Classification of Range-Doppler Radar Echoes for Condition Monitoring in Industrial Processes: Shallow Learning versus Deep Learning

The process industry (PI) is extremely diverse. In consequence, a plethora of technical processes are applied, which often involve liquids, being processed, stored and transported in chemical as well as biological reactors and tanks. Tailored to such applications, we have previously proposed a novel approach for condition monitoring (CM). Herein, radar systems are used and their functionalities are augmented, which allows for precise and accurate level measurements of liquids, alongside the distinction between different process conditions (PCs). For this purpose, range-Doppler radar measurements are classified based on manually extracted distinctive, highly interpretable features and state-of-the-art techniques from shallow learning (SL). The objective of this contribution is to compare our approach with techniques from deep learning (DL). For data acquisition, typical PCs have been selected: filling, draining, heating, boiling and stirring (with and without foaming). The measurements have been conducted using a frequency-modulated continuous wave (FMCW) Doppler radar system, operating in the V-/W-band. As shown, with deep neural networks (DNNs) and appropriate data preprocessing, the accuracy can be increased significantly, by up to 2% and more, despite only very limited training data.