Component and Factor Analysis of Pork Farm Odour using Structural Learning with the Forgetting Method

In pig farming, odour measurement and reduction are necessary for a cleaner environment, lower health risks to humans, and higher quality of pig production. There have been many studies on the modelling of pork farm odour by analysing the chemical components in odorous air. It is suggested that the component analysis approach should be extended to factors such as temperature, relative humidity, and airflow speed. However, previous component and factor analysis did not examine which components or factors contribute significantly to a complex odour. A relative contribution analysis of potential odour components and factors to the perception of odour would allow identification of significant odour components and major contributing factors. Odour reduction practice for pork farms could then be directed towards the significant components and factors, thus improving the efficiency in developing odour measurement and reduction technologies. It is generally accepted that neural networks have several advantages over conventional techniques, for instance, their ability to automatically learn the relationship between the inputs and outputs without any previous knowledge of the system being studied, their powerful generalisation ability, and their capability of handling non-linear interactions. However, typical neural network models suffer from the so-called 'black box' problem, i.e. it offers no information about the system other than the input/output relationship. In this paper, existing methods for odour strength prediction were reviewed, and a neural network based multi-component multi-factor odour model was developed. To reveal the relative contribution of the inputs, the neural network was trained using an algorithm called structural learning with forgetting. By applying the structural learning with forgetting based algorithm, unnecessary neural connections faded away and a skeletal network emerged. The resulted skeletal network enabled an analysis of the contribution of components and factors. The effectiveness of the proposed approach was demonstrated by simulation and comparison studies.