Browse Topic: Machine learning

Items (1,112)
This paper explores the integration of two deep learning models that are currently being used for object detection, specifically Mask R-CNN and YOLOX, for two distinct driving environments: urban cityscapes and highway settings. The hypothesis underlying this work is that different methods of object detection will work best in different driving environments, due to the differences in their unique strengths as well as the key differences in those driving environments. Some of these differences in the driving environment include varying traffic densities, diverse object classes, and differing scene complexities, including specific differences such as the types of signs present, the presence or absence of stoplights, and the limited-access nature of highways as compared to city streets. As part of this work, a scene classifier has also been developed to categorize the driving context into the two categories of highway and urban driving, in order to allow the overall object detection
Patel, KrunalPeters, Diane
The integration of artificial intelligence (AI) and machine learning (ML) into automotive safety-critical systems presents unique challenges, particularly the “changing anything changes everything” (CACE) property inherent in many AI/ML models. CACE highlights the high degree of interdependence within AI/ML systems, where even minor adjustments can have significant, unforeseen impacts on system behavior, posing risks in safety-critical applications. This paper examines the intricate nature of the CACE principle and its implications for the development cycle of AI/ML-based applications. Through case studies and theoretical analysis, we highlight CACE-related challenges and discuss strategies to mitigate these risks in safety-critical environments. Our analysis aims to raise awareness of this often-overlooked challenge, offering insights for safer, more robust AI/ML deployment in the automotive industry.
Tong, WeiLi, GangS, RameshYang, TianbaoShuttlewood, BingMudalige, Pri
Precisely understanding the driving environment and determining the vehicle’s accurate position is crucial for a safe automated maneuver. vehicle following systems that offer higher energy efficiency by precisely following a lead vehicle, the relative position of the ego vehicle to lane center is a key measure to a safe automated speed and steering control. This article presents a novel Enhanced Lane Detection technique with centimeter-level accuracy in estimating the vehicle offset from the lane center using the front-facing camera. Leveraging state-of-the-art computer vision models, the Enhanced Lane Detection technique utilizes YOLOv8 image segmentation, trained on a diverse world driving scenarios dataset, to detect the driving lane. To measure the vehicle lateral offset, our model introduces a novel calibration method using nine reference markers aligned with the vehicle perspective and converts the lane offset from image coordinates to world measurements. This design minimizes
Karuppiah Loganathan, Nirmal RajaPoovalappil, AmanNaber, JeffreyRobinette, DarrellBahramgiri, Mojtaba
This paper introduces an innovative digital solution for the categorization and analysis of fractures in Auto components, leveraging Artificial Intelligence and Machine Learning (AI/ML) technologies. The proposed system automates the fracture analysis process, enhancing speed, reliability, and accessibility for users with varying levels of expertise. The platform enables users to upload images of fractured parts, which are then processed by an AI/ML engine. The engine employs an image classification model to identify the type of fracture and a segmentation model to detect and analyze the direction of the fracture. The segmentation model accurately predicts cracks in the images, providing detailed insights into the direction and progression of the fractures. Additionally, the solution offers an intuitive interface for stakeholders to review past analyses and upload new images for examination. The AI/ML engine further examines the origin of the fracture, its progression pattern, and the
Sahoo, PriyabrataRawat, SudhanshuGarg, VipinNaidu, GarimaSharma, AmitNarula, RahulBindra, RiteshKhera, PankajGoel, PoojaMondal, Arup
The surge in electric vehicle usage has expanded the number of charging stations, intensifying demands on their operation and maintenance. Public charging stations, often exposed to harsh weather and unpredictable human factors, frequently encounter malfunctions requiring prompt attention. Current methods primarily employ data-driven approaches or rely on empirical expertise to establish warning thresholds for fault prediction. While these approaches are generally effective, the artificially fixed thresholds they employ for fault prediction limit adaptability and fall short in sensitivity to special scenarios, timings, locations, and types of faults, as well as in overall intelligence. This paper presents a novel fault prediction model for charging equipment that utilizes adaptive dynamic thresholds to enhance diagnostic accuracy and reliability. By integrating and quantifying Environmental Influence Factors (EF), Scenario Influence Factors (SF), Fault Severity Factors (FF), and
Wang, HaoWang, NingLi, YuanTang, Xinyue
