Browse Topic: Machine learning

Items (1,646)
Due to the rapid growth in ride-hailing services and the steady expansion of urban rail transit networks, understanding the complex interaction between the two is important. Most existing studies only focus on the overall competition and cooperation effects. Few scholars discuss local spatial differences and how the built environment works in detail. To investigate the spatial heterogeneity of built environment factors between the ride-hailing and the subway systems by situation and location, our study categorizes 3.71 million ride-hailing trips into three types: Proximal-to-Station Trip (PST), Medial-to-Station Trip (MST), and Distal-to-Station Trip (DST). We employed the values of built environment factors per grid as independent variables, and the grid-based volume of each ride-hailing trip category (PST, MST, DST) as the dependent variable. Additionally, we conducted analysis mainly with the multi-scale geographically weighted regression model (MGWR). We found that ride-hailing trips and subway trips have competition, cooperation, or complementarity relationships, which depend on trip purposes and public transit accessibility. Besides, we think that this relationship relies on the built environment, and it also shows clear spatial heterogeneity. The impact of different built environment factors has significant variation and spatial heterogeneity: Catering, Shop, and Bus Stop continuously trigger the demand for ride-hailing services, while Residence generally reduces the use. In PST, Enterprise, Medical, and Residence inhibit ride-hailing due to the high subway accessibility. In DST, these functions conversely boost ride-hailing usage due to limited bus coverage. These findings provide crucial insights for smart transportation management, multi-modal, and location-specific strategies to boost sustainable urban transportation development.
Zhao, Shiya, Xiao, Lingling
Accurate and real-time traffic state prediction is fundamental to optimizing urban traffic control and improving the service quality of online mapping platforms. While data-driven deep learning methods have shown promise in handling large-scale data, most existing approaches suffer from limitations due to their reliance on simplistic modeling components that cannot fully capture complex and multi-scale spatiotemporal dependencies. Unlike prior studies that predominantly utilize static graph structures or sequential processing, this research addresses the challenge of modeling dynamic correlations by proposing a Spatiotemporal Parallel Fusion Network. The theoretical framework of this study integrates graph spectral theory with advanced attention mechanisms to model both the physical and semantic topology of traffic networks. The proposed methodology features a multi-source spatial modeling framework that fuses a static connectivity graph, an inverse distance weighted graph, and a flow similarity graph via multi-relational graph convolution. This approach allows for the simultaneous capture of physical connectivity and dynamic traffic patterns. Furthermore, utilizing a parallel architecture, the temporal modeling component integrates multi-head self-attention with two-dimensional causal dilated convolutions. A cross-attention fusion module is further introduced to dynamically regulate spatiotemporal interactions. Comprehensive tests conducted on four practical traffic datasets indicate that the introduced model significantly outperforms thirteen cutting-edge benchmark methods. The study confirms that the proposed multi-graph strategies and parallel architectures effectively reduce mean absolute error and root mean square error, validating the efficacy of the proposed solution for the advancement of intelligent transportation systems.
Chen, Lulu
Although autonomous vehicle (AV) technology is regarded as a transformative force for enhancing road safety, collisions remain inevitable during the transition phase toward large-scale commercial deployment, making it crucial to investigate the causal mechanisms of accident severity for optimizing AV systems and improving transportation infrastructure. To address the dual challenges of extreme class imbalance and the lack of interpretability in existing studies utilizing real-world data, this paper proposes an interpretable accident severity analysis model integrating Random Over-Sampling Examples (ROSE) with the LightGBM algorithm and SHAP (SHapley Additive exPlanations) framework. Utilizing real-world testing data from the California Department of Motor Vehicles (DMV), this study first establishes a rigorous “Split-then-Sample” strategy to effectively mitigate class imbalance while eliminating data leakage risks, and subsequently constructs the ROSE-LightGBM classifier. Experimental results demonstrate the model’s superior robustness in identifying minority class (injury) accidents, significantly outperforming baseline models such as Random Forest and XGBoost with a Recall of 80.00% and an F1-score of 77.42%. SHAP-based multi-dimensional analysis further reveals that severe vehicle damage, high point-of-interest (POI) complexity, and collisions involving vulnerable road users (VRUs) are dominant factors escalating accident severity, whereas high-level sensor configurations equipped with LiDAR and physically separated roads exhibit significant risk-mitigating effects; additionally, the study quantifies the non-linear threshold relationships between environmental features and accident risks. The analytical framework proposed in this paper not only provides an effective paradigm for modeling imbalanced data but also offers actionable, scientifically interpretable insights for the formulation of AV safety strategies.
Hou, Yi, Wu, Shaoping, Dong, Chunjiao, Ma, Jihui, Lu, Zhenzhen, Liu, Zhe, Sun, Dongyuan
To investigate pilot fatigue during tunnel navigation, we developed a simulated experimental platform based on a large-scale ship maneuvering simulator (LSSM). EEG data were collected from 12 pilots while they navigated the tunnel using a saline-electrode wireless portable EEG device (EMOTIV FLEX). We processed the pilot’s EEG signal data using the toolbox, which was EEGLAB (via MATLAB), to extract the wave characteristics that reflect the fatigue/multi-branch longitudinal short memory (LSTM) performance (i.e., Alpha to Delta waves). We developed a multi-branch LSTM model on the foundation of an optimized attention mechanism. The conventional machine-learning models showed limited performance, with random forest achieving an accuracy of 82.75% and an AUC of 0.75, and SVM achieving an accuracy of 75.00% and an AUC of 0.70. The experiment findings revealed that using EEG temporal features in conjunction with the task performance of pilots in the LSTM model delivered an accuracy of 91.67% and an AUC of 0.95, and precision/recall levels above 89. The study offers an efficient and accurate technical solution to check the fatigue of pilots in tunnel navigation and will substantially contribute to the safety of the cargo shipments with the help of the inland waterways.
Yang, Yichen, Xu, Yanmin, Guan, Hongxu
Considering the huge influence of personal driving habits on the actual driving mileage of electric vehicles (EVs), together with the weak interpretability of most existing evaluation models, this research combines logistic regression and scorecard methods to build a new assessment model that focuses on energy consumption and driving behavior, and a complete closed-loop technical framework has been set up in the research process, covering key steps such as raw data preprocessing, feature mining and screening, model design and construction, as well as final score output. Taking real collected vehicle operation data as the basic research material, we extracted 17 feature indicators from three major dimensions, including driver pedal operation habits, real-time vehicle operating state, and actual energy consumption performance, and adopted the mutual information (MI) calculation method to select eight core features with strong differentiation for 100-kilometer power consumption. Since obvious nonlinear correlations exist between these core indicators and vehicle energy consumption, we adopted decision tree-based grouping processing, weight of evidence (WOE) encoding, and information value (IV) quantitative analysis to realize feature discretization and optimization, and then constructed a scoring model relying on a logistic regression algorithm. The verification results demonstrate that the proposed model has an Area Under the Curve (AUC) value of 0.83 when identifying driving behaviors with high or low energy consumption, proving its excellent differentiation ability. This model turns complicated driving behaviors into intuitive quantitative scores, effectively making up for the defects of traditional black-box models that cannot offer practical guidance for energy-saving optimization, and it can also provide solid theoretical reference and practical application support for promoting energy-saving driving behavior among EV users.
Deng, Shuxiang, Guan, Di, Wu, Yunzhu, Liang, Yongkai, Ma, Kunqi, Xu, Hang
With the rapid development of e-commerce and on-demand logistics, distribution centers are confronted with the dual challenge of surging cargo volumes and insufficient loading efficiency. In scenarios involving the parallel transportation of multiple vehicles, loading decisions across different vehicles exhibit mutual coupling characteristics. Without coordinated global planning, such unregulated decision-making not only diminishes the overall space utilization rate of the fleet but also gives rise to risks of uneven load distribution and compromised transportation stability. Existing research in this field, however, remains primarily focused on single-vehicle loading optimization, a limitation that renders it unable to meet the practical operational demands of modern logistics systems. To address these issues, this paper investigates the three-dimensional cargo loading problem from a multi-vehicle collaborative perspective and proposes a multi-agent reinforcement learning (RL) method based on Value Decomposition Networks (VDN), which is designed to realize the joint optimization of transportation efficiency and vehicle operational safety. Specifically, the paper models the multi-vehicle loading process as a Partially Observable Optimal Multi-Agent Markov Decision Process (OPOMDP). By adopting a centralized training and distributed execution framework, each vehicle is enabled to generate a globally coordinated loading strategy solely on the basis of its local observation data. On this foundational framework, two key mechanisms are devised to enhance the performance of the loading strategy. First, a stability cost metric, which is constructed based on dynamic shifts in the vehicle’s center of gravity, is introduced to quantify the rollover risk of the vehicle; this metric is then integrated into the reward signal of the RL model, thereby guiding the strategy to maintain transportation safety while improving the cargo loading rate. Second, a specialized placement reward mechanism is proposed to incentivize each agent to proactively explore optimal cargo rotation methods and compact spatial packing structures, which in turn reduces intra-layer voids and elevates the overall volume utilization rate of the cargo compartment. Experimental results demonstrate that the proposed method outperforms traditional heuristic algorithms and mainstream RL baseline algorithms by a significant margin in terms of loading success rate, volume utilization rate, and weight utilization rate. Meanwhile, the stability constraint mechanism embedded in the model effectively mitigates the vehicle rollover risk, ultimately achieving the synergistic optimization of loading efficiency and transportation safety.
