Distribution Transformer Parameters Detection Based on Low-Frequency Noise, Machine Learning Methods, and Evolutionary Algorithm

• Published in: Sensors (Basel, Switzerland) Vol. 20; no. 15; p. 4332
• Main Authors: Jancarczyk, Daniel, Bernaś, Marcin, Boczar, Tomasz
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 04.08.2020; MDPI
• Subjects: Acoustics; Algorithms; Classification; Cooling; Evolutionary algorithms; Fault diagnosis; Frequency spectrum; genetic algorithm; Genetic algorithms; LF noise; low-frequency noise; low-frequency sensor; Machine learning; Mathematical models; Methods; Neural networks; Noise; Noise spectra; Parameters; power transformer; Principal components analysis; Sound pressure; Vibration
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s20154332

The paper proposes a method of automatic detection of parameters of a distribution transformer (model, type, and power) from a distance, based on its low-frequency noise spectra. The spectra are registered by sensors and processed by a method based on evolutionary algorithms and machine learning. The method, as input data, uses the frequency spectra of sound pressure levels generated during operation by transformers in the real environment. The model also uses the background characteristic to take under consideration the changing working conditions of the transformers. The method searches for frequency intervals and its resolution using both a classic genetic algorithm and particle swarm optimization. The interval selection was verified using five state-of-the-art machine learning algorithms. The research was conducted on 16 different distribution transformers. As a result, a method was proposed that allows the detection of a specific transformer model, its type, and its power with an accuracy greater than 84%, 99%, and 87%, respectively. The proposed optimization process using the genetic algorithm increased the accuracy by up to 5%, at the same time reducing the input data set significantly (from 80% up to 98%). The machine learning algorithms were selected, which were proven efficient for this task.