A method for classifying mixing devices using deep neural networks with an expanded receptive field

The paper presents the results of research aimed at developing a method and software tools for identifying the class of a mixing device by its resistance coefficient through experimental data processing. Currently, the main methods for studying mixing devices are finite element methods, as well as procedures of estimating turbulent transfer parameters using laser dopplerometry and chemical methods of sample analysis. These methods require expensive equipment and provide results only for certain types of equipment. This makes it difficult to extend the inferences to a wider class of devices with different designs of mixing impellers. The proposed method involves processing the results of an experiment in which a point light source forming a beam directed vertically upwards is located at the bottom of a container filled with a transparent liquid. A mixing device with variable rotation frequency is placed in the container. When performing experiments in real conditions, small deviations in the size and location of the mixing device lead to difficult-to-predict fluctuations of the funnel surface. Therefore, the image of one marker describes a trajectory that is difficult to predict. It, under certain conditions, can intersect with the trajectories of other markers or be interrupted at the moment when the marker is closed by a stirrer blade passing over it. The resulting image of the markers is associated with a change in the rotational speed of the blade by a rather complex relationship. To identify this dependence, it is proposed to use deep neural networks operating in parallel in two channels. Each channel analyzes the video signal from the surface of the stirred liquid and the time sequence characterizing the change in the speed of rotation of the blades of the device. It is proposed to use neural networks of various architectures in the channels - a convolutional neural network in one channel and a recurrent one in another. The results of the operation of each data processing channel are aggregated according to the majority rule. The computational novelty of the proposed algorithm lies in the expansion of the receptive field for each of the networks due to the mutual conversion of images and time sequences. As a result, each of the networks is trained on a larger amount of data in order to identify hidden regularities. The effectiveness of the method is confirmed by testing it with the use of a software application developed in the MatLab environment.