Meng, Xiancheng, Xue, Guixiang, Yin, Haiyang, Sun, Yunjun
To address the energy management strategy of hybrid tracked vehicles, and in order to achieve a more rational power distribution among the engine-generator set and the power battery, thereby improving overall fuel economy and system operating efficiency, this paper proposes an integrated energy management control strategy based on a rule-based strategy and deep Q-networks (DQN). Considering the structural characteristics and operating features of the series-hybrid tracked vehicle power system, a relatively comprehensive vehicle dynamics and energy management simulation model is established on the MATLAB/Simulink platform. We introduce a deep reinforcement learning method. The agent’s observation inputs are the vehicle’s operating state and power demand, whereas fuel consumption, battery limits, and various system operating restrictions are folded into the optimization objective. By continuously interacting with the model environment, the agent can continuously learn energy allocation strategies for the powertrain. After training is completed, the trained agent is tested under specific operating conditions, and the proposed integrated control strategy is compared with traditional constant-voltage control and a multi-point speed control strategy. Simulation results show that, under the same test conditions, the integrated control method can optimally adjust engine speed, thereby reducing the number of operating points in the low-efficiency range. While ensuring vehicle dynamic performance and system stability, this strategy can reduce fuel consumption by at least 3.66%, thereby demonstrating its feasibility and effectiveness in energy management for series-hybrid tracked vehicles.
Yan, Zhipeng, Shi, Feina, Zhao, Zhiying, Niu, Liye, Geng, Jiangbo
Single-point incremental forming (SPIF) is highly flexible and successful in sheet metal forming, particularly in the fabrication of complex components without requiring molds and dies. Nevertheless, the accurate definition and definition of the best forming angle in SPIF is still challenging, especially in limited experimental conditions. This specific research focuses on regression modeling using machine learning algorithms in finding the best possible forming angles in SPIF, using a limited number of experimental trials. Some of the most significant factors in SPIF, namely wall angles, tools, and experimental conditions, become the inputs in regression models. Based on research, five individual regression models were made and compared using the leave-one-out cross-validation (LOOCV) technique for accurate assessment and evaluation despite the limited experimental conditions. Hyperparameter tuning was conducted using GridSearchCV, while extra optimizations were implemented in the bigger models, specifically in the ensembled models. Based on research findings, boosted models, particularly the use of XGBoost and CatBoost, were shown to outperform other regression models in modeling the SPIF process due to their prowess in modeling nonlinear interactions among other process parameters despite limited experimental conditions, which emphasizes the robustness and flexibility offered by machine learning techniques in modeling any process, notably in SPIF process modeling.
Vijayakumar, M. D., Pradeep Castro, Ponnuswamy, Kandavel, Narasimhan, Sreehari, Thirunangur Raghu
Consumable spare parts for civil aircraft are characterized by intermittent demand patterns. Constrained by long order replenishment cycles, traditional single-step demand forecasting struggles to effectively support inventory decisions. Moreover, research on intermittent demand forecasting primarily focuses on single-step, short-term cycle predictions, while multi-step forecasting using advanced deep learning methods remains in its early stages. To address this issue, this study introduces a multi-step forecasting approach based on the Transformer architecture, employing a MIMO strategy to achieve long time series forecasting for intermittent demand. The model is validated using a dataset from an airline maintenance base consisting of 360 intermittent demand items. A comparative analysis is conducted against three traditional time series models (SES, Croston, and SBA) and several neural network architectures (FNN, GRU, and LSTM). The performance of different methods was compared using six evaluation metrics. Results demonstrate that the Transformer consistently delivers outstanding performance across various forecast horizons. Notably, in long-horizon forecasting tasks, the model’s advantage in prediction accuracy becomes increasingly significant as the forecast step increases.
Zhang, Zhaomin, Zhao, Xin, Zhang, Jian, Fu, Weifang
With the continuous advancement of collaborative unmanned aerial vehicle (UAV) swarm applications, the importance of intent recognition for aerial swarms in UAV management and control has become increasingly evident. To address the shortcomings of existing intent recognition methods, this paper proposes a knowledge-driven aerial swarm target intent recognition method designed for few-shot scenarios. First, a target trajectory clustering algorithm is designed by combining trajectory similarity measurement (which considers directional features) with density peaks clustering. This algorithm groups targets exhibiting the same or similar motion trends, thereby transforming the swarm target intent recognition problem into a swarm member intent recognition problem, which reduces the complexity of intent recognition. Second, considering the behavioral dynamics inherent in aerial swarm collaborative applications, a target intent inference model is constructed based on a Dynamic Bayesian Network (DBN). Utilizing the characteristic information of swarm targets at different time points, this model achieves intent inference and recognition through a hierarchical process: “sub-intent inference → accumulation of historical sub-intent sequences → intent inference.” Finally, simulation experiments are conducted to verify the feasibility and effectiveness of the proposed method.
Jiang, Jianxiong, Gui, Yang, Sun, Peng, Yang, Xinyu, Yang, Guangwei, Zhong, Yun
To realize automated grasping of randomly placed C-type circlips in automotive and mechanical assembly processes, this paper proposes an identification method based on the improved YOLOv11 algorithm and an exclusive association strategy. First, a detail enhancement convolution (DEConv) module is introduced into the C3k2 module of the YOLOv11 network to enhance the ability to capture the edges and high-frequency details of the circlips; Second, the parallel patch-aware attention (PPA) module is adopted in the detection head instead of the traditional convolution, and through multi-branch fusion and channel-spatial attention mechanism, it adaptively focuses on the key areas of small targets; Finally, the normalized Wasserstein distance (NWD) loss function is used instead of the IoU loss to alleviate the gradient disappearance problem in the positioning of small targets and improve the model’s detection accuracy. Based on precise detection, this paper innovatively proposes a circlip-hole association strategy based on multi-level spatial constraints, achieving accurate pairing of randomly placed circlips through bounding box-inclusive attribution judgment, distance screening, and selection of the best circlip-pair. Experimental results show that compared with the original YOLOv11 model, our model improves the mAP50 and mAP50-95 significantly by 5.7% and 3.3%, respectively; its association accuracy reaches above 95%, and the average processing time is 142 ms. Both in terms of accuracy and speed, our model meets the requirements of industrial real-time grasping and provides reliable prior information for the subsequent precise clamping of the robotic arm.
Wang, Chuang, Li, Jiyan, Wang, Yongming, Tan, Libin
This paper presents GBC-YOLOv5s, a lightweight vehicle detection model designed for unmanned aerial vehicle (UAV) applications in the growing low-altitude economy sector. By utilizing the baseline YOLOv5s architecture, we integrated three structural enhancements to achieve an optimal trade-off between algorithmic precision and processing expenses. First, ghost convolution is used to exploit feature map redundancy, substantially cutting both parameter count and computational overhead. Second, a Bi-Level routing attention (BRA) mechanism is integrated. This allows the model to capture global context along with fine-grained local features of vehicle targets effectively. Third, standard upsampling is replaced with content-aware reassembly of features (CARAFE). This better preserves feature details, which proves particularly important for detecting small objects. Testing was conducted on a custom UAV-based vehicle dataset. GBC-YOLOv5s achieves 86.6% average precision (AP) with only 4.7 M parameters. This represents a 34.7% reduction from the original YOLOv5s with 7.2 M parameters. The accuracy drop is merely 0.2%. Ablation experiments validate both the individual impact and synergistic nature of each proposed module. Results show that GBC-YOLOv5s reaches a practical balance between performance and efficiency. It offers a workable solution for real-time vehicle detection on resource-limited UAV platforms and intelligent traffic monitoring applications.
