Towards Trustworthy Degradation Prediction: An Interpretable Deep Learning Approach with Sparse Feature Extraction and Temporal Fusion