In this paper, a sequential deep neural network in conjunction with a joint time-frequency domain data representation is explored for the problem of cattle behaviour classification. The experimental evaluation is based on a real-world dataset with over 3 million samples, collected from sensors with a tri-axial accelerometer, magnetometer and gyroscope, attached to the collar tags of 10 beef steers. The experimental result demonstrate that the time–frequency domain data representation allows to efficiently trade-off a large reduction of model size and computational complexity for a very minor reduction in classification accuracy. This shows the potential of this classification approach to run on resource-constrained embedded and IoT devices. Most importantly, the proposed behaviour classification method achieves a high classification performance with an $F_1$ Score of 94.9% for 3 behaviour classes, and 89.3% for 9 behaviour classes. This is in comparison to the current state-of-the-art with an $F_1$ Score of 94.3% (for two classes) and 88.7% (for 8 classes).

在本文中，针对牛的行为分类问题，探索了结合时频域联合数据表示的顺序深度神经网络。实验评估基于包含超过 300 万个样本的真实世界数据集，这些样本是从带有三轴加速度计、磁力计和陀螺仪的传感器收集的，并附在 10 只牛肉的领标签上。实验结果表明，时频域数据表示允许有效地权衡模型大小和计算复杂度的大幅降低，而分类精度的降低则非常小。这显示了这种分类方法在资源受限的嵌入式和物联网设备上运行的潜力。最重要的是，$F_1$3 个行为类别的得分为 94.9%，9 个行为类别的得分为 89.3%。这与当前最先进的技术相比$F_1$ 得分为 94.3%（两个班级）和 88.7%（8 个班级）。