In view of the dynamic nature of air target track， time sequence variation and multiple intentions in battlefield， this paper proposes a recognition method based on end-to-end generic attribute learning， which is used as the basic framework of intelligent multi-intention recognition model. By integrating the time sequence features and attribute characteristics of the target track， the unified input coding information is compressed and modified to encapsulate the knowledge and experience of experts as labels， and the commander's thinking mode of wartime situation judgment is learned to eliminate the interference factors brought by its concealment， deception and confrontation， so as to obtain the complex tactical intention of the specific target. Through the simulation experiment， the common multi-classification evaluation system is used to analyze the influence of the end-to-end training method on the results， and the comparison analysis with the relevant methods shows that the proposed algorithm is more effective and valuable for multi-intention recognition， and can be used to support the establishment of the correlation between non-cooperative targets and defense important places in the combat planning system.