Vehicle ADAS Systems majorly comprises of two functions: Driving and Parking. The most common form of damage to the vehicle which goes unnoticed with unidentified cause are parking damages. A vehicle once parked at a certain location may get damaged without knowledge of the user. In this work developed a solution that not only pre-warns the driver but also prepares the vehicle beforehand if it suspects a damage may occur. This eliminates the latency between damage and information capture, detects small damages such as scratches, classifies the type of damage and informs the user beforehand. This is solution is different from our competitors as the existing solutions informs the user about the scratches/damages, but these solutions are expensive, have high response time, and the damage information is captured after the damage has occurred. The solution consists of the following check blocks: Precondition, Sensor Control and Action Module. The Precondition Module observes the vehicle
Debnath, SarnabPatil, PrasadBelur Subramanya, SheshagiriGovinda, Shiva Prasad
As global warming and environmental problems are becoming more serious, tires are required to achieve a high level of performance trade-offs, such as low rolling resistance, wet braking performance, driving stability, and ride comfort, while minimizing wear, noise, and weight. However, predicting tire wear life, which is influenced by both vehicle and tire characteristics, is technically challenging so practical prediction method has long been awaited. Therefore, we propose an experimental-based tire wear life prediction method using measured tire characteristics and the wear volume formula of polymer materials. This method achieves practical accuracy for use in the early stages of vehicle development without the need for time-consuming and costly real vehicle tests. However, the need for improved quietness and compliance with dust regulations due to vehicle electrification requires more accuracy, leading to an increase in cases requiring judgment through real vehicle tests. To address
Ando, Takashi
The proliferation of intelligent technologies in the future battlefield necessitates an exploration of crew workload balancing strategies for human-machine integrated formations. Many current techniques to measure cognitive workload, through qualitative surveys or wearable sensors, are too brittle for the harsh, austere operational environments found in military settings. Non-invasive workload estimation techniques, such as those that analyze physiological effects from video feeds of the crew, present a way forward for workload-aware Soldier-machine interfaces that could trigger events – such as task reallocation – if limits on crew or individual workload are exceeded. One such technique that is being explored is the use of facial expression analysis for workload estimation. We present the performance results of regression and classification models developed from supervised machine learning algorithms that predict pNN50, a common heart rate variability metric used as a physiological
Mikulski, ChristopherRiegner, Kayla
Reproducing driving scenarios involving near-collisions and collisions in a simulator can be useful in the development and testing of autonomous vehicles, as it provides a safe environment to explore detailed vehicular behavior during these critical events. CARLA, an open-source driving simulator, has been widely used for reproducing driving scenarios. CARLA allows for both manual control and traffic manager control (the module that controls vehicles in autopilot manner in the simulation). However, current versions of CARLA are limited to setting the start and destination points for vehicles that are controlled by traffic manager, and are unable to replay precise waypoint paths that are collected from real-world collision and near-collision scenarios, due to the fact that the collision-free pathfinding modules are built into the system. This paper presents an extension to CARLA’s source code, enabling the replay of exact vehicle trajectories, irrespective of safety implications
Ai, YanAdhikari, BikramPark, Chung-KyuKan, Cing-DaoWijesekera, Duminda
Personalization is a growing topic in the automotive space, where Artificial Intelligence can be used to deliver a customized experience in features like seat positioning and climate control. Considering that the leading cause of accidents is driving at an inappropriate speed, personalizing the speed limit for a driver can greatly improve vehicle safety. Current speed limits apply to all drivers, irrespective of skill, including special speed limits when there are adverse weather conditions. As these speed limits do not consider an individual’s skill and capabilities, the limit could still be inappropriate for a given driver in that specific driving context. Therefore, we propose a system that can profile the driver’s style to recommend a personalized speed limit, based on both the environmental context and their skill in that environment. The system uses a neural network to classify the driver’s behavior in specific environments by monitoring the vehicle data and the environmental
Perumal, RathapriyaChouhan, MadhvendraRangarajan, Rishi
To obtain real-time tire wear status during vehicle operation, this paper proposes a tire wear detection method based on signal analysis. Firstly, PVDF piezoelectric thin film sensors are pasted in the center of the airtight layer of tires with different degrees of wear to collect tire stress data under different working conditions. Secondly, filter and extract the time-domain and frequency-domain feature information of the collected data to construct a feature dataset. Finally, a deep regression model is established to train the feature dataset and achieve real-time detection of tire damage status. The results indicate that the prediction algorithm based on signal analysis and feature extraction achieves a maximum error of 0.3mm in tire wear detection, demonstrating high accuracy in tire wear detection. Providing tire information for safe driving of vehicles has high industrial application value.
