To solve the existing problems in bearing fault diagnosis, such as the difference of domain distribution, the mismatch of label space and the deficiency of labeled samples in the target domain., we propose a feature-level meta-transfer learning framework (FMTLF) for few-shot cross-condition fault diagnosis of bearings. Specifically, FMTLF attempts to calculate the prototype of each subdomain and combines it with domain adversarial training. It not only helps the model to accumulate meta-knowledge of the source domain, but also effectively learns discriminative patterns of the target domain. A feature-level discard-excitation module is designed to perform discard and excitation on the features for the setting to improve the summarization ability. Comparison experiments on Case Western Reserve University public datasets verification of validity of FMTLF and the feasibility.