Federated learning (FL) enables Industrial Internet of Things (IIoT) devices to collaboratively train models without exposing their raw data. However, synchronous FL is sensitive to heterogeneous and time-varying computation and communication conditions because each round is delayed by the slowest selected client. Existing latency-aware selection methods can mitigate straggler delays but may repeatedly favor resource-rich devices, thereby limiting the participation of slower clients. GATES is introduced as an adaptive groupwise client-selection framework that balances round-completion latency and long-term participation requirements. At each communication round, clients are sorted according to their estimated computation and upload latency and then partitioned into groups with similar latency profiles. Client-specific virtual queues track accumulated participation deficits, while a joint selection metric evaluates each group according to its completion latency and participation demand. The resulting design reduces the client-selection candidate space and limits latency dispersion within synchronous rounds. Experimental results demonstrate that GATES achieves accuracy comparable to FedAvg and FedProx while consistently reducing per-round training latency under representative non-IID conditions.

Federated Learning,Industrial Internet of Things (IIoT),Edge Intelligence,Resource Heterogeneity,Distributed Machine Learning,Client Selection