Elsevier

Biosystems Engineering

Volume 173, September 2018, Pages 124-133
Biosystems Engineering

Special Issue: Engineering Advances in Precision Livestock Farming
Research Paper
Feed-forward and generalised regression neural networks in modelling feeding behaviour of pigs in the grow-finish phase,☆☆

https://doi.org/10.1016/j.biosystemseng.2018.02.005Get rights and content

Highlights

  • Use of artificial neural networks for predicting feeding behaviour is proposed.

  • Neural networks are effective in mapping non-linear inputs and outputs.

  • Feed-forward neural networks were shown to be the most accurate forecasting model.

  • Implementing machine learning could play an important role in management decisions.

Feeding patterns of pigs have been investigated for use in management decisions and identifying sick animals. Development of models to predict feeding behaviour has been limited due to the large number of potential environmental factors involved and complex relationships between them. Artificial neural networks have been proven to be an effective tool for mapping complicated, nonlinear relationships between inputs and outputs. However, they have not been applied to feeding behaviour prediction. In this study, we compared the use of feed-forward (FFNN) and generalised regression neural networks (GRNN) in forecasting feeding behaviour of pigs in the grow-finish phase, using time of day and temperature humidity index as inputs. Models were calibrated on data from 1923 grow-finish pigs collected from 2008 to 2014, and their predictive ability was tested using data from four additional grow-finish groups collected from 2014 to 2016. Results indicated that FFNN trained with the Levenberg–Marquardt (LM) and scaled conjugate gradient (SCG) algorithms were the most accurate forecasting models. In three of the four validation groups, models trained with LM and SCG algorithms exhibited strong performance, with correlations between predicted and observed feeding behaviours ranging from 0.623 to 0.754. Large deviations between predicted and observed behaviours in the fourth validation group were probably the result of an outbreak of pneumonia, which demonstrates the potential for the model to be used in automated detection of disease outbreak and other stress events. This work is the first step in developing a fully automated system for detecting changes in feeding behaviour.

Introduction

Feeding behaviour of grow-finish pigs can be used to inform producers of both health status and stress level. Many parameters have been studied to better understand feeding behaviour of pigs, including feed intake, meal length, meal interval, number of meals, and total time spent eating (Morgan et al., 2000, Nienaber et al., 1990, Nienaber et al., 1991, Quiniou et al., 2000). Most of these measurements have been obtained from feeding systems that allow only one pig to feed at any given time, which is not representative of commercial production where pigs typically feed in a group setting (Brown-Brandl, Rohrer, & Eigenberg, 2013).

Feeding behaviour is dependent on several environmental and genetic factors, including but not limited to temperature, humidity, gender, breed, and time of day. Deviations from normal feeding behaviour may indicate that grow-finish pigs are experiencing a stressful event, such as illness, issues with feed quality, or heat-related stress. Models of feeding behaviour could be used as a management tool to assess stress levels within a population and to identify sick animals.

Several different approaches have been used to analyse and model feeding behaviour of pigs. Linear regression and analysis of variance models have been used extensively (Brown-Brandl et al., 2013, Nienaber et al., 1990, Nienaber et al., 1991, Quiniou et al., 2001). However, application of these methods is limited due to complex, non-linear relationships between multiple input variables (Comrie, 1997). Gaussian models (Morgan et al., 2000), three-process random models (Berdoy, 1993), and logistic models (Tolkamp & Kyriazakis, 1999) have also been applied to predict feeding behaviour. There are two major drawbacks to these types of models. They tend to be very complex, and they require prior knowledge of relationships between input variables, i.e. a predefined functional form for the model.

Artificial neural networks (ANN) have emerged as a powerful tool in applications where complexity of relationships between inputs and outputs makes formulating a comprehensive mathematical model nearly impossible (Hecht-Nielsen, 1989). An ANN is a set of computing systems that imitates learning abilities of neurons in the brain. Artificial neural network models have the ability to handle large amounts of noisy data, without requiring prior information on model form. An additional advantage of ANN models over other statistical methods is that they require less training data (Paola & Schowengerdt, 1995). The ability to learn by example makes a neural network a very flexible and powerful tool.

This study focused on application of ANN models for prediction of feeding behaviour patterns of pigs during the grow-finish phase. Abilities of feed-forward neural networks (FFNN) and generalised regression neural networks (GRNN) to predict feeding behaviour of grow-finish pigs throughout the year, using time of day and temperature humidity index (THI) as inputs, were compared.

Section snippets

Feed-forward neural network (FFNN)

Feed-forward neural networks are one of the most popular ANN models used in engineering applications. The network architecture and learning algorithm of a FFNN can be viewed as a generalisation of the well-known least-mean-square (LMS) algorithm (Haykin, 2007). Figure 1 shows the architecture of a typical FFNN, which includes three layers: an input layer, a hidden layer which is responsible for performing intermediate computations, and an output layer. The input signal propagates through the

Neural network training and parameter optimisation

Temperature humidity index and time of day (time period) were used as inputs in development of several ANN models to predict feeding behaviour of grow-finish pigs. Environmental temperatures are known to affect feeding behaviour of pigs (Quiniou et al., 2000). Ideally, barn temperatures would have been used in the development of our ANN models. However, barn temperature data were only available for part of this study. Thermal conditions inside the barn were approximated by THI due to its strong

Conclusions

Artificial neural network models have become increasingly popular in many different fields due their ability to elucidate complex, non-linear relationships between parameters. The focus of this study was to identify an accurate and efficient neural network model for predicting feeding behaviour of grow-finish pigs based on time of day and THI. Four different ANN models were trained using electronic feeder data from 1923 grow-finish pigs, and their performance was assessed using data from four

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Acknowledgements

The authors would like to thank K. Simmerman for excellent technical assistance and the U.S. Meat Animal Research Center swine crew for outstanding husbandry.

References (29)

  • B.J. Tolkamp et al.

    To split behaviour into bouts, log-transform the intervals

    Animal Behaviour

    (1999)
  • A. Abraham

    Artificial neural networks

  • A. Blum
    (1992)
  • T.M. Brown-Brandl et al.

    Development of a livestock feeding behaviour monitoring system

    Transactions of the American Society of Agricultural Engineers

    (2011)
  • Cited by (0)

    Mention of a trade name, proprietary product, or specified equipment does not constitute a guarantee or warranty by the USDA and does not imply approval to the exclusion of other products that may be suitable.

    ☆☆

    The USDA is an equal opportunity provider and employer.

    View full text