Browse Topic: Artificial intelligence (AI)

Items (2,670)
The inconsistency in bearing data distributions under diverse conditions often affects the representations of the faulty data and leads to indistinct decision boundaries and even negative transfer resulted from overlapping class distributions, greatly limiting the accuracy of the diagnosis model. To cope with the challenge, a pseudo-label-guided dual-supervised alignment (PDSA) method is developed for bearing fault diagnosis across diverse operating scenarios in this paper. To address the fixed alignment strategy issue, an adaptive distribution alignment layer is incorporated to ResNet18 to achieve dynamic data distribution alignment under varying condition, To enhance classification performances, a dual-supervised mechanism, comprising shallow-layer supervised contrastive learning is introduced through target domain pseudo-labels in target domain and deep-layer regularization class consistency. Experiments on two publicly available bearing datasets demonstrated this model realizes refined class-level alignment, strengthens fault states representation, and shows notable superiority in both accuracy and robustness.
Sun, HaoRen, ShijinGu, Zhangqing
During the cutting process of low-stiffness structural components, the coupling effect between dynamic deformation and cutting forces presents a significant challenge in accurately predicting machining-induced deformations, thereby complicating quality control in the manufacturing of such parts. To address this issue, a cutting force-structural coupling simulation method that combines experiment and finite element is proposed, which takes into account the low-stiffness characteristics of structural components. Focusing on thin-plate parts as the research object, an orthogonal experimental scheme is designed considering workpiece thickness that serves as an indicator of rigidity. A milling force prediction model correlated with workpiece thickness is established. Based on the predicted cutting forces, a multi-analysis-step simulation method is introduced to analyze the machining deformation of structural parts. Additionally, a theoretical analytical model for the machining deformation of thin-plate workpieces is developed. A comparison between the theoretical and simulation results shows a relative error of less than 1.03%, validating the accuracy of the proposed simulation method. Finally, the exponential regression model for the machining deformation is constructed using training data obtained from the simulations. The prediction error of the regression model is less than 15%. The findings of this study are also applicable to predicting machining deformations in other large and low-stiffness structural components.
Zhao, YongshengGao, PengfeiXu, JingjingLiu, Zhifeng
The morphological characteristics of ternary phase diagrams play a pivotal role in optimizing material properties and facilitating the design of novel alloys. In this study, machine learning (ML) is used to predict the number of phases in ternary alloy systems. A new feature descriptor for phase diagram prediction is proposed in ML, which includes the characteristics of element properties, thermodynamic properties of materials and CALPHAD parameters. Initially, this study constructed a dataset comprising various feature descriptors and validated their correctness employing ML models such as LRC, SVM, RFC, Bagging and GBDT. Subsequently, comparing the performance of different models, and the better-performing models Bagging and GBDT were selected for further prediction studies. The models were fine-tuned using grid search and random search methods to optimize their predictive performance. Ultimately, by predicting phase diagram data for multiple ternary systems at different temperatures, the accuracy rate near the temperature range of the given experimental data was approximately 82%. This demonstrates phase diagram descriptors in conjunction with machine learning to predict ternary phase diagram proposed in this study is practicable. The predicted data also provide guidance for experimental determination of phase diagrams and lay the foundation for future material design and optimization.
Fan, HanchaoSu, YuJin, ZongxiaoLi, JunLee, SoowohnTang, JianguoFu, HuaqingDu, Zhi
With the development of controlled nuclear fusion technology, the tokamak device, as the most promising magnetic confinement fusion reactor for advanced engineering applications, requires remote maintenance of its internal components, which has become a key factor affecting both operational efficiency and safety. As a critical component directly exposed to high-temperature plasma, the divertor target plate needs to be periodically replaced and carefully maintained to ensure stable and reliable reactor operation. However, this region is subject to extreme conditions, including high temperature, high vacuum, and intense radiation, making conventional manual maintenance infeasible. This necessitates the development of intelligent and automated teleoperation systems. To address the automated assembly and disassembly requirements of divertor target plates, this study designs an integrated target plate actuator comprising key functional units: a positioning module, a screwing module, a quick-change module, and a passive compliance structure. The actuator achieves rapid and precise alignment with target plate holes, accommodates bolts of different specifications, and exhibits excellent impact resistance. Furthermore, stiffness and mechanical analyses, supported by finite element simulations, verify the actuator’s safety and reliability under high loads and impact forces. To further enhance operational performance, a segmented disassembly and assembly control strategy based on reinforcement learning is proposed, enabling the actuator to adaptively handle torque variations and ensure precise and stable bolt operations. The results demonstrate that the proposed actuator and control strategy significantly improve the accuracy, stability, and efficiency of target plate operations under complex working conditions, providing a reliable solution for automated divertor maintenance in tokamak devices.
Zang, XizheYu, XingzuCao, Zhangbin
Under China’s intelligent manufacturing strategy, manufacturing enterprises are expected to achieve digital and networked operations by 2025, with full digital transformation by 2030. Intelligent factories, the core of this transformation, rely on interconnected, integrated, and data-fused systems. This paper focuses on the micro-assembly intelligent workshop at the Nanjing Research Institute of Electronics Technology, which produces micro-circuit modules for large-scale complex electronic systems. The workshop combines discrete and process manufacturing modes, presenting unique challenges for digital management. A digital management platform based on a five-layer architecture (device, network, data, application, and decision layers) is proposed to address multi-dimensional business needs, including production scheduling, logistics, execution, and decision optimization. A hierarchical workflow structure of the workshop, consisting of a main workflow and several sub-processes, is in-depth studied and designed. The platform is constructed based on requirements analysis and workflow design of the workshop and integrates systems such as MES, APS, WMS, and SCADA, supported by AI-driven big data analytics. This study offers a practical framework for advancing digital transformation in the electronics industry.
Zhang, JianWang, JiafengGuo, Yongzhao
Aiming at the problems of seed cane pile-up and unstable seed supply efficiency in the sugarcane seed production line caused by the seed supply device, a stable seed supply control system was designed, which consists of a seed collection box, an elastic seed-clearing plate and an electrical control system, etc. The EDEM-RecurDyn coupling simulation was adopted to analyze the seed supply process, and the optimal elastic seed-clearing plate structure was designed. Using the single factor test and Box–Behnken experimental design analyzed the effects of the seed supply belt speed, the speed of the first conveyor belt, the number of sugarcane seeds in the collection box and the seed cutting efficiency on the supply efficiency. Establish a quadratic regression model for the efficiency of seed supply and determine the optimal parameter combination: the seed supply belt speed of 0.097 m/s, first conveyor belt speed of 1.639 m/s, and the number of sugarcane seeds is 14. Using the number of sugarcane seeds as the input quantity for the controller, the real-time data is fed back by the TOF sensor. The controller automatically adjusts the seed-cutting efficiency to maintain the continuity and stability of the seed supply process of the seed supply device. The test results show that after applying this system, the seed supply efficiency reached 1.77 setts/s, which was 6% higher than that of the fixed-parameter system. This research can provide technical support for the stable seed supply of integrated equipment for sugarcane seed production.
