Detection of dairy cattle Mastitis: Modelling of milking features using deep neural networks

Abstract

Dairy cattle Mastitis is one of the most notable and costly diseases in dairy industry worldwide. The total Mastitis cost to dairy industry in New Zealand is up near $280 million a year; this cost includes drop in milk production, cattle treatment and other costs. This research includes the examination and analysis of data collected from a commercial robotic dairy farm, in order to design and build a computational model that can help efficient and accurate detection of Mastitis in dairy cattle herds. Accurate Mastitis detection helps cut treatment costs, control the disease, retain milk production levels and maintain milk quality grade. In addition to cutting financial costs, efficient detection helps cows by protecting them and relieving pain caused by the disease. Computational models can help achieve these by helping farmers to adopt timely and suitable cattle treatment regime, and by preventing healthy cows from being infected. For this study, robotic data have been collected for 12 months from a barn-based dairy farm at Makikihi in South Canterbury - New Zealand. At data collection time, that farm was the largest dairy farm in the world in terms of the number of milking-robots under one roof (24 milking robots in one barn). The collected dataset contains sensor data of more than 1,900 cows being milked more than 1.1 million times during the time of data collection. Having about 29,000 milking instances fully labelled (healthy/sick), a deep neural network (DNN) was used to build, train and validate a classification model using variable combinations, including variables that have not been studied before. The model has shown the ability to perform detection tasks with a high level of accuracy; with Specificity (Sp) of 99%, and Sensitivity (Se) of 97%. With this high and stable Sp, and the relatively high Se, the proposed model avoids the problem of false positive alerts. By using deep neural networks to build a Mastitis detection model, this study exploits the main characteristic that gives deep learning predominance compared with other techniques - representation learning, which means that the trained models can extract patterns that used to be ignored by other techniques, to present a robust definition of Mastitis, using real-world sensor data, generated by milking robots in a commercial dairy farm, including data for previously unexploited features. The results of this study allow viewing dairy cattle Mastitis detection from a different angle, which brings about a broader understanding of some of the signs and symptoms of Mastitis, leading to better control and management of the disease.