Currently, the early warning of electrical fires primarily relies on evident fault characteristics, making it challenging to promptly respond during the initial stages of risk development. The pre-overheating symptom phase in electrical circuits was focused on, and an enhanced AE (autoencoder) framework that integrates a two-layer LSTM (long short-term memory network) with OCSVM (one-class support vector machine) was proposed for detecting early signs of circuit overheating.By analyzing the underlying disaster mechanisms and current waveforms associated with three types of circuit conditions—overload, harmonic distortion, and aperiodic currents—a comprehensive dataset was constructed. The AE-2LSTM architecture was employed to effectively capture deep temporal features in current time series data, enabling high-precision signal reconstruction. Leveraging the significant disparity in reconstruction errors between normal and anomalous samples, the OCSVM module was introduced to perform single-class anomaly detection using only normal training samples. Experimental results demonstrate that the proposed method achieves an anomaly recognition accuracy of 98.16%, effectively addressing practical challenges in electrical fire monitoring, such as the scarcity of abnormal data and severe class imbalance. This approach provides a robust scientific foundation and technical support for the early warning, proactive prevention, and control of electrical fires.