Prognostics and Health Management (PHM) technologies have emerged as a key enabler to provide early indications of system faults and perform predictive maintenance actions. Implementation of a PHM system depends on accurately acquiring in real time the present and estimated future health states of a system. For electronic systems, built-in-test (BIT) makes it not difficult to achieve these goals. However, reliable prognostics capability is still a bottle-neck problem for mechanical systems due to a lack of proper on-line sensors. Recent advancements in sensors and micro- electronics technologies have brought about a novel way out for complex mechanical systems, which is called embedded diagnostics and prognostics (ED/EP). ED/EP can provide real-time present condition information and future health states by integrating micro-sensors into mechanical structures when designing and manufacturing, so ED/EP has a revolutionary progress compared to traditional mechanical fault diagnostic and prognostic ways. But how to study ED/EP for complex mechanical systems has not been focused so far. This paper explores the challenges and needs of efforts to implement ED/EP technologies. In particular, this paper presents a technical framework and roadmap of ED/EP for complex mechanical systems. The framework is based on the methodology of system integration and parallel design, which includes six key elements (embedded sensors, embedded sensing design, embedded sensors placement, embedded signals transmission, ED/EP algorithms, and embedded self-power). Relationships among these key elements are outlined, and they should be considered simultaneously when designing a complex mechanical system. Technical challenges of each key element are emphasized, and the corresponding existed or potential solutions are summarized in detail. Then a suggested roadmap of ED/EP for complex mechanical systems is brought forward according to potential advancements in related areas, which can be divided - nto three different stages: embedded diagnostics, embedded prognostics, and system integration. In the end, the presented framework is exemplified with a gearbox.