Fault diagnosis technology plays a pivotal role in ensuring production safety and enhancing operational efficiency of industrial equipment. However, traditional fault diagnosis methods often suffer from insufficient feature extraction capabilities and poor model generalization when handling real-world scenarios with limited data and complex, variable working conditions, resulting in unsatisfactory diagnostic accuracy and reliability. To address these challenges, this study proposes a deep transfer learning-based fault diagnosis framework integrating multi-scale feature extraction and multi-layer domain adaptation (MS-ML). Firstly, vibration signals are processed through a multi-scale convolutional neural network (MSCNN) to extract hierarchical features, capturing both localized details and global characteristics across different scales. This architecture significantly enhances the model's capability to discern intricate fault patterns. Subsequently, a multi-layer domain adaptation strategy is employed to minimize distribution discrepancies between source and target domain data, thereby improving the model's generalization performance under diverse operating conditions. The proposed method is rigorously validated on two benchmark datasets: the Jiangnan University dataset and the Paderborn University dataset. Comparative experiments with existing methods demonstrate that our approach not only achieves superior fault diagnosis accuracy but also exhibits robust cross-domain adaptability, maintaining stable performance in transfer scenarios. These findings provide both theoretical foundations and technical advancements for optimizing intelligent diagnostic systems in industrial applications.