A Deep Learning Based Framework For Physical Assets' Health Prognostics Under Uncertainty For Big Machinery Data

Abstract

The ongoing development in sensor technology has allowed engineers to monitor complex systems through multisensorial data, generating thousands of data-points in time. This big machinery database is commonly stored to later be used by engineers for reliability purposes through traditional Prognostics and Health Management (PHM) techniques. However, most part of this valuable information is often wasted since PHM methods frequently rely on expert knowledge for their implementation, as well as a good understanding of the physics of failure that govern the system. Hence, to estimate reliability related parameters, such as the State of Health (SOH) or the Remaining Useful Life (RUL) of electrical and mechanical components, data-driven approaches can be applied to complement PHM methods. In this context, the purpose of this thesis is to develop and implement a novel Deep Learning (DL) framework for the health state estimation of systems and components, based on big machinery data. Accordingly, the following specific objectives are defined: Develop an architecture capable of extracting temporal and spatial characteristics from the data. Propose a health state estimation framework, and validate it using two benchmark datasets: C-MAPSS turbofan engine, and CS2 Lithium-Ion Batteries datasets. Finally, give an estimation of the uncertainty propagation for the health state prognostics yield by the proposed framework. This thesis proposes a DL framework, which integrates the advantages of spatial management from Convolutional Neural Networks, along with the sequential analysis capabilities from Long-Short Term Memory Recurrent Neural Networks. Dropout is used as a regularization technique, as well as a Bayesian Approximation for the estimation of the uncertainty of the model. Henceforth, the proposed architecture is named CNNBiLSTM. For the C-MAPSS dataset, four different models are trained, one for each sub-dataset, aimed to estimate the RUL. All four models yield state-of-the-art results for the Root Mean Square Error (RMSE) on their prognostics, showing robustness in the training process and small uncertainty for the test RMSE as well as for the RUL prediction. Similar results are obtained for the CS2 dataset, where the model trained using all battery cells estimates the State of Charge and SOH of the batteries with a lower RMSE than the state-of-the-art results, and a small uncertainty over its estimated values. Results yielded by the trained models show that the proposed DL framework is adaptable to different systems and can successfully obtain abstract temporal relationship from the sensorial data for reliability assessment. Furthermore, models show robustness during the training process, as well as an accurate output estimation with a small uncertainty