eXplainable artificial intelligence for non-visual multiclass recognition of EHDA Modes

Electrohydrodynamic atomization (EHDA) is a versatile technology applied to different fields ranging from process industries to materials science and medicine. Depending on the operating conditions, EHDA provides different spray modes, which are mostly recognised either by high-speed imaging or, less frequently, by current measurements. While high-speed imaging is very successful for lab experiments, it may be difficult to apply in field applications with limited optical access to the spray. To this scope, this study specifically uses frequency-domain analysis of emitted electric current signal data to propose an eXplainable Artificial Intelligence (XAI)-based approach for multi-class recognition of EHDA modes, improving the accuracy of electric current-based classification and allowing an online control of the spray performances. To this scope, a new dataset of experimental data for various liquid types with different chemical-physical properties has been built. The dataset is used to tune the XAI-based method through a supervised learning approach. By combining advanced feature engineering and a one-dimensional convolutional neural network (1D-CNN), the proposed approach achieves accurate classification, making possible the identification of dripping, intermittent, cone-jet, and the challenging multi-jet modes, without the need for visual data. The use of post-hoc XAI techniques ensures transparency, confirming that the model bases its decisions on frequency patterns aligned with the physics of the process. The proposed method demonstrates robustness and a certain adaptability, being capable of classifying with appreciable accuracy EHDA modes for liquids with physical properties different from those used for its training, marking a significant advancement in EHDA process control. This innovation lays the foundation for integrating AI-based classification into closed-loop systems for real-time optimization, addressing both academic and industrial challenges in process efficiency and automation. • XAI framework for non-visual EHDA mode recognition. • Frequency-based signal analysis enables multi-class spray regime detection. • 1D-CNN achieves high accuracy on diverse EHDA datasets. • XAI ensures transparency, aligning decisions with process physics. • Robust approach for real-time EHDA control and automation.