SpecFusionNet: An ultra-lightweight dual-branch network with cross-attention for millisecond-level voiceprint recognition in power transformers

Abstract

Real-time condition monitoring of power transformers through on-device voiceprint analysis is indispensable for grid stability. However, this critical application is frequently hindered by the excessive model size and inference latency of conventional deep learning approaches. To resolve this, this study introduces the Spectral Fusion Network (SpecFusionNet), a novel dual-branch architecture that synergistically integrates a Convolutional Neural Network with an adaptive Fast Fourier Transform-Transformer to efficiently model complementary voiceprint features. Evaluated on a comprehensive dataset covering six operational states, SpecFusionNet demonstrates exceptional performance, achieving a state-of-the-art accuracy of 99.68% with a model footprint of only 0.27 MB and a single-sample inference time of 0.38 ms. Exhibiting profound efficiency gains—being 99% smaller than MobileNet and 3.1 times faster than a Temporal Convolutional Network, demonstrating significant advantages in efficiency. Consequently, SpecFusionNet establishes a practical and powerful methodology, effectively bridging the gap between high diagnostic accuracy and the demands of resource-constrained applications.