Deep Transfer Learning for Bearing Prognostics: A CNN Adapted Regression Approach to Predict the Remaining Useful Life of Self-Aligning Double-Row Ball Bearings

Self-aligning double-row ball bearings are widely used in rotating machinery, and unexpected bearing failures cause costly downtime; therefore, accurate Remaining Useful Life (RUL) prediction is essential for predictive maintenance. However, bearing degradation signals are non-stationary and run-to-failure datasets are typically limited, which complicates robust data-driven prognostics. Existing methods often train on the entire bearing lifetime, weakening the input–target relationship, while fixed-threshold onset detection is prone to false alarms. This study aims to develop a data-efficient deep learning framework that reliably estimates bearing RUL while explicitly focusing learning on the progressive degradation interval. The study uses a run-to-failure dataset generated on an accelerated test rig developed at the University of Ferrara. It includes six experiments (E1–-E6) sampled at 25.6 kHz, with 5-s vibration recordings collected every 5 minutes under radial loads of 3–-5 kN. The data processing method applies standardization techniques to show degradation patterns before extracting the health indicator for detecting initial signs of degradation. The Short-Time Fourier Transform (STFT) produces spectrograms to show changing time-frequency signal characteristics in the data, preserving the time–frequency evolution essential for capturing non-stationary degradation. A gradient-based criterion is employed to check the slope of the standardized RMS by filtering short-term spikes before passing the data to pre-trained convolutional neural networks (CNNs) adapted for regression task (ResNet, Xception, DenseNet). Transfer learning from ImageNet reduces domain-specific data dependency, making the framework effective under limited run-to-failure data. Experimental validation confirms average absolute RUL errors of 2.7% for DenseNet-201, 2.9% for Xception, and 4.2% for ResNet-18. Onset-aware training reduces the mean absolute error (MAE) by 94% over full-sequence learning, and external transfer evaluation on the PHM benchmark reduces MAE (from 30.8 to 20.1). Validated under controlled constant-speed test-rig conditions, this approach provides accurate bearing RUL predictions and shows strong potential to support predictive maintenance and reduce unexpected failures.