Efficient Cloud-Sourced Transport Mode Detection Using Trajectory Data: A Semi-Supervised Asynchronous Federated Learning Approach

The Internet of Things enables collaborative efforts in pattern recognition tasks within intelligent transportation systems, such as transport mode detection (TMD). However, collecting individual trajectories, like GPS records, typically involves privacy issues. To address this, federated learning (FL) frameworks have recently been applied. In such frameworks, users retain their private client datasets and are responsible for training local models. Only the updated client model parameters are sent to a central server, where these parameters are aggregated to update a global model. Then, the server broadcasts the updated global model to all users for the next round of local training. In this way, users can contribute to the global model without sharing private data. However, traditional FL frameworks are inefficient as the server has to wait for multiple users to upload their model parameters for synchronous parameter aggregation and consistency updates. This process also faces risks from unreliable clients. Furthermore, private client datasets are often unlabeled, posing challenges for local model training. Therefore, this article proposes a semi-supervised asynchronous FL framework for both point-level and segment-level TMDs. Specifically, the proposed framework incorporates model splitting techniques, model shift penalties, and entropy-based aggregation strategies to address model complexity, model drift, and data imbalance, respectively. Moreover, a CNN-based deep learning model with multiple encoders is proposed, and a pseudo-labeling-based approach is applied to utilize unlabeled datasets. The case study demonstrates that the proposed model achieves satisfactory performance on a real-world dataset, and the proposed FL framework is robust under varying hyperparameter configurations.