Fault detection in insulators based on ultrasonic signal processing using a hybrid deep learning technique

Abstract

Identifying problems in insulators is a task that requires the experience of the operator. Contaminated insulators generally do not represent a system failure, however, due to higher surface conductivity, they may suffer from electrical discharges and may result in irreversible failures. The identification of possible failures in inspections can help to forecast faults to improve reliability in the power grid. Based on this need, this article presents a study on fault prediction in distribution insulators, through a laboratory evaluation in a contaminated insulator, where 13.8 kV (root mean square) was applied considering an ultrasound detector connected to a computer for data set acquisition. In the sequence, a time series prediction, using a hybrid deep learning technique defined as wavelet long short-term memory (LSTM), was performed. The hybrid LSTM was proposed considering feature extraction through the wavelet energy coefficient. Finally, for a complete evaluation, deeper LSTM layers were included, and both the training method and the hardware configuration were evaluated. The wavelet LSTM algorithm showed interesting accuracy results when compared to classic prediction algorithms like the non-linear autoregressive exogenous model.