In the current development of intelligent driving technology, high-level autonomous driving systems are undergoing significant transformations, transitioning from traditional rule-based approaches to AI-driven end-to-end methods. This paper presents a comprehensive study on high-confidence simulation testing methods for AI-based end-to-end autonomous vehicles, focusing on innovative methodologies that enhance the accuracy and robustness of testing. The study emphasizes three core aspects: the use of 3D Gaussian-based synthetic data to generate high-fidelity test scenarios that improve model performance in complex and rare long-tail driving situations; advanced physical-level sensor modeling that simulates sensor behavior under various environmental and weather conditions, including rain, fog, and high-glare sunlight; and the development of an axle-coupled Vehicle-in-the-Loop (ViL) system by China Automotive Engineering Research Institute (CAERI) to precisely simulate longitudinal and lateral road loads without the need for vehicle steering modifications. The proposed methods address key challenges in replicating real-world driving conditions in a controlled environment, ensuring reliable performance validation for autonomous systems. The results demonstrate that these comprehensive simulation approaches significantly enhance the reliability, consistency, and validity of test outcomes, supporting more robust development and deployment of intelligent driving technology. This research introduces innovative practices that lay a strong foundation for the future development of high-confidence simulation systems.