Xianyi, XieYang, HaoJin, Lisheng
The mechanical properties of materials play a crucial role in real life. However, methods to measure these properties are usually time-consuming and labour intensive. Small Punch Through (SPT) has non-destructive characteristics and can obtain load-displacement curves of specimens, but it cannot visually extract the mechanical properties of materials. Therefore, we designed a proprietary SPT experiment and fixture, built a finite element method (FEM) model and developed a multi-fidelity model capable of predicting the mechanical properties of steel and aluminium alloys. It makes use of multi-fidelity datasets obtained from SPT and FEM simulation experiments, and this integration allows us to support and optimize the predictive accuracy of the study, thus ensuring a comprehensive and reliable characterization of the mechanical properties of the materials. The model also takes into account variations in material thickness and can effectively predict the mechanical properties of materials
Zou, JieChen, YechaoLi, ShanshanHuayang, Xiang
Image-based machine learning (ML) methods are increasingly transforming the field of materials science, offering powerful tools for automatic analysis of microstructures and failure mechanisms. This paper provides an overview of the latest advancements in ML techniques applied to materials microstructure and failure analysis, with a particular focus on the automatic detection of porosity and oxide defects and microstructure features such as dendritic arms and eutectic phase in aluminum casting. By leveraging image-based data, such as metallographic and fractographic images, ML models can identify patterns that are difficult to detect through conventional methods. The integration of convolutional neural networks (CNNs) and advanced image processing algorithms not only accelerates the analysis process but also improves accuracy by reducing subjectivity in interpretation. Key studies and applications are further reviewed to highlight the benefits, challenges, and future directions of
Akbari, MeysamWang, AndyWang, QiguiYan, Cuifen
Rechargeable lithium batteries are widely used in the electric vehicle industry due to their long lifespan and high energy density. However, after long-term repeated charging and discharging, various electrochemical reactions inside lithium batteries can lead to performance degradation and even cause battery fires. Estimating the health status and predicting the remaining life of lithium batteries can provide insights into their future operating conditions, which is crucial for achieving fault warnings and ensuring the safe operation of battery-related equipment. In terms of predicting the health status of lithium batteries, this paper proposes a method based on an improved Long Short-Term Memory (LSTM) for health status estimation. This method first employs nearest neighbor component analysis to eliminate feature redundancy among the multidimensional health factors of the battery. Then, a differential grey wolf optimization algorithm (DEGWO) is used to globally optimize the
K, Meng Zi
The ISO TR 5469 Technical Report provides a framework to classify the AI/ML technology based on usage level and the properties and requirements to mitigate cyber and functional safety risks for the technology. This paper provides an overview of the approach used by ISO TR 5469 as well as an example of how one of the six ISO TR 5469 desirable properties (resilience to adversarial and intentional malicious input) can be analyzed for adversarial attacks. This paper will also show how a vehicle testbed can be used to provide a student with an AI model that can be used to simulate a non-targeted cyber security attack. The testbed can be used to simulate a poisoning attack where the student can manipulate a training data set to deceive the AI model during a simulated deployment.1 The University of Detroit Mercy (UDM) has developed Cyber-security Labs as a Service (CLaaS) to support teaching students how to understand and mitigate cyber security attacks. The UDM Vehicle Cyber Engineering (VCE
Zachos, MarkSeifert, Heinz
Lane-keeping is critical for SAE Level 3+ autonomous vehicles, requiring rigorous validation and end-to-end interpretability. All recently U.S.-approved level 3 vehicles are equipped with lidar, likely for accelerating active safety. Lidar offers direct distance measurements, allowing rule-based algorithms compared to camera-based methods, which rely on statistical methods for perception. Furthermore, lidar can support a more comprehensive and detailed approach to studying lane-keeping. This paper proposes a module perceiving oncoming vehicle behavior, as part of a larger behavior-tree structure for adaptive lane-keeping using data from a lidar sensor. The complete behavior tree would include road curvature, speed limits, road types (rural, urban, interstate), and the proximity of objects or humans to lane markings. It also accounts for the lane-keeping behavior, type of adjacent and opposing vehicles, lane occlusion, and weather conditions. The algorithm was evaluated using
Soloiu, ValentinMehrzed, ShaenKroeger, LukePierce, KodySutton, TimothyLange, Robin
Road safety and traffic management face significant challenges due to secondary crashes, which frequently cause increased traffic, delays, and collisions. Traditional methods for anticipating secondary crashes often overlook the importance of different road types, resulting in suboptimal predictions and response plans. This research presents a novel method that combines a hybrid machine-learning model with a functional class-based weighting strategy to classify secondary crashes. The functional classes in the dataset are categorized as interstates, arterial roads, collector roads, and local roads. The dataset also includes comprehensive crash narratives and various road attributes. Each functional class is assigned a weight reflecting its proportional importance in the likelihood of a subsequent crash, based on historical data and road usage patterns. This weighting technique is integrated into a hybrid model architecture that trains a Random Forest (RF) model on structured data to
Patil, MayurMarik PE, Stephanie
Precise state estimation during a lateral maneuver is not just a theoretical concept but a practical necessity. The performance of the Kalman filter is directly impacted by the comprehensive research and innovative approaches to counter nonlinearity and uncertainty. The use of machine learning in control theory is one such development that has significantly enhanced the effectiveness of our work. This paper provides an enhanced adaptive Kalman filter architecture with a neural network for a rapid obstacle avoidance maneuver. The proposed design exemplifies not just its effectiveness in terms of better state estimation in the presence of complex nonlinear vehicle dynamics and disturbances but also its potential downsides sometimes. Simulation results verify the same by ensuring a significant improvement to the traditional design, demonstrating better accuracy and the need for such advances in vehicle dynamics and control.
