Multi-Frequency Fiber-Optic Sensing Integration With Ddata-Augmented Models for Partial Discharge Pattern Recognition

To address challenges in partial discharge (PD) pattern recognition, including insufficient time-frequency feature extraction, limited samples, and poor model generalization, this paper proposes a hybrid method integrating multi-frequency fiber-optic sensing, Adaptive Optimal-Kernel Time-Frequency Analysis (AOK), Diffusion Probabilistic Model (DDPM) data augmentation, and an improved Convolutional Neural Network (CNN). A tri-frequency Fabry-Perot (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{F}-\mathrm{P}$</tex>) sensing array is constructed to synchronously acquire ultrasonic signals from five typical PD types (MT, AC, SD, SD-p, SD-p-w). The AOK algorithm generates <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$300 \times 375$</tex>-pixel high-resolution timefrequency matrices, effectively suppressing cross-term interference and enhancing time-frequency focusing. For smallsample scenarios, DDPM is introduced to augment original timefrequency data with physically consistent samples. Through 1000step noise scheduling, 200 samples per class are generated. t-SNE visualization confirms the generated data preserves characteristic overlaps (e.g., between SD-p and SD-p-w) while improving diversity. An improved CNN-SVM hybrid model is designed, mapping three-channel time-frequency matrices to RGB-like images. This architecture optimizes decision boundaries by combining deep feature extraction with support vector machine classification advantages.