This study delves into the difficulties involved in managing heat, within lithium ion (Li ion) batteries for vehicles emphasizing the impact of temperature on battery performance and safety. It stresses the importance of maintaining temperatures to extend battery lifespan and boost efficiency. By employing machine learning (ML) methods to predict battery temperatures and optimize thermal management systems this research aims to enhance battery safety, performance and longevity. The study compares models like Recurrent Neural Networks (RNNs) Multi Layer Perceptrons (MLPs) and Linear Regression to evaluate their effectiveness in forecasting battery temperature under conditions. The results suggest that incorporating ML into battery management systems can result in efficient and safer operations for automotive applications. A detailed analysis of the three ML models-MLP, RNN and Linear Regression-is carried out for predicting battery temperatures. The main objective is to assess and contrast their accuracy in forecasting based on data. Simulation outcomes such as Mean Absolute Error (MAE) Coefficient of Determination (R2) and Root Mean Squared Error (RMSE) offer insights into each models performance across criteria. This study contributes towards advancing methodologies for predicting battery temperatures thereby improving decision-making processes, for energy management systems and related uses.
In this study, we explore the influence of environmental and soil factors on crop suitability using statistical models and machine learning techniques. We employ a Multinomial Logit Model (MLM) and predictive models including Random Forest (RF), Gradient Boosting (GB), and Light Gradient Boosting Machine (LightGBM) to assess their impact on rice and maize production. Our findings indicate that nitrogen, rainfall, and humidity significantly enhance rice yields, whereas temperature and soil pH negatively affect it. For maize, nitrogen is beneficial, while potassium, temperature, rainfall, and soil pH are detrimental. The models also highlight the paramount importance of rainfall and humidity in crop selection, with both factors having substantial importance scores across RF, GB, and LightGBM. This data-driven approach achieves over 99% accuracy in crop classification via RF, suggesting a robust framework for agricultural decision-making that can significantly improve crop productivity and sustainability.
The rapid advancements in robotics and artificial intelligence have paved the way for the development of autonomous systems with remarkable capabilities. This research presents the design and development of an autonomous omnidirectional mobile robot. Our primary objective is to create a versatile and agile robot capable of navigating complex environments efficiently and autonomously. The robot's design features a mechanical base equipped with four omnidirectional Mecanum wheels, four DC motors, a motor driver, a single-board computer-based control system, an artificial intelligence visual sensing module, an infrared tracking module, an ultrasonic sensor, and a rechargeable solar-powered battery. The Mecanum drive system utilizes wheels with rollers obliquely attached to their circumference, positioned at 45 degrees to the wheel plane and 45 degrees to the axle line. This unique configuration enables the robot to move seamlessly in various directions, including forward, backward, sideways, diagonally, and rotation. Moreover, the robot incorporates a fully automated autonomous driving system capable of environment sensing, obstacle detection, and independent operation without human intervention. This project not only contributes to the development of advanced robotics but also serves as a valuable educational tool for students. It provides practical experience in mechatronics, artificial intelligence, and smart robotic systems, thereby enhancing their hands-on skills and preparing them for the dynamic engineering landscape.
Although Melanoma has been classified as the deadliest kind of skin cancer, with an early prognosis, the chance of treatment goes up. To identify and detect melanoma in digital and dermoscopic images, a deep learning-based model that uses Inception-v3 architecture has been developed in this study. The Hebbian principle and the multi-scale processing method were both used in the architectural design of Inception-v3 to guarantee excellent optimization. This system employs parallel computation throughout many GPUs to use RMSprop as an optimizer. Network weights were fine-tuned throughout the training phase of the model, which feeds mistakes from each iteration back into the Inception-v3 network via the backpropagation approach. Upon finishing the training phase, the model will use the lesion image as input to the pipeline to predict a mole's diagnosis. The PH2 dataset contains 200 dermoscopic images, from which the model achieves accuracy, specificity, and sensitivity of 88.55%, 86.94%, and 95.00%, respectively. Additionally, the model's accuracy of 83.49% was assessed using a digital dataset of 170 photos from UMCG dataset. In both cases, the model is considered as a binary classifier and evaluated using a five-fold cross-validation approach.
The development of modern automotive powertrain control software is subject to various requirements ranging from safety, comfort, emission reduction, performance to efficiency. Consequently, the software engineering process is both complex and time consuming. To reduce these efforts, we explore the usage of reinforcement learning within the powertrain control software engineering process. We aim to automate parts of the software engineering process, especially the parameterization of embedded software. We will introduce the overall concept and current state of our evaluation based on an (initial) use-case and will discuss the associated limits and boundaries.
