A semi-supervised autoencoder with an auxiliary task (saat) for power transformer fault diagnosis using dissolved gas analysis

This paper proposes a semi-supervised autoencoder with an auxiliary task (SAAT) to extract a health feature space for power transformer fault diagnosis using dissolved gas analysis (DGA). The health feature space generated by a semi-supervised autoencoder (SSAE) not only identifies normal and thermal/electrical fault types, but also presents the underlying characteristics of DGA. In the proposed approach, by adding an auxiliary task that detects normal and fault states in the loss function of SSAE, the health feature space additionally enables visualization of health degradation properties. The overall procedure of the new approach includes three key steps: 1) preprocessing DGA data, 2) extracting two health features via SAAT, and 3) visualizing the two health features in two-dimensional space. In this paper, we test the proposed approach using massive unlabeled/labeled Korea Electric Power Corporation (KEPCO) databases and IEC TC 10 databases. To demonstrate the effectiveness of the proposed approach, four comparative studies are conducted with these datasets; the studies examined: 1) the effectiveness of an auxiliary detection task, 2) the effectiveness of the visualization method, 3) conventional fault diagnosis methods, and 4) the state-of-the-art, semi-supervised deep learning algorithms. By examining several evaluation metrics, these comparative studies confirm that the proposed approach outperforms SSAE without the auxiliary task, existing methods, and state-of-the-art deep learning algorithms, in terms of defining health degradation performance. We expect that the proposed SAAT-based health feature space approach will be widely applicable to intuitively monitor the health state of power transformers in the real world.