Low-cost sensor arrays are an economical and efficient solution for large-scale networked monitoring of atmospheric pollutants. These sensors need to be calibrated in situ before use, and existing data-driven calibration models have been widely used, but require large amounts of co-location data with reference stations for training, while performing poorly across domains. To address this problem, a meta-learningbased calibration network for air sensors is proposed, which has been tested on ozone datasets. The tests have proved that it outperforms five other conventional methods in important metrics such as mean absolute error, root mean square error and correlation coefficient. Taking Manlleu and Tona as the source domain and Vic as the target domain, the proposed method reduces MAE and RMSE by 17.06% and 6.71% on average, and improves R2 by an average of 4.21%, compared with the suboptimal pre-trained multi-source transfer calibration. The method can provide a new idea and direction to solve the problem of cross-domain and reliance on a large amount of co-location data in the calibration of sensors.