Considering the difficult maneuverability of the hovercraft and the driving burden of the pilot, an automatic tracking system including an intuitive human-computer interface and an adaptive chattering-free full-order terminal sliding mode (ACFTSM) controller is designed for humans (non-engineers in particular). The human-computer interface is easy to be understood and used. For the design of the controller, four degrees of freedom (DOF) model of hovercraft is established. And based on the four DOF model, the desired velocities are proposed by a TSM manifold to achieve the finite-time trajectory tracking. Then, the desired velocities can be tracked by the designed CFTSM controller which has superior properties, such as fast and finite-time convergence, highly robust to the disturbances, and low sensitivity to the system parameter variations. Moreover, both the chattering and singularity problems of TSM are resolved by the CFTSM method. By the well combining with radial basis function neural networks in this paper, the new method, named ACFTSM, also has an ability to deal with the strong nonlinearity and uncertainty of hovercraft's complex model. The controller guarantees the fast and finite-time convergence of velocities and position tracking errors. And numerical simulations are implemented to demonstrate the robustness and superiority of the designed controller.
19 October 2022 in Measurement Science and Technology
Oxygen content is one of the most critical factors for high-efficiency combustion. Online measurement of oxygen content from flame images is important but still challenging. For construction of an oxygen content prediction model, most current feature extraction methods are not straightforward. Additionally, there are always sufficient data for common operating conditions in practice, while only limited data for other operating conditions. The data collection process for model training is costly and time-consuming. To tackle the problem, this work presents an augmented flame image soft sensor for automated combustion oxygen content prediction. A convolutional neural network (CNN) regression model is designed to predict the oxygen content directly from flame images, without a single feature extraction process. Moreover, a regression generative adversarial network with gradient penalty is proposed to generate flame images with oxygen content labels. It overcomes the imbalanced and insufficient data problem arising in the CNN regression model training. The proposed soft sensor is compared with several common regression methods for oxygen content prediction. Experimental results show that the proposed method can predict the combustion oxygen content with high accuracy from flame images although the original datasets are imbalanced.
This work described in this paper aims to enhance the level of automation of industrial pearl classification through deep learning methods. To better extract the features of different classes and improve classification accuracy, balanced training datasets are usually needed for machine learning methods. However, the pearl datasets obtained in practice are often imbalanced; in particular, the acquisition cost of some classes is high. An enhanced generative adversarial network, named the multiview Wasserstein generative adversarial network (MVWGAN), is proposed for the imbalanced pearl classification problem. For the minority classes in the training datasets, the MVWGAN method can generate high-quality multiview images simultaneously to balance the original imbalanced datasets. The augmented balanced datasets are used to train a multistream convolution neural network (MS-CNN) for pearl classification. The experimental results show that MVWGAN can overcome the imbalanced learning problem and improve the classification performance of MS-CNN effectively. Moreover, feature visualization is implemented to intuitively explain the effectiveness of MVWGAN.