The internet has become a significant part of our lives, impacting society in various ways. An increasing number of people are using the internet, bringing both positive and negative aspects. Malicious actors continuously seek to exploit vulnerabilities, leak sensitive information, compromise systems, and manipulate data. Some attacks emphasize the importance of robust privacy features in web browsers, which users frequently neglect. This paper closely examines the security levels of popular web browsers, focusing on elements, such as cookies. There is existing literature on browser security, discussing challenges, security improvements, and the concern of fingerprinting. However, few studies focus on cookies, a crucial feature to online privacy for everyday users. This project assessed browser security by examining cookies and performance. Cookies from diverse websites were tested and analyzed across various browsers, highlighting the need for strong defenses against cookies to safeguard user privacy and enhance our comprehension of browser security.
Security is important in the Internet of Things (IoT) due to the widespread adoption of IoT devices in home and enterprise networks. Multi-factor authentication (MFA) is an effective approach to prevent unauthorized use of devices. However, many IoT devices do not have input devices such as a keyboard, mouse, or touchscreen, which makes it challenging to enable multi-factor authentication. This paper proposes a novel approach utilizing Audio Data Transfer (ADT) to enable a second channel for MFA on IoT devices. The effectiveness of this dual-channel MFA system, tailored for IoT’s unique constraints, is thoroughly analyzed through various tests, demonstrating its potential to enhance IoT security. The proposed approach requires a user to be proximate to the device, limiting the transmission distance between the user and the device. However, the proposed approach is accessible, cost-effective, and enhances security in IoT.
Wave behavior underlies useful technologies such as active sonar, passive sonar, and echolocation. For instance, wave behavior is used to determine the sonar signature of underwater vehicles for stealth designs. To obtain a sonar signature, the fastest method is to simulate a sonar testbed using physics software. However, these software rely on the finite element method or finite difference method, which are slow due to their dependence on explicit computations. To overcome this issue, researchers have begun developing physics-informed neural networks (PINNs) to learn wave behavior. As PINNs rely on soft computing and act as operators, they can learn from noisy data and make quick predictions. Our PINN differs from other PINNs that are trained on static simulation grids with open boundaries. The goal of our PINN is to predict the reflection of a pressure wave off of a dynamic obstruction in an arbitrary grid, mimicking in part, the behavior of active sonar. Given a random source location, static obstructions, and a dynamic obstruction, the PINN predicted the wave evolution for 500 timesteps in 1.54 seconds and was over 1200 times faster than the finite difference method. The PINN was trained for 180 epochs using supervised and unsupervised learning and reached a mean squared error of 3E-5. Results show that the PINN demonstrates visual and numerical accuracy and avoids learning unwanted artifacts like spurious waves, due to its ability to generalize. For future work, we will add more degrees of freedom to define complex dynamic obstruction shapes like submarines.
This paper assesses the long-term creep effect on a recently designed Microwave Cavity Flow Meter. This sensor, designed for precise flow rate measurement in nuclear reactor coolant cycles, relies on membrane deformation in the cavity for flow sensing. Operating in challenging conditions with high temperatures of similar to 650 degrees C and a constant fluid load of similar to 100 MPa, its long-term performance is impacted by the unpredictable creep effect. Long-term creep deformation and damage exhibit significant variance across multiple parameters. Therefore, a probabilistic creep model would be beneficial for addressing this variability. We propose a probabilistic creep model for recalibration of the microwave cavity flow sensor based on the Monte Carlo method and the Wilshire-Cano-Stewart (WCS) creep model.
The globalization in the semiconductor domain, primarily to reduce the cost of fabrication, has engendered the integrated circuits (ICs) to be compromised. These semiconductor chips are vulnerable to various attacks such as - hardware Trojan (HT) intrusion. These stealthy HTs can compromise the integrity and functionality of the devices by leaking critical information, compromising the performance, or causing a complete device shutdown. Detecting HTs before deploying ICs in critical infrastructure, is essential to safeguard against these threats. Significant research is done in this area; however, a fool proof technique of HT detection is still required. With this background, this paper presents a technique for the efficient detection of hardware Trojans within ICs, leveraging both power side channel analysis and machine learning (ML) algorithms. The framework proposed in this paper measures the variations in the power consumption of a ring oscillator physical unclonable function (ROPUF) and trains the collected dataset to execute the ML algorithms. The ROPUF is programmed on an Artix-7 FPGA mounted on a Nexys 4 Digilent board, implemented with a Trojan. Our approach uses k-nearest neighbors (knn), logistic regression (LR), and decision tree (DT) as the three ML algorithms to detect hardware Trojan, assessed on four parameters - accuracy, precision, recall, and F1 score. We also present confusion matrix of each executed algorithm. The results demonstrate that LR gives the best precision of 100%, while best accuracy and recall is presented by DT at 98%.