Sudhakhar, Monish Dev
Path tracking control, which is one of the most important foundations of autonomous driving, could help the vehicle to precisely and smoothly follow the preset path by actively adjusting the front wheel steering angle. Although there are a number of advanced control methods with simple structure and reliable robustness that could assist vehicles achieving path tracking, these controllers have many parameters to be calibrated, and there is a lack of guidance documents to help non-professional test site engineers quickly master calibration methods. Therefore, this paper proposes a parameter virtual calibration method based on the deep reinforcement learning, which provides an effective solution for parameter calibration of vehicle path tracking controller. Firstly, the vehicle trajectory tracking model is established through the kinematic relationship between the vehicle and the target path, combined with the Taylor series expansion linearization method. Next, a vehicle path tracking
Zhao, JianGuo, ChenghaoZhao, HuiChaoZhao, YongqiangYu, ZhenZhu, BingChen, Zhicheng
Technology development for enhancing passenger experience has gained attention in the field of autonomous vehicle (AV) development. A new possibility for occupants of AVs is performing productive tasks as they are relieved from the task of driving. However, passengers who execute non-driving-related tasks are more prone to experiencing motion sickness (MS). To understand the factors that cause MS, a tool that can predict the occurrence and intensity of MS can be advantageous. However, there is currently a lack of computational tools that predict passenger's MS state. Furthermore, the lack of real-time physiological data from vehicle occupants limits the types of sensory data that can be used for estimation under realistic implementations. To address this, a computational model was developed to predict the MS score for passengers in real time solely based on the vehicle's dynamic state. The model leverages self-reported MS scores and vehicle dynamics time series data from a previous
Kolachalama, SrikanthSousa Schulman, DanielKerr, BradleyYin, SiyuanWachsman, Michael BenPienkny, Jedidiah Ethan ShapiroJalgaonkar, Nishant M.Awtar, Shorya
In the modern automotive industry, improving fuel efficiency while reducing carbon emissions is a critical challenge. To address this challenge, accurately measuring a vehicle’s road load is essential. The current methodology, widely adopted by national guidelines, follows the coastdown test procedure. However, coastdown tests are highly sensitive to environmental conditions, which can lead to inconsistencies across test runs. Previous studies have mainly focused on the impact of independent variables on coastdown results, with less emphasis on a data-driven approach due to the difficulty of obtaining large volumes of test data in a short period, both in terms of time and cost. This paper presents a road load energy prediction model for vehicles using the XGBoost machine learning technique, demonstrating its ability to predict road load coefficients. The model features 27 factors, including rolling, aerodynamic, inertial resistance, and various atmospheric conditions, gathered from a
Song, HyunseungLee, Dong HyukChung, Hyun
This paper presents a Digital Twin approach based on Machine Learning (ML), aimed at creating software-based sensors to reduce the auxiliary devices of the vehicle and enabling predictive maintenance, thus reducing carbon footprint. The solution is applied to the electric Lubrication Oil Pump (eLOP), a crucial component within a vehicle's powertrain system. The proposed eLOP Digital Twin integrates ML-based sensors to estimate critical parameters such as temperature, pressure and flow rate, reducing the reliance on physical sensors and associated hardware. This approach minimizes manufacturing complexity and cost, enhancing energy efficiency during both production and operation. Furthermore, the Digital Twin facilitates predictive maintenance by continuously monitoring the component's performance, enabling early detection of potential failures and optimizing maintenance schedules. This leads to lower energy consumption and reduced emissions throughout the component's lifecycle. The
Khan, JalalD'Alessandro, StefanoTramaglia, FedericoFauda, Alessandro
Vehicle restraint systems, such as seat belts and airbags, play a crucial role in managing crash energy and protecting occupants during vehicle crashes. Designing an effective restraint system for a diverse population is a complex task. This study demonstrates the practical implementation of state-of-the-art Machine Learning (ML) techniques to optimize vehicle restraint systems and improve occupant safety. An ML-based surrogate model was developed using a small Design of Experiments (DOE) dataset from finite element human body model simulations and was employed to optimize a vehicle restraint system. The performance of the ML-optimized restraint system was compared to the baseline design in a real-world crash scenario. The ML-based optimization showed potential for further enhancement in occupant safety over the baseline design, specifically for small-female occupant. The optimized design reduced the joint injury probability for small female passenger from 0.274 to 0.224 in the US NCAP
Lalwala, MiteshLin, Chin-HsuDesai, MeghaRao, Shishir
Advancements in sensor technologies have led to increased interest in detecting and diagnosing “driver states”—collections of internal driver factors generally associated with negative driving performance, such as alcohol intoxication, cognitive load, stress, and fatigue. This is accomplished using imperfect behavioral and physiological indicators that are associated with those states. An example is the use of elevated heart rate variability, detected by a steering wheel sensor, as an indicator of frustration. Advances in sensor technologies, coupled with improvements in machine learning, have led to an increase in this research. However, a limitation is that it often excludes naturalistic driving environments, which may have conditions that affect detection. For example, reductions in visual scanning are often associated with cognitive load [1]; however, these reductions can also be related to novice driver inexperience [2] and alcohol intoxication [3]. Through our analysis of the
Seaman, SeanZhong, PeihanAngell, LindaDomeyer, JoshuaLenneman, John
