Non-intrusive load monitoring (NILM) is of great importance for enhancing demand-side management of electricity and establishing efficient and reliable power systems, with event detection being one of the most critical components. However, existing event detection methods suffer from low detection accuracy in complex events such as high volatility, long transitions over time, and nearly simultaneous occurrences, as well as issues with generalization across different usage scenarios. To address these issues in event detection, this paper proposes an event detection method based on the Isolation Forest algorithm and a dual-layer filtering mechanism. First, normalization is applied to preprocess the data, obtaining normalized current root mean square (RMS) values to reduce unexpected external interference. Then, the Isolation Forest algorithm is used for preliminary anomaly detection, identifying suspicious events with high anomaly scores. Finally, a dual-layer filtering mechanism is employed: an adaptive time-duration filtering window and an adaptive current difference threshold window, which filter out false events, significantly improving the detection rate. The proposed method does not require pre-training of the model. Testing experiments on the Private, BLUDE, and PLAID datasets show that the proposed method can accurately detect various complex events, outperforming other advanced algorithms in terms of detection performance. Additionally, the method exhibits strong generalization capabilities, making it suitable for a wide range of application scenarios.