Li, ShangpingXu, HechangOuyang, RunhongLi, Kaihua
SAE TOMORROW TODAY - SDVs, AI, and the Next Era of Automotive Innovation135787/28/2026
What does it really mean to build a software-defined vehicle? As AI reshapes the automotive industry, SDVs may become the foundation for the future rather than the destination. Listen in as we sit down with Jeffrey Chou, Founder and CEO of Sonatus, a leading provider of intelligence-driven SDV solutions, to explore why SDVs are best understood as a platform for innovation -- one that is scalable, upgradable, and proven at scale. This conversation dives into the evolution of SDVs, from over-the-air updates and AI-powered diagnostics to intelligent infrastructure services that could one day allow vehicles to share computing power, storage, and data with the world around them. You'll also get insight on how Sonatus scaled its software, the cultural shift required to bring Silicon Valley and automotive engineering together, and why collaboration -- not competition -- will define the future of mobility. If you're interested in automotive OS, AI, SDVs, or the future of vehicle architecture, this episode offers an insider's perspective on where the industry is headed next. We'd love to hear from you! Share your comments, questions and ideas for future topics and guests podcast@sae.org. Don't forget to take a moment to follow SAE Tomorrow Today-a podcast where we discuss emerging technology and trends in mobility with the leaders, innovators and strategists making it all happen--and give us a review on your preferred podcasting platform. Follow SAE on LinkedIn, Instagram, Facebook, X, and YouTube. Follow host Grayson Brulte on LinkedIn, X, and Instagram.
Patterson, Lori
This study details the development and experimental validation of a high-fidelity one-dimensional (1D) simulation model for a two-speed transmission designed for off-road vehicles, such as tractors and backhoe loaders used in agricultural and civil engineering applications. The model, implemented in the AMESim platform from Siemens, integrates physics-based loss sub-models for all major components, including gears, bearings, seals, and fluid drag (churning) losses. After development, the model was rigorously validated against test bench data, with efficiency measurements taken across various speed, torque, and oil level combinations, demonstrating a strong correlation with experimental results. A detailed analysis enabled the quantification of the contribution of each loss mechanism, identifying the countershaft gears and input shaft bearings as the primary contributors. Furthermore, a Machine Learning (ML)–based calibration framework, employing Bayesian Optimization, was implemented to reduce discrepancies between simulation and experiment and to generate a synthetic dataset for the creation of fast-executing surrogate models. The study concludes that the proposed methodology constitutes an effective tool for efficiency analysis and optimization during early design stages, establishing a foundation for future integration with ML techniques and the development of digital twins.
Ferreira, Tiago SimaoFallahi, FarzadKedziora, SlawomirHichri, BassemKiefer, Jean-Daniel
Extruded Rails are critical energy-absorbing components in automotive structures designed to mitigate impact loads during the frontal collisions. Traditional crashworthiness design relies heavily on computationally expensive finite element simulations and iterative design exploration. This work proposes a machine learning–driven framework for rapid front extruded rails design using a trained geometric deep surrogate model. A design-of-experiments (DoE) was conducted by varying geometric parameters including width, height, and wall thickness of a thin-walled extruded rail structure. For each design variant, LS-DYNA simulations were performed to obtain performance metrics such as mean crush force and peak crush force. These simulation results were used to train an AI surrogate model capable of predicting crash responses directly from geometric parameters. The proposed approach significantly reduces computational cost by replacing repeated high-fidelity crash simulations with machine learning surrogate predictions. By enabling fast and accurate evaluation of crash response metrics, the workflow shortens design cycles and supports sustainability-driven crashworthiness assessment by reducing simulation resource usage. The framework establishes a scalable, simulation-driven engineering pathway across vehicle platforms and provides a foundation for future closed-loop, AI-assisted crash design workflows.
Kumar, ManikSrinivasan, Sriram
Accurate prediction of ground settlement induced by rectangular pipe jacking, a prevalent trenchless technology in urban infrastructure development, remains a significant challenge. This study addresses this by developing and evaluating a robust machine learning (ML) framework. Leveraging 104 sets of field monitoring data from the Liuye Avenue West Extension rectangular pipe jacking project in Hunan, China, key construction parameters including jacking force, advance rate, and grouting pressure were utilized as inputs to predict ground settlement. A Particle Swarm Optimization (PSO) algorithm was integrated for automated hyperparameter tuning of six distinct ML models: standalone Least Squares Support Vector Machine (LSSVM), Backpropagation Neural Network (BPNN), Random Forest (RF), and their respective PSO-optimized counterparts. Comprehensive performance evaluation using Mean Squared Error (MSE), Mean Absolute Error (MAE), and Coefficient of Determination (R^2) revealed that the PSO-LSSVM hybrid model achieved superior predictive accuracy and generalization capability. Specifically, on the test dataset, the PSO-LSSVM model yielded an MSE of 0.367, MAE of 0.424, and an R^2 of 0.941. These findings demonstrate that the proposed PSO-enhanced LSSVM model significantly outperforms baseline models, offering a highly effective and reliable tool for predicting ground deformation in similar complex pipe jacking projects.
Hu, ShiweiHu, RongZhang, HongChen, YiHu, Da
To address the detection and monitoring needs of fatigue damage in ferromagnetic materials, this paper proposes a nondestructive testing method based on the evolution of magnetic hysteresis characteristics. By constructing a hysteresis loop measurement system, the variation patterns of coercivity (Hc) in Q235 steel specimens under cyclic loading were investigated, revealing three-phase characteristics of fatigue damage: the initial linear growth phase (N ≤ 8,000), the rapid rise phase (8,000 < N ≤ 12,000), and the stable oscillation phase (N > 12,000). Experimental results demonstrate that the relationship between coercivity and damage degree (D) can effectively characterize the processes of crack initiation, propagation, and instability, with significant inflection points observed at D = 0.6 and D = 0.8. The quantitative model based on coercivity provides a novel method for early warning and condition assessment of fatigue damage, offering advantages such as non-contact operation and high sensitivity. This study provides theoretical foundations and technical support for the health monitoring of engineering structures.
Chen, LiDing, Keqin
Nowadays, as computer technology makes quick progress, innovative algorithms like deep learning are getting used more and more in underground engineering and lots of other fields. When working on rectangular pipe jacking tunnel projects, accurately predicting the magnitude of pipeline settlement is really key to keeping the work moving smoothly. But traditional ground settlement prediction methods mainly rely on empirical formulas and numerical simulation software. When applied to tunnels with complex geometries, though, these methods usually don’t work as well as needed. To fix this problem, our study came up with a new model called PSO-LSTM-Self-Attention Mechanism (shortened to PSO-LSTM-SAM), specifically designed to predict pipeline settlement caused by rectangular pipe jacking work. This model takes the data collected from construction monitoring and uses that as the input for time series modeling work. That allows for in-depth analysis of real-time settlement data, and as a result, it can make more precise predictions of long-term pipeline settlement. To verify the effectiveness of the PSO-LSTM-SAM algorithm, the researchers compared its prediction results with those from a conventional LSTM network, an LSTM-SAM network, and a PSO-SVR network. They also checked the model’s performance by looking at pipeline settlement predictions from different monitoring points, using data from the Changsha Railway Transit Line 6 project. The results show that the PSO-LSTM model, with the self-attention mechanism added in, greatly boosts how accurate tunnel settlement predictions are, and the model fits the data better, too. This proves that the PSO-LSTM-SAM model works well: by using the strengths of deep learning, it offers a new way to predict pipeline settlement when building rectangular pipe jacking tunnels.
