Edge intelligence propelled federated learning as a promising technology for embedding distributed intelligence in the Internet of Things (IoT) ecosystem. The multidimensional data generated by IoT devices is enormous in volume and personalized in nature. Thus, integrating federated learning to train the learning model for performing analysis on source data can be helpful. Despite the above reasons, the current schemes are centralized and depend on the server for aggregation of local parameters. So, in this paper, we have proposed a model that enables the sensor to be part of a defined cluster (based on the type of data generated by the sensor) during the registration process. In this approach, the aggregation is performed at the edge server for sub-global aggregation, which further communicates the aggregated parameters for global aggregation. The sub-global model is trained by selecting an optimal value for local iterations, batch size, and appropriate model selection. The experimental setup based on the tensor flow federated framework is verified on MNSIT-10 datasets for the validity of the proposed methodology.