计算机科学
微控制器
人工智能
GSM演进的增强数据速率
计算机视觉
边缘检测
边缘计算
模式识别(心理学)
图像处理
人工神经网络
嵌入式系统
边缘设备
随机存取存储器
语音识别
作者
Rolando Ilagan,Ammara Mehmood,Muhammad Asif Zahoor Raja
标识
DOI:10.1109/jiot.2026.3736584
摘要
Precision Livestock Farming on open pasture relies on continuous behaviour monitoring through worn trackers, but this requires compact models that run on IoT devices with battery and memory limits. A search over architectures and inputs found that pairing the body-parallel accelerometer axis with Overall Dynamic Body Acceleration (ODBA) in a Long Short-Term Memory (LSTM) network achieves 76.40% pooled accuracy (80.48% after k = 10 voting) over six behaviour classes under leave-one-animal-out cross-validation. In the 64-unit input sweep, this pairing outperformed a three-axis + ODBA counterpart by around 10 percentage points on raw accelerometer output. However, that advantage does not survive per-channel standardisation or carry to a second, collar-mounted dataset, so we report it as specific to this configuration rather than a general result. Engineered jerk channels and tuning, selected on the evaluation folds, raised accuracy to 81.92% at 23.4 kB (against 75.92% in the same run), or 84.78% without class rebalancing at a reduced walking recall (we recommend the class-balanced variant). The 32-unit reference model occupies 22.4 kB of weights and runs in 140.7 ms on a 64 MHz ARM microcontroller; the smallest variant tested, an 8-unit model, quantises to a complete 1.35 kB INT8 image running in 30.4 ms.
科研通智能强力驱动
Strongly Powered by AbleSci AI