Abstract:  To address the issue of difficulties in weighing yaks, a dynamic weight measurement algorithm based on GAN and BPNN is developed by integrating IoT and AI technologies. The STM32 microcontroller is used to acquire raw pressure sensor data from 80 yaks while the yaks are walking smoothly. Subsequently, 3 000 pieces of fake data were generated using a GAN network, and regression prediction is performed using a BPNN neural network to achieve dynamic weight measurement of yaks. Under normal walking conditions, the position judgment is carried out by using the pair‑projecting infrared devices, and using these for data collection. Then, collected raw pressure data were submitted to the prediction model for regression prediction. The experimental results show that resulting in an average weighing time of approximately 4 seconds per yak and the average absolute error between the predicted and actual weights of the yaks is 0.92%. Surpassing the best accuracy (±5%) achieved by experienced technicians in weight estimation. The yak dynamic weighing algorithm based on BPNN and GAN can quickly obtain accurate and automated yak body weight data. It meets the actual application requirements, provides technical support for yak automated weighing, and has significance for achieving yak precision farming.

Key words: yak, dynamic weighing, BP neural network, generative adversarial network, predictive model

摘要： 针对牦牛体重称重难的问题，结合物联网和人工智能技术开发一种基于生成对抗网络GAN和反向传播神经网络BPNN的动态体重测量算法。在牦牛平稳行走状态下，利用STM32单片机获取80头牦牛的原始压力传感器数据。利用GAN网络生成3 000条模拟数据，并使用BPNN神经网络进行回归预测，对牦牛体重进行动态测量。在平稳行走状态下，使用对射红外装置进行位置判断，借此进行数据采集工作，并将采集的原始压力数据交由预测模型进行回归预测。试验结果表明，平均每头牛称重时间约为4 s，预测结果与牦牛真实体重的平均绝对误差为0.92%。优于经验丰富的技术人员估重的最佳精度(±5%)，能够满足实际生产需求。试验采用的基于GAN生成对抗网络和BPNN神经网络构建的牦牛动态称重算法能够快速、精确、自动地获取牦牛的体重数据。符合实际应用需求，为牦牛自动化称重提供技术支持，对实现牦牛精准化养殖有着很强的现实意义。

关键词: 牦牛, 动态称重, BP神经网络, 生成对抗网络, 预测模型