当前的基于贝叶斯网络的目标意图识别方法主要聚焦于建立更合理的网络结构来提高识别准确率，忽视了网络参数精度对准确率的影响。由于目标所处环境的复杂性，获取的目标状态数据具有一定的不确定性，这造成用于贝叶斯网络参数学习的样本数据中包含有不确定信息。然而，目前的贝叶斯网络参数学习方法均没有考虑样本中不确定信息的问题，影响了目标意图识别的参数学习精度，使得准确率也随之降低。为了解决此问题，该文提出了一种不确定样本下的贝叶斯网络参数学习方法，在不损失样本数据信息的前提下直接利用不确定样本进行参数估计，以提高参数学习的精度。首先，从贝叶斯网络精确推理的角度，结合消息传播推理和判别学习方法，建立了不确定样本下的条件对数似然函数，并作为参数学习的目标函数；为了缓解小样本集下的过拟合问题，依据最大熵原理构建了参数的L₂范数正则化项；然后采用梯度下降法对目标函数进行优化求解以得到参数估计值。通过基于朴素贝叶斯分类的目标意图识别的实例测试，结果表明：所提方法有效提高了参数学习的精度和目标意图识别的准确率，并增强了小样本集下目标意图识别的泛化性能，缓解了过拟合问题。

Current target intention recognition methods based on Bayesian network mainly focus on establishing a more reasonable network structure to improve the recognition accuracy, ignoring the influence of network parameter precision. Due to the complexity of the target's environment, the obtained target state data is uncertain, which causes the sample data used for Bayesian network parameter learning to contain uncertain information. However, in the current target intention recognition based on Bayesian network, the problem of uncertain information in the sample is not considered in the learning of network node parameters, which affects the precision of parameter learning and reduces the accuracy of target intention recognition. In order to solve this problem, a Bayesian network parameter learning method under uncertain samples is proposed, which can directly use uncertain samples to estimate parameters without losing sample data information, so as to improve the accuracy of parameter learning. Firstly, from the perspective of Bayesian network precise reasoning, combined with message propagation reasoning and discriminant learning methods, the conditional log-likelihood function under uncertain samples is established and used as the objective function of parameter learning. In order to alleviate the over-fitting problem under small sample sets, the L₂ norm normalization term of parameters is constructed according to the maximum entropy principle. Then the gradient descent method is employed to optimize the objective function to obtain the estimated parameters. Through the example test of target intention recognition based on naive Bayes classification, the results show that the proposed method effectively improves the precision of parameter learning and the accuracy of target intention recognition, and enhances the generalization performance of target intention recognition under small sample sets, thus the over-fitting problem is alleviated.