Smart factory maintenance: building a predictive model of pathogen contamination duringthe food fabrication process

• Sébastien Delmotte ^[1] ; Alexandre Feugier ^[2] ; Jean-Alain Brossier ^[2] ; Florent Arrignon ^[1]
  1. [1] MAD-Environnement, Nailloux, France
  2. [2] Royal Canin, Aimargues, France
• Localización: Digital Maintenance in the Digital Twin Era: Proceedings of the 64th ESReDA Seminar& Doctoral Workshop / coord. por Aitor Goti Elordi, Antonio J. Guillén, Juan Chiachío Ruano, Manuel Chiachio Ruano, 2024, ISBN 978-84-1325-228-5, págs. 147-166
• Idioma: inglés
• Enlaces
  • Texto Completo Libro (pdf)
• Resumen

  • One of the major aims of food safety is to prevent pathogens (mostly bacteria, e.g. Salmonella enterica sp.) to transmit to final products during the food fabrication process. Detection of contamination is thus of crucial importance at every stage of the process. The sanitation (maintenance) is both preventive and performed at a given frequency or corrective when a contamination is detected. Optimization of sanitation faces two main issues: (i) the apparentrandom occurrence of contamination due to multiple sources and inherent to complexity biological processes; (ii) the delay between sampling and results of microbiological analysis (1 to 2 days), which implies to remove or to recall a part of the production once a contamination is detected. This can lead to major financial loss. Therefore, prediction of contamination is fundamental to optimize prevention and to reduce the costs of a contamination and is a main driver of the factory management. To this aim, we have firstly analysed the daily data of salmonella contamination from 5 factories in the world over 8 years. For each factory, the sampling points covered all the different functional zones from raw materials storage to palletization of final product. After data cleaning and processing of descriptive statistics, we modelled contamination using survival analysis and transposing basic reliability concepts in calculatingthe Mean Time between Contaminations (MTBC). A Weibull distribution (2 parameters) was retained among several models as the best fit to data. The reliability diagram was obtained and we characterized the functioning durations without contamination at the scales of functional zones and factories. These results were used to: (i) compare the different zones inside factories and to compare factories to each other; (ii) optimize the frequencies of maintenance. Such indicators can be used to assess the maintenance efficiency over time. Finally, we built for one factory an artificial neural network (multilayer perceptron) to predict the occurrence of a contamination in the final product knowing the contaminations in the different zones occurring the previous week or the previous day and we implemented it into a dedicated visualization software displaying the factory plan, the contamination points and the final predictions. Thissoftware can be connected to the real-time measures of contamination. It constates a first step to the digital twin of the factory dedicated to risk management. management.