Structural health monitoring (SHM) systems employing ultrasonic guided waves have the capability to effectively monitor aerospace structures and detect instances of material degradation. In recent years, an increasing number of machine learning methods regarding SHM have emerged, leveraging their capacity to detect anomalies. This can serve as a foundation for predicting maintenance requirements and improving the operational efficiency of aircraft. However, large amounts of labeled sensor datasets for pristine and representative damaged specimens are required for supervised machine learning methods to generate reliable predictions. This is a major challenge, especially for a broad range of damages and flaws in aircraft materials. Additionally, the application of federated learning in operational aircraft fleets based on SHM data aims to enhance prediction accuracy by enabling collaborative model training across decentralized aircraft. Incorporating physical knowledge into the loss function of the neural network within a decentralized federated learning architecture can lead to a reduction in the amount of labeled datasets required, allowing for collaborative learning throughout aircraft fleets while preserving data privacy. As an initial step towards integrating ultrasonic guided waves propagation behavior into the neural network's training process, multiple Physics-Informed Neural Networks embedded into a federated learning architecture are introduced as a semi-unsupervised federated learning approach incorporating the linear second-order partial differential acoustic wave equation.