Observer Kalman filter identification and multiple-model adaptive estimation technique for classifying animal behaviour using wireless sensor networks

Abstract

The identification of a mathematical model capable of describing the behaviour of animals given input such as feed has great potential for behavioural control purposes. Such models will allow to make predictions which are fundamental to any closed loop control such as control of the feeding. This paper investigates the problem of mathematically modelling animal behaviour. An observer Kalman filter identification method was successfully applied to input–output data and two models representing the hypotheses that animals are actively feeding and the hypotheses that animals are inactive were identified. The input and output of each of the identified models were feed dry matter offer and the pitch angle of the neck, respectively. The pitch angle of the neck of the animal was successfully measured and aggregated by a ZigBee-based wireless sensor network. Two fourth-order models describing the dynamics of an animal in the active and inactive behaviour modes showed good performance in terms of prediction error, cross-correlation between the residual and the output as well as cross-correlation between the residual and the input with 99% confidence interval. A multiple-model adaptive estimation approach was applied to determine the likelihood of each of the two models being the correct model for a specific input of dry matter feed. The average classification success rate was 87.2% for the whole experiment.