The Engine Bleed Air(EBA) System is a critical component prone to failure in civil aviation aircraft. Failures of the EBA system can impact flight operations’ efficiency and pose potential safety risks, so it is necessary to conduct fault diagnosis and prediction. However, the existing data-driven methods for baseline prediction of EBA systems struggle to effectively capture the performance degradation process from real-world data, which can lead to inaccurate prediction and poor stability. So the existing methods still have limitations such as missed detections and false positives in practical applications. In response, we propose the diffusion baseline framework, renowned for its significant advancements in deep probability generation, to analyze EBA dataset. We introduce a concise overview of diffusion principles and its training process, employing this framework to model the probability distribution of sensor relationships under normal aircraft conditions. Our study includes a comparative analysis against conventional time series prediction methods and alternative probability prediction approaches on real aviation industry datasets, demonstrating the superior performance of probability prediction facilitated by the diffusion framework in EBA system. Subsequently, based on the latest EBA dataset, we conduct fault diagnosis and establish health indices based on distribution patterns, elucidating the correlated probability relationships within high-dimensional temporal data. This findings highlight the efficacy of the diffusion method in probabilistic modeling of real-world industrial high-dimensional data time-series, there by enhancing the effectiveness of fault diagnosis and prognosis.