Chen, YiWeng, XiaoxuanZhang, HongLi, YongsuoHu, Da
Trajectory tracking control serves as the core operational component of autonomous vehicles, directly determining driving safety and passenger comfort by ensuring control precision and stability. To enhance the tracking accuracy and stability for autonomous vehicles, this study proposes a coupled lateral–longitudinal trajectory tracking controller based on multi-agent reinforcement learning. The framework first establishes a Model predictive controller (MPC) derived from vehicle dynamics, formulating the lateral control process as a Markov decision process. A reward function incorporating lateral error, heading error, and steering angle is designed, followed by the construction of a Deep Q-Network (DQN) Agent to optimize the prediction horizon of the MPC. Subsequently, a position–velocity dual-loop PID controller is developed for longitudinal control, with its parameter optimization strategy learned through a Deep Deterministic Policy Gradient (DDPG) Agent. The Extended State Observer (ESO) is incorporated to perform steering angle compensation for internal modeling errors and external disturbances. Co-simulation experiments are conducted in CarSim and MATLAB/Simulink, and the results demonstrate that the coupled controller achieves superior tracking accuracy and stability in both overtaking and lane-changing scenarios compared with the decoupled controller.
Kun, FengJinxiang, ZhaiLi, Wenli
In this study, an efficient method for concurrent thermomechanical performance and weight optimization under modal constraints is proposed to address the coupled design challenges of thermomechanical characteristics (thermal capacity, thermal deformation, and modal) and structural weight in straight-ribbed brake discs. Based on high-fidelity computer-aided engineering (CAE) simulations of brake disc thermomechanical behavior, a neural network (NN)-based surrogate model and a ResNet-guided geometric feature recognition (RGFG) model for automatic modality recognition were developed, and integrated with a particle swarm optimization (PSO) framework for optimal solution exploration. When applied to a passenger vehicle brake disc case study, the surrogate model of NN demonstrates remarkable accuracy: it shows more than 95% agreement with the CAE results in thermal capacity prediction, the prediction accuracy of thermal deformation exceeds 90% compared to CAE results and 83.4% compared to test result, thereby validating the method’s effectiveness. Compared with conventional CAE approaches, the surrogate model of NN achieves a subsecond prediction speed, significantly reducing computational costs. The surrogate model of RGFG achieves a test accuracy exceeding 95%. Furthermore, the proposed optimization framework offers valuable insights for the inverse design of brake discs.
Han, SimiaoJiang, DaxinHan, ChaoWang, JindaSui, Qinghai
SAE TOMORROW TODAY - Why the Biggest Challenge for SDVs Isn't the Technology135777/20/2026
Software-defined vehicles (SDVs) are transforming the automotive industry ... but are OEMs focused on the right priorities? With over 35 years of award-winning automotive software expertise, Elektrobit's comprehensive SDV ecosystem empowers OEMs, Tier 1s, ODMs and Big Tech to build future-ready solutions with speed and confidence -- driving faster innovation and seamless integration across the vehicle lifecycle. Listen in as we sit down with Dr. Moritz Neukirchner, Head of Cross-Portfolio Growth and Alliances, to discuss how automotive operating systems are reshaping the future of mobility and why software is becoming the defining factor in vehicle innovation. From over-the-air updates and AI-defined vehicles to open-source software, you'll learn why many OEMs are rethinking their SDV strategies after years of overambitious goals. This conversation also explores why organizational change, not technical capability, is now the biggest hurdle to building scalable, customer-focused SDVs. If you're interested in automotive OS, AI, SDVs, or the future of vehicle architecture, this episode offers an insider's perspective on where the industry is headed next. We'd love to hear from you! Share your comments, questions or ideas for future topics with Grayson on Twitter or send them to podcast@sae.org. Follow SAE on LinkedIn, Instagram, Facebook, Twitter, and YouTube.
Patterson, Lori
It is very hard to position helicopters in complex environments, and this severely limits their ability to navigate on their own. This paper proposes a navigation algorithm that uses a combination of different sensors and deep learning. It uses a special type of deep learning called ResNet50 and a special type of machine learning called LSTM. This algorithm extracts features of the environment and uses a Kalman filter to estimate the state of the system. The system is made more robust by merging information from multiple levels. The algorithm’s ability to maintain stable navigation in the face of faulty sensors is noteworthy, as is its use of an adaptive inference strategy that dynamically adjusts computational load. This strategy strikes a balance between performance and resource consumption. Experiments show that the plan works well in places where GPS is not available. This makes it much better for the helicopter to fly by itself, and it can be used in places like the army, for looking at places from the sky, and for helping people in danger.
Yang, Ming
This paper solves the problem of resource and energy constraints on orbit computing for LEO satellites. By combining MADDPG reinforcement learning and Lyapunov optimization, the paper proposes a computing framework and implements an adaptive task offloading model for space flight using a multi-agent deep actor critic algorithm, MADDPG. The joint optimization mechanism is implemented by multi-agent dynamic task offloading. Through the transformation from the state with long-term constraints into optimization of the status of queue stability, the load scheduling under threshold energy in accordance with the characteristics of energy constraints was realized by introducing Lyapunov virtual queues into the process of policy evaluation of deep reinforcement learning. The experimental results show that the proposed framework enables a lightweight preliminary calculation, balanced energy consumption to reduce resource allocation, and realizes the stable queues through adaptability of tasks under energy balance conditions, which can provide high-efficiency computing assistance and support for space orbit tasks such as monitoring remote sensing of Earth.
Yan, MingZhao, LiangXu, LexiZhou, XiaofeiHawbani, AmmarSun, Yunhe
Aiming at the problems of traditional physical model methods in aircraft endurance prediction, an end-to-end prediction model based on depth deterministic policy gradient (DDPG) is proposed. The model realizes continuous mapping from flight parameters to range index through Actor-Critic dual network architecture, and combines experience playback mechanism and soft update strategy of target network to effectively suppress training oscillation and improve convergence stability. UAV Delivery Aircraft Versus hybrid dataset was used to verify model performance in test samples. The results show that the MAE of the model is 9.2 km, which is 42.1% lower than that of DQN; the prediction accuracy of the model is the best (MAE 7.3 km) in cruise phase, which is due to the dynamic compensation of time series difference error to wind speed disturbance; in environmental disturbance test, the error increment (50.0%) is significantly lower than that of DQN (78.0%) at low temperature (-5 ° C), which highlights its robustness to battery voltage sag. The model provides real-time and reliable decision support for aircraft endurance management in high-dynamic airspace.
Bai, RongqiangChen, Li
In recent years, drone technology has seen widespread application in both civilian and military fields. By 2025, China will introduce supportive policies from multiple dimensions, including industrial development, technological innovation, and application promotion, to significantly increase the number of UAVs in use and their frequency. However, drones are prone to malfunctions due to factors such as bad weather and electromagnetic interference, which may result in serious consequences, including property damage and casualties. Therefore, improving the accuracy of fault detection and the response time of drones is of great significance. Although current research has made progress, there are still deficiencies: First, most of them rely on a single or limited data source, resulting in incomplete information and vulnerability to interference, which leads to low detection accuracy and reliability; Second, traditional methods are mostly based on fixed thresholds or simple rules, lacking real-time dynamic monitoring and adaptive analysis capabilities, making it difficult to issue timely warnings of potential faults. To this end, this study proposes a multi-scale time series prediction model based on multimodal and multi-branch, integrating multimodal data, constructing a dual-branch architecture, and combining deep learning and attention mechanisms to enhance the anomaly detection effect of unmanned aerial vehicles. A dual-branch anomaly detection model based on 1DCNN-BiLSTM and continuous wavelet transform is proposed, including a trajectory prediction difference branch and a full time series data branch. In the dual-branch output stage, the attention gating mechanism is utilized to fuse features and improve the detection performance. The experimental results show that this model performs excellently in both normal trajectory prediction and anomaly detection, providing an effective solution for drone anomaly detection.
