Improving Machine Learning Diagnostic Systems with Model-Based Data Augmentation - Part B: Application

Data augmentation can be used to train more robust machine-learning classifiers. Classically, synthetic data from a data augmentation are used to augment measurement datasets and use them for training of machine learning (ML) algorithms. However, the synthetic data often do not represent the measurements perfectly. This leads to insufficiently trained ML-models for real world application. In this paper, ML models are trained using only synthetic data. These ML models are then transferred to the available measurements utilizing transfer learning. This approach is showcased for the detection of a power and distribution transformer fault and benchmarked with state-of-the-art diagnostic systems. The performance of all diagnostic systems is analyzed by limiting the amount of fault-condition measurements available for the training process and by comparison of learning curves. It is shown that the model-based data augmentation combined with fine tuning is capable of improving the accuracy for the analyzed diagnostic task.