Deep learning has achieved significant success in wearable sensor-based human activity recognition. The inclusion of attention mechanisms allows to focus on more important parts of the feature map to further improve recognition accuracy. However, previous works on attention mechanisms typically involve a large number of parameters, posing a significant challenge for resource-limited devices. To address this issue, we have designed a multi-scale efficient spatial attention mechanism that enhances feature representation while reducing the number of parameters. Our method involves dividing the input into multiple branches, applying lightweight spatial attention to each branch, and then concatenating the branches to obtain the final feature maps. This approach provides spatial feature representations in different subspaces with low computational complexity, making it more suitable for deployment on resource-constrained devices. Our experimental results demonstrate that the proposed method, when integrated with ResNet, achieves advanced performance when compared with the state-of-the-art methods on two publicly available datasets: PAMAP2 and WISDM. Specifically, our method achieves an accuracy of 94.41% on PAMAP2 and 98.10% on WISDM.