Ultrasonic Material Recognition of Surrounding Objects for Robot Based on CNN-LSTM-GRU Method

In extreme environments which are dusty, dark, of direct sunlight, or filled with flammable gases, robots can not use common perception methods such as lidar and camera to obtain external environmental information, and also can not identify the surrounding environment's material information accurately. Therefore, ultrasonic technology is applied to cope with this problem for its adaptability in severe environments. In this article, we propose a non-contact material recognition method for robots based on deep learning models and data augmentation. We present an empirical mode decomposition-white gaussian noise (EMD-WGN) data augmentation procedure. We propose a convolutional neural network-long short-term memory network-gated recurrent unit network (CNN-LSTM-GRU) for various material identification based on ultrasonic echo signals, and the proposed model combines deep feature extraction and sequence's long-term dependency efficiently. The performance of the proposed model is compared with other deep learning models, such as LSTM, CNN-LSTM, LSTM-GRU, etc. The experimental results show that the CNN-LSTM-GRU model is more effective and efficient than others by using the same hardware and dataset. The proposed non-contact material recognition method demonstrates higher accuracy and efficiency in ultrasonic material identification compared with the existing methods.