A novel deep joint distribution alignment network for unsupervised remaining useful life prediction via variational auto-encoder

Accurate remaining useful life (RUL) prediction is a key component of prognostics and health management (PHM). It is critical for enhancing industrial system reliability, preventing catastrophic equipment failures, and thereby ensuring overall system safety. While unsupervised domain adaptation (UDA) is widely used for cross-domain RUL prediction, achieving accurate conditional distribution alignment (CDA) in continuous regression tasks remains challenging due to unreliable target pseudo-labels and complex industrial noise. To address these issues, a novel deep joint distribution alignment network via variational auto-encoder (JDAN-VAE) is proposed. Instead of deterministic feature extraction, the VAE maps raw inputs into a continuous probabilistic latent space. This regularized manifold synergizes with maximum mean discrepancy (MMD) constraints, preventing the model from overfitting to domain-specific noisy outliers and facilitating smoother distribution alignment. Concurrently, an auxiliary stage classifier (ASC) is trained using discretized source RUL labels to generate reliable pseudo-labels for the target domain. A joint alignment strategy is then executed, matching the marginal distribution via multi-kernel MMD (MK-MMD) and the conditional distribution via local MMD (LMMD). Extensive experiments on two distinct industrial benchmarks, namely the C-MAPSS aircraft turbofan engine dataset provided by NASA and the bearing vibration dataset provided by Xi’an Jiaotong University, demonstrate that JDAN-VAE effectively overcomes severe domain shifts and consistently outperforms state-of-the-art methods.