Federated learning (FL) has attracted growing attention for decentralized training of diagnosis models across clients while preserving privacy. Industrial big data are inherently not independent and identically distributed, which deteriorates the performance of global FL diagnosis model on individual clients. To solve this problem, a novel personalized FL framework, namely federated diverse knowledge learning (FDKL) is proposed in this paper. In the FDKL method, a trainable client-specific knowledge memory module is designed to store the bias of feature representation locally. Meanwhile, the client-general knowledge regularization module is developed to guide local feature encoder learn global unbiased knowledge. Two modules work synergistically and eventually customizes a personalized model for each client. Extensive experiments conducted on a self-build propulsion shaft test rig verify the effectiveness and superiority of the proposed method.