Augmentation and classification of motor imagery electroencephalogram signals for human–robot collaborative disassembly

As an essential step in remanufacturing end-of-life (EoL) products, disassembly is performed to retrieve high-value parts and materials for use in subsequent remanufacturing processes. Human–robot collaboration (HRC) supported by a brain-machine interface (BMI) can provide an intelligent and versatile solution to address the requirements of disassembly. An intuitive control function of a robot enabled by BMI and human brainwaves (e.g. motor imagery electroencephalogram (MI-EEG)) will be useful for the human to remotely guide the robot to conduct disassembly operations under complicated conditions. To realise such a BMI-enabled HRC disassembly system, it is critical to develop an effective classifier of MI-EEG signals. Nevertheless, the major challenges are that it is possible only to acquire limited MI-EEG signals for system training, and there are weak features and a low signal-to-noise ratio in the signals causing the classification accuracy to deteriorate. To tackle these challenges, in this research, two novel models are developed. The novelties of the models are reflected from the following aspects: (i) to overcome the difficulty in collecting a large number of high-quality MI-EEG signals, a deep convolutional generative adversarial network (DCGAN)-based optimisation model, namely, SAN-DCGAN, has been designed to carry out signal augmentation. The SAN-DCGAN model incorporates new improvements, including a redesigned DCGAN, soft thresholding with an attention mechanism component (SA component), and a spectral normalisation mechanism (SN). (ii) A multi-branch and multi-scale convolutional neural network (MM-CNN)-based model is developed to classify the augmented MI-EEG signals to support robot control. Experiments were conducted to validate the models. Experimental results show that with MI-EEG signal augmentation, the average classification accuracy reached 81.52%, which is higher than the results obtained using other existing models. The research was successfully demonstrated in case studies for disassembling bolts, rigid busbars, and flexible cables from electric vehicle batteries.