Zhang, Qingpei, Yu, Tianhao, Zhao, Zheming
In the field of unmanned aerial vehicle (UAV) signal recognition based on deep learning, complex electromagnetic environments pose severe challenges. Among them, interference suppression and recognition robustness under low signal-to-noise ratio (SNR) conditions are major research focuses. To tackle these issues, this paper proposes a robust radio frequency (RF) signal recognition method for UAVs. The method is built on an improved ResNet18 network. It separates the video transmission signals and control signals of UAVs, and adopts mixed-SNR training. These designs greatly strengthen the model’s ability to perform reliable recognition. First, the short-time Fourier transform (STFT) is used to convert UAV I/Q signals into time-frequency images. A refined time-frequency masking strategy is then adopted. This strategy separates and suppresses interference from control signals according to their bandwidth differences from video transmission signals, thus improving signal quality. Second, a squeeze-and-excitation (SE) channel attention module is embedded into each residual block of ResNet18. This structure enables the model to concentrate on critical features. The classification head is also rebuilt into a three-layer fully connected network to strengthen nonlinear transformation and complex feature interaction. Given the performance variation of the model across different SNR levels, a mixed-SNR training strategy is adopted. Fine grid search experiments are conducted to identify the optimal data distribution for maximizing recognition performance. Experimental results on the public DroneRFb-DIR dataset show that the improved ResNet18 model achieves a 3.07% increase in classification accuracy. The time-frequency masking method brings an extra performance gain of 1.08%. Although the mixed-SNR strategy causes a slight accuracy drop at the high SNR of 20 dB, it improves recognition accuracy by 8% to 53.01% within the low-SNR range of −10 dB to 10 dB. It also effectively reduces misclassification between different UAV types and enhances the model’s robustness. The proposed method provides a feasible solution for accurate and stable UAV signal recognition in complex electromagnetic scenarios.
Yu, Jiankun, Chen, Yongyou
LiDAR place recognition plays a crucial role in SLAM systems and robot localization. To address the drawback of existing hand-crafted descriptors—namely, their limited topological representation capability in complex scenes—and the high training cost and strong dependence of semantic-assisted methods, our study proposes a label-free hybrid geometric graph-based LiDAR place recognition algorithm, termed Geometric Place Recognition (GeoPR). GeoPR exploits the intrinsic geometric structural information of point clouds. Our algorithm adopts a preprocessing module to directly extract stable geometric primitives from raw point clouds as geometric nodes. On this basis, we design a hybrid geometric graph descriptor by introducing logarithmic distance weighting and Gaussian soft binning to enhance robustness against discretization errors and dynamic disturbances. During the matching process, our research adopts a two-stage “global search - geometric verification” framework: using a KD-tree for rapid candidate screening, followed by a geometric consistency verification module based on correspondences to eliminate false positive loop closures. Experiments on public datasets show that, without any training data or GPU acceleration, GeoPR achieves significantly higher precision and recall than mainstream hand-crafted methods, while its scene adaptability is comparable to advanced deep learning methods. Therefore, this work provides an efficient and reliable place-recognition solution for automatic driving and lightweight mobile robot platforms.
Cao, Bingyin, Zhang, Xudong
In recent years, Reinforcement Learning (RL) has been increasingly applied to the collision avoidance systems of Unmanned Aerial Vehicles (UAVs) due to its high performance in complex tasks and relatively low online computational cost. Despite these advantages, the lack of interpretability in RL and decision-making processes poses significant challenges to safety verification, which severely limits the practical deployment of RL-based autonomous flight systems in safety-critical scenarios. Ensuring safety during both training and testing remains the key for real-world applications. To address this issue, a Run Time Assurance (RTA) filter framework for a UAV collision avoidance system based on RL is proposed. The framework continuously monitors the avoidance actions generated by RL and ensures that unsafe actions can be identified and corrected in real time. The framework provides security while retaining the performance advantages of RL. First, a discrete kinematics model of the UAV and a collision avoidance environment for a three-dimensional multi-intruder UAV are established. The Proximal Policy Optimization (PPO) algorithm is used to train the UAV avoidance policy. Secondly, the RTA filter with safe constraints is designed based on this model. To reduce the position error between the discrete kinematics model and the continuous dynamics model of the UAV, optimization is performed using the Sequential Least Squares Programming Method (SLSQP). The RTA is applied to the training and testing phases of UAV collision avoidance systems to monitor and correct unsafe actions online. Finally, three training and testing methods are set: no RTA, RTA without punishment, and RTA with punishment. Comparing different relaxation coefficients allows for analyzing their impact on filter performance. Simulation results demonstrate that the RTA effectively enhances safety throughout training and testing. Compared with the no RTA method, the collision rate is reduced by 7.04%, and the success rate is increased by 2.68%. In extreme collision avoidance scenarios, the average collision rate obtained by different reinforcement learning algorithms is reduced by 21.46%, and the average success rate is increased by 15.54%.
Ma, Zan, Yan, Liqin, Bai, Jie, Liu, Yubin, Sun, Shuguang
The Unmanned Aerial Vehicle (UAV) swarm require efficient scheduling to maintain a high mission success rate in dynamic environments. At present, most methods are designed for static targets. But real buoys move and drift. The recommendation of “short-term drift theory” leads to the failure of multi-scheduling. We propose a Dynamic Hierarchical Group Control-Deep Reinforcement Learning (DHGC-DRL) framework based on hierarchical deep reinforcement learning. The framework follows a divide-and-conquer strategy and splits the problem into two parts. The first part is a global task assignment. The second is local path planning. Using the Coordinated Dynamic Task Allocation (CDTA) transformer in upper-layer building policies affect dynamic requests and quickly extracts information from the environment. This enables the model to find the best task assignment. Combinatorial Multi-Armed Bandit (CMAB) and the transformer module are used in the lower layer. These blocks modify the path and help the UAV adapt to the movement of the buoys. This makes the planning result more accurate. We also propose a coordination mechanism for the reward function. The mechanism uses a joint reward function, which allows two layers to exchange messages with each other. This can achieve real-time and fast solutions. The algorithm is tested by using the Double-Level Deep Reinforcement Learning (DL-DRL) baseline in parallel. The DHGC-DRL algorithm achieves faster convergence and improved training efficiency. The mission success rate increases by 25%. Path feasibility improves by 14.3% and overall performance is improved by 8.2% compared to the baseline.
Zhang, Xin, Du, Xuefeng
Temperature management in both residential environments and commercial vehicle cabins are rapidly evolving toward intelligent, interconnected climate-control ecosystems. However, current systems operate in isolation, with no mechanism for synchronizing thermal comfort preferences between a user’s smart home and the cabin environment of a truck used for either personal mobility, last-mile delivery, or commercial logistics. This disjointed control results in inconsistent comfort experience, unnecessary HVAC energy consumption, and the inability to anticipate user needs based on transition patterns (e.g. leaving a cool home for a hot truck cabin) between home and vehicle. To address this technical challenge, this research paper provides bidirectional communication temperature-synchronization framework in which a smart-home HVAC app communicates seamlessly with the truck’s cabin climate-control system, integrating user-specific comfort profiles(such as temperature preferences, real-time cabin and home temperature, humidity levels, HVAC operating states, outside weather forecasts, trip start times, thermal-soak estimates etc.) and environmental prediction models to maintain consistent thermal comfort across both environments. Based on user’s comfort profile, our framework derives the Comfort Deviation Index (CDI), Transient Temperature Gradient (TTG) and applies predictive algorithms to precondition the cabin or home environment accordingly. The AI technique Long Short-Term Memory (LSTM) network is used for forecasting temperature transitions across home-to-vehicle. Our technique becomes even more pronounced in small-haul trucks that operate under frequent stop-and-go conditions, where rapid cabin temperature fluctuations are common. There is currently no existing research or framework that enables seamless temperature synchronization between a user’s home environment and the truck cabin. A smart-home–to-cabin synchronization approach would fill this gap by intelligently accounting for outside weather, trip duration, thermal soak, and energy-efficiency constraints. Additional challenges include ensuring secure communication over cloud and edge networks, managing latency constraints for real-time actuation, and preventing conflicts between home HVAC automation and automotive HVAC control logic
Prajapati, Saurav, Sagar, Kshitij Singh, B. S., Sanchitha
Automotive radar systems do not authenticate the signals they receive, allowing an external transmitter to create phantom objects that can trigger automatic emergency braking (AEB). This paper quantifies the threat using real radar measurements from the public RadarScenes dataset and evaluates a software countermeasure that requires no hardware modification. Computer-generated phantom radar returns were added to 200 motorway sequences and adjusted so the OS-CFAR detector accepted them as real objects. In the undefended system, radar detections were tracked as obstacles and braking was initiated when an object appeared within 30 m, resulting in unnecessary braking in 87.0 percent of attacked sequences. The proposed two-layer filter combines kinematic plausibility gating with camera-radar cross-validation. Layer 1 checks track history and physical movement over three consecutive frames, while Layer 2 reduces confidence when a clear-weather radar detection lacks camera confirmation. With both layers active, false braking fell to 3.5 percent, with no genuine detections missed across 200 genuine sequences. SHAP analysis of a Random Forest classifier identified kinematic implausibility as responsible for 62.3 percent of correct suppressions and produced a machine-readable evidence record to support cybersecurity documentation under ISO/SAE 21434 and safety documentation under ISO 21448. These results apply to an offline replay experiment using synthetic injections. Validation with real attack hardware, synchronized camera data, and vehicle testing remains future work.
