In recent decades, fault diagnostics and failure prognostics have shown substantial promise for health monitoring and risk management in complex engineering systems, including smart factories, power plants, space systems, and heavy machinery. However, various uncertainties—such as those from model, data, process, environmental factors, and the inherent nature of engineered systems—significantly affect the credibility and applicability of these methods. Therefore, accurately quantifying the effects of these uncertainties is essential and one of the most widely held concerns to ensure trustworthy decision-making based on diagnostic and prognostic results. This special issue highlights recent advances in uncertainty-aware diagnostics and prognostics for managing the health of engineered systems, featuring eight collected papers.
Shen et al. presented an unsupervised domain adaptation transformation reconstructed gated recurrent unit framework considering prediction uncertainty for machinery prognostics under variable lubrication conditions. Their method integrated a domain adaptation layer and uncertainty quantification with recurrent neural networks. Lifecycle accelerated tests of the rolling bearing from PRONOSITA and ABLT-1A cross-validated the feasibility and effectiveness of the proposed machinery prognostics framework.
Jiang et al. leveraged the swin transformer deep learning model and acoustic emission technology for bearing fault diagnosis under low-speed and heavy-load conditions. Their method demonstrated improved sensitivity and accuracy in early fault identification and prevention of critical mechanical failures.
Zeng et al. introduced a data augmentation method based on image translation for Bayesian inference-based damage diagnostics, addressing uncertainties due to limited monitoring data in structural health assessment. By using CycleGAN to translate monitoring data across different miter gates, this approach augmented actual data with synthetic data, enhancing the volume of monitoring data available for the target miter gate. Their method demonstrated the potential of combining synthetic data generation with probabilistic model updating in structural health monitoring (SHM).
Najera-Flores et al. addressed the need for continuously updated predictive models in SHM by developing an uncertainty-aware, structure-preserving machine learning model to detect domain shifts based on real-time measured data. The model enforced distance preservation of the original input state space and then encodes a distance-aware mechanism via a Gaussian process kernel. A numerical nonlinear structure being subjected to damage conditions was used to demonstrate how the trained model can be used to detect domain shifts robustly with no prior knowledge about the underlying structural changes.
Qian et al. presented a conditional invertible neural network-based Bayesian model updating method to address the challenge that physics-based analysis may not accurately reflect the underlying true physics due to various sources. The method updated an existing corrosion simulation model based on the observations to make the prediction closer to field observations. Results demonstrated that the proposed model updating approach can provide more accurate inference results with a reduced computational cost in comparison to the classical approximate Bayesian computation approach.
To improve the ability to generalize from discrete data labels to the uncertain continuous spectrum of possible faults in diagnostics, Zhou et al. proposed an interpretability-enhanced probabilistic fault diagnosis framework that incorporates Bayesian variational inference into the deep learning architecture. This framework combined convolutional neural network feature extraction with a Bayesian approach to probabilistically interpret hidden data correlations. Validated on a lab-scale gear setup, the framework achieved nearly 100% accuracy in classifying known faults and 96.15% accuracy in classifying unseen fault severities.
To minimize the need for complex time-series feature engineering typically found in machine learning models, Mallamo et al. introduced a more interpretable flight regime identification approach that can handle complex and diverse flight profiles by leveraging general flight maneuvers to extract robust features that drive state of health and remaining useful life predictions. The method achieved an optimal balance between reducing redundant data and preserving variability, ensuring computational efficiency while maintaining robust feature extraction. Its performance was demonstrated and validated using the N-CMAPSS benchmark dataset, simulating flight profiles for a 737-800 equipped with a CFM-56 engine.
Kamariotis et al. provided a comprehensive review on the consistent classification and treatment of uncertainties within the fields of structural dynamics, system identification, and SHM. They demonstrated selected available strategies by means of their implementation on a benchmark shear frame with nonlinear dynamic behavior, and investigating the uncertainty quantification challenges in five key downstream SHM tasks: model inference, environmental and operational variability characterization, model updating, state estimation/virtual sensing, and inverse problem formulation for damage identification. The review identified open challenges in SHM, including the definition of prior assumptions affecting the uncertainty quantification, scalable uncertainty quantification solutions for complex problems, and strategies for reducing uncertainties.
Finally, the guest editors would like to appreciate all the efforts of the authors and reviewers in the development of this Special Section. We would also like to express our sincere gratitude to Professor Michael Beer, Editor-in-Chief of the ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part B for the incredible support throughout the preparation of this Special Section.