Classification of Partially Labelled Partial Discharge Datasets Using Semi-Supervised Learning

Classification of Partially Labelled Partial Discharge Datasets Using Semi-Supervised Learning

연구 분야: Artificial Intelligence

학회: 2024 10th International Conference on Condition Monitoring and Diagnosis (CMD)

Condition monitoring (CM) of high-voltage assets is crucial in avoiding power failures and electrical outages of the power system. One of the most effective methods of monitoring such assets is detecting and analysing partial discharges (PD). Managing the increasing amount of PD data produced by condition monitoring systems (CMS) is only possible if computers help human operators. In this context, machine learning (ML) algorithms are key in exploiting these data to improve the electrical system performance. In particular, supervised learning (SL) algorithms are helpful in the classification and detection of PD. However, these algorithms require labelled data sets, which are scarce because labelling the data requires human experts. This paper compares different algorithms and their effectiveness in labelling data using semi-supervised learning (SSL). Corona, interfacial, and surface PD were utilised to evaluate the effectiveness of these approaches. The data analysis was separated into feature extraction and model training. A representation of the data suitable for SSL was achieved by combining the representation of time-frequency maps and temporal-spectral response with principal component analysis for dimensional reduction. Then, representative instances were obtained for each group using k-means and ‘manually’ labelled. Two different SSL algorithms were trained. With ~1 % of labelled data, both SSL models reached an F1 score of 0.91, close to the F1 score of 0.93 obtained for a support vector machine (SVM) trained on a fully labelled dataset.