To address the inconsistent bearing fault data distribution that leads to the difficulty of feature offset and distinctive feature extraction, a bearing fault diagnosis method based on time-frequency filter and offset attention neural network is proposed, which processes the fault signal from offline and online parts. In the offline part, a time-frequency filter is proposed to extract the distinctive features from time domain and frequency domain; A spatial sampling method considering both global and local features is proposed. In the online part, an offset attention neural network is proposed. Compared with self attention, offset attention is more conducive to the extraction of offset features, so as to reduce the impact caused by inconsistent data distribution. Experiments on the bearing datasets of Xi'an Jiaotong University (XJTU) and Case Western Reserve University (CWRU) have achieved 100% accuracy, which proves that the proposed method can efficiently extract the distinctive features of fault signals, and effectively suppress the influence of feature offset. The comparative experiment on the bearing dataset of CWRU proves the superiority of the proposed method. In addition, experiments are also carried out on the dataset of gas turbine main bearing collected in the industrial field, and the results show that the proposed method has practical significance.