Aerospace self-lubricating bearings are critical components in aircraft transmission systems, where wearinduced degradation under high-load and dynamic conditions poses significant challenges to operational safety and system longevity. In recent years, deep-learning methods have shown promise in wear prediction by leveraging abundant monitoring data from sensor networks. However, these methods often struggle to detect early-stage degradation and rely on labor-intensive feature engineering, limiting their effectiveness in handling noisy, high-dimensional data. To overcome these issues, this article proposes an improved variational autoencoder (VAE) and graph attention network method for wear prediction based on acoustic emission (AE) signals. First, a clustering-guided contrastive VAE (CGC-VAE) model is proposed to process noisy, high-dimensional AE signals. The CGC-VAE employs K-means clustering to segment wear stages, combined with Gaussian mixture model (GMM) regularization and contrastive learning, to extract low-dimensional, discriminative latent features. Subsequently, a temporal graph attention network (T-GAT) is proposed to construct a dynamic graph based on temporal proximity and feature similarity, which can effectively model the spatiotemporal relationships of latent features. It employs graph attention mechanism and a long short-term memory (LSTM) layer for accurate wear prediction. Finally, experimental validation on aerospace self-lubricating bearing datasets, covering full lifecycle and partial wear scenarios, demonstrates the superior accuracy and adaptability of the proposed method.