In-line real-time water activity prediction based on soft sensors: a case study in a pet food industry

Abstract

Water activity is considered an important parameter of quality, which represents the amount of water available for biochemical and chemical reactions, which enables the growth of microorganisms. It also contributes to food sensory characteristics, like texture and flavor. Water activity measurements can be performed by different equipment and methodologies. Using traditional equipment, a sample need to be collected and placed in a closed chamber to attain equilibrium. This procedure does not provide real-time values and may be associated to increased costs due to production out of specifications if the observed water activity is not in the desired range. The goal of this work is to fill the gap to monitor water activity value in-line in order to increase the response time, control process and make any nonconformity in the production readily verified. As a consequence, a reduction of costs is expected. This work proposes a methodology based on soft sensors to predict and control water activity from moisture content values. As identified in the literature, there is a relationship between water activity and moisture. However, it is not simple to characterize it once it is unique for each food. For that, different machine learning (ML) techniques (SVM, LS-SVM, MLP, GPR and LR) are adopted to map and learn this relationship The Statistical Process Control (SPC) methodology was also proposed as tool to monitor the soft sensor accuracy and to indicate when the underlying model may be retrained. The proposed soft sensors were applied to the case of a pet food industry. The results for all ML models were compared in order to guide the selection of which one would be adopted. All models had good performance, but GPR presented the best balance between model accuracy and training time.