Abstract: With the rapid progress in superheavy element research in China, the continuous-wave superconducting heavy-ion accelerator CAFE2 has become a critical platform for superheavy element synthesis. As China’s first continuous-wave superconducting heavy-ion accelerator, CAFE2 is notable for its high energy output and operational stability. However, one of the primary challenges during its operation is the failure of superconducting radio-frequency (SRF) cavities. Among these, quenching represents the most serious failure mode, as it often results in simultaneous failure of multiple cavities, significantly impacting experimental outcomes. The similarity between quenching RF signal waveforms and other fault patterns makes accurate fault identification essential for ensuring the accelerator’s reliable performance. During quench events, the loaded quality factor (Q_L) of SRF cavities exhibits significant changes, which provides a key metric for detecting quenching. This study aims to develop a dynamic Q_L model based on the cavity’s differential equations, enabling efficient identification and classification of quenching events. Extensive experiments were conducted on the CAFE2 system at the Institute of Modern Physics, Chinese Academy of Sciences, and the model was validated by analyzing the operational data of SRF cavities. Our research revealed three distinct quenching mechanisms through the analysis of fault signals collected from the CAFE2 system, using the dynamic Q_L model. To further enhance fault recognition accuracy, a set of expert-derived features and applied machine learning techniques were designed to automate the classification of these quenching mechanisms. By constructing a random forest classification model, high-precision classification of the various quenching types was achieved and the efficiency of fault diagnosis was greatly improved. Real-time RF signal collection from SRF cavities was involved in our methodology, focusing on key features such as cavity pressure, forward power, and reflected power. These signals were processed using the dynamic Q_L model to identify characteristic Q_L changes indicative of quenching. Additionally, large-scale fault data analysis enabled us to differentiate quenching from non-quenching events, reducing false positives and enhancing the accelerator’s operational safety. The results demonstrate that the dynamic Q_L model is effective in accurately detecting and classifying SRF cavity quenching events. This work provides both theoretical insights and practical guidelines for the operation and maintenance of superconducting accelerators. By improving fault detection accuracy and informing operational strategies, our findings contribute to the continued advancement of superheavy element research. Moving forward, we aim to further refine the model and explore additional machine learning-based methods to address emerging challenges in accelerator operations.