Smart grids have attracted more and more attention due to the characteristic of itself that provide much more interactive solution comparing from the conventional grid, such as enabling demand side management (DSM) and demand response and so on. In smart grid, Wide Area Networks (WAN) is a part of the largest communication network that provides communication to the energy generation domains such as solar panels. Internet-of-Thing (IoT) can be adopted to provide communication in WAN. However, IoT may suffer in term of quality of services as well as energy efficiency because of how far the data must transmit since WAN cover the largest area. Therefore IoT-Edge Computing can be adopted to process the data nearer to the sources of data generation before transmitting into the cloud services. Energy efficiency in IoT is critical especially when smart grid is considered. Specifically, when a fault occurs in a grid, the sensors and the technologies in the areas affected by the fault will be isolated and have the shortage of power supply. In this paper, we proposed power control method to curb the problem state before. To find the best action, deep reinforcement learning (DRL) is adopted. DRL has the ability to learn make decision in a dynamic environment through a combination of deep learning and reinforcement learning techniques. Simulation results show that with the proposed method, the energy efficiency manage to elevate.
Managing operating theatres (OT) is critical in achieving an effective health care delivery system. This includes the Master Surgery Scheduling Problem (MSSP), which assigns surgery groups such as medical specialties to OT time slots. Constructive heuristics for the MSSP have been incorporating greedy approaches to generate the solution. However, greedy methods tend to ignore hidden factors and could lead to solution infeasibility. Hence, motivated by other timetabling problems, this study incorporated the Saturation Degree ordering into a ranking-based constructive heuristic for solving the MSSP to improve solution feasibility. The proposed algorithm was benchmarked against the greedy and random approaches, as well as a regret-based approach from the literature and found that it has reduced repair mechanism usage by 57.43%. Furthermore, the ranking-based algorithm avoided infeasible solutions in all runs, whereas the greedy and regret-based approaches failed to do the same when using a greedy repair mechanism. Overall, this work has proved the efficacy of the Saturation Degree ordering in enhancing solution feasibility.
Multi-stage classification of Alzheimer's disease (AD) refers to classifying the disease into its multiple stages. Aside from a binary classification task that classifies between the normal control (NC) and AD stages only, an additional prodromal stage known as Mild Cognitive Impairment (MCI) is also being classified. MCI is the stage between the healthy subjects known as the NC class, and the patients with heavy symptoms, in the AD class. In other words, MCI subjects only have slight or mild symptoms of Alzheimer's, thus leading it to be a challenge for detection. Classification models usually perform well in binary classification tasks, but not in multistage classification tasks, due to the faint difference in their features. Thus, this research proposes the incorporation of an attention mechanism into the classification model to increase its multi-stage classification performance. The attention mechanism facilitates the classification task by identifying the important features in MRI images so that the model can better differentiate the multiple classes. The MRI data used in this study is obtained from the Open Access Series of Imaging Studies (OASIS) database. The experimental results show that the attention-incorporated model has achieved an improved classification performance as compared to the normal model without attention. The generalizability of the enhanced model is also improved as observed from the training-classification gap results. Hence, the exceptional performance of the attention mechanism positions it as a solution to boost and enhance multi-stage AD classification.
Sun Dried coconut meat, also known as Copra is used mainly on coconut oil extraction and its by-products for livestock use. The grade of copra is set by the Philippine National Standard for Copra and one of the criteria for the grading/classification is dryness of the copra meat as indicated by the meat color. This study investigates the use of Faster Region-Based Convolutional Neural Network with Inception v2 model in detecting and classifying the level of dryness of copra images as either Optimally Dry, Under-Dry or Over-Dried levels of copra meat. The researchers trained the model and tested it using a validation set which yielded an accuracy rate of 90% for Under Dry, 92% for Optimally Dry copra and 81% for Over Dry, with an overall accuracy rate of 87.67%.
With the development of modern industry 4.0, intelligent path planning is an essential research direction of schedule systems for an automated guided vehicle (AGV), which has been widely used in logistics distribution centers of enterprises. This work implements the improved Q-learning algorithm to solve the typical obstacle avoidance problems in path planning. Specifically, the conventional Q-learning algorithm has shortcomings including low operational efficiency and slow learning speed. The improved Q-learning algorithm is successively proposed by adding a learning process based on the original Q-learning algorithm, which enables AGV to find obstacles and target locations within the shortest time, thus improving the efficiency of path planning. Finally, the simulation experiments are carried out in the grid environment with MATLAB. In comparison to the conventional Q-learning algorithm, the improved Q-learning algorithm has faster convergence and higher learning efficiency, improved by 20%.
