针对异常行为识别中人体动态骨架特征表达能力不充分的问题，提出了一种基于改进时空图归一化流的无监督异常行为识别方法，利用Transformer和卷积块注意力模块，在全局域和时空域中提高模型的特征表达能力，提升异常行为识别算法性能。首先，将Transformer模块引入归一化流的仿射层，在全局层面增强动态骨架特征信息的有效性；然后，分别在空间与时间图卷积模块中引入卷积注意力，有效地提升动态骨架特征的空间和时间表达能力；最后，在ShanghaiTech数据集和UBnormal数据集上进行仿真验证，识别精确度分别达到86.4%和70.2%，证明了方法的有效性。

In order to solve the problem of insufficient feature extraction of human dynamic skeleton features in abnormal behavior recognition, an unsupervised abnormal behavior recognition method based on enhanced spatiotemporal graph normalization flow was proposed. Transformer and convolution block attention module were employed to enhance the feature expression capability of the model and the performance of the abnormal behavior recognition algorithm in the global and spatiotemporal domains. Firstly, the Transformer module was incorporated into the affine layer of the normalized flow to augment the efficacy of dynamic skeleton feature information at the global level. Subsequently, the convolution attention was introduced into the convolution module of space and time graphs respectively to effectively enhance the spatial and temporal representation of dynamic skeleton features. Finally, simulation verification was conducted on the ShanghaiTech and UBnormal datasets, and the recognition accuracy attains 86.4% and 70.2% respectively, thereby demonstrating the effectiveness of the method.