Auxiliary discrimination technology for suspicious data of <italic>γ</italic> dose rate based on decision tree algorithm——a case study of Xiaomaiyu station near the Fuqing coastal NPP

γ radiation air absorbed dose rate (γ dose rate) is a rapid and sensitive parameter for indicating human-induced nuclear activities. With advantages of long-term, high-frequency, real-time, in-situ, and online monitoring capabilities, γ dose rate has been widely applied in radiation monitoring networks worldwide. To satisfy the need for real-time analysis of massive γ dose rate data from hundreds of observation stations in China, this study proposed an artificial intelligence algorithm—the decision tree algorithm implemented in Python. Taking the Xiaomaiyu station near Fuqing nuclear power plant as an example, we utilized γ dose rate data (17520 records) along with auxiliary data on monsoon, tidal height and precipitation from 2018 to 2019 as the training set. The decision tree algorithm was applied to discriminate the suspicious data in the test set of γ dose rate data in 2020. We successfully identified 6 suspicious records out of 8784 data. These 6 suspicious records were classified into “normal data” rather than “abnormal data” after the manual verification. To validate the reliability of this AI method, 6 sets of normal data over three standard deviations were selected for manual verification. It was found that the combined effects of rainfall and tidal height resulted in significant increases in γ dose rate (over three standard deviations). These 6 sets of data were classified into “normal data” by the decision tree algorithm. In summary, the decision tree algorithm reduced the workload by 94.8%, compared with the 115 suspicious data identified by traditional anomaly data discrimination methods (over three times standard deviation). This study proposed minute-level and real-time assisted discrimination technology, which could be applied in the national and provincial γ dose rate dissemination systems to enable instantaneous alarms for suspicious data and enhance discrimination efficiency by orders of magnitude. Given the limitations in the quantitative understanding of γ dose rate variability, the volume of training datasets, and the number of monitoring stations, further efforts are needed to improve these AI-based approaches. These improvements will benefit the identification of suspicious data and provide more robust, reliable, and efficient support for nuclear safety regulation in the era of big data in combination with the manual verification techniques. These advancements will also offer technical support for the timely response to public concerns over nuclear issues in the future.