Traffic flow detection methods tend to be diversified and intelligent, and the data anomalies become a prominent problem due to the influence of equipment, network and environment. Recovering abnormal data to obtain complete and accurate traffic flow data is crucial for traffic flow research work. To address this issue, traditional and mainstream traffic flow data recovery methods require a large amount of complete historical traffic flow data as a priori information for learning the trend of traffic flow data. Large amount of complete historical traffic flow data is hard to obtain in real traffic scenarios because traffic flow data collection is always accompanied by data anomalies. In this paper, we analyze the causes of abnormal data, classify the types of abnormal data, and propose identification methods applicable to different types of abnormal data based on the collection principles of freeway multi-source traffic data. We propose an auxiliary discrimination mechanism-oriented generative adversarial network (ADM-GAN) model to overcome the difficulties of traffic flow data recovery. Different from previous studies, we develop an auxiliary discrimination matrix to increase the utilization of original data, enhance the recovery accuracy, and improve the computational efficiency. We evaluate our model under different data missing rates (10%, 20%, 30%, 40% and 50%) by manually introducing missingness to the actual traffic flow data on the G50 freeway near Suzhou, China. The experimental results show that the proposed model outperforms other comparative methods. Under different missing rates, the recovery results obtained by ADM-GAN method are better than those obtained by comparative methods, and the lower the missing rate, the more obvious the advantages.