Early Detection of Transformer Winding Faults Using Multimodal Vibration and Acoustic Sensor Data Fusion with Deep Learning Acoustic Sensor Data Fusion with Deep Learning

Abstract

Early changes in transformer winding geometry often appear as weak mechanical and acoustic signatures before they develop into insulation breakdown, severe inter-turn short circuit or service failure. This paper presents a non-invasive vibro-acoustic deep learning framework for early winding-fault detection. Synchronized vibration and acoustic recordings were obtained from a controlled 15 kVA, 11/0.415 kV, 50 Hz oil-immersed laboratory transformer under healthy operation and four emulated early-fault states. Vibration windows were represented as continuous wavelet transform scalograms, while acoustic windows were represented as Mel-spectrograms. A dual-branch CNN-BiLSTM model with additive attention was used to learn each sensing channel before feature-level fusion and five-class classification. The dataset comprised 600 raw recording runs and 7,200 synchronized windows. The split was performed at raw-run level before segmentation-based learning to prevent leakage among overlapping windows. Across five independent training runs, the fusion-attention model achieved 98.92 ± 0.24% accuracy, 98.61 ± 0.27% precision, 98.47 ± 0.29% recall and 98.54 ± 0.26% F1-score on the held-out test set. The results indicate that paired vibration and acoustic evidence improves early recognition of winding looseness, axial displacement, radial deformation and incipient inter-turn weakness under the controlled test conditions. The method provides a laboratory-validated alarm layer whose substation use should be checked against established maintenance evidence from frequency-response analysis, dissolved-gas analysis, thermal inspection and expert review