Deep learning model for flow regime identification in mixing tanks from pressure signals

Abstract

Identifying the flow regime (laminar, transition, or turbulent) is crucial in stirred tank mixing processes, affecting mixing efficiency and overall performance. Pressure is often the dominant driving force acting on fluid elements, and its fluctuations provide valuable information for classifying different flow regimes. This paper introduces a data-driven method based on a Convolutional Deep Learning (CDL) neural network to identify flow regimes in a baffled stirred tank. The algorithm was trained on time-pressure signals from force sensors placed in the baffles and flow regime predictions derived from the Newtonian power number, serving as training labels. Once trained, direct measurements of Reynolds number parameters, such as temperature-dependent density and fluid viscosity, which can vary significantly in industrial mixing tanks, are not required to determine the flow regime. The dataset was split into two subsets to implement and validate the CDL model, with 57.14% used for training and validation and 42.86% reserved for testing. The results show that the model successfully classifies pressure signals across all three flow regimes. The proposed non-invasive method provides reliable, fast, and low-cost in-line monitoring of tank mixing, enabling precise flow regime identification without measuring the fluid’s viscosity or density.