The distribution discrepancy between source domain data and target domain data will be aggravated due to time-changing operation conditions of the mechanical equipment in practical. Therefore， the performance of the intelligent fault diagnosis model is weakened. A novel method based on deep contrastive transfer learning is proposed for mechanical fault diagnosis under variable working conditions. Multilayer convolution block is used as the prepositive feature extractor to extract representative features from raw vibration data， which can improve the performance of fault classifier and domain discriminator. The feature extracted from prepositive feature extractor is transmitted to feature fusion device， and the convolution features can be refined and connected by the local and global receptive fields in prepositive feature extractor which can strengthen the feature-expressed capacity of the model. The refined features are utilized for fault classifier and domain discriminator to diagnose mechanical fault under different conditions. The Wasserstein distance is applied in fault classifier for measuring the discrepancy between source and target domain data.Based on the mutual information noise contrastive estimation， the mutual information domain discriminator is proposed to distinguish working conditions. All of them can raise the transfer capacity of the proposed method. Experiments on bearing and gear demonstrate that the proposed method diagnoses mechanical faults under variable working conditions based on transfer tasks effectively.