Traditional methods for developing and evaluating autonomous driving functions, such as model-in-the-loop (MIL) and hardware-in-the-loop (HIL) simulations, heavily depend on the accuracy of simulated vehicle models and human factors, especially for vulnerable road user safety systems. Continuation of development during public road deployment forces other road users including vulnerable ones to involuntarily participate in the development process, leading to safety risks, inefficiencies, and a decline in public trust. To address these deficiencies, the Vehicle-in-Virtual-Environment (VVE) method was proposed as a safer, more efficient, and cost-effective solution for developing and testing connected and autonomous driving technologies by operating the real vehicle and multiple other actors like vulnerable road users in different test areas while being immersed within the same highly realistic virtual environment. This VVE approach synchronizes real-world vehicle and vulnerable road user
Chen, HaochongCao, XinchengGuvenc, LeventAksun Guvenc, Bilin
Several challenges remain in deploying Machine Learning (ML) into safety critical applications. We introduce a safe machine learning approach tailored for safety-critical industries including automotive, autonomous vehicles, defense and security, healthcare, pharmaceuticals, manufacturing and industrial robotics, warehouse distribution, and aerospace. Aiming to fill a perceived gap within Artificial Intelligence and ML standards, the described approach integrates ML best practices with the proven Process Failure Mode & Effects Analysis (PFMEA) approach to create a robust ML pipeline. The solution views ML development holistically as a value-add, feedback process rather than the resulting model itself. By applying PFMEA, the approach systematically identifies, prioritizes, and mitigates risks throughout the ML development pipeline. The paper outlines each step of a typical pipeline, highlighting potential failure points and tailoring known best practices to minimize identified risks. As
Schmitt, PaulSeifert, Heinz BodoBijelic, MarioPennar, KrzysztofLopez, JerryHeide, Felix
Artificial Intelligence has gained lot of traction and importance in the 21st century with use cases ranging from speech recognition, learning, planning, problem solving to search engines etc. Artificial Intelligence also has played a key role in the development of autonomous vehicles and robots ranging from perception, localization, decision to controls. Within the big AI umbrella there is machine learning which is all about using your computer to "learn" how to deal with problems without “programming". Deep learning is a branch of machine learning based on a set of algorithms that learn to represent the data directly from the input such as an image, text, Sound, etc. Within deep learning there are Convolutional Neural Networks and Recurrent Neural Networks (CNN/RNN). The study here used convolutional neural network approach to perform image/object recognition. Given that the objective of the autonomous or semi-autonomous vehicle is to promote safety and reduce number of accidents, it
Mansour, IyadSingh, Sanjay
Rear impacts make up a significant portion of crashes in the United States. To date, regulations on rear impacts have focused on fuel system integrity and seat performance, while most research has focused on seat performance in relation to occupants’ injuries, with some analyses of crash severity and seat belt effects. The performance of seats and seat belts may vary depending on the size of the occupant. Understanding how occupant characteristics, as well as crash scenarios, affect injury outcomes can show opportunities for further enhancements in rear impact occupant protection. This paper presents analyses using survey weighted logistic regression models to understand the factors affecting serious injury outcomes (i.e., MAIS 3+) in rear impacts, exploring the potential for improving occupant outcomes. Three separate models are evaluated, focusing on 1) overall injury level, 2) head, neck, and cervical-spine injuries, and 3) thorax, abdomen, thoracic- and lumbar-spine injuries for
Greib, JoshuaJurkiw, ReneeKryzaniwskyj, TanjaOwen, SusanVan Rooyen, PaulWhelan, StaceyWilliamson, John
The accurate prediction and evaluation of load-deflection behavior in bump stoppers are critical for optimizing driving performance and durability in the automotive industry. Traditional methods, such as extensive experimental testing and finite element analysis (FEA), are often time-consuming and costly. This paper introduces a machine learning-based approach to efficiently evaluate load-deflection curves for bump stoppers, thereby streamlining the design and testing process. By leveraging a comprehensive dataset that includes historical test results, material properties, and geometrical dimensions, various machine learning methods, including Gradient boosting, Random forest & XG Boost were trained to predict load-deflection behavior with high accuracy. This approach reduces the reliance on extensive physical testing and simulations, significantly enhancing the design optimization process, leading to faster development cycles and more precise performance predictions. A case study is
Hazra, SandipTangadpalliwar, Sonali
The current leading experimental platform for engine visualization research is the optical engine, which features transparent window components classified into two types: partially visible windows and fully visible windows. Due to structural limitations, fully visible windows cannot be employed under certain complex or extreme operating conditions, leading to the acquisition of only local in-cylinder combustion images and resulting in information loss. This study introduces a method for reconstructing in-cylinder combustion images from local images using deep learning techniques. The experiments were conducted using an optical engine specifically designed for spark-ignition combustion modes, capturing in-cylinder flame images under various conditions with high-speed cameras. The primary focus was on reconstructing the flame edge, with in-cylinder combustion images categorized into three types: images where the flame edge is fully within the partially visible window, partly within the
Wang, MianhengZhang, YixiaoDu, HaoyuXiao, MaMao, JianshuFang, Yuwen
Machine learning has witnessed widespread adoption across various domains, bringing about transformative changes in decision-making, trend prediction, task automation, and personalized experiences. Despite the remarkable predictive capabilities of machine learning models, the associated uncertainty in their predictions remains a critical concern. Uncertainty estimation plays a pivotal role in ensuring robust decision-making, going beyond mere outcome prediction to quantify the model's confidence and potential error. This paper first presents a review of existing uncertainty quantification techniques in machine learning, including Monte Carlo dropout and ensemble methods, highlighting their advantages in addressing uncertainty as well as their limitations. Then, it presents an efficient and fast novel technique for uncertainty quantification using a combination of the ensemble technique and Gaussian process regression providing an accurate estimation of uncertainty bounds. Due to its
Chavare, SudeepMourelatos, Zissimos P.
