Deep learning frameworks(such as deep convolutional networks) require data to have a regular shape. However, discrete features extracted from heterogeneous data cannot be collected in a regular shape to convolute. In this article, a Two-Dimensional Discrete Feature Based Spatial Attention CapsNet(TDACAPS) is proposed to convert one-dimensional discrete features into two-dimensional structured data through Cartesian Product for surface electromyogram(sEMG) signal recognition. sEMG signal varies from person to person is the main signal source of prosthetic control. Our model transforms multi-angle discrete features into structured data to find the inherent law of sEMG signal. Due to uneven information distribution of structured data, this model combines capsule network with attention mechanism to place emphasis on abundant information regions and reduce ancillary information loss. Extensive experiments show our model yields an improvement for sEMG signal recognition of almost 3% than capsule network and other neural networks under different conditions. Our attention mechanism that employs overlapping pooling to search feature map weight is preferable to the squeeze-and-excitation module, convolutional block attention module and others. Moreover, we validate that our model has great expansibility on Wine Quality Dataset and Breast Cancer Wisconsin.

深度学习框架（例如深度卷积网络）要求数据具有规则的形状。但是，不能以规则的形状收集从异构数据中提取的离散特征进行卷积。本文提出了一种基于二维离散特征的空间注意力帽网（TDACAPS），通过笛卡尔积将一维离散特征转换为二维结构化数据，用于表面肌电信号（sEMG）的识别。sEMG信号因人而异，是假体控制的主要信号源。我们的模型将多角度离散特征转换为结构化数据，以找到sEMG信号的内在规律。由于结构化数据的信息分配不均，该模型结合了胶囊网络和注意力机制，将重点放在丰富的信息区域上，减少了辅助信息的丢失。大量实验表明，在不同条件下，我们的模型对sEMG信号的识别性能比胶囊网络和其他神经网络提高了近3％。我们的注意机制采用重叠池来搜索特征图权重，优于挤压和激励模块，卷积块注意模块等。此外，我们验证了我们的模型在葡萄酒质量数据集和威斯康星州乳腺癌上具有很大的可扩展性。我们的注意机制采用重叠池来搜索特征图权重，优于挤压和激励模块，卷积块注意模块等。此外，我们验证了我们的模型在葡萄酒质量数据集和威斯康星州乳腺癌上具有很大的可扩展性。我们的注意机制采用重叠池来搜索特征图权重，优于挤压和激励模块，卷积块注意模块等。此外，我们验证了我们的模型在葡萄酒质量数据集和威斯康星州乳腺癌上具有很大的可扩展性。