Transformers are essential in modern power systems, and effective fault diagnosis is crucial for ensuring grid safety and stability. While Dissolved Gas Analysis (DGA) is a widely used technique for fault diagnosis, traditional methods, such as the IEC three-ratio approach, often fall short in handling complex fault scenarios. Recent advancements in Convolutional Neural Networks (CNNs) have shown promise in diagnosing transformer faults due to their powerful feature extraction capabilities. However, challenges remain in applying CNNs to DGA data, particularly regarding feature extraction limitations and data imbalance issues. This paper introduces a novel approach combining the Synthetic Minority Over-sampling Technique (SMOTE) with an enhanced one-dimensional CNN (1D CNN) for transformer fault diagnosis. The proposed method begins with using SMOTE to balance the dataset by generating additional minority class samples, addressing the issue of data imbalance. An improved 1D CNN architecture is then designed to enhance feature extraction capabilities by incorporating multiple convolutional layers and replacing traditional pooling layers. Specifically, the CNN structure includes four Conv1D modules with varying channel numbers, followed by a global average pooling layer to control the size of the output feature maps. Experimental results demonstrate that the proposed CNN model significantly outperforms traditional methods and existing deep learning models in terms of diagnostic accuracy, particularly after data augmentation. The combination of SMOTE and the improved 1D CNN effectively addresses the challenges of feature extraction and data imbalance, showcasing its efficacy in transformer fault diagnosis.