Federated Learning (FL) has been considered as an appealing framework to tackle data privacy issues of mobile users (MUs) for mobile crowdsensing (MCS). However, due to heterogeneous features of MUs and dynamic of wireless channel state, this have a significant effect on the global model performance and the training efficiency, where the MU scheduling problem becomes crucial. In this paper, we focus on unknown MU scheduling problem for mobile crowdsourcing system in wireless network to enable FL, where MUs' quality of local model knowledge are unknown a priori. Our objective is to find a scheduling strategy from the perpective of maximizing MUs' quality of local model and improving the FL communication efficiency. We model MU scheduling problem as an integrated Combinatorial Multi-Armed Bandit (CMAB) program taking into account reputation mechanism, and then propose a discounted-UCB and reputation-based scheduling algorithm to balance between exploration and exploitation, as well as to reduce the FL communication latency. Moreover, we validate the superiority of the proposed scheduling algorithm via experiments conducted on an image classification task. The results show that the proposed scheduling algorithm can well promote the convergence performance of federated learning and save the training latency, compared with the state-of-the-art scheduling strategy.