Composite material remaining useful life estimation using an IoT-compatible probabilistic modeling framework

Monitoring real-time failure in fiber-reinforced composites using nondestructive evaluation (NDE) is a fairly difficult task due to the associated evolving material state uncertainties, as well as due to data management mining and signal processing challenges. In this context, a number of statistical, probabilistic and physics-based models have been proposed to make predictions of remaining useful life (RUL) in the built environment. In parallel to modeling efforts, the data quality from sensing, NDE and testing methods is compromised by a variety of intrinsic (hardware) and extrinsic (noise) factors that make the damage assessment process both challenging to visualize and computationally expensive to analyze. Most importantly, the uncertainties in the recorded data, also hinder efforts to create data-driven methods in the framework of what is currently called digital twin modeling. To address such issues more efficient data workflows and processing procedures are currently sought to assist with both damage monitoring assessment, as well as modeling and predictions. The objective, therefore, of this manuscript is to present a novel data-driven probabilistic modeling methodology that is capable of producing RUL estimates for composites. The framework consists of two parts: the hardware/software integration that allows the establishment of data streaming and handling procedures that ultimately feed in real time the probabilistic approach. The proposed probabilistic modeling approach is based on the use of an outlier analysis combined with information quality metrics to identify a given set of degradation states. Predictions of RUL are then made by extracting and processing features from NDE dataseis which are then used in unsupervised clustering. Such information is subsequently leveraged to train a support vector machine (S VM) alongside a hidden Markov chain model (HMM). The signal classifications from SVM combined with the HMM are then used as inputs for an adaptive neuro-fuzzy network that produces the RUL predictions. An application of the proposed approach in mechanical testing experiments of an aerospace-grade composite material is presented. Extensions that can make this approach applicable to industrial applications as well as in digital twin implementations are discussed.