In complex engineering systems with thousands of components, preventive maintenance and repair maintenance can be costly and disruptive. While many Condition-Based Monitoring (CBM) and Health and Usage Management Systems (HUMS) have already been proposed for predictive maintenance and prognostics, they routinely assumed exact knowledge of the interconnections among various components and those between components and sensors. In addition, they often assumed physics-based models for propagation of failures over time. Although such schematics information may be available at a small component level, in complex engineering systems with many component suppliers, such information may not be complete or even available at all. Then, model-free online learning using outputs from sensors scattered across the system to automatically detect, identify and predict failures and remaining useful life (RUL) at component, module and system levels is the only solution. This motivates our current work. Based on our previous works on deep learning, state tracking, data fusion and optimization, we propose a CBM/HUMS processing system that is capable of using data-driven analytics and artificial intelligence for predictive maintenance and prognostics that can work with any existing complex systems without the need for system schematics.