This study investigates the application of a neural network to predict first-year student retention in engineering. The proposed algorithm predicts whether a student will return for the second academic year based on various factors, including: enrollment, ACT scores, and GPA. It examines the effects of parameters on the model's performance, such as the cohort, the semesters taken into consideration (fall only vs. fall and spring), and the academic standing of the students utilized for training and testing. The model's overall accuracy ranges from 66.7% to 95.2%, with models trained on students in strong academic standing having the highest accuracy. This variance in accuracy hampers the model's capacity to learn and forecast retention for at-risk students. To address this issue, we suggest future workarounds, such as employing data oversampling techniques or gathering data from additional years. This study adds to the research literature on the application of machine learning to the prediction of student retention. The results can help identify students at risk of not being retained and help them promptly increase their academic success toward graduation.
Background Study: Horizontal UX (User Experience) refers to UX that can be present in more than one view or page of a given app or website. For example, in an e-commerce website, promotional coupons, header, footer which surface at multiple places in a user’s journey, are examples of such horizontal UX modules. Applications that deliver such cross functional horizontal UX widgets (components or modules) are referred to as horizontal web applications [1]. These applications rely heavily on the effective and efficient delivery of JavaScript, Cascading Style Sheet (CSS) resources, besides other media forms like - images, audio, video to ensure fast load and interaction times [2].Problem Statement: While there are established techniques for optimizing JavaScript and CSS resources in web applications, there is a notable gap in the literature regarding the intricacies of their application in the context of horizontal web applications. This lack of reference poses challenges for developers aiming to enhance performance in these specific scenarios.Proposed Model/Technique: This paper visits a set of optimization techniques for these web resources (assets) in the context of large-scale horizontal web applications. However, general practices for optimizing horizontal web applications is beyond the scope of this paper. These methods are derived from extensive experience in developing production systems at eBay Inc., and are designed to address the unique performance considerations of horizontal UX components.Results: The paper evaluates those optimization methods specific to web resources, in the context of horizontal web applications, highlighting their effectiveness in reducing resource sizes and load times. Additionally, it discusses the learnings and trade-offs encountered when implementing these optimizations, providing insights for developers in the field.
A versatile microwave matching layer that can adapt itself to changing scenarios is presented in order to deliver more power to the human body. The matching layer consists of an antireflection coating slab together with a frequency selective surface which can be fine-tuned to work around 2.5 GHz. An analytical approach to design this versatile matching layer has been offered under certain restrictions. Underlying governing mechanisms of impedance matching have been discussed.
There is a dearth of coursework on designing and implementing services that incorporate machine learning models in university curricula despite the growing industry demand. We describe the design of a course titled "ML Production Systems" which covers the implementation, deployment, monitoring, and updating of machine learning models as part of user-facing web services. The course is designed around a semester-long project to implement and deploy a home sale price prediction service. The course is a required course in a Master’s in Machine Learning program and graduate Machine Learning Engineering certificate program. The course was taught in an online synchronous format to twenty-nine students in Fall 2023; fifteen of the students had more than one year of professional experience as software engineers, while the remaining fourteen students were accelerated Master’s students. Analysis of open-ended written student feedback indicated that students in both groups had a positive experience and found the course to be valuable. Assessment of learning outcome achievements with a final exam and project completion rates indicated that nearly all of the students successfully achieved the learning outcomes. No statistically-significant differences in achievement between the working professionals and accelerated Master’s students were detected. We believe that this course and its successful delivery will serve as a blueprint for other faculty who may want to implement similar courses.
Anthropogenic activities release pollutants into the air, which can negatively affect human health and the environment. One such pollutant is nitrogen dioxide (NO 2 ), which can contribute to smog formation, decreased crop growth and yield, and respiratory damage. This study aimed to find a relationship between land use/land cover (LULC) classifications and NO 2 levels in the air. We used Google Earth Engine (GEE) to collect LULC and air quality data using the Google Dynamic World and the Sentinel-5P NRTI NO 2 datasets. We focused on Pasadena, California, as it provided a good demonstration of an urban area surrounded by greenery, allowing for an adequate analysis of both forms of landscape and their impact on air quality. Random forest (RF) and decision tree (DT) classifiers were used on the provided datasets, with the estimated probability of complete coverage for each LULC type being the input features and the NO 2 density being the output label, measured in mol/m 2 . Our output labels were then discretized, classifying the categories into high and low NO 2 . The machine learning classifier found a correlative relationship between LULC and NO 2 levels, as signified by our modeled accuracy outputting a value of 85%, with an average f1 score of 86%. We performed 10-fold cross-validation to enhance the reliability of model evaluation. The results from this study suggest that machine learning models can be used to predict the changes in air quality based on changes in LULC from anthropogenic activities. With future studies confirming this relationship, inner-city green spaces may benefit mental and physical well-being.
