Supervised Domain Adaptation for Remaining Useful Life Prediction Based on AdaBoost With Long Short-Term Memory

With the advent of high-quality deep learning algorithms, several methods have been proposed for addressing domain adaptation (DA) problems in remaining useful life prediction. The majority of these methods are unsupervised DA techniques, with popular adversarial approaches known to exhibit superior performance among them. However, we have identified that adversarial approaches have an instability problem; that is, their performance is highly dependent on the initial weights of the deep-learning network. Furthermore, unsupervised DA methods can only achieve limited performance improvements when the domain shift is substantial. To address these challenges, this study proposes a supervised DA method based on AdaBoost with long short-term memory (LSTM) as the base estimator. The proposed approach demonstrates effectiveness, particularly when the target-domain data are significantly smaller than the source-domain data. The proposed methodology was tested on a publicly accessible dataset, and when compared to previous unsupervised domain adaptation prediction methods, it achieved state-of-the-art prediction performance on most domains.