Pu, ZhenglinZhang, Lin
Aiming at problems such as low efficiency and poor accuracy in fault identification for traditional small satellites, this paper proposes a multi-model fusion method based on machine learning. By constructing a telemetry data preprocessing module based on the Data Generation Adversarial Network, it effectively deals with outliers and fills missing values. Combining single model methods such as polynomial curve fitting, the grey model, and the ARMA model, and introducing the Long Short-Term Memory network and Gated Recurrent Unit to fuse with these models enhance the ability to process complex data features. The prediction results of each model are fused using machine learning methods, and finally, the fused value is taken as the final prediction result. The numerical simulation results show that this prediction method can predict the anomalies of different types of satellite telemetry parameters and has achieved good results.
Liu, BiyanChen, YeGuo, Qi
This paper constructs a reinforcement learning framework based on the PPO algorithm for drone air combat to solve 1v1 pursuit-evasion in 2D beyond-visual-range air combat. Firstly, the mission scenario is modeled, defining key roles of ATA and AA. Then, state transition models of pursuer and evader are built based on flight kinematics. To handle reward sparsity in policy network training, a dense reward function combining distance and angle rewards is designed to guide the agent in learning tail-chasing and interception strategies. Using the Actor-Critic architecture, deep neural networks implement the decision-making and evaluation modules. The PPO algorithm trains the pursuing drone in a simulation. Results show that after ~5 million steps, the agent learns a stable strategy, completing tasks promptly and generalizing well in unseen scenarios. This research offers ideas for drone combat and guidance, and supports autonomous decision-making in complex air battles.
Yu, KangjieGong, ZhengHu, RunchangLiu, Huixiang
This paper presents an innovative study in exploring, evaluating, and implementing deep-learning architectures for the calibration of multimodal sensor systems. The aim of this paper is to leverage the use of sensor fusion to achieve dynamic, real-time alignment between 3D LiDAR and 2D camera sensors. Static calibration methods are tedious and time-consuming, which is why we propose utilizing conventional neural networks (CNNs) coupled with geometrically informed learning to solve this issue. We leverage the foundational principles of extrinsic LiDAR–camera calibration tools such as RegNet, CalibNet, and LCCNet by exploring open-source models that are available online and compare our results with their corresponding research papers. Requirements for extracting these visual and measurable outputs involved tweaking source code, fine-tuning, training, validation, and testing of each of these frameworks for equal comparisons. This approach aims to investigate which of these advanced networks produces the most accurate and consistent predictions. Through a series of experiments, we reveal some of their shortcomings and areas for potential improvements. We find that LCCNet yields the best results among all the models that we validated.
Karramreddy, Venkat Sai RaxitMitchell, Liam
Simulation plays a significant role in the validation and verification of Automated Driving Systems (ADS). In a scenario-based validation strategy, the road and the actions of the traffic participants must be captured in a portable and flexible format for simulation. XML-based parametric models constitute a common combination upon which the static and dynamic aspects of the environment are captured. Although there are plenty of tools for generating these XML files there are few alternatives to verify their content. This paper suggests a method for converting and simplifying a synthetic road network into a graph for which the Chinese Postman Problem is solved. The resulting sequence can be converted back into a route that can be sampled to verify the drivability of the whole network. Once the network is verified, it can be safely used for simulation, increasing the speed at which ADS systems are developed. The graph representation can also be used to provide interactive feedback to LLMs (Large Language Model), which are increasingly used for automatic generation of roads and scenarios.
Vargas Rivero, Jose RobertoKern, AndreasMenken, StefanHarth, MichaelKuipou, Franck Russel
Traffic flow prediction is of great significance for improving the operation efficiency of the transportation system, optimizing travel experience and reducing traffic congestion. Traditional traffic flow prediction methods are difficult to capture the spatio-temporal nonlinear characteristics of traffic flow due to its simple model and insufficient feature extraction ability. Therefore, an intelligent traffic flow prediction system based on deep learning is proposed, constructs a deep learning model based on graph convolution and fusion of attention mechanism LSTM. Based on this, a traffic flow prediction system is implemented. Experiments show that, on the PeMSD4 and PeMSD4 datasets, the error of the model in RMSE and Mae indicators is significantly reduced compared with the traditional methods, which provides an efficient solution for traffic flow prediction and congestion analysis, and has both theoretical innovation and engineering practical value.
Tang, ZhanLu, XiaoyuYang, NianXiang, XiaohongHou, XiangPeng, Xiaoli
In order to reduce traffic accidents caused by cars straying from lanes, a lane line recognition and deviation warning system based on machine vision is designed. It mainly includes image preprocessing, lane line detection, and the design of a deviation warning model. “In this study, an ROS-based intelligent vehicle-mounted camera is adopted for road image collection. To reduce the computational load of data processing while guaranteeing the algorithm’s accuracy and reliability, grayscale conversion and region of interest (ROI) extraction are implemented to finish the image preprocessing stage. Additionally, a fusion strategy of global and local thresholds is introduced to enhance both the operational speed and detection accuracy of the algorithm” use the Canny operator for the edge feature extraction; and complete the fitted lane lines with the improved Hough transform. Finally, based on the Kalman filter and camera viewpoint conversion coefficient algorithm, the lane line offset is detected in real time, and the deviation is judged in combination with the monitoring interface. Simulation experiments show that the system is able to effectively recognize the lane line and judge the deviation status under the condition of setting the offset threshold of 70 pixels, which significantly improves the accuracy and real-time performance of the lane deviation warning and provides effective technical support for reducing traffic accidents.
Wang, XufengZhang, ChunshuWang, YanChen, YihuiJi, Rui
Public transportation serves as a crucial component of urban mobility, contributing to the alleviation of urban congestion, reduction of travel expenses, and mitigation of air pollution. Nonetheless, the dynamic passenger demand and the complex traffic conditions render traditional bus timetables inadequate, leading to ineffective allocation of public transportation resources. Consequently, it is essential to create bus timetables that are responsive to actual traffic scenarios and fluctuating passenger demand. This study regards the bus timetable planning problem as a Markov decision-making process within a discrete time framework, proposing a deep reinforcement learning-based optimization model for bus timetables. In particular, the model is designed to account for both bus companies and passengers, incorporating a state space and reward calculation method that emphasizes passenger comfort. Then Deep Q-Network (DQN) methodology is employed to issue instructions on whether a bus departure at each time, and bus timetable is generated gradually over time. Experimental results indicate that the proposed approach significantly reduces bus travel costs and enhances the overall travel experience for passengers in comparison to traditional methods.
Xu, JieXia, DongYang, JianxiWang, Bing
In recent years, with the low-altitude economy developing rapidly, the operation and management of low-altitude airspace has gradually become a hot topic. Unmanned aerial vehicles (UAVs) constitute a fundamental component of the low-altitude airspace ecosystem, significantly influencing its structure and functionality. The technological advancement of UAVs has fundamentally transformed the operational paradigm for low-altitude airspace management. This paper presents a comprehensive review of UAV-supported technologies in the context of low-altitude airspace operations and management. It systematically analyzes key technologies and applications of UAVs in areas such as airspace capacity and safety assessment, trajectory planning, and standardized flight management. Drawing from kinematic analysis and traffic flow theory, UAV density control and collision risk prediction offer quantitative insights into airspace capacity evaluation. Additionally, probabilistic analysis and simulation techniques enhance the accuracy and efficiency of safety assessments. In trajectory planning, multi-objective optimization algorithms tailored to operational scenarios—such as logistics delivery and agricultural operations—have significantly improved the utilization of airspace resources. Concurrently, collision avoidance techniques leveraging graph search, numerical optimization, and machine learning ensure flight safety in complex environments. Standardized flight management relies on pilot qualification review, airworthiness certification, and planning standardization, while discussing airspace segmentation strategies based on geofencing and intelligent control systems. Future developments in UAV-supported technologies are expected to trend toward higher precision, intelligence, and regulatory integration. By incorporating cutting-edge fields such as deep reinforcement learning and digital integration, these technologies are poised to further enhance the efficiency and safety of low-altitude airspace management, thereby providing robust technical support for the sustainable growth of the low-altitude economy.
