Improving cattle health by developing novel data fusion and machine learning approaches to Internet of things livestock data

The worldwide demand for meat and animal products is expected to increase by at least 40% in the next 15 years. In the last 20 years numbers of cows have decreased by approximately 25% and herd size has gradually increased from 89 to 170 (AHDB, 2014). The cost of endemic cattle diseases (BVD, mastitis, respiratory disease, Johne's ,Bovine tuberculosis etc.) is above £1Billion/year with UK sales of veterinary medicines for farmed animals are over £290M per annum (CHAWG, 2014). Most intervention to disease on farms is reactive rather than proactive (Ruston et al., 2015). Indicators such as raise in temp, change in activity or behaviour are early indicators of disease and can be measured with various senor technologies. An effective, automated precision monitoring solution would be of huge benefit for the early detection of disease in cattle, however, there are no algorithms for cattle health yet that have high predictive value. Key reasons are :
a) There are diverse systems utilised such as wearable sensors (measuring body temperature, activity, pH, animal movement, locations), non-wearable sensors (measuring ambient temperature and humidity), automatic milking systems collecting production data, weighing platforms, measures within milk, as well as a raft of health, fertility and production data. The heterogeneity of data type, quantity and quality creates serious challenges for data aggregation and management, different data fusion methods need to be explored and developed and evaluated
b) multifactorial diseases, though one can monitored no of variables but there are complex and often correlated patterns, all need to be accounted for effective machine learning (ML) solution. Figure below shows pattern of temperature and activity of cows monitored and features high correlation between these)
c) data representativeness; since most models are trained on only subset of data and in human telemetry domain there are several approaches ensembles, to gather the diversity of the seen data enabling ongoing learning and thus giving good and improved performance (Fischer et al., 2015). None yet been tried for animal health sensing data
The above issues currently greatly limit the chances of achieving a performant ML-based health monitoring solution for the cattle.
Methods: Proposes research will use existing available unique dataset of animal IoT data (see section 5) The data will include 10 farms and over 1000 animals.
We will use various data fusion techniques to combine information from a multi-sensor data array to validate signals and create features (Dong and HE, 2007). We will then use signals pre-processing techniques such as FFT and others developed as part of our current project EL4L (el4l.com). We will then evaluate various machine learning algorithms and ensemble method (such as KNN, neural networks, Random forests etc.) utilizing Microsoft Azure Machine Learning Studio already utilized by the groups. Models will be validated on 'new' data. Key Questions are:
1. What methods are best for data fusion (signal level fusion, feature level fusion or decision level fusion for predicting cattle health (disease event, high somatic cell counts) and production (milk, weight gain) and what are penalties of those (with respect to performance, hardware implementation, software implementation)? Year 1
2. What features are important and have higher predictive value for early prediction of disease i.e single features, fused features ? How early can we predict health event on cattle farm? (related to 1) Year 2
3. Does using ensemble methods (using online and offline machine learning classifier) gives higher predictive value for this use case? (year 2-3) and validation study to test this on new data