Generic Model-based Framework for Predictive Particle Monitoring using Advanced Image Analysis and Deep Learning

Particle processes has gained significant importance in chemical and biochemical engineering in the last two decades. Especially within fermentation, flocculation, crystallization, there has been an increased industrial focus on optimizing and enhancing controllability of these processes. At the same time, both optics and image-analysis algorithms have improved significantly. It is now possible to analyze sample particle populations in real-time, by automatically taking out samples from the production tank to a mono-layer lab-on-a-chip device. Here, microscopy images are taken and analyzed using advanced image analysis. In this work, a deep neural network is used to estimate the birth and growth rates of the given particle process in real-time, using the raw image, the results from the image analysis, and the measured and controlled process variables. When using deep neural networks, there is a greater risk of overfitting [1]. To accommodate this, it has previously been suggested to add random noise to the input data [2]. Here we utilize the prior knowledge on the inherent sampling error from the image analysis, and show how it is possible to reduce the risk of overfitting the neural network model and at the same time account for the measurement uncertainty already during the model generation.