Remaining Useful Life Prediction for Electromechanical Actuator With Scale-Aware Domain Adaptive Deep Transfer Learning

As the aviation industry advances toward electrification, accurate Remaining Useful Life (RUL) prediction for electromechanical actuators (EMAs) is critical for ensuring flight safety. Traditional RUL models often struggle with generalization across varying operating conditions due to data distribution discrepancies. While transfer learning methods improve generalization, existing approaches typically measure data distribution at a single scale, limiting their effectiveness for EMA data with strong scale dependence. In response to this challenge, we propose a Scale-Aware Domain Adaptive Deep Transfer Learning (SDADTL) model for RUL prediction of EMA. The proposed framework begins with the establishment of a robust framework Leveraging a pre-trained Long Short-Term Memory (LSTM) model. This network captures complex temporal dependencies within the source domain data, providing a strong foundation for predictive modeling. Subsequently, a hybrid loss function is introduced, combining Mean Squared Error (MSE), Maximum Mean Discrepancy (MMD), and Correlation Alignment (CORAL). This loss function effectively mitigates domain discrepancies by aligning feature distributions between the source and target data, thereby ensuring the seamless transfer of knowledge. Then, the pre-trained LSTM model undergoes further training on the target domain data to fine-tune its parameters for the specific characteristics of the system. Extensive experiments conducted on several benchmark EMA datasets demonstrate that the proposed approach outperforms existing domain adaptation techniques, leading to a significant improvement in the precision and reliability of RUL predictions.