Abstract: 【Objective】This study aimed to establish a contactless detection method of hog respiratory rate using millimeter-wave radar combined with deep learning， thereby providing an intelligent and precise health monitoring solution for large-scale hog farming， enhancing breeding efficiency， reducing disease risks， and advancing the modernization of management for livestock industry.【Method】Respiratory radar data of hogs were collected using an AWR1843BOOST radar. A dual alignment calibration method （DACM） was developed to achieve phase alignment and unwrapping， specifically aimed at mitigating the effects of multipath interference and noise. Signal denoising and smooth tracking were performed through Kalman filtering. Subsequently， a model was constructed using EfficientNetV1 model as the backbone， integrating convolutional block attention module （CBAM） and dilated convolution （DilatedConv） module， and incorporating a lightweight Transformer module to achieve an efficient fusion of local multi-scale features and global dependencies. Finally， the respiratory rate prediction values were generated through global average pooling and a fully connected layer.【Result】Compared to the MobileNetV3 and ResNet1D models， the EfficientNet-CBAM-Transformer model demonstrated a superior predictive performance； on the validation set， the model achieved a mean absolute error （MAE） of 1.06 bpm， a mean absolute percentage error （MAPE） of 1.33 bpm， a root mean square error （RMSE） of 5.5%， and a coefficient of determination （R²） of 0.91. Compared with the original EfficientNetV1， the MAE， RMSE， and MAPE were reduced by 31.17%， 28.49%， and 29.49% respectively， while the R² increased by 7.06%. By utilizing automated searches to uniformly scale network depth， width， and input resolution， the EfficientNet-CBAM-Transformer model attained enhanced feature representation capabilities under the same computational budget， and the prediction results of EfficientNet-CBAM-Transformer were closer to true values， with the R² of 0.86 in the fitted regression equation. The numbers of sample groups with true or predicted values of MAE greater than 1.00 bpm was reduced to six groups， showing a decrease of 50% compared to the baseline EfficientNetV1 model， thus indicating a better detection accuracy of the improved model.【Conclusion】The EfficientNet-CBAM-Transformer model based on radar echo data provides the best regression prediction， whose detection accuracy meets the requirements for hog respiratory rate detection in actual production， thereby offering an efficient， non-invasive， and intelligent technology for real-time monitoring of hog health and technical support for future development of portable devices and online intelligent detection systems.