This paper presents a novel unsupervised incremental approach for real-time bearing health monitoring using online Principal Component Analysis (PCA) for feature fusion. The method addresses the need for memory-efficient, embedded monitoring systems that can operate without prior training data or full historical records. Time-domain features are extracted from streaming vibration signals and fused using an adaptive online PCA algorithm with exponential weighted moving averages (EWMA) and continuity corrections to ensure temporal consistency. A health index is constructed from the first principal component using a standardized cumulative sum (CUSUM) approach, enabling robust detection of gradual degradation patterns. The methodology is validated on the XJTU-SY bearing dataset comprising 15 run-to-failure experiments under varying operating conditions. The algorithm has been successfully implemented on an STM32 microcontroller, demonstrating its practical viability for embedded prognostic applications in resource-constrained environments.