Pattern Recognition Method for Detecting Partial Discharge in Oil-Paper Insulation Equipment Using Optical F-P Sensor Array Based on KAN-CNN Algorithm

• Published in: Journal of lightwave technology Vol. 43; no. 12; pp. 6004 - 6012
• Main Authors: Liu, Hong, Zhang, Zhixian, Song, Ruimin, Shu, Zhiqing, Wang, Jianxin, Tian, Haoyuan, Song, Yuxuan, Chen, Weigen
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.06.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Acoustics; Acousto-optics; Algorithms; Artificial neural networks; Convolution; direct-coupled F-P; Discharge; Discharges (electric); Fault detection; Fiber optics; KAN+CNN; Kernel; partial discharge; Partial discharges; Pattern recognition; Refractive index; Resonant frequency; Sensitivity; Sensor arrays; Sensors
• ISSN: 0733-8724; 1558-2213
• DOI: 10.1109/JLT.2025.3552628

Ultrasound detection can promptly identify partial discharge (PD) faults inside power transformer by using fiber-optic Fabry-Perot (F-P) array sensors. This paper presents a new design scheme for a wideband acousto-optic direct-coupled F-P sensor. Indirect-coupled F-P sensors have a narrow, highly sensitive response region, and direct-coupled F-P sensors have a wide band response from 1 to 200 kHz. We constructed a PD database by simulating different PD signals in oil tank. We proposed learnable KAN kernel to replace traditional CNN kernel and introduced KAN as a replacement for the ReLU function. Furthermore, we conducted pattern recognition research with a convolutional depth of only 6 layers. The test accuracy on oil tank dataset reached 98.7% for KAN Convolution + ReLU and 99.5% for KAN Convolution + KAN. The validation accuracy achieved by inputting PD signals measured from a real 220 kV power transformer into the two models was 97.5% and 98.75%, respectively.