Abnormal Current Detection Method in Electrical Lines Based on AE-2LSTM-OCSVM Model

Electrical fires represent a significant threat to both lives and property in everyday work and life. Fires resulting from electrical faults exhibit characteristics such as high concealment, prolonged development cycles, and substantial harm. If abnormal data are not identified promptly, considerable losses may ensue. However, current electrical fire warning methods primarily focus on detecting flames and smoke, rendering them ineffective in small spaces. Furthermore, the deep learning models employed necessitate a balanced dataset of normal and abnormal samples for effective training, which presents certain limitations. To address these challenges, this paper combines the autoencoder (AE) module, the long short-term memory (LSTM) module, and the one-class support vector machine (OCSVM) module to propose an abnormal current detection method utilizing the AE-2LSTM-OCSVM model. First, the current data are segmented and denoised. Subsequently, the AE-2LSTM model serves as the reconstruction framework to derive the reconstruction error for both normal and abnormal current data. Ultimately, this reconstruction error is utilized as input for the OCSVM, allowing for the construction of a decision boundary that delineates normal current data, which in turn facilitates the detection of abnormal current data. A Simulink simulation was conducted utilizing three current waveforms that are likely to induce electrical fires, resulting in the acquisition of abnormal current data to validate the method proposed in this article. The experimental results demonstrate that the method presented herein can effectively identify abnormal currents that pose electrical fire hazards.