Gong, LeiMa, ZhenxiaoLuo, Qin
As a special vehicle, motor caravans have high customer demand and expectations for product quality under current market conditions. At the same time, customers generally have strong demands for functional differentiation and modification. To meet the requirements, manufacturers need to redesign and construct the production process platform, including redesigning and transforming various functional unit modules on the vehicle. For example, the flexibility of production process platform systems, standardization of electrical interfaces, and modularization of functional units, etc. In the implementation process, by embedding flexible architecture into the existing universal process system, motor caravans modification can be flexibly organized according to customer orders and requirements while ensuring streamlined production. In the implementation process, the focus of the new installation process system is to match the electrical systems inside the vehicle, ensuring not only the matching between different unit systems, but also the compatibility between the new installation system and the original vehicle system. When formulating the modification process, it is necessary to meet the actual usage requirements and environmental conditions of the motor caravans, while also considering the speed of production organization. The beginning of in car modification often means rebuilding local systems, and whether the compatibility and compatibility of the overall system modules are complete is the ultimate goal pursued by modification production enterprises. In the research process, methods such as comparison, literature review, and examples were used to attempt to illustrate the role of flexible process system architecture in promoting motor caravans modification, especially personalized and differentiated modification, in the context of rapid development of the contemporary automotive industry. Especially in today's rapidly developing advanced control technology of artificial AI, breaking down the whole into smaller parts and standardizing them one by one is of great significance for improving production efficiency and enhancing the automation level of process systems.
Li, Sheng
Aiming at the problem of insufficient modeling of spatio-temporal heterogeneity in road traffic accident prediction, a dual task machine learning framework integrating geographical environment, location attributes and time periodicity is proposed. The dataset used in this study was derived from traffic accident records of Nanchang during 2019–2023. Firstly, geographical identifiers are generated by rounding and aggregating latitude and longitude coordinates. At the same time, the location type is processed by a one-hot encoding, so as to carry out spatial clustering analysis of accident hotspots. Compared with the North-South pattern, the contribution of geographical features shows a strong East-West trend. The kernel density heatmap identified Zone A and zone B as dual core high-risk areas. Secondly, the sinusoidal/cosine function is used to encode the time feature circularly, which effectively captures the daily change of the accident. The quantitative analysis of random forest regression model showed that time characteristics accounted for 89.2% of the variance of accident frequency interpretation, significantly exceeding the contribution of geographical factors (10.2%) and location attributes (0.6%). After hyperparameter optimization, the accuracy of XGBoost classifier in predicting serious accidents is 75.97%, and the AUC value is 0.8412, which has strong robustness, and provides reliable support for dynamic risk assessment of traffic management system.
Luo, JiangZhang, YuxinLi, XinWu, Ronghai
The collection of road high-frequency data often involves inputs from multiple sensors, such as stress and strain, and sampling of these data features a high sampling rate of up to 2,000 Hz. High-frequency sampling enables capturing of the internal stress and strain of the pavements when vehicles are passing and facilitates the analysis of the pavement structure and prediction of its long-term service performance. However, while the sensors are continuously collecting data, the time the vehicles pass is discrete and unpredictable, resulting in a large number of low information density or irrelevant data. Even when the massive high-frequency data are collected, challenges remain in data transmission, storage, and analysis—the challenges are attributable not only to the massive quantity and complexity of data from multiple sensors, but also to the inconsistent data formats, misaligned timestamps, and multi-sensor data fusion difficulties. In response to the challenges specified above, a new approach combining traditional road observation data with deep learning models is proposed here to efficiently process and analyze massive sensor data. This method not only improves the data processing efficiency but also provides new insights into innovation of road engineering technologies.
Gang, JianZhang, YueChen, YinghaoZheng, XiaoyanWang, TaojieLiu, YilinGuan, WeiWu, Jiangfeng
Zero-gravity seats alleviate prolonged sitting fatigue by optimizing human body pressure distribution, but the correlation mechanism between body size parameters and pressure distribution remains unclear. This study proposes a deep learning model based on multimodal data fusion, combining pressure matrices and postural angle data to construct a convolutional neural network (CNN) with a height prediction error ⩽3 cm. Experiments collected pressure and posture data from 100 participants with diverse anthropometric percentiles. Through the fusion of features and the optimization of the model, the study managed to quantify how height and weight impact pressure gradients. The results indicate that the model achieved a prediction R2 value of 0.73, which confirms that there is a strong correlation between pressure distribution and body size parameters. The findings offer theoretical and technical support for the adaptive adjustment systems within intelligent cabins.
Bi, TengfeiNie, JiachengDu, ChangjiangJi, YuechenWang, SongSun, Jiawei
With the rapid development of the low-altitude economy—represented by drone logistics, aerial inspections, and air taxis—air traffic has exhibited new characteristics including diverse forms, high density, and significant speed differences. To address these changes, the traditional air traffic control system requires upgrades, particularly in dynamic aircraft scheduling. This study proposes an air traffic control model (DS-ATM) tailored to this domain, built on the Deepseek large model. By integrating spatiotemporal graph neural networks with multi-objective reinforcement learning algorithms, the model achieves real-time path planning and conflict resolution in complex airspace environments. Validated using public datasets such as OpenSky Network, NASA UTM Dataset, and METAR meteorological data, experimental results demonstrate its significant advantages in reducing conflict rates and scheduling delays.
Li, RuiZhao, FangyuShe, YueLi, Wujie
The virtualization of powertrain systems is a key enabler for modern powertrain development. While physics-based 0D/1D simulation models provide accuracy and interpretability, these models are typically computationally demanding, prolonging the development process and usage throughout the V-cycle. Moreover, achieving real-time-capable simulation models through model simplifications remains challenging, as it often leads to significant losses in accuracy. In contrast, data-driven approaches can achieve high computational efficiency without significantly compromising model accuracy. This opens the possibility for not only online control applications, such as model predictive control or reinforcement learning, but also for computational expensive offline control prototyping using ultrafast-running data-driven digital twins. This work focuses on the elaboration of a scalable methodology for the development of ultrafast-running powertrain models for stationary and transient engine operation. This includes the efficient generation of training data with great variance, data analysis, and preparation, an optimized partitioning method using the Jensen–Shannon distance, feature engineering, model training of a multilayer perceptron (MLP), a long short-term memory (LSTM), and gated recurrent unit (GRU) network, followed by the model evaluation using test data and the concluding model deployment. In order to demonstrate the concept, a calibrated 0D/1D model of a dual-fuel marine main engine provided by WinGD Ltd. for a pure car and truck carrier is utilized as the reference physics-based model. The case study provides a comprehensive examination of the development of ultrafast-running data-driven fuel consumption models in both stationary and transient engine operation. The results show that the proposed methodology yields robust results and minimizes the loss of accuracy to 1.80%–2.14% for the MLP predicting the steady-state fuel consumption and to 0.67%–0.96% (GRU) and 1.52%–1.68% (LSTM) for predicting the transient fuel consumption, while achieving a multiple 104-fold reduction of the real-time factor (RTF) on an identical CPU.
