Deep Learning-based high-resolution spectrogram reconstruction for robust bearing fault diagnosis

Abstract

Accurate and reliable bearing fault diagnosis is critical for the safe and efficient operation of rotating machinery. Conventional time–frequency (TF) analysis techniques, such as the short-time Fourier transform (STFT), are limited by their fixed window sizes and the inherent trade-off between time and frequency resolution, often leading to suboptimal fault representation under varying operating conditions. To address these challenges, this paper proposes a deep learning-based framework that reconstructs high-resolution, noise-robust TF representations directly from raw vibration signals. The proposed model integrates a denoising encoder with a learnable TF transformation network to adaptively capture fault related frequency patterns without manual parameter tuning. The proposed model’s performance is evaluated using multiple CNN-based classifiers on two benchmark datasets: Case Western Reserve University (CWRU) and HUST. Experimental results demonstrate that the proposed approach achieves superior classification performance, with improvements of up to $1.52\%$ in accuracy and F1-score over HRSpecNet, and over 13% compared to TFA-Net. These results confirm the proposed model’s ability to generate discriminative, high-resolution TF representations suitable for intelligent bearing-fault analysis and real-time industrial predictive maintenance.