Sehgal, Vishal, Sekaran, Nikhil
The brake pedal assembly is a critical load bearing automotive component that must perform reliably under severe service conditions, especially during emergency braking events. Structural assessment is commonly carried out using computationally intensive non-linear finite element analysis, which can limit rapid design iterations during early product development. This study presents a physics informed surrogate modelling framework for predicting displacement and von Mises stress responses of a brake pedal directly from unstructured three dimensional mesh geometry. A structured dataset was created through parametric morphing and a Design of Experiments approach by varying pedal arm angle, pedal moment length, flange/web thickness, and combined geometric conditions. The simulation database was developed using conventional solvers and the generated cases were used to train Physics Artificial Intelligence based on the Transformer Neural Simulator architecture. The trained model showed close agreement with conventional solver results for both peak response values and contour distributions. For representative testing cases, predicted displacement and stress responses remained within acceptable engineering tolerance compared with the finite element baseline. After training, the PhysicsAI workflow reduced the prediction turnaround time compared with the conventional solver workflow. The results suggest that Design of Experiments driven datasets combined with geometric deep learning can provide an efficient workflow for rapid brake pedal structural assessment during concept development and iterative design studies.
N., Rengarajan, Praveen, Kannan, Philip Thomas, Febin
Accurate representation of strain-rate dependent stress–strain behavior is essential for reliable structural analysis and crashworthiness assessment in automotive engineering as vehicle structures and safety-critical components experience a wide range of loading conditions, from quasi-static deformation during forming and durability events to high strain-rate loading during crash and impact scenarios. Conventional material characterization approaches rely on extensive experimental testing to generate strain-rate dependent constitutive data, including specialized high strain-rate techniques that are costly, time intensive, and require dedicated infrastructure, which often limits the availability of comprehensive material data during early vehicle development phases and supplier evaluation. This paper presents a data-driven methodology for predicting strain-rate dependent stress–strain behavior of automotive steels using limited experimental input data. The proposed approach integrates physically meaningful material properties with a machine-learning framework to construct a generalized stress prediction model that remains consistent with classical material mechanics principles. Fundamental parameters governing material behavior, including strain, strain rate, yield stress, density, Poisson’s ratio, and elongation limit, are used as inputs to capture the combined influence of deformation state, loading rate, and intrinsic material characteristics. A Random Forest regression algorithm is employed to model nonlinear relationships without relying on predefined analytical hardening or rate-sensitivity formulations. The methodology is developed and validated using experimentally obtained stress–strain data from multiple automotive steel grades under quasi-static and dynamic loading conditions, and model performance is evaluated through correlation analysis between predicted and experimental stress responses across the full strain range. The results demonstrate strong agreement with experimental observations and robust generalization capability across different materials and strain-rate regimes, with minor deviations observed at very high strain levels attributed to limited data density near fracture conditions. The proposed framework significantly reduces dependence on extensive physical testing while maintaining accurate representation of material behavior, providing a practical and efficient alternative for material characterization that supports faster material selection, reduced testing effort, and improved robustness of automotive computer-aided engineering simulations.
Shedolkar, Pravinkumar, Dudhewar, Swapnil, Shukla, Anurag, Mhaskar, Dnyaneshwar, Chavan, Avinash, Kakade, Ajit
Oil canning, characterized by localized buckling and surface waviness in thin metal panels, is a critical quality parameter in the automotive industry. Since outer panels serve both structural and aesthetic functions, even minor surface distortions can adversely affect perceived vehicle quality. Traditionally, evaluation of oil canning relies on extensive nonlinear Finite Element simulations. For example, assessing a single hood design variant may requires thousands of nonlinear Finite Element analyses. Considering multiple design iterations across major vehicle panels including hood, roof, doors, fenders, and trunk, the total number of simulations can exceed tens of thousands, leading to substantial computational cost and development time. Consequently, this conventional approach is both time intensive and resource demanding. To address these limitations, advanced Machine Learning techniques are employed to significantly reduce simulation effort as well as accelerate design exploration. Initially, a deep learning based one-dimensional Reduced Order Model is developed to predict load–displacement response and oil canning propensity for a given panel. Indentation points on the hood panel are selected using a space filling Latin Hypercube sampling method. The Machine Learning based one-dimensional Reduced Order Model is trained by using 300-500 indentation points and validated using approximately 40-50 unseen indentation points. The verification is performed on multiple hood geometries. The predicted results demonstrate strong correlation with Finite Element generated load–displacement curves with that of ML based 1-dimensional reduced order model. Furthermore, to enable rapid design exploration across different panel geometries, a dataset of 40 different hood designs is utilized. A geometric deep learning approach is applied to cluster panel geometries, and the resulting representations are used within a 3-dimensional deep neural network tool Ansys SimAI™ to predict oil canning propensity and quantify it using the canning ratio. Validation study is carried out on unseen panel geometries, and the Machine Learning predicted oil canning zones show good agreement with Finite Element identified critical regions. The proposed Machine Learning framework reduces prediction time from hours to minutes, offering a scalable and cost-effective solution for early-stage design evaluation and mitigation of oil canning in automotive panels.
Sarnobat, Rajvardhan Vilasrao, Kumar, Dinesh, Jawale, Akshay, Balaji, Ananya, Kulkarni, Shreya
Children dying in hot cars is not a new problem. Since 1998, more than 1,000 children have died this way in the United States alone — roughly 37 every year. That number has barely been reduced, which tells something alarming: the solutions built so far aren't solving it completely. Door-based reminders are a good example of why. They trigger when a rear door opens, but opening a door doesn't confirm a child was left behind. Single-sensor setups have their own gaps too; no single sensor handles every real-world scenario reliably on its own. Multi-sensor systems get closer, but most use rigid threshold-based rules that fall apart the moment conditions deviate — a sleeping infant barely moves, a child under a blanket reads completely differently on a thermal sensor. This paper proposes Smart-ROPS (Smart Rear Occupant Protection System), an embedded ML architecture that targets these specific failure modes without needing expensive hardware or complex software. Four sensor types feed the system: millimeter-wave radar, CO₂, thermal infrared, and ultrasonic. Data from these passes through three layers — the Perception Layer handles initial sensor fusion, the Processing Layer runs a decision tree as the primary classifier, and when confidence is low, a shallow neural network steps in for ambiguous cases. If a risk is confirmed, the Intervention Layer manages the response: starting with a smartphone alert and cabin ventilation, escalating to emergency services if needed. This is presented as an architectural design contribution — hardware prototyping and real-world testing are left for future work.
Shrimal, Harsh
Overtaking another vehicle on a two-lane road can turn dangerous within a few seconds, especially when the gap to an oncoming vehicle is short or that vehicle is moving fast. Many driver-assistance systems rely on a simple, fast rule called the Overtake Safety Margin formula to decide whether an overtaking maneuver is safe, without running a full traffic simulation. This study checks the formula's reliability using a physics-based simulator that models three vehicles: the car overtaking, the car being overtaken, and the oncoming car. One million generated overtaking scenarios were evaluated using both the OSM formula and the kinematic simulator to identify cases where the formula classified a maneuver as SAFE while the simulator predicted an UNSAFE outcome. These mismatches, called blind spots, represented 3.98% of the evaluated scenarios and were concentrated in high oncoming-speed, short-gap situations, pointing to a specific weakness rather than random error. To catch these blind spots automatically, a small and efficient machine-learning model, a micro-transformer, was trained on the simulator's results. It reached 96.3% overall accuracy and correctly flagged 86.9% of the formula's blind spots, outperforming LightGBM, which detected 73.2%. The model's confidence scores were well calibrated, so they can be trusted to trigger an extra safety check when needed. An uncertainty-based search then used these confidence scores to find risky scenarios efficiently: 187 of the 500 most uncertain cases turned out to be unsafe, about 9.4 times more effective than random selection. Finally, the 30 highest-priority cases were replayed in the CARLA driving simulator, and 27 were confirmed to reproduce unsafe outcomes. Together, these results show that pairing a fast rule-based formula with a calibrated learning-based monitor can catch dangerous situations the formula alone would miss, supporting the safety documentation required for automotive safety-of-the-intended-functionality standards.
