Abstract
The objectives of this paper are to evaluate the detection performance of a previously developed multivariate spatial dynamic linear model (DLM), which aim to predict outbreaks of either diarrhea or pen fouling amongst growing pigs, and to discuss potential post processing strategies for reducing alarms. The model is applied to sensor based water data from a commercial herd of finisher pigs (30–110 kg) and a research facility herd of weaner pigs (7–30 kg). Performance evaluation is conducted by applying a standardized two-sided Cusum, on the forecast errors generated by the spatial model. For each herd, forecast errors are generated at three spatial levels: Pen level, section level, and herd level. Seven model versions express different temporal correlations in the drinking patterns between pens and sections in a herd, and the performances of each spatial level are evaluated for every model version. The alarms generated by the Cusum are categorized as true positive (TP), false positive (FP), true negative (TN), or false negative (FN) based on time windows of three different lengths. In total, 126 combinations of herds, spatial levels, model versions, and time windows are evaluated, and the performance of each combination is reported as the area under the ROC curve (AUC). The highest performances are obtained at herd level given the longest time window and strongest temporal correlation (AUC = 0.98 (weaners) and 0.94 (finishers)). However, the settings most suitable for implementation in commercial herds, are obtained at section level given the medium-length time window and strongest temporal correlation (AUC = 0.86 (weaners) and 0.87 (finishers)). The combination of a spatial DLM and a two-sided tabular Cusum has high potential for prioritizing high-risk alarms as well as for merging alarms from multiple pens within the same section into a reduced number of alarms communicated to the caretaker. Thus, the spatial detection system described here, and in a previous paper, constitute a new and promising approach to sensor based monitoring tools in livestock production.
Highlights
For more than 20 years the development of sensor-based detection models within the field of livestock science has been subject to an increasing scientific focus
The alarms generated by the Cusum are categorized as true positive (TP), false positive (FP), true negative (TN), or false negative (FN) based on time windows of three different lengths
Both a review of clinical mastitis (CM) detection models (Hogeveen et al, 2010), as well as a recent review paper focusing on livestock related sensor-based detection models in scientific literature from 1995 to 2015 (Dominiak and Kristensen, 2017), show that it is exceedingly difficult to achieve detection performances so high that the number of false alarms will be acceptable in a real-life production herd
Summary
For more than 20 years the development of sensor-based detection models within the field of livestock science has been subject to an increasing scientific focus. In future development of sensor-based detection systems, it may be highly relevant to focus on both achieving very high detection performances and implementing different methods for prioritizing, sorting, or categorizing the generated alarms. Such methods have not had primary focus throughout the scientific literature (Dominiak and Kristensen, 2017). Only three methods; Fuzzy logic (de Mol and Woldt, 2001), Naïve Bayesian Network (NBN) (Steeneveld et al, 2010), and Hidden phase-type Markov (Aparna et al, 2014), are described as alarm-reducing methods in peer-reviewed papers from 1995 to 2015
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
