To address the issues in the mechanical condition monitoring of air circuit breakers using acoustic signals—specifically, the dependence on manual parameter setting and poor interpretability of modal decomposition methods, as well as the limited applicability of short-time analysis techniques—this paper proposes a sound event detection model combining improved symplectic geometric mode decomposition (ISGMD) and a time-frequency attention (TFA) mechanism. The method involves synchronously collecting acoustic signals, main shaft angular displacement, and contact voltage signals during circuit breaker operation to perform time-frequency correlation analysis on closing/opening events. ISGMD is utilized to adaptively decompose the acoustic signals, overcoming interference from invalid components and the limitation of unclear physical meaning. Subsequently, S-transform is applied to construct time-frequency spectrograms, highlighting the time-frequency distribution patterns of the signals and thereby building the dataset required for subsequent model training. Finally, a deep learning network is constructed by embedding the time-frequency attention mechanism into the feature extraction module. This enables the network to dynamically focus on frequency intervals associated with the closing/opening events. Combined with the bidirectional long short-term memory (Bi-LSTM) network to deeply explore long-term dependencies in the sequences before and after sound events, the model achieves accurate localization of the boundaries of closing/opening events, effectively reducing the probabilities of false alarms and missed detections. The results indicate that the proposed method achieves an accuracy, recall, and F1-score of approximately 93%. For data from different microphone positions and distances, the root mean square error (RMSE) is less than 0.44 ms; for different devices, the RMSE is below 0.57 ms, demonstrating good generalization capability and stability. ISGMD provides interpretable signal decomposition from the perspective of physical mechanisms, while deep learning drives the automatic learning of complex event features from the data level. The synergistic approach formed by these two approaches achieves millisecond-level localization of sound events, providing reliable support for the intelligent diagnosis of the mechanical condition of circuit breakers.