Combining Machine Learning and Simulation for Structural Health Monitoring in Urban Air Mobility

Abstract

For Urban Air Mobility to become a reality, the estimated per passenger cost per mile must decrease by an order of magnitude, to be comparable with the cost for a small sedan. NASA reports that this will be achieved through increased autonomy, technology improvements, and increased operational efficiency. To increase operational efficiency, these vehicles will have to employ Health Utilization Monitoring Systems (HUMS) to help maintenance personnel in identifying and predicting damaged components. HUMS data can add an extra layer of protection against mechanical system failures. In particular, Structural Health Monitoring (SHM) can significantly reduce maintenance and operational costs. One complexity of SHM is in correlating sensor data to damage characteristics. Finite element analysis (FEA) can be combined with machine learning to bridge this gap. This presentation describes a method for detecting cracks in an outer panel based on sensor strain gage data. In this method, diagnostic capabilities are developed based on Artificial Neural Network (ANN) algorithms, and finite element analysis (FEA) results of damaged airframe structure generate the data required for ANN training. Parametric FEM models are employed to consider various crack configurations. The proposed method is applied on a critical zone of an eVTOL fuselage by first creating an FEA model of the fuselage panel, including rivets, stringers, and frames. Next different damage configurations are assessed using the contour integral method and measuring the different stress distributions caused by a progressive crack in critical areas. The FEA approach also suggests the optimal sensor grid design that should be implemented to accurately capture crack characteristics. Sensor outputs, in terms of strains, are measured at different locations and subsequently used in the ANN for predicting crack location and size. Training with the ANN is designed to avoid overfitting on the data and yielded reasonable accuracy in predicting the test data.