Embedding and Extracting Domain Knowledge in Machine Learning for Condition-based MaintenanceTools Rengasamy, Divish (2022) Embedding and Extracting Domain Knowledge in Machine Learning for Condition-based Maintenance. PhD thesis, University of Nottingham.
AbstractAs sensors become ubiquitous in condition-based maintenance (CBM) for the aerospace sector, there is a significant increase in data available to diagnose and prognose aerospace components' health and longevity. Machine Learning (ML), and more specifically, Deep Learning (DL), are popular tools deployed to analyse big data for CBM. For safety-critical systems such as aerospace, the ML models' predictive power to accurately diagnose fault and prediction of Remaining Useful Life (RUL) is crucial. Furthermore, the ability to explain and justify ML models' output is essential in decision making for CBM. In this thesis, DL models' predictive capability is improved by developing novel approaches that embed knowledge implicitly and explicitly into the loss function for condition-based maintenance. A proposed regression-based weighted loss functions that implicitly embeds knowledge allows the DL model to learn and focus on hard-to-learn instances, while minimising worst-case predictions and providing improved predictive performance. The proposed regression-based weighted loss function, along with an existing classification-based weighted loss function, namely, Focal Loss, are compared to traditional loss functions using aerospace gas turbine engines' RUL prediction and fault detection of the air pressure system, respectively. Furthermore, an asymmetric loss function that biases the DL models to favour early predictions in RUL is proposed. The asymmetric loss functions is tested on the same aerospace gas turbine engines dataset. The weighted loss functions and asymmetric loss function is able to increase the predictions accuracy over traditional loss functions. In addition to embedding knowledge, this thesis also introduces a new framework to extract knowledge from a learned model to explain the ML model's outputs by identifying important features driving the predictions. We proposed a new ensemble feature importance framework that fuses multiple ML models and their feature importance calculation approaches using both crisp and fuzzy decision fusion to create a more accurate and interpretable post-hoc explanation for the ML models. Subsequently, a fuzzy ensemble feature importance (FEFI) framework is proposed to overcome the shortcomings of crisp value-based ensemble feature importance. Both the new crisp and fuzzy frameworks are investigated using synthetic data under different conditions and also using a real-world case study of factors that affect creep rate in additive manufactured materials. Experiment results reveal that for our case studies FEFI framework is qualitatively more accurate in determining the feature importance compared to the crip valued-based ensemble feature importance framework and also traditional approaches.
Actions (Archive Staff Only)
|
Tools
Tools
