［Objective］The turnout switch machine fault diagnosis techniques based on deep learning is summarized, the application status of various deep learning methods in turnout switch machine fault diagnosis is analyzed, their advantages and limitations are explored, and future research directions are proposed. ［Method］The importance and challenges of turnout switch machine fault diagnosis are first introduced, followed by a comparative analysis of the characteristics of model-driven and data-driven diagnostic approaches. Then, fault diagnosis methods based on deep neural networks, autoencoders, convolutional neural networks, recurrent neural networks, and hybrid multi-deep models are elaborated in-depth, with comparisons of their respective performance characteristics. The limitations of current research, including the need for large amounts of labeled data, model complexity, and interpretability are discussed. Several future research directions are proposed. ［Result & Conclusion］Turnout switch machine fault diagnosis techniques based on deep learning demonstrate strong capabilities in feature extraction and data processing, significantly improving diagnostic accuracy and efficiency. However, current deep learning methods face challenges such as the requirement for large datasets, high model complexity, and limited interpretability. Future research should focus on data preprocessing techniques, multi-source information fusion, diagnosis methods for imbalanced and small sample scenarios, transfer fault diagnosis, and interpretable deep diagnostic models to enhance the wide application and intelligence level of deep learning in turnout switch machine fault diagnosis.