Weller, LouisZanelli, AlessandroYang, QiruiBrutsche, MartinGrill, MichaelKulzer, André Casal
Uncertainty quantification (UQ) is increasingly recognized as essential when machine learning (ML) is employed in domains that are safety-relevant, cost-intensive, or legally binding, such as the product engineering of battery electric vehicle (BEV) energy systems. UQ methods aim to estimate the aleatoric, epistemic or both uncertainties associated with the predictions of a machine learning model. However, the landscape of UQ methods is diverse and rapidly evolving, with no single approach proving optimal across all tasks. Consequently, the selection of methods in practice is often driven by experience, constrained by limited comprehensive knowledge, time, and implementation capacity. This paper introduces an application-oriented process model supporting data scientists in selecting UQ methods in ML by adapting the SPALTEN [1] problem-solving methodology and the Algorithm Selection Process Model (ASPM) into an Algorithm Selection Process Model for Uncertainty Quantification (UQ-ASPM). This model can be integrated into the modeling phase of a data mining process, such as the Cross Industry Standard Process for Data Mining (CRISP-DM). Ethnographic observations and expert interviews conducted within the research environment of BEV energy system development were analyzed using inductive qualitative content analysis to identify practical barriers, motivations, and requirements. The resulting process translates task requirements and boundary conditions into UQ-specific criteria, primarily including the source of uncertainty, integration depth, and output type. It employs a funnel-like narrowing from method families to candidate algorithms and utilizes a transparent evaluation matrix with weighted criteria, consequence analysis, and learning through a continuous information pool. An illustrative predictive-maintenance example demonstrates the instantiation of the process when an existing deterministic ML model must be retained. The contribution is made at a meta-level, facilitating structured navigation of the method space rather than providing a direct comparison of individual UQ algorithms.
Holderied, NiklasHörtling, StefanBause, KatharinaDüser, Tobias
This paper investigates the integration of Artificial Intelligence (AI) within radar-based perception for Advanced Driver Assistance Systems (ADAS) under safety considerations aligned with ISO 26262 [1] for functional safety and ISO 21448 (SOTIF) [2] for performance-related safety of the intended functionality. The study evaluates a hybrid architecture in which AI-based perception modules are combined with deterministic supervisory mechanisms to maintain safety compliance. A simulation-based case study using CARLA with radar sensor modeling is presented to compare a deterministic radar perception pipeline with an AI-enhanced approach under nominal and degraded environmental conditions. Performance is evaluated using precision, recall, and F1 score metrics. Results indicate improved recall and F1 score under adverse scenarios for the AI-based perception module, accompanied by a moderate increase in false positives. The paper discusses architectural constraints required to limit non-deterministic behavior, including confidence gating, deterministic supervision, and scenario-based validation. The findings are limited to simulation and are intended to provide preliminary insights into the technical and safety implications of incorporating AI-based radar perception within ISO 26262-compliant ADAS architectures.
Jain, Yesha
The aim of this work is to develop a modular, real-time-capable digital twin of an electric powertrain based on machine learning (ML)-based model structures and a systematic, component-oriented architecture with a focus on efficiency estimation in test bench environments. The further goal here is to enable virtual testing, which can be used for frontloading and thus both prevent errors and increase the speed of product development. Based on a comprehensive set of measured and derived test bench data, a multi-stage procedure is implemented that integrates data acquisition, physically informed feature selection, modeling at the component and subsystem level, and hybrid coupling strategies. The digital twin captures inverter, electric machine, and mechanical transmission stages and generates consistent predictions of key variables such as torque, speed, power factors, and subsystem as well as overall drivetrain efficiency. The methodology enables a systematic comparison of black box, dark grey box, grey box, and bright grey box architectures with respect to prediction accuracy, information content, and real-time capability. The methodology provided uses new model structures that explicitly integrate physical dependencies while also using ML models to map nonlinear effects. The hybrid architectures presented have been shown to significantly reduce the measurement effort while achieving nearly identical model quality and surpassing purely physics-based models in terms of accuracy, robustness, and real-time capability. For the final bright grey-box architecture, average relative efficiency errors below 1 % are achieved while maintaining real-time execution rates. The study shows that bright grey box-models in particular offer a best-case compromise between the requirements of information content, error quality, and synchronization rate, thus representing a methodological advance over conventional digital twins, which are often created at the component level. The shown methodology provides an implementable framework for digital twins of electric powertrains in industrial test environments.
Kopp, LennartProksch, DanielOckert, NielsKarthaus, CarstenKley, Markus
Recent advancements in Vision-Language Models have opened new possibilities for bridging the gap between Systems Engineering artifacts and automated code generation. Traditional Large Language Models are primarily trained on textual data and generic code repositories, which limits their ability to interpret graphical engineering artifacts such as Simulink block diagrams or system architecture models. In safety-critical domains like the automotive industry, these graphical models are central to development workflows and must remain closely aligned with textual requirements and implementation code to ensure traceability, compliance, and functional correctness. This paper proposes a Vision-Language Model-centered multimodal training framework for code generation that integrates textual requirements, graphical model-based artifacts, and annotated source code into a unified learning process. By leveraging models which combine vision encoders with language backbones, the approach enables the model to jointly learn the structural semantics of engineering diagrams and the linguistic and syntactic patterns of requirements and code. This alignment allows the model to generate code that is not only syntactically correct but also semantically consistent with both textual specifications and graphical designs. We evaluate the approach on a representative automotive dataset consisting of requirements, Simulink block diagrams, and C/C++ implementations. Preliminary results demonstrate that incorporating visual model representations significantly improves code correctness, requirement alignment, and structural consistency compared to text-only baselines. These findings highlight the potential of Vision-Language Models to enable more accurate, adaptive, and domain-compliant code generation, paving the way for the integration of VLMs into future model-based software development workflows.
Padubrin, MarcelKulzer, Andre CasalGuerocak, Erol
Accurate tire models are a key enabler for vehicle dynamics simulation, control design, and lap time optimization, particularly in the context of Formula Student race cars, where vehicle setups and tire characteristics differ significantly from production vehicles. State-of-the-art tire models, such as Pacejka’s Magic Formula, generally provide high prediction accuracy. However, their predefined functional structure and large number of coupled parameters are designed for broad applicability across many tire types rather than for specific racing tires. This often results in limited interpretability, nontrivial parameter identification, and unnecessary model complexity for specialized applications such as Formula Student. This paper presents a data-driven approach for deriving compact and physically interpretable tire force models using symbolic regression. The proposed method employs an intelligent tree search to systematically explore the space of mathematical expressions and identify models that optimally balance prediction accuracy and structural simplicity. In contrast to black-box machine learning approaches, the resulting models consist of explicit mathematical expressions that enable physical interpretation and efficient evaluation. The methodology is applied to experimental tire test bench data, focusing on the lateral force – slip angle relationship at constant vertical load. In a first step, the symbolic regression algorithm is utilized to derive a set of candidate mathematical expressions. These models are subsequently benchmarked against 200 independent data sets comprising various tire types and vertical loads. The evaluation reveals that the identified models approximate the measured tire behavior with accuracy comparable to, and in many cases exceeding, the Magic Formula, while exhibiting lower model complexity. The results demonstrate that symbolic regression can uncover alternative tire models that better represent the characteristics of Formula Student racing tires than conventional approaches. Owing to their compact structure and physical consistency, the derived models are particularly well suited for real-time vehicle simulations, parameter studies, and control-oriented applications in Formula Student vehicle development.
