Abstract: It is shown that hard samples worsening the performance of a system model has a decisive impact on the growth in the system's boundary decision capability. Due to the diversity of dangerous goods and the uncertainty of their instance forms as welll as the feature of "long-tail distribution" of the baggage X-ray image data generated by the on-site security system, the intelligent baggage detection model, which is trained based on the data set of finite sample collection, has the problem of low detection accuracy when applied to the on-site baggage security system. In this paper, we analyze the closed-loop data analysis process of the intelligent baggage security check, and propose a discrete reinforcement search method for selecting the hardest sample set from the X-ray images of dangerous goods instances generated during the operation of the baggage security system. Then, the hardest sample set is used as new training dataset to complete the learning update of the intelligent baggage detection model, thus realizing the continuous growth of the performance of the intelligent baggage security check software.