Accuracy for aircraft noise identification model generalized by applying swarm learning

To conduct detailed studies on aircraft noise and manage noise maps around airports, it is necessary to measure aircraft noise under various conditions. In such measurements, it is necessary to obtain not only the noise signals but also information that identifies the noise, such as the type of aircraft and flight mode, and so on. To automate this measurement, we have been developing a technique for identifying aircraft noise using machine learning based on measured acoustic information. To enhance the generalizability of this technique, it is necessary to create an integrated model by accumulating aircraft noise data measured by various organizations. However, this type of information is subject to strict security and involves a large amount of data. To solve this problem, we applied Swarm learning, which accumulates only the weight information learned by machine learning in each organization, to create an integrated model. We compared the accuracy of models within each organization, models trained with the accumulated data from all organizations, and models that integrated only the weights extracted from each organization's model using Swarm learning. As a result, the models using accumulated data and those using Swarm learning had almost the same accuracy.