The capture of a target spacecraft by a chaser is an on-orbit docking operation that requires an accurate, reliable, and robust object recognition algorithm. Vision-based guided spacecraft relative motion during close-proximity maneuvers has been consecutively applied using dynamic modeling as a spacecraft on-orbit service system. This research constructs a vision-based pose estimation model that performs image processing via a deep convolutional neural network. The pose estimation model was constructed by repurposing a modified pretrained GoogLeNet model with the available Unreal Engine 4 rendered dataset of the Soyuz spacecraft. In the implementation, the convolutional neural network learns from the data samples to create correlations between the images and the spacecraft’s six degrees-of-freedom parameters. The experiment has compared an exponential-based loss function and a weighted Euclidean-based loss function. Using the weighted Euclidean-based loss function, the implemented pose estimation model achieved moderately high performance with a position accuracy of 92.53 percent and an error of 1.2 m. The in-attitude prediction accuracy can reach 87.93 percent, and the errors in the three Euler angles do not exceed 7.6 degrees. This research can contribute to spacecraft detection and tracking problems. Although the finished vision-based model is specific to the environment of synthetic dataset, the model could be trained further to address actual docking operations in the future.
In one, docking is defined as “when one incoming spacecraft rendezvous with another spacecraft and flies a controlled collision trajectory in such a manner to align and mesh the interface mechanisms”, and defined docking as an on-orbital service to connect two free-flying man-made space objects. The service should be supported by an accurate, reliable, and robust positioning and orientation (pose) estimation system. Therefore, pose estimation is an essential process in an on-orbit spacecraft docking operation. The position estimation can be obtained by the most well-known cooperative measurement, a Global Positioning System (GPS), while the spacecraft attitude can be measured by an installed Inertial Measurement Unit (IMU). However, these methods are not applicable to non-cooperative targets. Many studies and missions have been performed by focusing on mutually cooperative satellites. However, the demand for non-cooperative satellites may increase in the future. Therefore, determining the attitude of non-cooperative spacecrafts is a challenging technological research problem that can improve spacecraft docking operations. One traditional method, which is based on spacecraft control principles, is to estimate the position and attitude of a spacecraft using the equations of motion, which are a function of time. However, the prediction using a spacecraft equation of motion needs support from the sensor fusion to achieve the highest accuracy of the state estimation algorithm. For non-cooperative spacecraft, a vision-based pose estimator is currently developing for space application with a faster and more powerful computational resource.
คณะวิศวกรรมศาสตร์
This research suggested natural hemp fiber-reinforced ropes (FRR) polymer usage to reinforce recycled aggregate square concrete columns that contain fired-clay solid brick aggregates in order to reduce the high costs associated with synthetic fiber-reinforced polymers (FRPs). A total of 24 square columns of concrete were fabricated to conduct this study. The samples were tested under a monotonic axial compression load. The variables of interest were the strength of unconfined concrete and the number of FRRlayers. According to the results, the strengthened specimens demonstrated an increased compressive strength and ductility. Notably, the specimens with the smallest unconfined strength demonstrated the largest improvement in compressive strength and ductility. Particularly, the compressive strength and strain were enhanced by up to 181% and 564%, respectively. In order to predict the ultimate confined compressive stress and strain, this study investigated a number of analytical stress–strain models. A comparison of experimental and theoretical findings deduced that only a limited number of strength models resulted in close predictions, whereas an even larger scatter was observed for strain prediction. Machine learning was employed by using neural networks to predict the compressive strength. A dataset comprising 142 specimens strengthened with hemp FRP was extracted from the literature. The neural network was trained on the extracted dataset, and its performance was evaluated for the experimental results of this study, which demonstrated a close agreement.
คณะเทคโนโลยีสารสนเทศ
Facial Expression Recognition (FER) has attracted considerable attention in fields such as healthcare, customer service, and behavior analysis. However, challenges remain in developing a robust system capable of adapting to various environments and dynamic situations. In this study, the researchers introduced an Ensemble Learning approach to merge outputs from multiple models trained in specific conditions, allowing the system to retain old information while efficiently learning new data. This technique is advantageous in terms of training time and resource usage, as it reduces the need to retrain a new model entirely when faced with new conditions. Instead, new specialized models can be added to the Ensemble system with minimal resource requirements. The study explores two main approaches to Ensemble Learning: averaging outputs from dedicated models trained under specific scenarios and using Mixture of Experts (MoE), a technique that combines multiple models each specialized in different situations. Experimental results showed that Mixture of Experts (MoE) performs more effectively than the Averaging Ensemble method for emotion classification in all scenarios. The MoE system achieved an average accuracy of 84.41% on the CK+ dataset, 54.20% on Oulu-CASIA, and 61.66% on RAVDESS, surpassing the 71.64%, 44.99%, and 57.60% achieved by Averaging Ensemble in these datasets, respectively. These results demonstrate MoE’s ability to accurately select the model specialized for each specific scenario, enhancing the system’s capacity to handle more complex environments.
คณะวิศวกรรมศาสตร์
Artificial intelligence for agriculture and environment is a collection of significant models for enviromental friendly Thailand development. The models create with machine learning and deep learning by Near infrared spectroscopy research center for agricultural and food products, including: Determining the nutrient needs (N P K) of durian trees by measuring durian leaves using a non-destructive technique using artificial intelligence, Identification of combustion properties of biomass from fast-growing trees and agricultural residues using non-destructive techniques combined with artificial intelligence, and Evaluation of global warming due to biomass combustion using non-destructive techniques using artificial intelligence. The basic technology used is Near infrared Fourier transform spectroscopy technology which measurement and output display can be done quickly without chemical, no requirement for special expert, and measurement price per sample is very low. But the instrument cannot be produced in Thailand.