Sekaran, Nikhil, Sehgal, Vishal
Finite Element Analysis (FEA) plays a vital role in accelerating product development and reducing overall engineering costs. Accurate simulation of powertrain components requires extensive preprocessing of solid models. Detailed contact definitions and bolt representations are required to properly evaluate screw joints. Pipe joints present similar challenges, as they often incorporate various types of nuts and connectors. As a result, engineers spend a significant amount of time manually modelling these joints. Automating the modelling process is difficult due to the high variability in bolt sizes and pipe joint designs. This paper proposes AI-based models to automatically classify these components and extract their distinctive geometric features. Because bolts and pipe joints have wide feature variations, developing a generalized AI model requires a large dataset which is often hard to obtain. To address this limitation, this paper explores strategies for generating smaller, specialized AI models tailored to each unique joint type. This study describes the development of deep learning neural network models capable of learning distinguishing features of different joint configurations from limited training data. A multilayer graphical neural network architecture is employed, enabling efficient extraction and representation of these unique geometric characteristics. The proposed models can robustly generalize different joint types while maintaining high classification accuracy and reliable feature detection despite constrained training data availability. Automated workflows have been implemented in commercially available preprocessing tools to validate the proposed AI methodology and demonstrate its effectiveness in practical engineering applications. The integration of AI model automation in preprocessing tools significantly reduces engineering effort and repetitive manual preprocessing steps. This work highlights the potential of hybrid AI models to accelerate CAE workflows and provides scalable framework for future automation in FEA.
Jayachandran, Janarthanan, Agarwal, Deepak
Long-term orbital missions require highly independent life-support systems, maintaining ecological stability under limited external support and subsystem attenuation. Conventional strategies are typically based on fixed threshold rules or deterministic mass-balance models and cannot adapt to disturbances or multi-objective trade-offs between resource safety and environmental load. This paper proposes a reinforcement learning framework for managing micro closed-loop ecological systems under reliability decay and environmental constraints. A reinforcement learning controller based on Proximal Policy Optimization (PPO) is trained to learn stable regulation policies with random disturbances. A dynamic state model is constructed to describe the coupled and time-sequential elements, ranging from water storage, oxygen concentration, energy reserve, subsystem reliability to ecological load. Degradation propagation is integrated into the transition dynamics through reliability-dependent recycling efficiency. Simulation experiments demonstrate that the proposed policy maintains all resource variables within the safety margins while reducing ecological imbalance compared with traditional strategies. Consequently, the proposed framework provides an expandable control architecture for independent regulation of space station life-support systems.
Zhang, Heyang, Wu, Meixin, Xie, Junze
Civil aircraft engine thrust reversers are easy to have overspeed faults when they are being deployed, therefore this condition is very important for flight safety and maintenance work efficiency. The low sampling frequency, about 1Hz, of onboard data recording systems, for example WQAR, together with external environment interferences such as sudden strong winds, therefore causes a quite large number of wrong and missed alarms in fault data. Therefore, the labeled data that can be obtained are not only few in number but also have low quality. In this article, a fault diagnosis framework which is based on semi-supervised learning is put forward to solve the problems of high labeling expenses and fault diagnosis under the condition of small samples. This framework makes use of a small amount of labeled data and a large amount of unlabeled data. It uses Semi-Supervised Support Vector Machines, Semi-Supervised Neural Networks, and Semi-Supervised Random Forests for dynamic self- training to realize classification of three kinds of states: “Normal State,” “Low-Speed Thrust Reverser Deployment Exceedance,” and “Medium-Speed Thrust Reverser Deployment Exceedance”. The experiment outcomes show that the Semi-Supervised Neural Network model which uses self-training method has the highest classification accuracy on the test set, achieving 94.8%. Further robustness analysis, such as label noise and feature noise disturbance experiments, indicates that this model displays the smallest performance lowering under noisy environments, hence it represents the best overall performance. This research gives an effective method for the diagnosis of engine thrust reverser function problems in an effective and strong way.
Yuan, Xiaojiang, Liu, Nan
Traffic oscillations degrade the operational efficiency and safety of vehicular platoons. This study proposes a predecessor-following platoon with periodically assigned feedback gains and investigates its frequency-selective disturbance attenuation mechanism. An auxiliary second-order undamped periodic model is first established, and the Bloch–Floquet formulation is used to derive the nominal dispersion relation and bandgap boundaries. Closed-loop disturbance attenuation is then evaluated separately using spacing-error transfer matrices that retain positive velocity feedback and actuator-response lag. Frequency-domain transmission analysis and time-domain simulations show that, under nominal homogeneous and fully controllable connected and automated vehicle (CAV) assumptions, the periodic gain configuration creates tunable attenuation bands and provides stronger downstream attenuation than a homogeneous-gain configuration. The results clarify the relationship among periodic gain assignment, bandgap formation, periodic-cell string stability, and finite-length traffic-wave suppression. The effects of vehicle heterogeneity, sensing delay, and mixed traffic remain topics for future investigation.
Yang, Xiujian, Fan, Yanwen, Zhuang, Qingyuan, Qiao, Feiyu
Large-scale satellite constellation on-orbit missions are characterized by multiple objectives, multiple constraints, and strong dynamism. Traditional ground control and management modes have become difficult to meet the requirements for autonomous real-time decision-making by satellites. This paper addresses the need for autonomous and rapid decision-making capabilities on-board satellites, and proposes a satellite autonomous task planning and decision-making method based on an attention pointer network, addressing issues such as limited on-orbit computing resources and insufficient generalization of planning methods. Firstly, by analyzing and mathematically describing the characteristics and constraints of satellite tasks, a Markov model for satellite autonomous task planning problems is constructed. Secondly, the pointer network is combined with a reinforcement learning framework, and an attention mechanism is introduced to perform probabilistic learning on the execution selection of satellite tasks, generating an optimized sequence of satellite task planning results. Finally, simulation experiments show that after training, this method exhibits certain generalization capabilities and significantly superior computational efficiency compared to traditional intelligent optimization algorithms such as genetic algorithms. It can adapt to the single-satellite autonomous dynamic planning requirements under large-scale constellation control and management, providing technical support for autonomous operation and online intelligent task planning of satellites in orbit.
Feng, Xiaoen, Li, Yuqing, Qiu, Jianbin, Li, Meicheng
Space membrane structures, prized for their lightness, affordability, and ability to fold and unfold, hold immense promise for missions like deep space exploration, high-definition Earth observation, and advanced satellite communications. However, these structures are typically large-size, ultra-lightweight, highly flexible, and have very low damping. This leads to complex dynamic behavior featuring significant geometric nonlinearity, strong coupling between rigid and flexible motions, and uncertain parameters. Consequently, creating accurate yet simple mathematical models for designing controllers is exceptionally difficult, and traditional control methods that rely on such models often fall short. To tackle this problem, this study focuses on a membrane antenna spacecraft equipped with cable actuators, aiming to suppress its vibrations during attitude maneuvers. Using the Deep Deterministic Policy Gradient (DDPG) algorithm, an intelligent vibration controller based on the Actor-Critic architecture is designed. This approach is data-driven, meaning it learns effective control strategies directly through interaction with the system’s environment, bypassing the need for a precise pre-defined mathematical model. Simulation results show that the controller demonstrated superior vibration suppression and strong adaptability across several key scenarios.
Zhang, Hua, Chen, Cheng, Liu, Xiang
To meet the demands of on-orbit monitoring and space situational awareness, this study proposes a Transformer-based relative motion pattern recognition method based on the self-attention mechanism, addressing the issues of difficulty in accurately characterizing the relative motion patterns of non-cooperative space targets under short-arc observation conditions, strong dependence on prior knowledge in traditional methods, and insufficient stability under noise interference. Firstly, this study constructes a non-cooperative spacecraft motion pattern action data set based on the Clohessy–Wiltshire (C-W) relative motion equation. By designing typical initial conditions of relative motion, trajectory samples including fixed-point following, hovering following, oscillating following, coplanar circling, non-coplanar circling, approaching and receding, etc. of multiple motion patterns are generated, and the relative position and relative velocity sequences in the local orbital coordinate system are used as model inputs. Subsequently, an motion pattern recognition network composed of an input representation module, a Transformer encoding module, and a classification output module is constructed. The trajectory sequence representation is achieved through linear embedding and position encoding, and the multi-head self-attention mechanism is used to capture the global dependencies and key dynamic features between different time steps, thereby completing the discrimination of the target motion pattern category. Finally, model training and testing are carried out based on the constructed data set, and experimental analysis is conducted from three aspects: training convergence, classification performance in the absence of noise, and robustness in the presence of noisy observations. The results show that the proposed method can effectively learn the discriminative features in short arc segment relative motion, and performs well in terms of accuracy, Macro-F1, and confusion matrix; in the presence of Gaussian noise interference, it outperforms the Multi-Layer Perceptron (MLP) baseline model while still maintaining clear category boundaries and high recognition stability.
