Unmanned aerial vehicles (UAVs) can be used to support data gathering, training of models, and wireless communication by acting as flying Base Stations (BSs). The usage of UAVs for wireless networks is quickly expanding in areas such as tracking as well as surveillance, defense, leading healthcare deliveries, telecommunications, and so on. As Deep Learning (DL)-assisted approaches require sending raw data from model training devices to UAV servers which is problematic because of device privacy concerns and UAVs' limited processing or communication resources. Federated Learning (FL), which is also called distributed deep learning, is presented as a solution, with the core notion of keeping original or raw data where it is created while transferring the only user's localized trained DL models to a centralized organization for aggregation. We will build a Synchronous Federated Learning (SFL) structure for multi-UAVs and also the comparative analysis of Asynchronous Federated Learning (AFL) and SFL.The SFL methodology will take some time to execute, but there will be no data loss or packet loss. AFL, on the other hand, takes less time but results in packet data loss. Simulation findings suggest that our proposed framework and technique improve global rounds and learning accuracy as compared to AFL.
The paper proposes the remote experiment that enables a student to conduct the physical control system experiment through the internet. The remote experiment is designed using MATLAB platform to create the user interface and to analyze the results. A DC motor controlled by PI controller is utilized as the physical system for demonstrating the effectiveness of the proposed system. Interface module with a dsPIC microcontroller is in charge of implementing a control loop and connecting between a user interface and a physical system. The experimental results reveal that there are a little delay times in live video monitoring but it does not affect the closed-loop control system and it can still be run in real time.
This paper proposes a robust combined step-size generalized variable step-size continuous mixed p-norm (CSS-GVSS-CMPN) algorithm to improve performance in an environment with impulsive noise. GVSS-CMPN algorithm is derived from a CMPN cost function and has outstanding performance. The proposed algorithm introduces a combined step-size (CSS) strategy to resolve the trade-off problem that occurs in the GVSS-CMPN algorithm. A variable mixing factor designed by using a sigmoidal activation function is applied to obtain a fast convergence rate and low steady-state error. A variable mixing factor is indirectly updated by using an inner parameter, and the update equation of the variable mixing factor is derived by optimizing an $\ell_{1}$ norm of a priori error. The sigmoidal activation function is modified and applied to the variable mixing factor for practical considerations. Simulation results show that the proposed algorithm has the best performance among compared algorithms in the environment with impulsive noise.
Telemedicine technology is important for rural area or underdeveloped countries. Panoramic dental x-ray image is used to diagnosis impacted tooth. A process to diagnosis impacted tooth is teeth detection. In this paper, we proposed a method for teeth detection by rotated tooth template. This method consisted of template preparation and image correlation computation step. Template of teeth is varied in the range of 0–360 degrees for detection impacted tooth. Correlation ratio between template and test image is computed. Correlation ratio over threshold is defined as teeth. To test the performance of the method, fifty images are used. Based on the results, both normal tooth and impacted tooth can be detected. This method prove that the rotated tooth template has potential to detection impacted tooth.
Two important aspects when training a classifier are being investigated, training time and classifier quality. Two well-known dimensionality reduction techniques are performed on three groups of data to create smaller data sets for classifier training. The first two techniques are the principal component analysis (PCA) and its modification called adaptive PCA $(\text{aPCA})$ . The third one is the linear discriminant analysis (LDA). Three groups of data sets are trained to create classifiers with and without being pre-processed with these feature reduction techniques. Training time and accuracy scores obtained from each classifier are compared, and the experiment results show that sample size and number of features are two important factors to be considered before selecting dimensionality reduction techniques. If the compatible data reduction technique and data set is chosen, the training time can be reduced without compromising the quality of the classifier.
