Expert knowledge integration to model complex food processes. Application on the camembert cheese ripening process M. Sicard a , C. Baudrit a , M.N. Leclerc-Perlat a , P.H. Wuillemin b , N. Perrot a,⇑ a INRA, UMR 782 Génie Microbiologique et Alimentaire, AgroParisTech, INRA, 78850 Thiverval-Grignon, France b Laboratoire d’Informatique de Paris VI (UPMC UMR7606), 75016 Paris, France article info Keywords: Complex systems Expert knowledge Cheese ripening Dynamic Bayesian network abstract Modelling the cheese ripening process continues to remain a challenge because this process is a complex system. There is still lack of knowledge to understand the interactions taking place at different level of scale during the process. However, knowledge may be gathered from scientific and operational experts’ skills. Integrating this knowledge with knowledge extracted from experimental databases may allow a better understanding of the whole ripening process. This study presents an approach adapted from cog- nitive science to elicit and formalise experts’ knowledge about the camembert-type cheese ripening pro- cess. Next, the collected data were unified in a mathematical model based on a dynamic Bayesian network. This formalism makes it possible to integrate this heterogeneous data. The established model presents an average adequacy rate of about 85% with experimental data. Ó 2011 Elsevier Ltd. All rights reserved. 1. Introduction The camembert cheese ripening is a complex system featuring numerous interacting variables that are responsible for the physi- cal, chemical, biological and structural changes. Like in many food processes, both automatic monitoring and human operators are necessary to control the cheese ripening process in order to main- tain product quality (Iiyukhin, Haley, & Singh, 2001). The quality depends on environmental factors (relative humidity, temperature, gas concentration in the chamber) and the interactions between inoculated micro-organisms and curd substrates resulting from variations in raw milk quality and cheesemaking conditions. Many studies have been carried out to try to understand this process from a microbial, physicochemical, biochemical or sensory point of view. Authors such as Leclercq-Perlat et al. (2004) studied microflora interactions to evaluate their synergistic effects on sub- strate consumption. Instrumental methods have also been devel- oped to provide an objective method of bio and physicochemical cheese characterisation during the camembert cheese ripening process (Martin-delCampo, Picque, Cosio-Ramirez, & Corrieu, 2007; Peres, Begnaud, & Berdague, 2002). Sensory properties have been studied to find relationships between cheese organoleptic properties and the modifications of microbiological, biochemical and physicochemical parameters (see for example Martin et al. (1999)). More recently, some authors have attempted to model part of the process. Hélias, Mirade, and Corrieu (2007) established a dynamic model of camembert mass loss during the ripening pro- cess, resulting from physical and biological phenomena. However, microflora growth prediction was not integrated into the model and knowledge is still lacking. Large databases are required to understand the numerous variables and their interactions. These databases are generally non-existent because of time limits, finan- cial constraints and scientific and technological obstacles. In facto- ries, the cheesemaker usually controls the ripening process through a limited number of instrumental measurements and empirical sensory perception (Perrot et al., 2004). Nevertheless, they are able to manage the complexity of the process and their assessment and reasoning play a decisive role (Lemoine, 2001). For example, Perrot et al. (2004) successfully built a decision sup- port system based on expert skills to control the ripening of a soft- mould cheese. The impact of the results was limited because microbial kinetics were not embedded. However this knowledge, such as microbial kinetics, may be extracted from the knowledge of cheese scientists working in the field of dairy science. Therefore, collecting expert knowledge is a challenging research technique to better understand food processes. New conceptual frameworks may be proposed to help integrate, synergise and uni- fy different types of fragmented knowledge for dynamic recon- struction. Nevertheless, the difficulty of the approaches is the knowledge acquisition bottleneck (Hoffman, Shadbolt, Burton, & Klein, 1995). It is often a difficult stage and more time consuming than building a working program. We propose concept and tools adapted from cognitive science to answer to this challenge in a first 0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.03.068 ⇑ Corresponding author. Tel.: +33 1 30 81 53 79; fax: +33 1 30 81 55 97. E-mail addresses: add@grignon.inra.fr (M. Sicard), cbaudrit@grignon.inra.fr (C. Baudrit), add@grignon.inra.fr (M.N. Leclerc-Perlat), pierre-henri.wuillemin@lip6.fr (P.H. Wuillemin), nperrot@grignon.inra.fr (N. Perrot). Expert Systems with Applications 38 (2011) 11804–11812 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa