Label adversarial domain adaptation network for predicting remaining useful life based on cross-domain condition

Abstract

Most data-driven methods for predicting remaining useful life assume that the data under different operating conditions follow the same distribution. However, this assumption rarely holds in real-world situation. Additionally, traditional methods do not fully utilize the hidden label information from the target domain or account for the transfer quality of source domain data. To address these issues, Label Adversarial Domain Adaptation (LADA) network is introduced in this paper. Specifically, LADA aims to filter the source domain data and maximize the use of hidden label information from the target domain. Firstly, a similarity measurement indicator based on the pearson correlation coefficient (PCC) and dynamic time warping (DTW) is employed to filter source domain data similar to the target domain data distribution. Then, in order to fully utilize the hidden label information from the target domain, the cloud model and golden section are utilized to create pseudo class labels. Furthermore, a feature difference module is established that minimizes the disparity between domain features. This is realized by using the maximum mean difference (MMD) and Kolmogorov–Smirnov (K–S) statistical test. The experimental results indicate that LADA has advantages in cross-domain RUL prediction.