Anselment, MarcelBorowski, JulianRudolph, Stephan
Kolmogorov-Arnold Networks (KANs) are a novel mathematical method to generate data-driven AI surrogate models. Compared to neural networks based on the MLP standard (Multi-Layer Perceptron), these offer further mathematical interpretability and thus allow improved validation of AI for industrial applications. In this paper, we use KANs to generate an AI vehicle model of a truck as a mathematically precise AI surrogate model. To do this, we combine the KAN approach with the approach of Neural Ordinary Differential Equations (Neural ODEs) to generate predictions for the time-series of the truck’s velocity. Furthermore, we compare the results of the AI based on KANs with the traditional approach using MLP in terms of model size, accuracy, and computational time in order to evaluate advantages and disadvantages of the KAN approach. The best AI-KAN vehicle model identified in this way is then embedded in a co-simulation via the Functional Mockup Interface standard, thus opening up a wide range of applications in AI-driven truck development.
Vaudrevange, Patrick K.S.Halverson, JamesRuehle, FabianFabcic, TomazDingler, ChristianPiskala Dilipkumar, SanthoshkumarIbrahim, MuhammedHerrnberger, MichaelKasper, JohannaTürk, LarsKeckeisen, Michael
The increasing complexity of modern software-intensive systems, particularly in the automotive domain, demands new approaches to bridge the gap between high-level engineering specifications and executable, safety-compliant code. This need is amplified by the rapid transition toward software-defined vehicles, where highly dynamic, updateable software functions significantly enlarge the scope and frequency of engineering activities and require scalable, transparent, and adaptive development processes. While recent advances in Large Language Models have demonstrated strong capabilities in automating tasks such as requirements analysis, code generation, and documentation, their deployment in safety-critical engineering workflows remains challenging due to the need for transparency, traceability, and controlled decision-making. This paper presents a modular multi-agent Large Language Model (LLM) pipeline that automates key steps of the systems engineering lifecycle - from requirement structuring and compliance checking to code and test generation - using specialized LLM agents orchestrated within a unified architecture. A central contribution of this work is the integration of a Human-in-the-Loop subsystem, which introduces configurable review checkpoints at critical stages such as requirements analysis, compliance assessment, code generation, and test creation. The human-in-the-loop module enables engineers to approve, reject, or modify intermediate results, ensuring human oversight, enhancing trustworthiness, and enabling adherence to functional safety standards. The system supports heterogeneous input formats and provides end-to-end traceability through structured outputs and detailed monitoring of performance metrics including model usage, token consumption, and automation efficiency. Initial evaluations indicate that the combination of multi-agent specialization and human-in-the-loop-guided oversight can significantly reduce engineering effort while maintaining the transparency and reliability required for regulated domains. By embedding controllable human supervision into the LLM-driven pipeline, this work offers a practical and scalable architecture for integrating Artificial Intelligence (AI) automation into safety-critical systems engineering processes, with particular relevance to automotive software development.
Padubrin, MarcelKulzer, André CasalGuerocak, Erol
The global automotive landscape is undergoing a significant paradigm shift driven by the rapid development cycles of emerging competitors, leaving traditional European OEMs with a critical time-to-market gap. To bridge this gap, automotive engineering must pivot from traditional hardware-based processes toward agile, digital data-driven methodologies. This paper presents a feasibility study on the implementation of data-centric approaches in component development, evaluated using the high-voltage wiring harness (HVWH) as a representative example. The HVWH serves as a practical validation case for the presented methodologies, covering both Artificial Intelligence (AI) based and deterministic methods. The study provides a detailed assessment of various AI-based and deterministic methodologies at specific stages of the product development process, targeting both product design and the product development process itself. The objective is to reduce time-to-market at the component-level by optimizing workflows, increasing process and development efficiency, and enabling knowledge reuse throughout the development process. Beyond individual method evaluation, the study examines how deterministic and AI-based approaches can be integrated into development workflows. For this purpose, process mining is first applied to identify general challenges specific to the HVWH development workflow and to derive use cases in which AI can contribute to reducing development time. From these, three use cases are selected for detailed investigation. For each use case, the necessary prerequisites, the applied methodology, the results and the limitations of AI integration are described and discussed. By integrating structured knowledge with automated workflows, the proposed frameworks allow for autonomous application of historical insights to current design parameters, streamlining the decision-making process. This semantic structure prevents the loss of critical engineering knowledge and enables continuous AI-assisted improvement across different vehicle generations. The study concludes that the proposed use cases provide a technically viable pathway to shorten development timelines, enabling European OEMs to match the speed of competitors while maintaining high standards of quality, functionality and safety.
Bode, Jana PascalKröll, SarahVohwinkel, NikolausPaetzold-Byhain, Kristin
In vehicle production, commissioning and testing processes of electric and electronic components are essential for value creation and quality assurance. The emergence of software-defined vehicles, however, leads to an increased scope and complexity of these processes as software functions depend on electric and electronic components for perception, execution, and processing tasks. In this context, this paper tackles a common challenge: Software that is deployed in vehicle production to implement commissioning and testing processes is developed upon specifications that define prerequisites, procedures, and target results in natural language. Therefore, extensive human interpretation and manual translation into executable code are needed being susceptible to errors as well as time-consuming. The large number of vehicle configurations and rapid changes in vehicle software further complicate the development of commissioning and testing software, particularly as verbose textual dependency descriptions risk impairing comprehensibility. Machine-processable specifications facilitating automated validation and code generation or direct execution could consequently ensure consistency, reduce manual effort, and accelerate the development process. For this purpose, we examine the processability of commissioning and testing specifications in natural language by proposing a pipeline designed to systematically transform these specifications into a machine-processable format. In particular, we introduce a unified schema that serves as an input format for the large language models tasked with the transformation. Subsequently, several large language models are evaluated in practical trials, based on their ability to translate commissioning and testing specifications into a machine-processable notation. In summary, this study aims to enable more efficient and data-driven software development based on textual requirements. This work offers valuable insights into the suitability and applicability of large language models within the planning of automotive commissioning and testing processes, targeting enhanced automation and efficiency.
Köhler, KatjaEl Asad, AimanHahn, MichaelReuss, Hans-Christian
This study investigates the feasibility of identifying individual e-bike riders based on CAN data using machine learning techniques. Datasets from 12 test riders performing various predefined cycling tasks on a dynamometer test bench are collected and used to ensure controlled and reproducible conditions. The recorded CAN data includes various sensor signals, such as power output, cadence, torque, and the used support mode. After pre-processing, two different methods of feature extraction are tested and compared, one based on snapshots of the data and one based on driving events such as braking and accelerating, measured by calculating statistics of the riding data over sliding windows. A range of machine learning models is employed to classify riders based on their distinct riding patterns using the extracted features. The evaluated models comprise KNN, Random Forest and Naïve Bayes. The findings demonstrate the efficacy of machine learning in differentiating riders, with Random Forest and KNN achieving the highest and most robust accuracy among the tested models. The KNN-model achieves up to 99% accuracy, the Random Forest up to 75%. The paper analyzes the influence of different signals, feature extraction methods and model parameterizations on the results. The results show that machine learning-based rider identification using CAN data is a viable approach for enhancing e-bike security, authentication, and personalization. Potential applications include theft prevention, automatic user recognition for personalized assistance settings, and access control. Future research could include the exploration of the impact of additional sensor data, real-world outdoor conditions, and deep learning approaches, with the aim to further enhance identification accuracy and efficiency.