This paper presents a comprehensive comparative performance assessment of Ray Tracing (RT) and Geometry-Based Stochastic Model (GBSM) channel modeling techniques in the context of Reconfigurable Intelligent Surface (RIS)-assisted Vehicle-to-Vehicle (V2V) communications at millimeter-wave (mmWave) frequency. Given the critical importance of accurate channel modeling for the development and optimization of next-generation vehicular networks, this paper explores the effectiveness of RT and GBSM methodologies in capturing the complex electromagnetic interactions facilitated by RIS. In a realistic V2V scenario, the proposed model takes into account environmental effects including reflection, scattering, and transmission through the arbitrary scatterers. To capture the effect of RIS scattering on the dynamic V2V channel characteristics, we have coupled the environmental scattering from the RT approach with the RIS scattered path-loss model. Important channel characteristics against various communication environment parameters, are thoroughly investigated, including the power delay profile (PDP), channel impulse response (CIR), and the frequency correlation function (FCF), while taking into account scenarios with and without RIS. The results highlight the performance comparison between the detailed environmental representation offered by RT and the statistical flexibility of GBSM, as well as the substantial gain in CIR achieved through the implementation of RIS. Our findings aim to guide the selection of appropriate channel modeling techniques for the design and analysis of efficient RIS-assisted V2V communication systems, paving the way for their successful integration into future intelligent transportation networks.
In the smart grid, the prosumers can sell unused electricity back to the power grid, assuming the prosumers own renewable energy sources and storage units. The maximizing of their profits under a dynamic electricity market is a problem that requires intelligent planning. To address this, we propose a framework based on Proximal Policy Optimization (PPO) using recurrent rewards. By using the information about the rewards modeled effectively with PPO to maximize our objective, we were able to get over 30% improvement over the other naive algorithms in accumulating total profits. This shows promise in getting reinforcement learning algorithms to perform tasks required to plan their actions in complex domains like financial markets. We also introduce a novel method for embedding longs based on soliton waves that outperformed normal embedding in our use case with random floating point data augmentation.
Voronoi diagrams are used in a wide range of applications, and many of those applications are in three dimensional space. Two important benchmarks you can measure for Voronoi solver algorithms are run time and memory usage. Run time is important due to the potential costs of computation, and memory usage allows for larger areas to be analyzed. Run time can be addressed via parallelization, but memory usage is dependent on data structure. In this paper we compare the run time and memory usage of a previously published 3D Voronoi solver implementation that utilized an array data structure with a new novel implementation that utilizes an oct-tree data structure.
The Internet-of-Things (IoT) constitute a network of millions of devices interconnected through the internet to facilitate data transfer amongst themselves. This network, along with the devices involved in it, is vulnerable to various cyber threats, due to its high utilization. Consequently, notable research is being conducted to enhance the security of IoT devices and networks. The state-of-the-art, however, still lacks a reliable solution against insider threats in the IoT network. This paper introduces a zero-trust architecture implemented by blockchain and ring oscillator physical unclonable functions (ROPUFs) to bolster the security of IoT devices and networks. The zero trust policies are inspired by the tenets outlined by the National Institute of Standards and Technology. The unique challenge response pairs generated by ROPUFs are utilized to validate the FPGAs. Blockchain features, such as smart contracts, tracking and tracing capabilities, and immutability, are employed for user authentication and access control. To assess the successful implementation of the proposed technique, a case study of an image file transfer between two users is presented. This scenario is effectively simulated using the Ganache framework provided by the Truffle suite. For implementing the ROPUF, Artix 7 Xilinx FPGA mounted on Nexys 4 Digilent board is used. The blockchain smart contract for this investigation is developed using the Solidity language.
This paper presents an intelligent system for monitoring energy usage in smart buildings. It functions as a bridge between the smart grid and smart buildings. A web-based platform is designed to monitor sensor data and track power consumption on the grid. AI technology provides tips for energy conservation based on factors such as power usage patterns and weather conditions. The adaptability of our approach allows for the tuning of automated protocols based on evolving power usage patterns and grid conditions. Data is transmitted between the client devices and the central server via a Zigbee network organized in a star formation, allowing for mutual communication. Our implementation results in the improvement of the overall energy efficiency of supported infrastructure while ensuring continuous and optimized operation in case of a bulk grid blackout.
