Year of Publication


Degree Name

Master of Science in Electrical Engineering (MSEE)

Document Type

Master's Thesis




Electrical and Computer Engineering

First Advisor

Dr. Kevin D. Donohue


Globally, dairy farming is a $700 billion industry, with more than 9 million dairy cows in the United States alone. Depriving cows of required activities such as sleep has been shown to negatively impact reproductive efficiency, decrease the volume of milk produced, and increase the risk of culling. Overcrowded herds can decrease individual animal health, demanding the need for automatic behavior detection that would provide insight into their state of health.

Using electroencephalography (EEG), electrooculography (EOG), and electromyography (EMG) to characterize the phases of sleep is a technique which has been used for decades. While these techniques are considered the gold standard for determining sleep states, they are not well suited for industrial applications such as monitoring sleep quality of a large herd of cattle. Previous studies have instead explored the viability of using accelerometers to capture motion information that may give insight into the quality of sleep of an individual animal. In these studies, a researcher assigns a true state classification based on a visual observation of the cows behaviors. These behaviors can be ambiguous, resulting in errors in the true state classifications. In this study, > 300 hours of EEG, EMG, and EOG signals were collected and scored. The data was segmented into 30 second windows and assigned one of four vigilance states: wake, drowsing, Rapid Eye Movement (REM) sleep, and non-Rapid Eye Movement (NREM) sleep. These more robust labels coupled with features extracted from synchronously gathered accelerometer data form the basis of the analysis performed.

This thesis evaluates the effectiveness of using accelerometer based features to predict the vigilance state of dairy cows. Three classification algorithms were used: Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Radial Basis Function Support Vector Machines (RBF-SVM). At best, classification between all four vigilance states had an accuracy of 47% ± 2%. While this outperforms a random guess, it falls short of the inter-scorer agreement of 75%. Grouping wake and drowse together and grouping REM and NREM together improved classification accuracy to 76% ± 2%. Leave one cow out cross validation revealed that these algorithms lacked generalizability, with some cows performing much worse overall. Leave one scorer out cross validation also showed a drop in classification accuracy between all four classes, but did not affect other classification strategies.

Additionally, a regressive approach to measuring EEG-EMG signal properties from accelerometer data was investigated. This involved predicting the relative power spectral density of different frequency bands of the EEG signal. This approach offers an alternative to classification for gaining information about brain state, and therefore quality of sleep in an animal. The same accelerometer based features used in classification were used to train a bagged trees regression model. Regression results are summarized by the correlation coefficient R. The Delta, Theta, Alpha, and Beta band R values were all > 0.4, however Gamma was worse at < 0.3. The same forms of cross validation done with classification were done with regression. Both resulted in lower correlations for Delta, Theta, Alpha, and Beta bands, and significantly lower correlation in the Gamma band.

These results indicate that using accelerometer data may be insufficient to accurately predict vigilance states, or infer brainstate spectral distributions. Additional sensors such as heart rate sensors may enhance the feature space, improving classification and regression results. However in the case of classification, the inter-scorer agreement defines the maximum attainable accuracy. Modifications to the methods of data collection which focus on improving inter-scorer agreement may improve the performance of the explored algorithms.

Digital Object Identifier (DOI)

Funding Information

I acknowledge the United States Department of Agriculture - National Institute of Food and Agriculture (USDA-NIFA) Exploratory grant (USDA NIFA Exploratory 558 Grant \# 2015-67030-24295) which funded the study (2016 - 2017) that generated the data studied in the current work.