Simmann, GabrielRauch, YannickBeißert, FlorianKriesten, Reiner
Electronic Control Units (ECUs) have played a pivotal role in transforming motorcars of yore into the modern vehicles we see on our roads today. They actively regulate the actuation of individual components and thus determine the characteristics of the whole system. In this, the behavior of the control functions heavily depends on their calibration parameters which engineers traditionally design by hand. This is taking place in an environment of rising customer expectations and steadily shorter product development cycles. At the same time, legislative requirements are increasing while emission standards are getting stricter. Considering the number of vehicle variants on top of all that, the conventional method is losing its practical and financial viability. Prior work has already demonstrated that optimal control functions can be automatically developed with reinforcement learning (RL); since the resulting functions are represented by artificial neural networks, they lack explainability, a circumstance which renders them challenging to employ in production vehicles. In this article, we present an explainable approach to automating the calibration process using residual RL which follows established automotive development principles. Its applicability is demonstrated by means of a map-based air path controller in a series control unit using a hardware-in-the-loop (HiL) platform. Starting with a sub-optimal map, the proposed methodology quickly converges to a calibration which closely resembles the reference in the series ECU. The results prove that the approach is suitable for the industry where it leads to better calibrations in significantly less time and requires virtually no human intervention.
Kampmeier, AndreasBadalian, KevinKoch, LucasLee, Sung-YongAndert, Jakob
The increasing regulatory complexity in automotive development places significant pressure on engineering teams to derive complete and correct requirements. This paper presents a multi-agent-based large language model (LLM) workflow designed to support requirement extraction from technical specifications and regulatory documents in compliance with automotive requirement guidelines. The approach structures the requirement derivation process across collaborating agents that interpret specification and regulatory text, generate candidate requirements for the early engineering activities, and cross-validate their outputs to improve consistency and traceability. To evaluate the applicability of the workflow in an industrial context, we applied it to the draft Euro 7 emissions regulation. The agents produced requirements for relevant functional domains, which were subsequently reviewed by domain experts at FEV. The evaluation focused on correctness, completeness, and coverage. Results indicate that the agentic workflow can achieve high alignment with expert expectations, demonstrates robust coverage of regulatory intent, and reduces manual effort in the early requirements engineering phase. The findings highlight the potential of structured multi-agent LLM systems to accelerate compliant software development processes and to enhance the reproducibility and quality of regulatory requirement interpretation in the automotive domain.
Abdalla, AbdelrahmanSchäfers, LukasSchmidt, FabianSchaub, JoschkaLee, Sung-YongAndert, Jakob
Semi-active suspension systems enhance ride comfort and handling performance by adaptively modulating damping characteristics. However, conventional model-based controllers often fail to maintain optimal performance under uncertain and time-varying vehicle conditions. This article proposes Bayesian Optimization–Tuned Proximal Policy Optimization with Non-Parametric Rewards (BO-NRPPO), a novel reinforcement learning (RL) framework that integrates Bayesian Optimization (BO) with Proximal Policy Optimization (PPO) and a non-parametric reward function (NRF). The proposed approach enables adaptive self-tuning, data-driven reward shaping, and uncertainty-aware policy learning. Moreover, a Trapezoidal Simple Moving Average (TSMA)–based reward normalization scheme is introduced to accelerate convergence and stabilize training. Simulation results across diverse driving scenarios demonstrate that BO-NRPPO outperforms the passive suspension, the classical Linear Quadratic Regulator (LQR), and PPO with parametric rewards. Specifically, compared to the passive suspension and the LQR baseline, BO-NRPPO achieves up to 6.63% and 5.14% improvements in handling stability, respectively. Concurrently, it delivers maximum enhancements of 46.96% and 42.55% in ride comfort over these two baselines. For real-world vehicle applications, this adaptive self-tuning capability significantly reduces the time-consuming manual calibration efforts typically required in chassis development. Furthermore, Hardware-in-the-loop (HiL) validation confirms its real-time applicability and robustness under uncertain driving conditions, highlighting its immense potential as a scalable intelligent suspension control solution.
Chen, GuoyingWang, XinyuWang, JiaqiZhan, XinwangBi, ChenxiaoCong, ShiqiHua, MinSun, TianjunGao, Zhenhai
Simulations can only be searched, reused and leveraged as training data for machine learning methods if suitable metadata are related. Manually obtaining these metadata is time-consuming and requires expert knowledge. Consequently, there often is a lack of metadata and this prohibits the reutilization of simulation data. Therefore, automated frameworks for metadata extraction are essential to obtain metadata information quickly, effortlessly and cost-efficiently. At present, there are no toolboxes for Finite-Element-Simulation data. Nevertheless, machine learning methods are a promising solution for this task. Training classical supervised machine learning methods for metadata generation often faces the lack of labeled data since manual labelling can be very costly. Therefore, rule-based extraction algorithms are used as an alternative for fundamental metadata extraction. For more enhanced tasks they are often not feasible. Active Learning is a suitable technique to overcome this contradiction. Here, as only necessary data points are labelled, it is possible to use machine learning methods for metadata extraction even for simulation models. In this work, the specific advantages and challenges of metadata are shown for Noise-Vibration-Harshness simulations. The focus of this contribution lies on the workflow for active learning applied on Finite-Element-Models, including data preprocessing and first training loops. Additionally, the benefits and challenges of high-level feature engineering on data size and extractor model performance are investigated. Moreover, the results show an extended Active Learning workflow which helps to investigate the given data, enhance the feature engineering and therefore the Machine Learning model quality. Finally, based on these results, recommendations for further development of these techniques are provided.
Luegmair, MarinusGröttrup, Sören
Achieving best-in-class Noise, Vibration, and Harshness (NVH) in electric powertrains demands a paradigm shift in development methodology. This paper presents a practice-oriented overview of simulation methods in NVH development methodology for electric drive units. This includes target cascading and multi-objective optimisation, and by attacking NVH at the source using KPIs early in the design cycle, significant reductions in development time and reliance on traditional testbed loops are realised. Machine learning (Neural Network) algorithms are utilized to find the best-in-class design, using multi-objective optimisation as well as refining simulation accuracy by adding tolerance effects while target cascading ensures alignment of system-level performance objectives down to subsystem contributions. Combined, these strategies enable rapid and robust NVH optimisation, using simulation for next-generation electric powertrain development. Several applications and real-life examples demonstrate how simulation helped with NVH issue identification or improvement.
Mehrgou, MehdiGarcia de Madinabeitia, InigoGraf, BernhardGojo, Josef
In this study, we propose a methodology for predicting the acoustic modes and natural frequencies of a sedan using artificial intelligence and demonstrate the feasibility of controlling its acoustic characteristics by modifying the hole distribution of the package tray. In typical sedan structures, the cabin cavity and trunk cavity are acoustically coupled through holes in the package tray. The distribution of these holes significantly affects the natural acoustic modes and frequencies of the vehicle. However, once the exterior shape of the vehicle is finalized during the design stage, options for structural modifications to mitigate noise issues caused by these modes become extremely limited. To address this challenge efficiently, we develop a deep learning-based neural network model trained on data derived from a simplified acoustic analysis model of a sedan that includes a package tray. Finite element analysis is performed to generate acoustic modes and natural frequencies, which serve as training data, for various hole distributions. The trained model is then used to predict acoustic natural modes and natural frequencies from unseen input images representing different hole configurations in the package tray. These predictions are made in a fraction of the time required for traditional simulation methods, thereby validating the model’s effectiveness. Furthermore, we demonstrate that the latent variables embedded in the trained model can be manipulated to control the acoustic modes and natural frequencies of the sedan. This indicates the potential for artificial intelligence-driven acoustic design optimization in early-stage vehicle development, offering both time efficiency and design flexibility without physical prototyping or extensive simulations.
Lee, Jin WooCho, JaehoNam, YounsicHan, Yongha
Items per page:
1 – 50 of 2670