Attention Driven Temporal Convolutional Network for Multi-Class Electromagnetic Interference Signal Classification

Abstract

Electromagnetic interference (EMI) produced by electrical discharges, switching operations and external disturbances is a problem of great importance and complexity for the reliable operation and condition monitoring of modern power systems. Correct signal classification is critical to the timely detection of insulation defect, partial discharge activity, arcing faults and other abnormal operating conditions of EMI signals. The temporal complexity of EMI signals, non-stationary behaviour and noise in EMI signals require alternative methods of analysis to that employed by traditional machine learning and feature-based methods, which are handcrafted. To solve these problems, the authors propose a deep learning framework in this paper, which combines Temporal Convolutional Networks (TCN) with an attention mechanism for multi-class EMI signal classification. The EMI dataset is first preprocessed, such as normalization and label encoding, to enhance data uniformity and model training. The preprocessed temporal sequences are then passed onto stacked dilated causal convolution layers in the TCN to learn both short-term and long-range temporal dependencies. An attention layer is then used to focus on the most informative signal segments to reduce irrelevancy in the temporal features. The transformed representations are then passed through fully connected layers and a Softmax classifier for final, categorical prediction. An EMI dataset of nine signal types, including Arcing, Data Modulation, Corona, Micro Sparking, Normal, Partial Discharge, Exciter Pulses, Process Noise and Random Noise, is used to evaluate the proposed framework. By integrating temporal feature extraction with an attention-based refinement mechanism, the proposed model achieved reliable EMI signal classification across diverse operating environments. The model attained an overall accuracy of 95.08%, with precision, recall, and F1-score values close to 95%. Additionally, AUC scores exceeding 0.94 for all classes highlight its strong classification and separation capability.