Deep learning–based welding defect classification often faces challenges such as limited training data, class imbalance, and high model complexity, which restrict real-time industrial applications. To address these issues, this paper proposes a hybrid CNN–Vision Transformer framework with GAN-based data augmentation for welding defect classification. First, welding images acquired using a wide dynamic range visual sensor are preprocessed through binarization, median filtering, morphological dilation, and cropping to enhance defect features. A Generative Adversarial Network (GAN) is then employed to generate synthetic samples and alleviate dataset imbalance. A lightweight CNN extracts local spatial features and reduces feature dimensionality, after which the resulting feature maps are converted into tokens and processed by a Vision Transformer encoder to capture global contextual relationships via self-attention. The proposed model classifies welding images into four categories: normal, burn-through, undercut, and welding collapse. Experimental results demonstrate that the hybrid architecture achieves improved classification accuracy and computational efficiency compared with conventional lightweight CNN models. In addition, the model attains 98.25% accuracy on the MNIST dataset, validating the effectiveness of the proposed framework.

Industry 4.0 has increased the demand for intelligent predictive maintenance systems capable of supporting real-time monitoring, early fault detection, and efficient decision-making in industrial environments. In this context, accurate prediction of machine failures has become essential for minimizing downtime, reducing maintenance costs, and improving operational reliability. This study employs the Predictive Maintenance Dataset from the UCI repository to develop and evaluate data-driven models for machine failure prediction and classification. The research pursues two primary objectives: first, to compare the performance of several machine learning algorithms in classifying machine failures, and second, to assess the effectiveness of deep learning approaches in achieving higher predictive accuracy. Among the machine learning models examined, the XGBoost classifier demonstrates the strongest performance. To further enhance prediction capability, this study adopts a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model, which integrates CNN’s strength in automatic feature extraction with LSTM’s ability to learn temporal dependencies from sequential data. Experimental results show that the proposed CNN–LSTM model outperforms traditional machine learning models as well as Artificial Neural Networks (ANN) in predicting machine failures. The main contribution of this study lies in the comparative evaluation of machine learning and hybrid deep learning techniques on an imbalanced predictive maintenance dataset. The findings confirm the potential of hybrid deep learning models for predictive maintenance applications and highlight their practical value in enabling proactive maintenance strategies, optimizing resource allocation, and enhancing the reliability of smart industrial systems.