Abstract: The aero engine is crucial for the safe flight of aircraft. Predicting its remaining useful life allows for timely maintenance, thereby preventing potential flight accidents. Deep learning’s powerful data mining capabilities provide a novel approach to predicting the remaining useful life of aero engines using vast amounts of historical data. However, traditional deep learning methods often only analyze temporal data correlations, overlooking the complex non-Euclidian spatial relationship between multiple sensor data. In addition, they seldom address the uncertainties in data or the prediction process, which impacts the reliability of their results. To solve these problems, we propose a probabilistic prediction method for aero engines that leverages Bayesian networks and the graph attention transformer. First, sensor data undergo preprocessing using a convolutional denoising autoencoder. Then, to extract the complex non-Euclidian spatial relationships between sensors, we construct a graph convolution network. In this network, sensor signal features serve as node characteristics, and the relationships between sensors are measured by cosine similarity. An attention mechanism assigns different weights to sensor nodes to improve the expressive ability of the graph convolution network. In this paper, the multi-head attention is integrated into the graph convolution network. Following the integration, node characteristics are aggregated with the assigned weights in the proposed graph attention network. Furthermore, to achieve joint extraction of the spatiotemporal relationships of sensor data, the graph attention mechanism is integrated into the transformer’s temporal-multi-attention module. This integration combines the spatial feature extraction strengths of the graph attention transformer with the temporal feature extraction abilities of the transformer model. The extracted spatiotemporal joint features are often utilized to predict the remaining useful life of the aero engine. At the same time, a Bayesian network quantifies the prediction uncertainty of the model using Gaussian mixture distribution and variational inference. Optimal approximate distribution parameters are obtained by minimizing the Kullback–Leibler divergence between the real posterior and approximate distributions. Consequently, the loss function of the proposed model network consists of two parts: minimum mean square error, which reflects the distance between the predicted remaining useful life and the actual remaining useful life, and distribution approximation error measured by Kullback–Leibler divergence. Unlike traditional point value predictions, our probabilistic prediction method yields not only an estimated remaining useful life but also a corresponding confidence interval, which provides a more reliable foundation for subsequent maintenance and decision-making. In addition, we validated the proposed method using an aero engine experimental dataset, comparing its performance against other methods through both comparative and ablation experiments. The results demonstrate the effectiveness of the proposed method and its superiority over existing methods.