Obstructed Material Classification Using mmWave Radar with Deep Neural Network for Industrial Applications

Abstract

Radar sensing technology uses radio electromagnetic (EM) waves to provide 3D space localisation and 4D motion sensing. The mmWave radar shows advantages in low cost, low power, environment robustness and capability in material classification. In this paper, the capability of mmWave radar to perform industrial multi-material classification with obstruction is studied by measuring the reflected radar signal. The classified materials are common engineering materials which include metal, polymer, ceramic, composite and natural. The experiment is conducted using the IWR1443BOOST mmWave radar sensor. From a series of experiment results, the received radar signal is the unique material signature of a target object. The relative power measured by IWR1443BOOST is correlated to the target object’s relative permeability and permittivity. This indicated the mmWave radar can easily pick up unique material properties as well as the physical structure of target object with minor assistance from deep neural network model. Three models which are linear classifier, fully connected neural network (FCNN) and convolution neural network (CNN) are trained and inference on the radar signal. CNN shows the most robust performance even under noise, while linear classifier converges fastest. All models achieved satisfactory accuracy with minimum amount of training epochs. This is because the radar signals are having clear discriminative distribution as proven in standard deviation against mean plot. The models also perform under 16 mm thick obstruction and can classify less than 5 mm thin material. From the experiment, the mmWave radar provides highly accurate multi-material classification with deep neural network. Due to its’ capability in wall-penetration and environment robustness characteristics, mmWave radar is a new alternative solution for industrial automation and sensing application.