Abstract: To achieve automatic and accurate identification of diseased chickens in complex intensive farming environments and to overcome the performance bottleneck caused by the susceptibility of single visual or acoustic modalities to noise and visual occlusion, a lightweight audio-visual multimodal fusion model (LMAV-YOLO) was proposed. The LMAV-YOLO model contained two lightweight feature extraction branches for visual and acoustic modalities.On this basis, a cross-modal attention fusion module was designed to adaptively weight and fuse the bimodal information according to feature quality, thereby enhancing the robustness of the model under environmental interference.Experiments on a self-constructed multimodal dataset showed that the LMAV-YOLO model achieved an mAP@0.5 of 92.3% and an identification accuracy of 94.6%, with only 3.67×10⁶ parameters, 7.82×10⁹ floating-point operations, and a model size of 14.7 MB. After INT8 quantization and TensorRT optimization, the inference speed on the Jetson Nano edge device reached 38.2 frames per second.Under strong noise and 40% visual occlusion interference, the accuracy drop of the model was controlled within 5 percentage points, outperforming the single-modal models. The LMAV-YOLO model combined high accuracy, strong robustness, and low computational requirements, and could be effectively deployed on edge devices, providing a feasible technical solution for real-time health monitoring in large-scale poultry farms.

Key words: complex farming environment, lightweight, audio-visual, multimodal fusion model, diseased chicken identification