Mao, Yicheng, Zhang, Zexu, Li, Yingshuo, Song, Zhuo, Zhang, Fan, Liang, Jintang
With the unprecedented surge in demand for flexible deployment and ultra-high throughput in next-generation wireless networks, airborne base stations (ABS), particularly unmanned aerial vehicles (UAVs), have become emerged as indispensable mobile communication infrastructure. To facilitate high-gain directional transmission within the millimeter-wave band, ABS frequently employ massive MIMO antenna arrays. However, under random disturbances caused by non-stationary turbulence, the system faces a severe physical bottleneck: weak anti-interference capability due to narrow beamwidth, which can easily lead to a sharp drop in link capacity. Addressing the two major shortcomings of existing deep learning prediction schemes—susceptibility to sensor noise interference and negative capacity performance shifts during stable periods, and response lag due to memory inertia during sudden turbulence periods—this paper proposes an ARET-ALSTM active compensation network based on SNR-aware multimodal attention, agile reset and adaptive event-triggered mechanisms. This network incorporates an event-triggered mechanism governed by a dynamic physical variance threshold, adaptively executing compensation during stable intervals to prevent excessive overshoot and redundant computational overhead. Simultaneously, it employs an agile feature reset mechanism based on an adaptive sliding window for concept drift detection, forcibly discarding old memories and reconstructing attention weights during sudden wind shear, achieving millisecond-level ultra-fast adaptation and directly driving network optimization with the goal of maximizing communication capacity. Simulation experiments show that, under different array configurations ranging from 2×2 to 16×16 and complex flow conditions, ARET-ALSTM successfully overcomes the physical capacity collapse limitation under large antenna scales, and its tracking agility and robustness are significantly better than benchmark algorithms such as Transformer, RNN and traditional Kalman filtering.
Tu, Jiajun, Zhou, Jihua, Li, Yanhang, Zhao, Tao, Huang, Hua, Wen, Yuanbin, Guo, Ziwen
Synthetic Aperture Radar (SAR) provides important all-weather images. These images help researchers observe the sea surface. However, separating the coast from the ocean in these images is hard. Near-shore areas face three main issues. The first issue is coherent speckle noise. The second issue is geometric distortion caused by the radar’s side-looking geometry. The third issue is feature aliasing caused by multiple scattering from coastal buildings. To solve these problems, this paper proposes the Enhanced Attention U-Net (EAU-Net). This model is a deep learning framework. It adds SAR scattering priors directly into the learning process. The architecture has three main parts. First, the encoder includes an adaptive local signal-to-noise ratio (SNR) module. This module filters multiplicative speckle dynamically. It stops the network from mistaking noise for real features. Second, the network bottleneck uses a lightweight Transformer block. This block learns global shape priors. It helps the network find the continuous coastline even when radar shadows cover parts of the image. Third, the decoder uses an adaptive edge refinement mechanism. A composite edge-aware loss guides this mechanism. It fixes rough edges and corrects pixel boundaries. This step removes jagged artifacts. Tests on a near-shore dataset show the reliability of this method. EAU-Net achieves a Dice Similarity Coefficient (DSC) of 0.9601 and an Intersection over Union (IoU) of 0.9234. The Average Symmetric Surface Distance (ASSD) is 5.7501 pixels. This paper compares EAU-Net with classical convolutional networks, global-attention architectures, and vision foundation models. The results show that EAU-Net preserves boundary precision better. It outperforms state-space models like VM-UNet and large-scale zero-shot models like SAM-Med2D. EAU-Net uses physical inductive biases to separate coastal building interference from marine targets. This provides a realistic foundation for downstream tasks like autonomous ship detection.
Bai, YuanJie, Zhong, Sheng
Accurate compensation of magnetic measurement errors is one of the key factors in ensuring the accuracy of geomagnetic navigation. Magnetic sensors are susceptible to temperature drift during use due to the disturbance of external ambient temperature, which affects the accuracy of magnetic field measurement. To improve the precision of the tri-axis geomagnetic sensor, a suitable thermal error compensation algorithm must be designed to process measured data correctly. Thus, a temperature compensation model and a high-precision geomagnetic directional technology were proposed in this paper. Firstly, a tri-axis geomagnetic sensor temperature compensation model based on particle swarm optimization-least squares support vector machines (PSO-LSSVM) was established. Then the inertial measurement unit (IMU) data is processed with zero drift correction and filtering. Finally, geomagnetic directional experiments were conducted with the gradient descent method. The experimental results show that the proposed temperature error compensation method can effectively alleviate the impact of thermal variations on the detector output and enhance the precision of the geomagnetic orientation system. It is expected to be applied to multiple navigations, aviation, and aerospace fields.
Fan, Chenyu, Guan, Feng, Liang, Zhuo, Zhang, Haifeng
The rapid evolution of satellite communication systems enables direct network access for terrestrial users via portable devices, but also introduces critical security challenges in access control. In particular, identity hijacking attacks allow malicious users to exploit legitimate access channels, leading to inefficient occupation of scarce satellite resources. Existing approaches mainly depend on trust evaluation or heuristic-based decision strategies, which are inadequate for adapting to dynamic environments. To address this issue, this paper proposes a trust-aware deep reinforcement learning (DRL) framework for satellite access control, which jointly considers user behavior evaluation and decision optimization. Specifically, a Bayesian trust evaluation model is first developed to quantify the credibility of user behavior by incorporating historical interaction records as prior information. Unlike conventional static trust schemes, an adaptive mechanism is introduced to enhance sensitivity to behavioral dynamics. Building upon the trust assessment, the satellite access process is formulated as a Markov Decision Process (MDP), where system states integrate channel conditions, user trust levels, and service demands. A Deep Q-Network (DQN) is then employed to learn optimal admission and scheduling policies that maximize long-term system utility under resource constraints. Simulation results demonstrate that the proposed method significantly outperforms conventional greedy and threshold-based strategies. In particular, it effectively mitigates the impact of malicious behaviors and improves system utility by over 80%, while achieving more efficient resource utilization in dynamic environments.
Zhou, Qihao, Liu, Wei, Zhang, Hongyuan, Chen, Te, Wang, Yuqi
Spacecraft operate in a highly complex in-orbit environment. Telemetry data from power systems is often plagued by multi-source disturbances such as plasma noise, electromagnetic interference, and communication packet loss, which places very strict demands on fault diagnosis model reliability. Using dynamic disturbance test manifolds and the comprehensive robustness quantification index (RQI) for evaluation, we found a critical vulnerability: baseline models such as the joint loss enhanced global sparse attention transformer (JLE-GSA-Transformer) perform well in resisting continuous signal disturbances, but their performance drops sharply when temporal structures are disrupted due to data missingness. This finding highlights that static diagnostic models have fragile robustness boundaries when operating under extreme non-stationary conditions. To solve this problem, this paper puts forward an adaptive optimization method that combines deep reinforcement learning (RL) with quantitative robustness feedback. By making use of the dynamic disturbance manifold, this method embeds the RQI into the reward function of the RL algorithm. Through continuous updates to the agent’s policy, it drives the adaptive adjustment of key parameters in the JLE-GSA-Transformer, mainly including the coding structure of global sparse attention (GSA) and the discriminative boundaries of temporal cosine-center joint loss. This closed-loop mechanism can significantly improve the stability, adaptability, and long-term reliability of the model when it works in complex environments with multi-source disturbances.
Liu, He, Wei, Cheng, Zhang, Zexu
This paper describes formal modeling and verification of generic algorithm for universal Time Division Multiplexing (TDM) communication, one of the orchestration algorithms used in PTB-FLA (Python TestBed for Federated Learning Algorithms), a light-weight framework for developing distributed applications, developed as a part of the TaRDIS (Trustworthy and Resilient Decentralized Intelligence For Edge Systems) project. Modeling and verification is done using UPPAAL model checker tool. When developing a distributed application, in early development phases a user of PTB-FLA can abstract away low-level details of message passing and some elements of orchestration, by choosing one of the available orchestration algorithms, therefore being able to focus on the actual distributed algorithm at the application level. The generic algorithm for universal TDM communication models peer data exchange between the nodes within a discrete time-slots, a frequent case in inter-satellite links. It is an orchestration algorithm intended to be used in satellite-swarm applications, such as distributed ODTS (Orbit Determination and Time Synchronization). Formal modeling and formal verification contributes significantly to increased reliability of orchestration layer, making development of a distributed algorithm easier by narrowing down possible sources of errors to application level. This paper gives overview of PTB-FLA framework, provides a description of the orchestration algorithm, shows its formal model in UPPAAL, and verifies some important algorithm properties, namely deadlock freedom (a safety property) and successful termination (a liveness property). This paper represents a continuation of one of the research lines related to PTB-FLA, which is formal verification of the framework’s generic orchestration algorithms, and other elements.
