Industrial food quality measurements open new possibilities for improved sustainability in food production processes, by optimizing the use of raw materials and reducing food loss and waste. There are usually many quality attributes and process parameters representing different stages of the production process, from raw materials and ingredients to end products. By combining all relevant data sources, we can develop digital systems for monitoring, optimizations, or decision support.
The road from sensor data to well-functioning systems is however not straightforward. The main challenges lie in preprocessing and combining data from different sensors, developing reliable models that relate quality attributes to controllable production parameters, and finally designing and implementing user-friendly systems based on such models. The aim of this project is to develop data analytical strategies and pipelines addressing all these challenges, based on use cases and data from Norwegian food companies.
The first part of the work addresses the challenge of combining data from different sensors along a continuous production process. A given raw material need time to travel through the process, and sensor data therefore need to be shifted in time to represent the same raw material at different points in the process. Determination of such time delays between sensors may be seen as an engineering task, but in many cases, it is more practical to use data to estimate the time lags. There are many possible data-driven methods available, from classical correlation metrics, via more flexible machine learning methods to advanced optimization frameworks. We have made an overview of a broad span of methods for estimating time lag between sensors, and compare their performance in different scenarios. The comparison is based on simulation studies as well as a real example from processing of poultry. The results will be a scientific paper, including recommendations on which method to use, and freely available MATLAB code for a selection of the best methods.
The second part of the work focus on methods for statistical process monitoring, more specifically on dynamic latent variable-based methods. A literature study has been conducted, and we aim to compare the usability of two different modelling strategies: either incorporating dynamics in the observable variables or in the latent space. The result will be a scientific paper, including evaluation of how the different strategies perform industrially.