< Back to previous page

Project

Precision Livestock Farming for sows and weaner pigs

This thesis described new methods of Precision Livestock Farming (PLF) developed to monitor various behaviours in pigs. The general objective of this thesis was to investigate whether technology can aid the monitoring ability of a stockperson and a farmer in individual animals and in groups. Specific focus of this thesis was to consider a type of algorithm which is the most suitable for Precision Livestock Farming applications on sows and weaner pigs. Regarding the PLF-methodology this thesis aimed to evaluate the relationship between complexity of animal behaviour being monitored, complexity of mathematical model developed for the monitoring purpose and potential of practical application of the model in a Precision Livestock Farming system on commercial pig farms. 

 

In this thesis we compared the performance of pairs of modelling techniques: more and less complex for monitoring of pigs behaviours of increasing complexity. Accuracies of classification of the behaviours, as it could be expected, were in general higher when more complex models (non- linear, more variables) were used. We got better classification accuracy of aggressive behaviour when Artificial Neural Networks were used in comparison to Transfer Function model (on same farm), the performance of Hidden Markov Models was better than of logistic regression for nest-building behaviour. Classification of postural behaviour was better with Linear Discriminant Analysis with multiple input variables than with Transfer Function only with one input variable and also more complex Transfer Function model had better performance than simpler linear regression for counting of piglets in a pen during farrowing.

 

Although, the accuracies of more complex models were better than of less complex models, the differences in accuracies between more and less complex models were relatively small. Specifically, the difference in accuracy between Artificial Neural Networks and transfer Function models (on same farm), for aggression detection was 8%, the difference in accuracy between Hidden Markov Models and logistic regression for nest-building classification was 3%, the difference in accuracy between Linear Discriminant Analysis and Transfer Function  model  for postural behaviour classification was 14% and finally the difference in accuracy between Transfer Function and linear regression model  for counting piglets in a farrowing pen was 6%. These comparisons reveal that good results can be obtained with simpler models also when more complex behaviours are modelled in real-time.

 

       Examples of aggression and postural behaviour monitoring with Transfer Function models described in Chapters 6 and 3 of this thesis showed clearly that using simpler models with few input variables gives a possibility to understand and interpret dynamics of those variables in relation to behaviour of interest. Validation of simpler Transfer Function model for aggression detection on dataset collected on second farm proved that these types of models generalise better on independent data. Sensitivity of ANN model on a second farm was much lower than on the first farm (8 and 96%). These results suggest that differences between both farms in terms of properties of animals (e.g. age, weight and size) and environment (e.g. camera placement, pen size) are significant enough to reduce the performance of complex ANN model but not simpler TF models. This suggests that ANN model selected on a dataset from the first farm was unnecessarily complex and overfit the data. Thus, generalized poorly to other data generated by the same underlying process.

 

Results presented in this thesis confirmed that PLF technology can aid the monitoring ability of a stockperson and a farmer in individual animals and in groups. It was possible to monitor a number of animal based variables in real-time with various sensors. Even a single camera hanged above a farrowing pen proved to be a sufficient sensor for this task. Specifically, progress of farrowing in sows was monitored by estimating the number of piglets in a pen with a standard error of 1.72 in the validation set.

 

Automated classification of postural behaviour in sows was developed on the basis of only accelerometer sensor data. The sensor was mounted on an ear of a sow. The developed method is applicable for automated labelling purposes, where there is a need for estimation of sows postures without the cost of time consuming labelling by human. More specifically 3 types of postural behaviours were considered: active, resting in a lateral position and resting in sternal position. The overall classification accuracy of these behaviours was 70% in the cross-validation dataset.

 

The same sensor technology as for the purpose of automated classification of postural behaviour in sows was used to classify nest-building behaviour. Classification made from again only accelerometer data allowed detection of nest-building behaviour with a sensitivity of 87%, a specificity of 85% and an accuracy of 86%. The most important application of the developed technique is farrowing prediction. The data generated by the system enable monitoring of the progress of a sow towards parturition without making frequent visits to the farrowing area.

 

Finally, aggressive behaviour in weaner pigs was monitored with a camera hanged above a pen. It seems to be possible to discriminate between aggressive interactions that led and did not lead to severe biting. Such continuous, real-time measurement of aggressive behaviour of pigs should in the first instance deliver objective and more accurate information that a farmer can use in order to reduce the aggression on his farm. In the second instance it should give a possibility to automatically lower the aggression level when real-time monitoring is implemented in a PLF system with control component.

 

The most important conclusion of the work is that using simple real-time models to monitor complex animal behaviour allows a good result with advantage of interpretation of dynamics of modelled variables. In models with fewer parameters it is possible to estimate these parameters in real-time while in models with many parameters this becomes impossible. It makes models with fewer parameters more flexible in real-time PLF applications in which models have to adapt to time varying behaviours of individual animals in real-time. Another benefit of using such models in PLF applications is that they are less computationally expensive and cheaper to implement in low cost technology. 

Date:22 Feb 2013 →  26 Jun 2018
Keywords:Preision Livestock farming, weaner pigs, sows
Disciplines:Control systems, robotics and automation, Design theories and methods, Mechatronics and robotics, Computer theory, Agriculture, land and farm management, Biotechnology for agriculture, forestry, fisheries and allied sciences, Fisheries sciences
Project type:PhD project