With the widespread adoption and proliferation of Internet of Things (IoT) technology, the human recognition activity (HAR) utilizing IoT devices, such as wearable sensors, can be applied across a range of diverse applications. Due to the complexity of activity recognition, most wearable activity recognition systems employ multiple homogeneous or heterogeneous sensors to capture excessive information. However, the increased sensor count and the way of multi-channel signal data pose significant challenges to tasks related to human activity recognition. Determining appropriate sensor channels that strike a balance between computational complexity and recognition accuracy emerges as a pivotal concern. In this part, we extend the sparse group lasso mechanism to address human activity recognition challenges and propose a hybrid attention-based multi-sensor pruning and feature selection deep neural network, abbreviated as HAP-DNN [1]. This architecture excels in conducting feature selection on the basis of sensor pruning. HAP-DNN comprises three modular and detachable components : the feature compression and reconstruction module for fusing and restoring sensor feature information, the feature weight calculation module for calculating sensor channel weights and feature weights, and the learning module for classification, akin to a filter-based feature selection method. Four public activity recognition datasets are used to verify our proposed architecture, and experimental results show that HAP-DNN achieves the best classification performance with the least number of retained feature channels.