Federated two-stage machine learning model for ultrasonic guided wave structural health monitoring of composite structures

Kurzfassung

Ultrasonic guided wave-based structural health monitoring (SHM) systems have the capability to effectively monitor the condition of composite aircraft structures. Recent advancements in the application of data-driven machine learning models to SHM data have improved the assessment of structural damage in composite materials. In perspective future applications, SHM systems will be applied to fleets of aircraft, potentially operated by different stakeholders. This implies that classic centralized machine learning concepts have to be adapted to SHM datasets distributed across several aircraft. Federated learning as a decentralized machine learning architecture enables collaborative cross-fleet damage detection and prediction of damage locations while preserving data privacy. This paper proposes a federated two-stage damage classification and localization method based on an ultrasonic guided wave experimental dataset from the Open Guided Waves platform of a composite plate. The first stage corresponds to a centralized unsupervised autoencoder for damage detection. The second stage utilize federated supervised regression models, incorporating convolutional neural networks (CNNs) and long short-term memory (LSTM) models for damage localization on the composite plate. The present study assesses the performance of two federated learning algorithms: FedAVG and FedProx. FedAVG aggregates model parameters from local clients by averaging them, whereas FedProx incorporates a regularization term to address heterogeneity among client data. This decentralized approach reduces data overhead while maintaining data privacy from the perspective of collaborative learning by transferring only the trained model parameters between centralized models and the decentralized server. A comparative analysis of the central versus the federated approach in terms of prediction accuracy and performance demonstrates the capability of the federated learning architecture to handle decentralized SHM datasets effectively, achieving comparative results despite the absence of locally available data.