Traction control plays a key role in improving vehicle safety, especially for driving scenarios involving low levels of tire-road friction. Over the past 30 years, academic and industrial research in traction controllers has mainly favored deterministic approaches. This paper introduces a traction control strategy based on a deep reinforcement learning agent tailored for straight-line acceleration maneuvers from standstill in low-friction conditions. The proposed agent is trained on two different electric vehicles, a front-wheel drive city car (from EU vehicle segment A), and a rear-wheel drive sedan (from EU vehicle segment D). The paper presents a deep reinforcement learning agent formulation suitable for training on different vehicles, assesses the performance of the resulting controllers in comparison with a benchmarking integral sliding mode controller, and evaluates their response to changes in vehicle mass, powertrain parameters and tire-road friction conditions. The assessment
Caponio, CarmineMihalkov, MarioHankovszki, ZoltanFuse, HiroyukiIvanov, ValentinSorniotti, AldoGruber, PatrickMontanaro, Umberto
The increasing complexity of software-defined vehicles (SDVs) necessitates robust and secure communication protocols to protect against cyber threats. This paper explores the utilization of Generative Adversarial Networks (GANs) to enhance the security of communication protocols in SDVs. GANs, consisting of a generator and a discriminator network, are employed to create and evaluate secure communication sequences, ensuring that unauthorized access and potential attacks are effectively mitigated. In this study, we develop a GAN-based framework that generates secure communication protocols tailored for the dynamic environment of SDVs. The generator is trained to produce communication sequences that are indistinguishable from authentic, secure sequences, while the discriminator is tasked with identifying any anomalies or potential vulnerabilities. By iteratively improving both networks, the framework learns to produce highly secure and resilient communication protocols. The performance of
Namburi, Venkata Lakshmi
In the automotive industry, the durability and thermal analysis of components significantly impact vehicle component robustness and customer satisfaction. Traditional computer-aided engineering (CAE) methods, while effective, often involve extensive design iterations and troubleshooting, leading to prolonged development times and increased costs. The integration of artificial intelligence (AI) and machine learning (ML) into the CAE process presents a transformative solution to these challenges. By leveraging AI and ML, the durability simulation time of automobile components is significantly enhanced. Altair’s Physics AI tool utilizes historical CAE data to train ML models, enabling accurate predictions of model performance in terms of durability and stiffness. This reduces the necessity for multiple simulations, thereby decreasing CAE model design and solution completion times by 30%. By predicting potential issues early in the design phase, AI and ML allow engineers to make informed
Patil, AmolSonavane, Pravinkumar
Topology reasoning plays a crucial role in understanding complex driving scenarios and facilitating downstream planning, yet the process of perception is inevitably affected by weather, traffic obstacles and worn lane markings on road surface. Combine pre-produced High-definition maps (HDMaps), and other type of map information to the perception network can effectively enhance perception robustness, but this on-line fused information often requires a real-time connection to website servers. We are exploring the possibility to compress the information of offline maps into a network model and integrate it with the existing perception model. We designed a topology prediction module based on graph attention neural network and an information fusion module based on ensemble learning. The module, which was pre-trained on offline high-precision map data, when used online, inputs the structured road element information output by the existing perception module to output the road topology, and
Kuang, QuanyuRui, ZhangZhang, SongYixuan, Gao
This paper presents a new regression model-based method for accurate predictions of stiffness of different glass laminate constructions with a point-load bending test setup. Numerical FEA models have been developed and validated with experimental data, then used to provide training data required for the statistical model. The multi-variable regression method considered six input variables of total glass thickness, thickness ratio of glass plies as well as high-order terms. Highly asymmetrical, hybrid laminates combining a relatively thick soda-lime glass (SLG) ply joined with a relatively thin Corning® Gorilla® Glass (GG) ply were analyzed and compared to standard symmetrical SLG-SLG constructions or a monolithic SLG with the same total glass thickness. Both stiffness of the asymmetrical laminates and the improvement percentage over the standard symmetrical design can be predicted through the model with high precision.
