Abstract: In the face of different bearing fault datasets, the small sample and differences in the distribution of each dataset lead to poor accuracy of the model for cross-dataset diagnosis. The proposed method introduces a fault detection approach that is based on the secondary-sequence meta-learning (SSML) algorithm, it is taken inspiration from the way humans gain knowledge. Firstly, the two independent datasets are sorted at the first sequence, which are divided into simple dataset and complex dataset correspondingly. Secondly, the simple dataset is divided into two datasets: one for meta-training and the other for meta-testing. Only the tasks in the meta-training set are sorted at the second sequence. The secondly sorted model is fine-tuned with a small amount of data in the meta-testing set to complete the fault diagnosis of the simple datasets. Finally, the model after secondary sequence is fine-tuned with a small amount of data from the complex dataset to complete the fault diagnosis of the complex dataset and realize the cross-data set fault diagnosis. SSML demonstrates superior performance compared to other models, this makes it very useful for diagnosing faults in complicated situations with small sample datasets and across different datasets.