Abstract: To ensure the long-term stable operation of vacuum medical systems, regular leak detection is required. To address the limitations of traditional leak detection methods that require shutdown and exhibit low efficiency, this study proposes an integrated ultrasonic leak detection and vacuum-driven self-impregnation in-situ repairing technology based on deep learning, which can quickly identify micro-leaks and seal them in place, thus ensuring the continuous operation of the systems. The method employs high-sensitivity piezoelectric sensors to capture structural acoustic signals in a negative pressure environment. A CNN-LSTM deep learning model is used to automatically extract spatiotemporal features, enabling high-precision identification and localization of micron-scale leak points. Subsequently, the vacuum pressure difference drives a low-viscosity sealant to form a forced viscous flow, achieving deep impregnation and solidified sealing of the leakage channels. Experimental results show that the as-proposed system achieves an average diagnostic accuracy of 97.1% under complex noise conditions. After repairing, the system vacuum is stably restored to 86.7 kPa, and the leakage rate decreases by more than 90%, significantly reducing the pump cycling frequency and energy consumption. This research realizes efficient online diagnosis and repair of medical vacuum systems without shutdown, providing a practical and extensible technical approach for intelligent maintenance of vacuum engineering equipment.

Key words: vacuum leak detection, ultrasonic inspection, CNN-LSTM, vacuum sealing, medical vacuum system