Djukic, Miodrag, Popovic, Miroslav, Li, Yuqing
Accurate performance prediction is essential for improving the reliability and service life of Hall thrusters in space applications. Because conventional scaling-based prediction methods are often limited when the operating state and magnetic-field structure become complex, this study develops a deep-learning framework for rapid prediction of key Hall-thruster performance indicators. Three representative neural-network architectures, namely Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and Bidirectional Long Short-Term Memory (BiLSTM), are constructed and comparatively evaluated. Among them, BiLSTM achieves the best overall performance, with a Mean Absolute Percentage Error (MAPE) of 6.44% on the test set. Additional validation based on real Hall-thruster experimental data gives a MAPE of 3.884%, indicating that the proposed model has good generalization ability and can provide an effective data-driven tool for Hall-thruster performance prediction.
Sun, Chang-You, Tian, Bin, Yu, Yun-Chao, Wang, Ke, Li, Shu-Feng, Xie, Kan, Yang, Su-Lan
The growing number of space objects raises collision risks and demand for on-orbit maintenance. Since most servicing tasks require observing the target, relative pose estimation is key for debris removal and satellite servicing. The complex orbital environment poses significant challenges for conventional feature extraction networks.. To address these challenges, this paper presents a multi-task convolutional neural network for space target pose estimation. The network uses a MobileNetV3-based encoder to extract image features and an Adaptively Spatial Feature Fusion (ASFF) module for adaptive multi-scale feature fusion. Although MobileNetV3 and ASFF have rarely been combined in the aerospace field, their combination offers significant advantages in low parameter count and reduced computational load, showing strong potential for spacecraft pose estimation. Two parallel prediction heads are employed: an EfficientPose head consisting of classification, bounding box regression, rotation matrix regression, and translation vector regression, together with a Heatmap head. This design leads to more robust and accurate pose estimation. The loss function follows a multi-prediction-head joint approach, where each component is weighted to form the final loss for multi-task optimization. The network is trained, validated, and tested on a simulation dataset of the DAWN satellite built with Unity. The dataset includes illumination variations, large distance changes, and multi-view target appearances, thus capturing the complex scenes typical of vision-based recognition tasks. Experimental results show that after 40 training epochs, the model achieves an average position error of about 0.05 m and a rotation error of about 0.8° on the test set. These results demonstrate the effectiveness and accuracy of the proposed network for space target pose estimation.
Wang, Wenzhi, Zhang, Zexu, Zhang, Fan, Song, Zhuo
Low earth orbit (LEO) satellite communication system is expected to play a key role in the 6G wireless networks. Due to the limited communication resource and inter-user interference (IUI), the user equipments (UEs) must be well scheduled to maximize spectral efficiency. However, rapid channel variations and high mobility in LEO system make the conventional user scheduling strategies relying on instantaneous channel state information impractical. To address this challenge, this paper proposes a deep reinforcement learning (DRL) framework based on batch normalization in CrossQ algorithm to optimize the downlink user scheduling in a LEO satellite massive multi-input multi-output (MIMO) system. The scheduling task is modeled as a Markov decision process (MDP), where the observation space integrates high-dimensional features including large-scale fading (LSF) coefficients, channel matrix, and temporal features capturing the orbital progress of the satellite. Unlike traditional DRL methods, the CrossQ-based algorithm applies batch normalization (BN) within the critic networks to eliminate the need for slowly-moving target networks, fostering training stability and sample efficiency. A continuous priority score vector is mapped to discrete scheduling decisions through a top-z selection logic, allowing the agent to effectively identify spatially orthogonal user subsets to suppress the IUI. To guide the agent toward the true optimization direction, a differential reward mechanism is designed. Simulation results demonstrate that the proposed CrossQ-based scheduler consistently achieves nearly optimal performance at different signal-to-noise ratios (SNR), reaching at least 91% and 94% of the Monte Carlo search ceiling under different user densities. Moreover, the trained model provides scheduling strategy in millisecond inference time, making it a practical and robust solution for future 6G non-terrestrial networks.
Deng, Chaojun, Zhong, Yun, Meng, Xiangxu, Li, Wenfeng, Zhao, Kanglian
Reliability testing and validation are essential parts of any automobile development project. These tests mainly depend on two things: how much of the product’s target life is covered and whether enough samples are tested to capture variations. Today, most projects face tight timelines and cost constraints, which makes it important to design an optimized reliability test plan. Accelerated testing is widely used to reduce testing time, but finding the correct acceleration factor for a complex system like an automotive engine is difficult. Existing studies describe methods such as simple stress tests, step-stress tests, and degradation tests to shorten test duration. However, these methods need more samples at different stress levels to establish a clear relationship between stress and life and then calculate the acceleration factor. While such models work well at the component level, they become much more complicated at the system level because different components can have different acceleration behaviors. In this paper, a method is proposed to derive the acceleration factor by correlating engine test bed data with field failure and usage data. Test engineers often use standard accelerated test cycles for engines used in similar applications. As a result, organizations usually have historical test data and field data from earlier, similar products. The proposed approach models accelerated test data and field failure data separately and then links them to determine the acceleration factor. This derived factor can later be used in future engine development projects to convert test cycles into equivalent field mileage coverage. Weibull analysis and regression techniques are used to support the case studies presented in this work.
Koulage, Dasharath Baliram, Mondal, Kanchan
Fuel adulteration increases operating costs, reduces vehicle efficiency, and contributes to air pollution. Fuel-quality verification typically requires laboratory infrastructure or dedicated sensors that are not available to most end users. In our previous work, we demonstrated that a software-only approach based on standard on-board diagnostics (OBD) signals and supervised machine learning can detect water adulteration in hydrated ethanol under real-world driving conditions with high accuracy. However, supervised classifiers remain challenging because fuel-trim dynamics vary across vehicles due to powertrain calibration, component aging, and operating environments, which can limit model generalization. This work addresses this limitation by introducing a detection framework that learns a vehicle-specific baseline of normal operation using gradient boosting regression (GBR) and flags anomalous events through residual-based cumulative sum (CUSUM) monitoring. The baseline model is trained using compliant-fuel data, and the CUSUM decision threshold is calibrated from temporally ordered normal-operation data using an expanding-window procedure with an initial warm-up segment. Across two vehicles, the proposed framework achieved recall values of 99.77% and 84.38%, with F1 scores of 99.88% and 91.53%, respectively, while maintaining zero false positives under compliant-fuel operation.
Marchezan, André Ricardo, Giesbrecht, Mateus
Autonomous driving systems are among the most complex software systems deployed in the commercial vehicle industry today. Modern autonomous driving platforms combine large-scale AI models, machine learning infrastructure, middle-ware, safety mechanisms and vehicle integration software into a tightly integrated software stack. This complexity introduces new cybersecurity challenges. Many of the threat modeling approaches used throughout the industry were developed around architectures with well-defined boundaries, limited software complexity and clearly documented interfaces. Autonomous driving systems operate differently.
Ma, Zhendong
The proliferation of small unmanned aircraft systems (sUAS) presents an asymmetric threat to ground maneuver forces operating in contested and gray-zone environments. The Bullfrog Autonomous Weapon Station (AWS) addresses this operational gap through a passive, AI-powered counter-UAS system employing computer vision and machine learning for autonomous detection, tracking, classification, and engagement. Field testing at Technology Readiness Experimentation (T-REX) 26-1 demonstrated 100% probability of defeat against Group 1 UAS targets with a mean engagement time of 6 seconds and 10 rounds per kill at ranges exceeding 160 meters. Operating in both autonomous and human-in-the-loop modes, Bullfrog achieved 99.45% operational availability while leveraging service-common M240B weapons and Modular Open Systems Architecture for rapid integration with Joint All-Domain Command and Control (JADC2) networks. At $300,000 per unit with $10 cost-per-engagement, Bullfrog demonstrates operational relevance, speed-to-field, and alignment with Army and Marine Corps autonomy priorities.
Cunningham, Jason, Clark, Alex
This paper details the successful scaling demonstration of a comprehensive supply chain screening process for commercial off-the-shelf (COTS) motherboard subassemblies used in tactical servers for naval applications. Our approach leverages Power Fingerprinting (PFP) technology, which uses unintended analog emissions and machine learning to provide independent, non-destructive, and scalable integrity assessment of microelectronics. The primary goal of the effort was to demonstrate the effectiveness and scalability of the PFP screening process without disrupting or delaying the manufacturing workflow. The screening successfully detected hardware and firmware modifications and identified two cases of abnormal behavior: unusual BIOS power reset and elevated CPU sensor readings on two motherboard subassemblies. Following our quality control forensic analysis, we determined the root cause of these anomalies and their potential impact on the host platform.
