Electrospray mode discrimination with current signal using deep convolutional neural network and class activation map

Abstract

The electrospray process has been extensively applied in various fields, including energy, display, sensor, and biomedical engineering owing to its ability to generate of functional micro/nanoparticles. Although the mode of the electrospray process has a significant impact on the quality of micro/nano particles, observing and discriminating the mode of electrospray during the process has not received adequate attention. This study develops a simple automated method to discriminate the mode of the electrospray process based on the current signal using a deep convolutional neural network (CNN) and class activation map (CAM). The solution flow rate and applied voltage are selected as experimental variables, and the electrospray process is classified into three modes: dripping, pulsating, and cone-jet. The current signal through the collector is measured to detect the deposition of electrospray droplets on the collector. The 1D CNN model is trained using frequency data converted from the current data. The model exhibits excellent performance with an accuracy of 96.30%. Adoption of the CAM configuration enables the model to provide a discriminative cue for each mode and elucidate the decision-making process of the CNN model.