A comprehensive review on quantum deep neural networks for prognostics and health management: Fundamentals, challenges and opportunities

초록

The domain of Prognostics and Health Management (PHM) targets accurate prediction of remaining useful life (RUL), efficient estimation of state of health (SoH), and assessment of anomaly indicators from multivariate data emanating from systems under functional as well as structural degradation that are prevalent across various engineering areas. Proactive assessment of SoH and prediction of RUL are core objectives of PHM and remain challenging as degradation dynamics is commonly nonlinear, (plausibly) non-stationary, multi-sensor based, often data-limited, and frequently affected by changing operating conditions. Quantum Deep Neural Networks (QDNNs), typically implemented through parameterized quantum circuits within hybrid quantum–classical workflows, have recently emerged as a potential alternative or complement to classical deep models due to their compact parameterization and expressive quantum feature spaces. This article presents a tutorial-style review of QDNNs for PHM. We first summarize the quantum-computing foundations required to interpret such models, including qubits, measurement, data encoding, parameterized quantum circuits, gradient estimation, and hybrid optimization. We then organize the PHM literature by architecture family: feedforward, recurrent, convolutional, generative, and attention-based quantum models. We discuss, for each family, its mathematical principle, typical encoding choices, datasets, reported metrics, and comparison baselines. The current evidence suggests that QDNNs can be competitive and sometimes more parameter-efficient than selected classical baselines; however, the literature remains heterogeneous, often simulator-dominated, and insufficient to establish systematic quantum advantage. We conclude by identifying the main open challenges as well as emerging opportunities for quantum enhanced prognostics. © 2026 Elsevier Ltd.

키워드