Optimization of a DBN-based fault diagnosis algorithm for Oil-immersed transformers
TL;DR
This work proposes an optimized Deep Belief Network architecture that outperforms traditional DBNs by approximately 5% and matches the diagnostic precision of state-of-the-art architectures (e.g., Transformers), while offering superior computational efficiency and robustness under limited data conditions.
Abstract
Existing deep-learning models for transformer fault diagnosis often struggle with feature decoupling, limiting their accuracy under multidimensional, nonlinear, and strongly coupled operating conditions. To address this, we propose an optimized Deep Belief Network (DBN) architecture. Unlike conventional DBNs, our model fuses nine key gas ratios using a four-layer topology equipped with modern regularization techniques. The model employs a two-stage learning mechanism: generative pre-training via Restricted Boltzmann Machines (RBMs) and discriminative fine-tuning via a back-propagation neural network (BPNN). To improve practical applicability under a transfer-learning framework, pre-training utilizes augmented simulation data from a multi-physics transformer model, while fine-tuning leverages publicly available dissolved gas analysis (DGA) data. Comparative experiments demonstrate that the proposed method achieves an overall accuracy of 95.3%. In composite fault scenarios, it outperforms traditional DBNs by approximately 5% and matches the diagnostic precision of state-of-the-art architectures (e.g., Transformers), while offering superior computational efficiency and robustness under limited data conditions.