Contrastive, Autoencoding, and Variational Representations for Telemetry-Driven RUL Prediction

Abstract

Predictive Maintenance in safety-critical industries relies on accurate Remaining Useful Life (RUL) estimation from multivariate telemetry. Still real-world datasets are often dominated by censored observations and frequently lack explicit failure annotations. These constraints limit the effectiveness of purely supervised learning and motivate the need for approaches that can leverage unlabeled data. This paper presents a Pseudo RUL Guided Semi Supervised Learning framework that combines unsupervised representation learning with physics and statistics based soft failure indicators to enable robust RUL prediction under scarce failure labels. Compact latent representations are learned from censored telemetry using three encoder families i.e. autoencoders, variational autoencoders, and contrastive learning. The learned representations are subsequently used as inputs to a lightweight regression model trained on the available labeled samples. In scenarios where no failures are recorded, soft-failure transitions are used to construct pseudo-RUL targets, allowing training to proceed even in fully censored settings. Experiments on three diverse multivariate time-series datasets demonstrate that the learned representations consistently reduce prediction error relative to raw features while also reducing model size.