Driver abnormal behavior detection enabled self-powered magnetic suspension hybrid wristband and AI for smart transportation

Abstract

With the development of intelligent transportation, a green, light, and comfortable behavior detection method that can protect driver privacy needs to be developed. This paper presents a self-powered behavior detection system based on a magnetic suspension hybrid wristband (MS-HW) and multi-scale convolutional channel attention residual network. The system consists of three modules: magnetic suspension electromagnetic generator module (MS-EMG), magnetic suspension triboelectric nanogenerator module (MS-TENG), and algorithm module (MCRnet). The whole wristband is a magnetic suspension double-layer tubular structure. Magnets and PTFE discs are attached as a moving stack, copper rings are evenly arranged on the outside of the inner tube, and coil groups are wound around the outer tube. During the reciprocating movement of the inner tube, the magnetic flux change of the coil generates electrical energy, and the charge transfer of the copper ring generates the triboelectric signal. Comsol simulation is carried out to optimize the configuration of the system. Then, we simulated a driving environment and collected the activity signals of 15 people. According to the characteristics of different action durations, many signal sampling points, and few channels, we propose a multi-scale convolutional channel attention residual network. Res Multiscale blocks in the network have multiple scaled convolutional kernels to collect signal features, satisfying different action durations. In the network, feature points continue to decrease, and the number of channels continues to increase. The efficient channel attention module (ECAblock) redistributes the weight of channels to further strengthen feature extraction. The stability of the whole network is guaranteed by the residual structure. Finally, the vibration table and network performance experiments are carried out to evaluate the power generation and sensing performance of the system. The output power reaches 0.39 mW, and the recognition rate of the network can reach 97.53 % on the test set.