With the rapid growth of railway passenger volume and heightened safety concerns from both operators and passengers, stricter demands have been placed on railway track maintenance and repair work. Current techniques for detecting track geometry faults largely rely on manual inspections or track inspection cars, which are often time-consuming and encounter limited occurrences. To address these shortcomings, this study integrates car-body vibration data with deep learning technology, aiming at enhancing the efficiency and accuracy of railway track maintenance work. By incorporating techniques such as sliding window sampling, data augmentation and attention mechanism, this method effectively resolves issues with imbalanced sample distribution and further enhances the model's classification performance. Experimental results demonstrate that this approach significantly improves the ability of deep learning models to detect and classify track geometry faults using car-body vibration data. The comprehensive application of car-body vibration data and deep learning technology provides an efficient, automated detection method for railway track maintenance, laying a data foundation for implementing preventive maintenance strategies.