On leveraging network-wide information from hotspot sensor networks using Multi-output Gaussian Process regression models

With the needs for full structural state awareness and health monitoring as well as emerging challenges of Urban Air Mobility (UAV) and Future Vertical Lift (FVL), Health and Usage Monitoring systems (HUMS) need to be more accurate, robust and reliable than ever before. In active-sensing guided-wave networks in particular, conventional Damage Index (DI)-based approaches have been the industry standard for decades because of their computational simplicity and ability to do the damage detection and quantification tasks. However, under specific circumstances, like for specific actuator-sensor paths within a network or due to varying operational conditions, DIs can suffer from various drawbacks that make them prone to inaccurate and/or ineffective damage quantification. This study builds on previous work by the authors where DIs were used to train single-output Gaussian Process regression models (SOGPRMs) for robust damage quantification, and the accuracy limit of SOGPRMs was shown to depend on the evolution of the chosen DI formulation with damage size. In this study, multi-output GPRMs (MOGPRMs) are used instead in order to leverage information about damage size from multiple actuator-sensor path DI values. It is shown that the proposed approach can overcome the different shortcomings of DI evolution with damage size in the different path by capturing the correlation between the DI evolution for different paths. The proposed framework is applied for an Al coupon with simulated damage, and the damage size quantification results are compared with those of SOGPRMs. It is shown that the information fusion approach exhibited by MOGPRMs gives more accurate damage size estimations compared to SOGPRMs.