Vibration spectrogram-based ResNet–LSTM for bearing fault detection and severity estimation

Abstract

Accurate and reliable fault diagnosis of rolling element bearings is essential for ensuring the safety and efficiency of rotating machinery. Traditional methods often rely on handcrafted features, which are limited in their ability to capture the complex and nonstationary nature of vibration signals. In this study, a deep learning framework that combines spectrogram representations of vibration signals with a hybrid ResNet–LSTM architecture is proposed to detect fault type and classify severity level, simultaneously. Time-domain vibration signals are first converted into spectrograms to preserve both temporal and frequency-domain characteristics. These spectrograms are then processed as sequences, enabling the ResNet component to extract discriminative spatial features while the LSTM captures temporal dependencies across frames. The proposed model is trained and validated under different sequence lengths, strides, and LSTM unit sizes. Evaluation of the model on test data demonstrates that the near-perfect accuracy (> 0.99) in fault detection and high accuracy (0.91–0.99) in severity level classification is achieved using the proposed framework. Evaluation of the model on the industrial case study further shows that the proposed model not only achieves high accuracy on controlled datasets but also provides reliable predictions under realistic operating conditions.