Intelligent monitoring of transformer insulation using convolutional neural networks

The ability to monitor and detect potential faults in smart grid system components is extremely valuable. In this paper, we demonstrate the use of machine learning techniques for condition monitoring in power transformers. Our objective is to classify the three different types of Partial Discharge (PD), the identify of which is highly correlated with insulation failure. Measurements from Acoustic Emission (AE) sensors are used as input data. Two broad machine learning based approaches are considered - the conventional method which uses a predefined feature set (Fourier based), and deep learning where features are learned automatically from the data. The performance of deep learning compares very favorably to the traditional approach, which includes ensemble learning and support vector machines, while eliminating the need for explicit feature extraction from the input AE signals. The results are particularly encouraging as manual feature extraction is a subjective process that may require significant redesign when confronted with new operating conditions and data types. In contrast, the ability to automatically learn feature sets from the raw input data (AE signals) promises better generalization with minimal human intervention.