Recently, machine learning techniques, particularly deep learning models, have been increasingly applied to the analysis, optimization, and prediction of system performance in industrial air compressor leak detection. In this study, we integrated such models for effective gas leak detection based on sensor-derived data, thereby supporting the development of more accurate and efficient factory leak sensing systems to reduce manufacturing costs. We investigated the use of Vision Transformer (ViT) models with transfer learning for gas leak detection in industrial air compressors, based on both audio and visual data. A labeled dataset was created using recordings of compressor sounds under leak and nonleak conditions. Using the intensity stereo localization method with microphone arrays, we estimated the sound intensity and source location. A multistage transfer learning strategy was adopted: ViT models pretrained on ImageNet were adapted with the Environmental Sound Classification (ESC-50) and fine-tuned on real leak data. Among four audio representations, Mel spectrograms achieved the highest accuracy (80%), making them most effective for ViT-based leak detection.