Degradation is a universal phenomenon, manifesting as the rusting of infrastructure, the capacity loss in batteries, or the aging of biological tissues. Historically, these processes have been studied within disciplinary silos—engineering, materials science, and medicine—each developing its own specialized terminology and techniques. However, as the world moves toward predictive maintenance and personalized medicine, there is a growing need for a unified mathematical toolkit. This paper synthesizes recent advances in data-driven degradation modelling, offering a cross-disciplinary roadmap for tracking and predicting the "Remaining Useful Life" (RUL) of complex systems.
A Unified Taxonomy of Degradation Modelling
The authors propose a comprehensive classification tree that organizes the diverse array of data-driven methods into three primary pillars: Statistical Inference, Dynamic Prediction, and Machine Learning. This structure helps researchers identify the most appropriate mathematical tools based on their specific data types and objectives.
- Statistical Inference: This "classic" approach relies on established probabilistic frameworks. It includes regression analysis for identifying simple trends, factor analysis for uncovering latent variables, and Bayesian statistics (often utilizing Markov Chain Monte Carlo methods) for updating predictions as new evidence becomes available.
- Dynamic Prediction: These models are designed to capture how a system's state evolves over time. Markov and Hidden Markov Models (HMMs) are particularly effective for systems that transition between discrete states, such as moving from a "healthy" to a "damaged" condition. This category also encompasses Gaussian Processes and time-series forecasting tools like Prophet.
- Machine Learning: Representing the "modern" frontier, this pillar encompasses supervised learning (e.g., Support Vector Machines), deep learning (e.g., LSTMs for time-series sensor data), and reinforcement learning, which is uniquely applied to optimize maintenance policies rather than just predicting failure.
Cross-Disciplinary Trends and the Engineering Explosion
A significant finding of this review is the "universal language" of decay. The mathematical logic used to model the wear of a railway track is often nearly identical to that used to track the progression of a chronic disease. Despite these similarities, the adoption rates of these methods vary significantly across fields.
Engineering has seen a dramatic explosion in publication volume, rising from just 2 papers in 2015 to 23 in 2023. This surge is largely attributed to the rise of the Industrial Internet of Things (IIoT) and the availability of high-frequency sensor data from assets like wind turbines and CNC machines. In contrast, medicine and materials science show more steady growth, with medicine favoring probabilistic Bayesian models to handle the inherent biological variability and uncertainty.
Grounding Theory in Standards
A unique contribution of this paper is the mapping of mathematical methods to international regulatory standards. For instance, methods in medicine are linked to ISO 10993 for biocompatibility and ISO 14971 for risk management. In engineering, these models support ISO 55000 for asset management and ISO 9001 for quality management. This alignment ensures that academic advancements remain relevant to industrial and clinical practice, bridging the gap between theoretical research and regulatory requirements.
Future Directions and Challenges
While data-driven methods offer powerful predictive capabilities, several hurdles remain for the field to address:
- Physics-Informed Machine Learning: Purely data-driven models can sometimes produce physically impossible results. Integrating physical laws (e.g., thermodynamics) into neural networks (PINNs) is a key area of active research.
- Uncertainty Quantification: In safety-critical applications, knowing the model's confidence level is as important as the prediction itself.
- Data Fusion and Imputation: Real-world data is often messy, with missing values from sensor failures or missed clinical appointments. Developing robust techniques for data integration and "imputation" is essential for reliable modelling.
By bridging the gap between disparate fields, this review provides a foundation for more robust, interpretable, and cross-applicable degradation models, paving the way for the next generation of predictive systems.
Summary updated