Uso de amostras sintéticas no treinamento de modelos para a identificação do estado de chaves seccionadoras em subestações de transmissão de energia elétrica

Abstract

Synthetic data generation is being increasingly explored in computer vision due to its capacity to produce large volumes of highly varied and automatically labeled data. This type of data minimizes costs and can fill gaps in real datasets. Synthetic data are especially relevant to enable applications that impose constraints for collecting real data in sufficient quantity and variety for machine learning. However, the data are generated from simplifications of reality, such as computer simulations and statistical models, creating a gap between the real and synthetic domains, consequently leading to difficulties in transferring the learned features. This work presents an experimental evaluation of synthetic data in a particular application: disconnect switch state recognition in transmission substations, a problem of great importance to the electric sector, but which is subject to the difficulties of a restricted environment full of equipment critical to power delivery. The data were generated by rendering virtual tridimensional models of substations and employing domain randomization, that is, the random manipulation of simulation parameters, including the act of opening and closing switches. We modeled the recognition task as a classification problem and evaluated six distinct convolutional network architectures, initially trained only on synthetic data and later with the addition of a small amount of real data. In tests with images captured by physical cameras installed in substations, most of the models trained solely on synthetic data achieved between 87% and 97% accuracy, demonstrating the existence of a synthetic-to-real domain gap that limits performance. Adding real data to the training set mitigated this effect, raising the accuracy to values between 95% and 99%. Furthermore, the proposed approach performed remarkably well in the recognition of rare events, achieving rates above 90%.