The work aims to design an on-line unsupervised integration algorithm, in order to solve the problems of poor model robustness, poor generalization and low loading rate in the existing 3D packing algorithm. In full consideration of the real-time premise of just-in-time cargo and with the container space utilization rate as the optimization goal, based on the end-to-end learning model framework of unsupervised deep fusion pointer network, the stacking process of online 3D packing was formulated as a Markovian decision-making process, to design reinforcement learning elements, and to give priority to the deep reinforcement learning algorithm. The decision-making actions of the agent were trained with the Monte Carlo tree search to generate an online three-dimensional boxing model with better learning ability. 125 randomly generated cargo data sets with different sizes and directions were tested under 7 constraint conditions. The experimental results showed that the average utilization rate of containers could reach 84.6%. The generalization of the algorithm is good, and the loading rate of the algorithm is much better than the current heuristic and depth learning method, providing theoretical basis and reference for on-line packing of cargo.