Detection of lameness of cow based on body weight using artificial neural network

Abstract

Present study describes the application of machine learning algorithms to detect the health status of Karan Fries crossbred cow, whether cow is lame or healthy. A total of 589 Karan Fries crossbred cows data from the herd maintained at National Dairy Research Institute (NDRI), Karnal were recorded for the health status (lame or healthy) as target variable and other non-genetic variables such as percent of body weight distributed to individual legs (using load cell platform), parity (1to10), status of pregnancy (non-pregnant, less than 90days, 91-180days and more than 181days pregnant), status of lactation (less than 60days, 61-120days, more than 121days and dry), and daily milk yield were used as input variables. These variables were used for the development of Multilayer Perceptron (MLP) and Radial Basis Function (RBF) neural networks (NNs). Simulation of network were carried out using varying data partition strategies for training and validation data sets (60:40, 70:30 and 80:20), number of nodes in hidden layers and different optimization algorithms. It was found that to predict the heath status of cows RBF neural architect with Leven-burg Marquardt (LM) optimization algorithm, gave the best performance in comparison to MLP with highest classification accuracy rate (83.19%) at 80:20 data partition strategy.