Yu, ChaoCleary, ThomasJoubaud, Laurentkister, EvanFisher, W Keith
In the automotive industry, there have been many efforts of late in using Machine Learning tools to aid crash virtual simulations and further decrease product development time and cost. As the simulation world grapples with how best to incorporate ML techniques, two main challenges are evident. There is the risk of giving flawed recommendations to the design engineer if the training data has some suspect data. In addition, the complexity of porting simulation data back and forth to a Machine Learning software can make the process cumbersome for the average CAE engineer to set up and execute a ML project. We would like to put forth a ML workflow/platform that a typical CAE engineer can use to create training data, train a PINN (Physics Informed Neural Network) ML model and use it to predict, optimize and even synthesize for any given crash problem. The key enabler is the use of an industry first data structure named mwplot that can store diverse types of training data - scalars, vectors
Krishnan, Radha
Taking a certain type of diesel engine turbocharger as the research object, a detailed study on the identification of turbocharger surge based on non-intrusive acoustic signals was conducted, and a novel turbocharger surge identification method based on multi-domain composite features of acoustic signals was proposed. The data related to the acoustic signals were collected through a series of supercharger surge reproduction experiments, and subsequently, a comprehensive database of these acoustic signals was established. Based on the multi-domain perspective of the time domain and frequency domain, 35 specific features were selected and extracted; the contribution of each individual feature to the occurrence of wheezing was calculated using the random forest algorithm, and the core contributing features were selected to be combined into a comprehensive multi-domain composite feature. This composite feature was then used for the recognition of turbocharger surge, serving as a highly
Zhu, JiaxuZheng, HongyuZong, Changfu
Electrochemical model of a fuel cell involves several parameters which influence its polarization curve. For a numerical fuel cell model to match experimental polarization curve, it is critical to find the right values of these parameters. It is hard to find the values of all the parameters experimentally, and hence parameter calibration is required. A fully automated workflow for calibration of fuel cell model parameters in a three-dimensional Computational Fluid Dynamic (CFD) simulation is created. The CFD model captures detailed electrochemistry and water phase change. The CFD polarization curve is generated by sequentially running a series of simulations starting from low current densities to high current densities. Experimental polarization curve is used as the validation target. An objective function is defined as the L2 norm of the difference between the experimental and the CFD generated polarization curve measured at various current densities. For calibration, eight fuel cell
Champhekar, OmkarJanakiraman, ArunGondipalle, SreekanthAjotikar, NikhilZehr, Randall
Fuel economy and the ability to maintain the state of charge (SOC) of the battery are two key metrics for the energy management of a full-power fuel cell hybrid vehicle fitted with a small-capacity battery pack. To achieve stable maintenance of SOC and near-optimal fuel consumption, this paper proposes an adaptive equivalent consumption minimization strategy (PA-ECMS) based on power prediction. The strategy realizes demand power prediction through a hybrid deep learning model, and periodically updates the optimal equivalent factor (EF) based on the predicted power to achieve SOC convergence and ensure fuel economy. Simulation results show that the hybrid deep learning network model has high prediction accuracy with a root mean square error (RMSE) of only 0.733 m/s. Compared with the traditional ECMS based on SOC feedback, the PA-ECMS effectively maintains the battery SOC in a more reasonable range, reduces the situation of the fuel cell directly charging the power cell in the high
Gao, XinyuJu, FeiChen, GangZong, YuhuaWang, Liangmo
The automotive subframe, also referred to as a cradle, is a critical chassis structure that supports the engine/electric motor, transmission system, and suspension components. The design of a subframe requires specialized expertise and a thorough evaluation of performance, vehicle integration, mass, and manufacturability. Suspension attachments on the subframe are integral, linking the subframe to the wheels via suspension links, thus demanding high performance standards. The complexity of subframe design constraints presents considerable challenges in developing optimal concepts within compressed timelines. With the automotive industry shifting towards electric vehicles, development cycles have shortened significantly, necessitating the exploration of innovative methods to accelerate the design process. Consequently, AI-driven design tools have gained traction. This study introduces a novel AI model capable of swiftly redesigning subframe concepts based on user-defined raw concepts
Yang, JiongzhiSarkaria, BikramjitKumaraswamy, PrashanthKailkere Srinivas, Praveen
Electric trucks, due to their weight and payload, need a different layout than passenger electric vehicles (EVs). They require multiple motors or multi-speed transmissions, unlike passenger EVs that often use one motor or a single-speed transmission. This involves determining motor size, number of motors, gears, and gear ratios, complicated by the powertrain system’s nonlinearity. The paper proposes using a stochastic active learning approach (Bayesian optimization) to configure the motors and transmissions for optimal efficiency and performance. Backwards simulation is applied to determine the energy consumption and performance of the vehicle for a rapid simulation of different powertrain configurations. Bayesian optimization, was used to select the electric drive unit (EDU) design candidates for two driving scenarios, combined with a local optimization (dynamic programming) for torque split. By optimizing the electric motor and transmission gears, it is possible to reduce energy
Chen, BichengWellmann, ChristophXia, FeihongSavelsberg, ReneAndert, JakobPischinger, Stefan
