While recent advancements in intelligent deep learning-based fault diagnosis methodologies have demonstrated considerable progress, the practical deployment of such techniques in industrial settings is impeded by challenges related to the acquisition of sufficient labeled data and the prevalence of non-stationary data environments. These factors render the extracted features susceptible to overfitting, consequently compromising the accuracy and efficacy of fault diagnosis. In this paper, we propose a novel framework that performs self-supervised contrastive learning in frequency domain to complish fault diagnosis with few labeled data. Furthermore, we investigate a frequency filter module designed to eliminate irrelevant spectral components and improve the efficacy of learned features. This module is utilized to process a limited amount of labeled data in the feature space, ultimately yielding the final frequency domain features. Experiments on two challenging bearing datasets demonstrate the superiority of the proposed framework over other self-supervised methods.