Hybrid wavelet transform and deep learning method for fault diagnosis of transformer vibration signals

Power transformers are critical components of power systems, and their safe and reliable operation is essential for both industrial and residential applications. Real-time condition monitoring and early fault detection are therefore of great importance, among which vibration-based analysis provides an effective non-invasive and online monitoring solution without direct electrical connections. This paper investigates deep learning-based fault diagnosis methods for transformer vibration signals, covering vibration mechanisms, data acquisition, and spectral analysis. An adaptive time-domain segmentation strategy is proposed to optimize the sample length and sliding step size for vibration signal modeling. Based on this strategy, an improved 1D-CNN is developed by incorporating small convolution kernels, BN, dropout regularization, and adaptive segmentation to enhance feature extraction capability. All simulations and model training are implemented using MATLAB and the PyTorch deep learning framework. Furthermore, a dual-branch CNN-based classification model leveraging continuous wavelet transform (CWT)-based time–frequency representations is constructed to address the complexity of transformer vibration signals. Unlike conventional single-label fault diagnosis approaches, this study introduces a dual-label encoding strategy together with a dual-branch CNN architecture, enabling independent learning of fault type and fault severity representations. This design reduces feature entanglement, enhances latent fault recognition capability under incomplete fault samples, and improves diagnostic interpretability. Experimental results demonstrate that the proposed method achieves over 99% accuracy in transformer fault diagnosis and more than 98% recognition accuracy on multi-point vibration data, indicating its effectiveness and practical applicability for transformer fault detection under limited prior knowledge.