Automatic Recognition of a Weakly Identified Animal Activity State Based on Data Transformation of 3D Acceleration Sensor

Smartbow ear-attached motion active sensor with a 3d accelerometer is used for animal activity tracking. Such technology is required to understand the welfare, nutrition scheme and management strategies for breeding cattle. The ear-tag with integrated sensor has no fixed location and orientation that leads to necessity to use the orientation independent features by solving a time series classification problem. In this paper we propose an accelerometer data transformation techniques based on Euler angle rotation and signal projection and show their equivalence relative to a reference coordinate system. The main aim is to increase a recognition accuracy for the weakly-identified states or actions. The previous research for the fitting of the calves has demonstrated certain difficulties by recognition of some rare states and actions, e.g. milk intake. The results show that an average area under the ROC- curve of 0.740 is achieved with improvement of 0.252 over classifications without data transformation.