Expertise-based decision support for managing food quality in agri-food companies

Highlights

• Collecting expert formal and informal knowledge through interviews.

• Original core ontology for causal relationships between situations of interest.

• Recommendations of actions based on the known-how to correct defects.

Abstract

In many agri-food companies, food quality is often managed using expertise gained through experience. Overall quality enhancement may come from sharing collective expertise. In this paper, we describe the design and implementation of a complete methodology allowing an expert knowledge base to be created and used to recommend the technical action to take to maintain food quality. We present its functional specifications, defined in cooperation with several industrial partners and technical centres over the course of several projects carried out in recent years. We propose a systematic methodology for collecting the knowledge on a given food process, from the design of a questionnaire to the synthesis of the information from completed questionnaires using a mind map approach. We then propose an original core ontology for structuring knowledge as possible causal relationships between situations of interest. We describe how mind map files generated by mind map tools are automatically imported into a conceptual graph knowledge base, before being validated and finally automatically processed in a graph-based visual tool. A specific end-user interface has been designed to ensure that end-user experts in agri-food companies can use the tool in a convenient way. Finally, our approach is compared with current research.

Introduction

In many agri-food companies, food quality is often managed using expertise gained from experience. For example, cheese-making chains that showcase their terroir are an economically and agriculturally important industry in France, there being around 17,900 milk producers, 1290 farm producers and 432 processing companies. Cheese-making companies with a “geographical indication”, such as the appellation d’origine protégée (AOP) or indication géographique protégée (IGP), market their products by promoting local resources produced in their terroir and communicating their expertise in terms of milk production and processing. Internal evolutions to appellations, especially in terms of turnover and difficulties encountered in the training of operators, greatly weaken the preservation and transmission of this expertise. This kind of problem is not restricted to cheese-making companies that showcase their terroir. In other agri-food companies, production line management in factories depends to a great extent on the operator’s experience. Consequently, overall quality enhancement may come from sharing collective expertise, which includes informal knowledge. Informal knowledge means knowledge that has not been acquired during learning classes, but rather through individual intentional or fortuitous experiences.

In this context, the development of knowledge engineering methods allowing knowledge bases to be exploited opens up new perspectives in terms of the preservation and data management of operational experience, by proposing complex automatic reasoning technics that go well beyond the description of standard processes (Buche et al., 2013a, Aceves Lara et al., 2017).

In this paper, we propose an original and complete methodology, as well as a dedicated software, for collecting formal and informal knowledge from operators and experts, collectively validating this knowledge, and codifying it in a well-founded language based on a core ontology that provides decision support. This decision support system (DSS) helps to control quality¹ and defects² of manufacture by recommending the most relevant technical action to take at the processing process level, with this process made up of several unit operations³. The DSS also allows all the defects and qualities impacted by a given action to be determined. These recommendations are based on formally representing the possible causal relationships linking defects/quality standards to actions by way of explanatory mechanisms.

Another type of application that this system could be put to is for training purposes. For example, it could help a new operator to get an overview of all the operations and get a better understanding of the different kinds of modifications that can be made to control a process (referred to here as levers).

A generic methodological approach for managing the different steps in the DSS design and implementation process has been developed in order to allow it to be used in different food environments (see Fig. 1). The first step involves defining the scope of the study (a processing process and a set of product quality standards or defects of interest) and collecting associated sources of information (technical reports, etc.). In the second step, the processing process is broken down into a set of unit operations and associated controlled parameters.⁴ In the third step, a systematic questionnaire is derived from the description of the process in order to collect expert knowledge on the potential impact that each unit operation may have on the product in terms of defects and quality standards. In the fourth step, expert knowledge is collected through two kinds of interviews: on the one hand, individual interviews, and on the other hand, collective and contradictory ones. Collective interviews are organized in order to resolve potential contradictions detected when pooling the data from individual interviews in order to obtain a consensus. The expert knowledge resulting from these interviews is then represented in the fifth step as a tree structure using mind mapping software. As mind map tools are only equipped with standard scripting mechanisms, our approach in the sixth step involves automatically translating the knowledge from the mind map software into the conceptual graph formalism (Chein and Mugnier, 2009), which allows specific automatic reasoning tasks to be performed.The tool runs on CoGui software, which is a conceptual graph editor that, firstly, permits the terminology, facts, rules and constraints of an application domain in a knowledge base to be managed, and secondly, allows this knowledge base to be queried and reasoned. Finally, the DSS designed in the seventh step is an end-user interface with associated programs based on CoGui API, ensuring that end users of the application can easily use it without knowing anything about conceptual graph formalism. This seven-step workflow is an iterative one, as the processing process and/or the expert knowledge on it may evolve.

The sections in this paper are dedicated to the following topics:

• the functional specifications of the desired system (Section 2),

• the methodology used to collect the expert knowledge on the processing process, and the use of mind mapping to structure knowledge (Section 3),

• the automatic translation of the mind map into the conceptual graph model (Section 4),

• a presentation of the decision support system (Section 5),

• the decision support system validation process (Section 6),

• a comparison with current research (Section 7).