How Can a Hybrid Network Predict Bearing Remaining Useful Life?

Rolling bearings are the heart of industrial machines, yet predicting when they will fail is tricky. Researchers have devised a hybrid method that blends AI and physics to estimate a bearing's remaining useful life (RUL). First, raw vibration signals are cleaned and broken into tiny waves using windowed empirical mode decomposition (EMD). These waves form TIR sequences that capture the signal's time variations. Next, a one‑dimensional temporal feature convolution‑reallocation network (TFCR) analyzes the sequences and gives an initial RUL estimate. Finally, a physics‑based Kalman filter refines the prediction, making it more stable and consistent with the bearing's real behavior. Tests on the XJTU‑SY dataset show this approach outperforms traditional methods, delivering more accurate forecasts that are less affected by load or speed changes. By merging signal analysis, deep learning, and physical modeling, machines can be monitored smarter, cutting downtime and maintenance costs.

The proposed method rests on three core pillars. 1) Windowed Empirical Mode Decomposition (EMD): EMD splits a nonlinear signal into Intrinsic Mode Functions (IMFs) representing oscillations at different time scales. A moving window captures local dynamics, reducing noise influence and improving component separation. 2) TFCR (Temporal Feature Convolution‑Reallocation) network: a 1‑D CNN with variable convolutional layers and a reallocation mechanism that reorganizes extracted features to highlight critical temporal patterns. Its configurable depth and width allow adaptation to datasets with varying sequence lengths. 3) Physics‑based Kalman filter: the filter employs a linear state‑space model describing bearing degradation (e.g., Lundberg fatigue law). The network's predictions serve as observations, while the filter corrects the estimate accounting for uncertainties in both model and measurement. This hybrid approach reduces prediction variance and ensures estimates obey mechanical energy conservation laws. Experiments on the XJTU‑SY dataset, which includes variable loads, speeds, and temperatures, show an average 15 % reduction in mean absolute error (MAE) compared to purely data‑driven methods, and a 30 % drop in prediction standard deviation. Moreover, Kalman integration enhances robustness against outliers and impulsive noise, making the system suitable for real industrial scenarios with rapidly changing operating conditions.