Aiming at the bottlenecks of high cost of accelerated tests and the lack of systematic modeling and uncertainty quantification in existing virtual methods for the whole machine life verification of high-reliability MEMS sensors, this paper proposes a high-confidence virtual evaluation method based on Physics of Failure (PoF). By obtaining local stress through multi-scale digital prototypes and thermal-vibration coupling simulation, a modular PoF model library is constructed to predict the single-point life; a hybrid probability distribution driven by physical sources is innovatively adopted to quantify manufacturing and environmental uncertainties; and the Time To Failure (TTF) of the whole machine is predicted based on the two-level mechanism of "multi-mode fusion within the device" and "first failure between devices". The prototype has been verified through a self-conducted combined thermal-vibration physical experiment using the MPU9250, demonstrating high accuracy (with a shell temperature deviation of ≤± 0.2℃and a modal error of <5%), with the whole machine Time To Failure (TTF) reaching 7961 h (about 1.82 years under 12 h daily operation), and accurate identification of weak links. The proposed method requires no large-scale physical tests, significantly improving evaluation efficiency and credibility.