Figure 11. Performance evaluation of the test sets shown in Table 5
after the combined dataset and the trained model weights are distributed
to the edge locations.
By applying the federated learning architecture of the classification
estimation model developed with the LSTM architecture, the weights of
the model are distributed over different locations. As a result of this,
as seen in Table 5, the improvement was followed and the following
results were obtained in this framework.
• The low number of datasets and low data diversity of local models
developed in distributed locations negatively affect the accuracy.
• An average increase in accuracy performance from 85% to 95% accuracy
was observed at each location.
• Classification accuracy drops to 45% at a temperature and humidity
outside of its own temperature and humidity range in each location’s own
dataset.
• Updating the weights of the locations centrally has enabled both the
widening of the temperature and humidity ranges perceived by the
locations, and higher accuracy with the increase in the number of
datasets.
Table 5. Evaluation of federated learning architecture of different
locations