On-Line Detection of Mastitis in Dairy Herds Using Artificial Neural Networks

Mastitis, one of the most significant diseases in dairy herds, is a highly complex sequence of events with various biological causes and associated physiological and behavioral effects that occur as bacterial infection progresses. The aim of the research is to develop a model for on-line detection of mastitis for robotic milking stations. The data include milking data collected over a four month time period by robots on four farms and monthly test-day milk data collected by veterinarian for determination of the incidence of clinical mastitis (i.e. healthy cows and sick cows). A major part of work involved data pre-processing that plays an important role in this model development. Since the milking machine operated continually and it failed from time to time, an approach for normalizing the variables using running means for herd and individual cows over their own history of milkings was used. The effect of biological differences on milk yield between cows was dealt with using relative differences among quarters instead of absolute values. The final selected variables were normalized peak electrical conductivity (EcMax), normalized quarter yield fraction (QYF) and maximum relative deviation of EcMax values among four quarters for a cow (EcDV). Parallel visualisation of data for each farm and combined datasets show that most of sick cows have high EcMax, high EcDV but low QYF, which was confirmed by correlation analysis. After data pre-processing, a linear discriminent function classifier was developed producing 81% accuracy for mastitic cases and 100% for healthy cases. In order to improve results, the dataset was further used for constructing artificial neural network (ANN) models. There were only 32 records for sick cows and 1026 for healthy cows corresponding to monthly veterinary diagnostic test. Due to the low number of mastitic cases, five random datasets were generated from original data each having most of the sick cows and a randomly selected portion of healthy cows. Five NNs (5 experts) were developed from these datasets and validated. All of these had 94% classification rate for mastitic cases and 100% for healthy ones. In order to further improve classification rate of sick cows, three self-organising maps were trained with three-fold cross validation. The three maps produced 95% average classification rate for mastitic cows and 97% for healthy cows. In the next stage of research, the results will be used to make a final diagnosis with an associated probability of uncertainty.