Unsupervised Hybrid Deep Learning for Unknown Bearing Fault Diagnosis and Severity Assessment
TL;DR
An unsupervised hybrid deep learning framework for unknown bearing fault diagnosis and severity assessment using vibration signals that combines Continuous Wavelet Transform, Convolutional Neural Networks, and Long Short-Term Memory autoencoders is presented.
Abstract
This paper presents an unsupervised hybrid deep learning framework for unknown bearing fault diagnosis and severity assessment using vibration signals. The proposed framework combines Continuous Wavelet Transform (CWT), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) autoencoders. Trained exclusively on 3,779 healthy data segments, the model detects anomalies via reconstruction error analysis evaluated across a total of 7,585 test segments. Experiments on the CWRU dataset show that the proposed hybrid model achieves competitive performance (AUC 0.990, accuracy 90.42%, and F1-score 89.57%) compared to spectral baselines, while uniquely preserving temporal dynamics—a critical advantage for non-stationary industrial environments. However, outer race faults were not reliably detected under the global threshold, which we report as a key limitation. A severity assessment and Health Index are also introduced for interpretable predictive maintenance.