A lightweight transformer winding condition assessment method with multi-scale image fusion and an improved attention mechanism

In recent years, vibration image analysis has emerged as a promising technique for assessing transformer winding conditions. This study proposes a lightweight assessment model for transformer windings, integrating an image fusion module and a recognition module to address the accuracy limitations of single-image analysis and the high computational demands of multi-scale analysis. First, a Parallel Efficient Mixed Attention Mechanism (PEMAM) is proposed, designed to enhance adaptability to transformer vibration signals while maintaining a low parameter count. This mechanism improves the feature extraction capability of the Image Fusion Framework based on a Convolutional Neural Network, significantly boosting the signal-to-noise ratio and enhancing resistance to distortion in fused images. Subsequently, multi-scale Markov field images, derived from the time and frequency domain features of vibration signals, are fused and fed into the PEMAM-enhanced recognition module for condition assessment. Experimental results indicate that the proposed method achieves 99.63 % accuracy in identifying transformer winding conditions while maintaining low model complexity and computational cost.