Domain Adaptation Method for Gear RUL Prediction Integrating Convolutional Attention and Adversarial Learning

In recent years, deep learning-based methods for predicting gear remaining useful life (RUL) have garnered widespread attention. However, these approaches continue to encounter various obstacles, such as fluctuating operational environments, interference from noise, and a lack of adequate lifespan data samples. To address these issues, this paper proposes an adaptive gear RUL prediction model that combines convolutional block attention mechanisms (CBAM) with adversarial learning. First, a network structure combining a CBAM and temporal convolutional network (CBAM-TCN) is designed. Then, a deep feature extraction network is constructed using CBAM-TCN. Finally, by incorporating Domain Adversarial Neural Networks (DANN), the feature extraction network is optimized to enhance the cross-domain invariance and trend of deep features. The prediction accuracy of the proposed RUL prediction method is validated through multiple cross-condition RUL prediction tasks, demonstrating superior performance compared to other RUL prediction methods.