Abstract: To enable real-time monitoring of the load states of engineering vehicles, we propose a group squeeze-and-excitation temporal convolutional network-sparse transformer (GSTCN-STransformer) load state recognition model. Based on real-time data including speed, load collected by onboard sensors and GPS, we construct the differential features, and then perform feature selection by maximum information coefficient (MIC) method. The GSTCN extracts local channel-correlated features, and then extracts the STransformer for global feature. Final classification is achieved via a fully connected network. Experiments conducted on real-world engineering vehicle transportation data demonstrate superior performance of the model over the comparison models in terms of accuracy, macro precision, macro recall, and macro F1 value. Additionally, gradient-weighted class activation mapping (Grad-CAM) provides visual analysis of the model's decision-making process.