Aguayo Gonzalez, Carlos R., Roberson, Ken
Unmanned Aerial Systems (UAS) pose a growing threat on the modern battlefield, demanding rapid detection and characterization capabilities for the warfighter. Existing single-model solutions are inadequate for Counter-UAS (C-UAS), as they struggle across varying ranges and cannot provide detailed contextual information beyond bounding boxes. We present ZEUS (Zero-shot Explainable Universal Segmentation), a multi-model detection and recognition system that integrates several machine learning approaches. ZEUS employs a high-performance UAS detector trained on synthetic, internally collected, and open-source datasets, with real-time capability demonstrated on edge hardware across both electro-optical and infrared modalities. For classification, ZEUS uses a zero-shot approach: detected UAS are segmented and compared against a library of 3D reference models rendered at various poses, enabling identification of new UAS types without retraining. This methodology additionally provides UAS pose and range estimates critical for threat assessment and engagement decisions.
Matousek, Gregory, Varberg, Nathan, Torrione, Pete, Brandon, Namdi, Inkawhich, Matt, Camilo, Joe
Artificial Intelligence (AI) and Machine Learning (ML) are increasingly transforming Computer-Aided Engineering (CAE) workflows by enabling faster design iterations and reducing computational costs. This paper presents the application of Ansys SimAI and Ansys GeomAI in modelling an automotive side impact scenario using high-fidelity data from LS-DYNA simulations. Two AI models are trained on datasets with systematically varied parameters: one encompassing pole impact position and door beam configurations, and another focusing on rocker panel reinforcements. Both models exhibit strong predictive performance, reliably capturing deformation patterns and force-time histories for previously unseen configurations. The datasets are subsequently merged to train a comprehensive surrogate model capable of simultaneously representing variations in pole position, door beam geometry, and rocker reinforcement design, demonstrating robust generalization across a multidimensional design space. To address the emerging bottleneck of geometry creation, GeomAI’s geometry exploration functionality is employed to generate new rocker reinforcement geometries from existing ones, which are then rapidly validated using the pre-trained surrogate model. The results confirm that LS-DYNA simulations can be leveraged effectively to build AI models that dramatically reduce design exploration time. With SimAI and GeomAI in the loop, CAE workflows can evolve from simulation-driven design toward AI-augmented autonomous engineering, where geometry generation, simulation, validation, and optimization converge into an intelligent closed loop.
Adya, Srikanth, Karadogan, Celalettin, Vasu, Shyam S., Lazarov, Nikolay, Husek, Martin, Haufe, André
The behavior of detergent–dispersant additives in lubricating oils depends upon their chemical composition and synthesis, particularly for lubricants used in rotary equipment under harsh conditions, in the upstream and downstream oil and gas industries. Alkylphenolate-based additives are extensively used, among others, due to their high alkalinity and good dispersion stability. However, their performance is extremely sensitive to synthesis parameters, making experimental optimization time-consuming and resource-intensive. In this study, an alkylphenolate-based additive (AKI-152) was synthesized and optimized systematically via a response surface methodology (RSM), derived from a central composite design (CCD). The influence of three important process variables, namely the alkylphenol/amino mass ratio, reaction temperature, and Ca(OH)2/CO2 mass ratio, was assessed with respect to the total base number (TBN), kinematic viscosity, and corrosivity of the final product. A quadratic regression model was developed and validated, showing good agreement between predicted and experimental results. Analysis of variance (ANOVA) indicated that the alkylphenol/amino ratio and Ca(OH)2/CO2 ratio exert the strongest influence on TBN, while temperature also contributes significantly within the investigated operating range. Numerical optimization identified a stable optimum region corresponding to TBN values of approximately 160 mgKOH/g, viscosity values close to 54 mm2/s, and corrosivity values below 0.76 g/m2 under the selected optimization criteria. No separate experimental run was performed to verify the selected numerical optimum, as the optimization was conducted within the experimentally investigated CCD space using the developed and statistically evaluated response surface models. These results demonstrate that the applied modeling approach can be effectively used to define suitable synthesis conditions for high-performance lubricant additives.
Gasimov, Rahib, Naghiyeva, Elmira, Sujayev, Afsun, Gadirov, Ali, Nasibova, Nigar
Estimating battery state of health (SOH) from field data is essential to ensure successful operation and increase the uptime of battery electric vehicles (BEVs). Most studies in the literature propose methods relying on datasets acquired under controlled laboratory conditions. However, SOH estimation becomes significantly more challenging when dealing with real-world data due to the increased variability and complexity of operating conditions. In this work, CAN telematics data, sampled at 1 Hz, were collected over approximately 20 months of operation from 10 electric commercial vehicles. During this period, a maximum battery degradation of 4% is observed within the fleet. Firstly, a model-based framework was introduced, in which a second-order battery equivalent circuit model (ECM) was coupled with an extended Kalman filter (EKF) to estimate the battery SOH. Results confirmed that the EKF is able to accurately capture the battery's physical behavior and degradation trend, yielding a maximum root mean square error (RMSE) of 1.23% when compared with the SOH signal provided by the onboard BMS. However, a Kalman filter requires accurate model parameter identification and high-frequency measurement data, leading to increased computational costs. To bridge these gaps, this paper utilizes the SOH estimates obtained from the EKF to train and validate a feedforward neural network (FNN) model, specifically designed to operate on aggregated metrics. The FNN model can provide accurate SOH estimates, with a RMSE as low as 0.26% during the testing phase. The approach proposed in this work combines the interpretability of model-based methods with the scalability and reduced data dimensionality of machine learning (ML) ones, making it more suitable for monitoring battery SOH in large fleets of BEVs.
D'Agostino, Valerio, Pulvirenti, Luca, Shanker, Anirudh, Cardone, Massimo, Rizzoni, Giorgio, Vitale, Francesco
Validation of brake systems is increasing in complexity due to electrification, software-defined architecture, integrated control modules, and higher functional safety requirements. Although physical testing remains the primary source of engineering evidence, interpretation of results including DVP&R/PVP&R compliance verification, anomaly detection, documentation, and milestone decision support continues to rely heavily on manual engineering analysis. This results in extended feedback cycles, inconsistent interpretation across teams, and limited traceability between raw data, reports, and governing specifications. To support engineers with more objective validation processes, there is a growing need for structured, data-driven intelligence that transforms dispersed test artifacts into actionable engineering decisions. This paper presents Data to Decisions, an AI-driven Test Intelligence Platform designed for integrated analysis of raw measurement data, test reports, validation plans, and specification requirements in brake system development. The platform has been applied to foundation brake systems (EPB and hydraulic calipers), brake control modules (IBC/EB100), and related actuation subsystems. It ingests heterogeneous inputs including DVP&R documents, customer specifications, test summaries, deviation logs, parameter files, and build configurations and converts them into structured, traceable validation datasets. A specification centered reasoning framework extracts governing limits, acceptance thresholds, instrumentation requirements, and staged validation criteria directly from source documents. Using natural language processing, rule-based logic, and pattern recognition models, the system evaluates both discrete and continuous data sets over an unlimited range of performance characterization metrics such as leakage, drag torque, piston travel, fatigue life, NVH behavior, structural durability, and actuator performance characteristics. Results are assessed against extracted specification limits to automatically identify compliance gaps, borderline conditions, parameter inconsistencies, and build-specific variations. All findings are traceable to original requirements and test evidence. The platform further enables closed-loop validation by linking physical test outcomes with virtual analysis results, supporting correlation studies and identifying opportunities for test optimization or targeted retesting. Automated generation of engineering and management level summaries reduces documentation effort while improving consistency and auditability. Pilot deployments demonstrate reduced manual data review effort, improved traceability of specification compliance decisions, enhanced anomaly detection, and faster decision making during and after DV and PV milestones. By combining rule-based validation logic with AI and Generative AI for document interpretation, pattern recognition, and automated summarization, the platform supports engineers in efficiently navigating large volumes of test data and specifications. This paper presents the system architecture, compliance evaluation methodology, and deployment results, illustrating how AI-enabled test intelligence can serve as a practical decision-support layer in modern brake system validation workflows.
Divakaruni, Saikiran, Willey, Joseph, Srivastava, Namrata, Sankar, Arya, Namala, Divya, Gowtham, Rahul Sangani
Items per page:
1 – 50 of 1646