Accurate lithium-ion battery state estimation is crucial for dynamic applications, such as Electric Vehicles (EVs). Artificial Neural Networks (ANNs) are widely explored in recent years due to its high adaptability for any non-linear system. However, ANNs take longer to train with large datasets, especially with complex architectures. This can be solved by using optimization algorithms prior to ANN training that sets the hyper-parameters in the optimal region, leading to high convergence rate, which is desired for EV based battery state estimators. This paper studies some of the recently used population-based optimization algorithms. This includes Firefly, Bat, Gravitation Search, Particle-Swarm, Antlion and Backtracking Search algorithms along with their variants. The performance of each of these algorithms are compared, in terms of optimizing the weights and biases for a Feedforward Neural Network used for lithium-ion battery State of Charge (SoC) estimation. The obtained results show that Antlion optimization Algorithm with Double-Step Search (ALO-DSS) produced low estimation error (0.0194 MSE), with less variation in the performance when run multiple times and a shorter computation time (343.6s) as compared to Firefly Algorithm (605.3s). Thus, ALO-DSS may be considered viable for optimization of ANNs for EV battery state estimation, especially for SoC estimation. Moreover, using optimization algorithms can save training time and computational costs, leading to higher feasibility in deploying ANNs for EV applications.
This research addresses the economic and ecological losses caused by diseases that damage potato leaves, such as Late Blight, Early Blight, Septoria leaf spots, curly leaves, and bacterial wilt. The study utilizes machine learning and deep learning techniques to swiftly and accurately identify these diseases at an early stage, reducing damage and losses to farmers. The research focuses on four categorization groups, including three leaf illnesses and one healthy leaf, with the main goal of providing early detection of disease. The study employs three deep learning models, VGGNet16, RenNet101, and modified AlexNet, with modified AlexNet proving to be the most accurate, achieving a training accuracy of 99.97% and a testing accuracy of 61%.
Edge computing (EC) is an act of bringing computational and storage capability near data sources. It helps to reduce response times and bandwidth requirements. However, the rapid proliferation of edge devices has expanded the attack surface and opportunity for adversaries to penetrate corporate networks. The limited computational abilities of edge devices and the heterogeneous nature of communication protocols further increase the security challenges of EC. Also, the trustworthiness of hardware devices is challenged due to security and privacy threats like trojan insertion, IP cloning, and hardware counterfeits. The application of Machine Language (ML) models in the edge computing paradigm creates a distributed intelligence architecture. Also, Field Programmable Gate Arrays (FPGAs) can exploit Physical Unclonable Functions (PUFs) characteristics to generate and store authentication keys. The PUF structure deployed with ML models in the edge layer can learn its complex input-output mapping from the Challenge and Response pairs (CRPs) to identify the suspicious and unknown responses. This article discusses the security and privacy issues in various layers of the EC architecture and proposes intrusion detection systems through the integration of FPGA-based edge sever and ML models. A PUF-assisted ML framework of the intrusion detection system is proposed to authenticate and detect potential attacks on the network.
The capacity of solar energy worldwide has grown significantly, from 40.277 to 580.159 MW over the last 9 years. The operation of solar panels is prone to defects due to changes in weather or the environment. Different types of defects can occur depending on the phase of the module, such as infant, midlife, or wear-out failure. Several types of defect images can be used to identify photovoltaic panel (PV) defects such as RGB, thermography, and Electroluminescence (EL) images. Recently, many researchers used EL images for PV defect classification due to their availability to create high-contrast patterns on PV modules, making it easier to detect defects, such as cracks or broken cells, that might be difficult to spot with the naked eye. Therefore, this work aims to analyze and classify PV defects in EL images using ShuffleNet, a Convolution Neural Network (CNN) based architecture. For comparison purposes, other CNN architectures, namely MobileNet and SqueezeNet will be implemented. The results show that the ShuffleNet architecture outperforms MobileNet and SqueezeNet architectures in terms of precision (92.53%), recall (92.24%), and F1-score (93.17%) in classifying EL PV module defect images.
Artificial Intelligence (AI) and computer vision have provided various ways to solve problems in our daily life. In this paper, a web-based safety eyewear detection system to detect the presence of safety eyewear in input images or video streams is developed using OpenCV, TensorFlow/Keras, and deep learning. This detection system is necessary to be deployed at risky workplaces such as construction sites to help reduce the risks of accidents and facilitate supervisors to detect workers who do not adhere to the regulations of wearing safety eyewear before entering a construction site. This paper uses a combination of transfer learning techniques using a pre-trained MobileNet architecture and Single Shot Detection framework to build a fast and efficient deep learning-based method for safety eyewear detection. With the help of Streamlit, the model is deployed into a web application to provide a user-friendly interface for the users. This web application can detect faces instantly by applying the safety eyewear classifier efficiently and quickly. Experimental results on the dataset collected demonstrated the superior performance of the proposed model with 98% accuracy.
