Mixup Domain Adaptations for Dynamic Remaining Useful Life Predictions

TL;DR

MDAN addresses dynamic RUL prediction under covariate shift by introducing a three-stage Mixup Domain Adaptation framework for time-series data. It first learns from the labelled source with mixup and self-supervised reconstruction, then creates an intermediate mixup domain to align source and target, and finally uses target-domain pseudo-labels with consistency regularization in the target. Across C-MAPSS and bearing datasets, MDAN achieves state-of-the-art performance in $12$ of $12$ domain-adaptation cases and demonstrates strong gains in the bearing experiments, all without a domain discriminator. The approach offers a simple, scalable path for robust transfer of RUL models under dynamic operating conditions and lays groundwork for future source-free and multi-source extensions.

Abstract

Remaining Useful Life (RUL) predictions play vital role for asset planning and maintenance leading to many benefits to industries such as reduced downtime, low maintenance costs, etc. Although various efforts have been devoted to study this topic, most existing works are restricted for i.i.d conditions assuming the same condition of the training phase and the deployment phase. This paper proposes a solution to this problem where a mix-up domain adaptation (MDAN) is put forward. MDAN encompasses a three-staged mechanism where the mix-up strategy is not only performed to regularize the source and target domains but also applied to establish an intermediate mix-up domain where the source and target domains are aligned. The self-supervised learning strategy is implemented to prevent the supervision collapse problem. Rigorous evaluations have been performed where MDAN is compared to recently published works for dynamic RUL predictions. MDAN outperforms its counterparts with substantial margins in 12 out of 12 cases. In addition, MDAN is evaluated with the bearing machine dataset where it beats prior art with significant gaps in 8 of 12 cases. Source codes of MDAN are made publicly available in \url{https://github.com/furqon3009/MDAN}.

Figures (9)

• Figure 1: Visualization of Mixup Domain Adaptation (MDAN): The first step learns the source domain using original source samples and mixup samples. An intermediate domain is established in the second step using the mix-up samples which linearly interpolates source and target samples. The last step is to extract the discriminative information of the target domain using pseudo labels and the predictive consistencies using mixup samples. Note that the mixup ratio controls the distributions of the mixed samples
• Figure 2: Source Domain Training: a model is trained to minimize the supervised learning loss, the mix-up loss and the self-supervised loss. The supervised learning loss is formulated as a MSE loss function while the mix-up mechanism is done in the feature level. The mixed samples and their corresponding labels are learned using the MSE loss function. The self-supervised loss is formulated as a controlled reconstruction learning process where the main goal is to reconstruct the masked inputs.
• Figure 3: Self-Supervised Learning: This phase is formulated as a controlled reconstruction learning process of randomly masked input samples. The goal is to predict the original input attributes given masked input features. The reconstruction process is done via the backbone network with the biLSTM encoder and the fully connected layer predictor.
• Figure 4: Intermediate Domain Training occurs with minimization of mix-up losses between the source domain samples and the target domain samples. The mix-up strategy dynamically interpolates samples of the input space and the feature space. The three backbone networks share the same network parameters.
• Figure 5: Target Domain Training is carried out by utilizing pseudo-labelled samples. Mixup samples are generated and optimized simultaneously with original pseudo-labelled samples.