Gear transmission is widely used in mechanical equipment. In practice, if the gearbox is damaged, it not only affects the yield rate but also damages other parts of machines; thus, increases the cost and difficulty of maintenance. With the advancement of technology, the concept of unmanned factories has been proposed; an automatic diagnosis system for the health management of gearboxes becomes necessary. In this paper, a compound fault diagnosis system for the gearbox based on convolutional neural network (CNN) is developed. Specifically, three-axis vibration signals measured by accelerometers are used as the input of the one-dimensional CNN; the detection of the existence and type of the fault is directly output. In testing, the model achieved nearly 100% accuracy on the fault samples we captured. Experimental evidence also shows that the frequency-domain data can provide better diagnostic results than the time-domain data due to the stable characteristics in the frequency spectrum. For practical usage, we demonstrated a remote fault diagnosis system through a local area network on an embedded platform. Furthermore, optimization of convolution kernels was also investigated. When moderately reducing the number of convolution kernels, it does not affect the diagnostic accuracy but greatly reduces the training time of the model.

This paper proposes a self-balancing wheeled robot named J4.α equipped with intelligent modules for automatic navigation. A mechanism involving dynamic mass was designed to control the linear motion of the robot. RGB-D camera was adopted to monitor the front environment; a deep convolutional neural network (ConvNet) was adopted to realize a certain level of socially compliant navigation. In combination with the topological localization scheme, J4.α can navigate between nodal locations without environment modification. To facilitate the work of the manipulator on-board, an auto-balancing module was designed to maintain force equilibrium. To further secure spatial reliability, motorized parking stands are also introduced. Prototypes were made and tests were carried out to examine the system dynamics. The experimental result confirmed the effectiveness of the intelligent navigation modules.

A state-of-the-art real-time rain detection and wiper control method is proposed in this article. Currently, commercial models adopt electronic sensors that can only sample the humidity of a small region of the windshield. The existing computer vision methods primarily focus on the detection and counting of raindrops and provide a recall rate of less than 70%. Here we adopted a holistic-view deep learning approach to build a visual classifier that is robust to large varieties of background scenes, illumination, and water forms. Specifically, Deep Residual Network (ResNet) was adopted as the visual classifier that distinguishes between rainy and fair street scenes and controls the wipers accordingly. To verify the practicality of the proposed deep learning framework, we tested the network on various embedded computing systems, including an embedded computing cluster. The results show that the deep learning rain detector outperforms previous state-of-the-art methods with higher rain recall and precision. It was also found that with the help of some graphic computation-enhancing components, commonly available embedded systems in the market can provide comparable performance to personal computers. While using the enhanced embedded systems to build a cluster, a performance superior to PC was witnessed. As the embedded system is cost-effective, small, and lighter, normalized performances for various aspects clearly show the competitive edge of the embedded systems and confirm the practicality of the proposed system. We also release the dataset of 160 k images used for training the visual rain detector.