进ACGAN和WOA-ShuffleNetV1方法。首先，结合频率感知模块、潜在空间优化和优化损失函数改进ACGAN，以提升生成样本的质量。然后，通过对比分析扩容比例、生成模型对分类结果的影响，验证改进ACGAN对多类不平衡小样本数据扩容的有效性。最后，采用鲸鱼优化算法优化ShuffleNetV1的超参数。实验结果表明，该方法在降低FLOPs、模型大小和参数量的同时，保持了较高的识别准确率，从而实现了轻量化与分类性能的平衡。

（4）基于PyOt5开发焊接缺陷智能检测系统，实现焊接质量的智能化评估。根据系统需求确定系统功能，集成了焊接缺陷检测智能助手、图像转换与数据增强、焊接缺陷检测以及结果展示等功能模块。

As a core process of modern industrial manufacturing, welding directly influences the mechanical properties and service safety of welded structures. However, factors such as material quality, process parameters and operational technology can cause defects such as porosity, crack, lack of penetration and slag inclusion. Ultrasonic testing has emerged as the primary method of evaluating welding quality thanks to non-destructive nature, highly sensitivity, strong penetration and real-time dynamic detection capabilities. Nevertheless, existing ultrasonic testing methods for recognizing defects have shortcomings such as low reliability of signal feature extraction, the small and imbalanced number of defect samples, and the poor generalization ability of traditional machine learning models. Thus, this paper introduces deep learning methods to the field of intelligent detection of welding defects, with a focus on the following aeras of research.

(1) Five welding specimens containing four typical defects—namely, slag inclusion, crack, porosity and lack of penetration—are designed and prepared. The defect echo signals are collected using the digital ultrasonic flaw detector CTS-4020. Based on the continuous wavelet transform, the Markov transition field and the Gram angle field, the one-dimensional ultrasonic echo signal is transformed into a two-dimensional image, thereby preparing a dataset of welding defect images. The dataset is evaluated using AlexNet, VGGNet and GoogleNet networks. And the dataset is also experimentally expanded by adding noise.

(2) To address the problems of a small number of welding defect samples, data imbalance and low recognition accuracy, an improved DCGAN-ResNet method for welding defect recognition is proposed. Firstly, DCGAN is improved based on Wasserstein distance with spectral normalization, and its effectiveness is verified using the enhanced dataset. Then, based on ResNet, the residual block is improved using group convolution and a channel attention mechanism to enhance the network's ability to recognize important features and perform nonlinear representations. Finally, hyperparameter optimization and cross-validation are used to reduce the risk of overfitting. Experimental results demonstrate that the improved DCGAN-ResNet method achieves superior performance in terms of feature extraction, recognition accuracy and efficiency for weld defect images.

(3) To address the needs of multi-class imbalanced data distribution and model lightweighting in ultrasonic testing of welding defects, an improved ACGAN and WOA-ShuffleNetV1 method is proposed. Firstly, ACGAN is improved by combining the frequency-aware module, latent space optimization and optimized loss function to enhance the quality of generated samples. Then, the effectiveness of the improved ACGAN for multi-class imbalanced small-sample data expansion is verified by comparing and analyzing the effects of the expansion ratio and the generation model on the classification results. Finally, the whale optimization algorithm is used to optimize the hyperparameters of ShuffleNetV1. Experimental results demonstrate that this method maintains a high recognition accuracy while reducing FLOPs, model size and number of parameters, thus achieving a balance between lightweight and classification performance.

(4) An intelligent welding defect detection system is developed based on PyOt5 to enable the intelligent assessment of welding quality. According to the system requirements, its functions include an integrated welding defect detection intelligent assistant, image conversion and data enhancement, welding defect detection and results display, etc.