Aprendizado profundo para estimação de instantes de chegada com aplicação à detecção e localização de descargas parciais em cabos isolados de média tensão

Abstract

Partial discharge is one of the main precursors of failure in insulated medium-voltage cables. Diagnosis based on its measurement can answer two questions: whether there is partial discharge activity in the cable and where the source is located. This thesis proposes treating detection and localization as parts of a single problem: estimating, in each signal, the set of arrival times of the partial discharge pulses, whose number is not known in advance and may be zero. To this end, the proposed deep learning model produces, for each sample of the signal, a value indicating the confidence that the corresponding instant is the arrival of a pulse. Under this formulation, detection occurs when at least one arrival time is found, and localization follows from single-ended time-domain reflectometry, from the difference between the arrival times of the direct pulse and its reflection. Training and internal evaluation use the author’s own datasets, published under an open license, comprising laboratory and field measurements. Robustness is further assessed on a dataset acquired and published by an independent institution, to which the formulation is applied without any readjustment. The estimation error is assessed under two aspects: identification, which measures whether the correct pulses were found, and timing accuracy, which measures the temporal error of each pulse found. The comparison is made against classical arrival time estimation procedures, calibrated on the same data and under the same protocol. In identification, the formulation achieves an F1 of 0.987, against 0.699 to 0.799 for those procedures, and keeps this advantage throughout the operating range of signal-to-noise ratio evaluated. In timing accuracy, the bias is sub-sample (0.16 sample) and the dispersion is 1.4 samples (14 ns), 32% lower than that of the best procedure compared. Under a change of measurement scenario, with different instrumentation and discharge sources, the formulation retains its performance, with an F1 of 0.962. Two case studies demonstrate the diagnostic outputs. In a field campaign of 8135 records, analyzed with no prior screening, all 1251 records containing partial discharge activity are detected, with false positives in 0.68% of the remaining records. In a laboratory test with a documented defect position, the source is located 1.70 m from the actual position, in a 144 m cable.