Over recent years, BorgWarner has intensified its efforts to explore and leverage trending technologies such as Artificial Intelligence (AI) and Machine Learning (ML) to enhance products and processes. This includes digital twin technology, which has potential use cases for system behavior analysis, product optimization and predictive maintenance. This paper outlines the development process of a digital twin for a commercial vehicle battery, which serves as a demonstrator and learning platform for this technology. In order to assess the feasibility as well as hard- and software requirements, a cloud-based digital twin demonstrator was developed, integrating vehicle telemetry data with physics-based battery electric and thermal models, and an aging prediction algorithm. The key components are an Internet of Things (IoT) gateway, simulation models, data processing and ingestion pipelines, a machine learning algorithm for anomaly detection, and visualizations of telemetry and simulation
Bongards, AnitaLiu, XiaobingBeemer, MariaGajowski, DanielRama, NeerajShah, KeyaFallahdizcheh, Amirhossein
Optimal control of battery electric vehicle thermal management systems is essential for maximizi ng the driving range in extreme weather conditions. Vehicles equipped with advanced heating, ventilation and air-conditioning (HVAC) systems based on heat pumps with secondary coolant loops are more challenging to control due to actuator redundancy and increased thermal inertia. This paper presents the dynamic programming (DP)-based offline control trajectory optimization of heat pump-based HVAC aimed at maximizing thermal comfort and energy efficiency. Besides deriving benchmark results, the goal of trajectory optimization is to gain insights for practical hierarchical control strategy modifications to further improve real-time controllers’ performance. DP optimizes cabin inlet air temperature and flow rate to set the trade-off between thermal comfort and energy efficiency while considering the nonlinear dynamics and operating limits of HVAC system in addition to typically considered cabin
Cvok, IvanDeur, Josko
Battery health status and driving rangeof electric vehicles (EVs) are critical factors in determining their market penetration. Choosing an optimal charging strategy—specifying how, when, and for how long to charge based on the driver’s travel behavior—can significantly mitigate battery degradation and extend battery life. This study introduces an EV powertrain system energy model designed to enhance the prediction accuracy of battery status under real-world driving conditions. By integrating with the Q-learning approach, this studyprovides tailored recommendations on charging behaviors, including charger type, start time, and charging duration. This study innovatively considers the rental costs caused by the battery capacity not being able to meet the daily driving range. Simulating a typical three-year usage scenario for an average driver in New England, the results indicate that thecharging strategy proposed by this study reduces battery degradation rates by 1.53‰, 3.57‰, and 7.68
Wang, JiayiJing, HaoOu, Shiqi (Shawn)Lin, Zhenhong
This study investigates the nonlinear correlation between laser welding parameters and weld quality, employing machine learning techniques to enhance the predictive accuracy of tensile lap shear strength (TLS) in automotive QP1180 high-strength steel joints. By incorporating three algorithms: random forest (RF), backpropagation neural network (BPNN), and K-nearest neighbors regression (KNN), with Bayesian optimization (BO), an efficient predictive model has been developed. The results demonstrated that the RF model optimized by the BO algorithm performed best in predicting the strength of high-strength steel plate-welded joints, with an R 2 of 0.961. Furthermore, the trained RF model was applied to identify the parameter combination for the maximum TLS value within the selected parameter range through grid search, and its effectiveness was experimentally verified. The model predictions were accurate, with errors controlled within 6.73%. The TLS obtained from the reverse-selected
Han, JinbangJi, YuxiangLiu, YongLiu, ZhaoWang, XianhuiHan, WeijianWu, Kun
With the improvement of autonomous driving technology, the testing methods for traditional vehicles can no longer meet autonomous driving needs. The simulation methods based on virtual scenario have become a current research hotpot. However, the background vehicles are often pre-set in most existing scenarios, making it difficult to interact with the tested autonomous vehicles and generate dynamic test scenarios that meet the characteristics of different drivers. Therefore, this study proposes a method combining game theory and deep reinforcement learning, and uses a data-driven approach to realistically simulate personalized driving behavior in highway on-ramps. The experimental results show that the proposed method can realistically simulate the speed change and lane-change actions during vehicle interaction. This study can provide a dynamic interaction test scenario with different driver style for autonomous vehicle virtual test in highway on-ramps and a more realistic environment
Qiu, FankeWang, KanLi, Wenli
This paper aims to forecast and examine traffic conflicts by integrating Random Forest (RF) alongside Long Short-Term Memory Network (LSTM). The paper begins with the Random Forest method, pinpointing essential elements affecting traffic conflicts, revealing that the speed difference between interacting vehicles and their leaders, as well as the average headway and distance have significant effects on the occurrence of traffic conflicts. The forecasted Time to Collision (TTC) metric demonstrates extraordinary accuracy, confirming the creation of a precise traffic conflict forecast model. The model expertly predicts the vehicle's trajectory. This model skillfully anticipates vehicle paths and potential traffic conflict, demonstrating strong alignment with actual traffic patterns and offering support for traffic management by highlighting imminent risks. Merging RF with feature selection and LSTM for temporal dynamics enhances the forecasting capability. Furthermore, it also illuminates
Cui, XinYuanShi, XiaomengShao, Yichang
Items per page:
1 – 50 of 1112