The federated learning method based on model aggregation can balance data and protect privacy, but the existing methods are difficult to achieve the effectiveness of open learning under data sharing. And it is difficult for the existed federated model to realize the asynchronous fusion on account of communication delay, which inhibits the optimal performance of the client. We propose an asynchronous federated learning paradigm to fuse the paramwters of federated center. The clients train their local network parameters based on the personal data and upload the updated parameters to the federated center. The federated center asynchronously fuse the uploaded parameters based on the sequential Kalman filtering. Numerical results show that the proposed method can achieve comparable performance to the centralized learning in fault diagnosis with the real data.