Representação e classificação de perfis térmicos de isoladores poliméricos para diagnóstico de poluição superficial

Abstract

Monitoring insulators in electric power systems is essential for reducing the risk of failures associated with surface pollution. In coastal environments, salt deposition and the presence of moisture can intensify surface leakage current and cause localized heating capable of leading the equipment to failure and causing interruptions in the power supply. In this context, infrared thermography is a promising non-invasive technique because it enables the thermal condition of the equipment to be assessed without direct contact or interruption of operation. This dissertation proposes a methodology for classifying the thermal profiles of polymeric insulators, aiming to distinguish between clean and polluted conditions. To this end, one-dimensional thermal profiles are extracted from thermograms, corrected relative to the ambient temperature, segmented into successive windows, and described using features related to thermal level, temperature dispersion, including variance and standard deviation, contrast, distribution shape, including skewness and kurtosis, spatial variation, and spatial dependence. Three forms of representing the information along the profile are evaluated: aggregated, sequential, and symbolic. The aggregated representation summarizes the features obtained from the windows, whereas the sequential and symbolic representations explicitly preserve the spatial order of these windows. For classification, a support vector machine, a random forest, a long short-term memory recurrent neural network, a hidden Markov model with Gaussian mixtures, and a discrete hidden Markov model are evaluated. The dataset comprises 344 thermal profiles distributed across 69 acquisition groups and includes clean insulators as well as insulators subjected to artificial and natural pollution. Feature selection, model tuning, and the initial performance estimation are conducted through nested validation with separation by acquisition groups, thereby preventing samples originating from the same acquisition from being distributed between the training and test sets. The results indicate that all evaluated models are capable of distinguishing between clean and polluted profiles. Subsequently, a robustness analysis is conducted in which the best models for each representation are evaluated over different repetitions of the training–test split, while keeping fixed the selected features, hyperparameters, and structures defined during nested validation. In this analysis, the support vector machine applied to the aggregated representation achieves the best performance, with a balanced x accuracy of 0.923 ± 0.058 and a multicriteria metric of 0.897 ± 0.067. It is concluded that the spatial segmentation of the thermal profile provides relevant information for identifying surface pollution and constitutes a promising approach for the diagnosis of polymeric insulators.