Browse Topic: Neural networks

Items (1,616)
To address the issues that multi-axle special vehicles are prone to rollover accidents during high-speed maneuvers due to their high load capacity and elevated center of mass (CoM), and that the traditional Time-To-Rollover (TTR) rollover early warning algorithm based on state estimation suffers from misjudgment and false alarm problems, this study takes a specific five-axle special vehicle as the research object to investigate rollover early warning algorithms. First, a vehicle dynamic rollover index accounting for unsprung mass was established, the index was validated via multi-condition simulations, and a rollover early warning threshold of 0.8 was determined in conjunction with the dynamic characteristics of multi-axle special vehicles. Subsequently, integrating vehicle state estimation results, a TTR rollover early warning algorithm based on state estimation was developed, and its early warning performance as well as inherent limitations were verified via simulations. Finally, a neural network-corrected TTR rollover early warning algorithm (NN-C-TTR) was proposed, the overall architecture of the algorithm and the neural network module were designed, and the algorithm was validated through both simulation tests and real-vehicle tests. The results demonstrate that the proposed algorithm effectively addresses the misjudgment and false alarm issues of the traditional TTR early warning algorithm, achieves accurate advance warning of rollover risks for multi-axle special vehicles, and exhibits outstanding reliability and engineering practicality.
Yin, Yongqi, Liu, Qiang, Yang, Yuxue, Zhu, Chengyong, Jiang, Xinhao, Zhang, Chunchao, Zhang, Zhengyi, Shu, Hongbin
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
Accurate identification of driving styles serves as a critical foundation for realizing customized driving experiences, optimizing autonomous driving systems, constructing efficient intelligent traffic management systems, and guiding the design of energy-efficient vehicles. This study aims to achieve efficient, stable, and accurate recognition of diverse driving styles. To this end, an improved driving style recognition method based on the Generalized Regression Neural Network (GRNN) is proposed. This method takes real-time vehicle driving data as the core input, enabling effective classification and precise identification of driving styles. First, the raw collected vehicle driving data were systematically preprocessed to eliminate redundant data and outliers, and three typical driving styles were classified and defined using the classic K-means clustering algorithm. Second, the standardized preprocessed dataset was rationally divided into a training set, an independent test set, and a validation set: the training set was used to optimize and train the GRNN recognition model, while the test set was adopted to quantitatively assess the overall recognition accuracy of the constructed model. Finally, the validation set was employed to comprehensively verify the generalization ability and stability of the model across different driving scenarios. Experimental results demonstrate that the proposed GRNN-based method achieves a driving style recognition accuracy exceeding 96%, with favorable generalization ability and stable performance. The research findings can provide reliable technical support and critical decision-making references for adaptive autonomous driving scenario application, intelligent traffic system optimization, and low-energy-consumption vehicle development.
Zhou, Weichen, Zhao, Shuang, He, Wenhao, Liu, Hai, Zhang, Hao
This paper proposes a reinforcement learning-based multi-section, coordinated variable speed limit control approach for expressways. The method handles joint speed-limited decision-making at coupled spatiotemporal traffic dynamics across multiple sections and lanes. The traffic state space is constructed based on indicators: lane flow, speed, occupancy rate, number of slow-moving vehicles, and minimum Time-to-collision (TTC). Lane-level speed limits for each control unit are defined as action variables. A weighted reward function is designed to consider operational efficiency, safety, rationality of speed gradient, and control response. The model integrates Convolutional Neural Networks (CNNs) and Gated Recurrent Units (GRUs) to capture spatial correlations and time-varying properties in traffic-state information. In addition, a risk-aware dual-policy output mechanism is introduced to adaptively balance safety-oriented and efficiency-oriented control objectives. Using the Houjie-Xinlian section of the Guangshen Expressway as an example, three control strategies are simulated on the SUMO platform: CGDQN, FCDQN, and an uncontrolled scenario. Experimental results show that the total reward for CGDQN exceeds that of FCDQN. CGDQN reduces vehicle queueing and slow-moving traffic across various scenarios, thereby decreasing the likelihood of traffic accidents. The results demonstrate that the proposed method achieves coordinated multi-section variable speed limit control on expressways under different conditions, and is both feasible and effective.
Zhang, Yuping, Xu, Tian, Ma, Lin, Lin, Siyu, He, Runjing, Yuan, Chenwei
In the process of design and verification of the aircraft ice protection system, a prediction model of ice protection system performance based on Severity Index(IPSI) is proposed, which quantifies the relationship between input and output energy of the ice protection system. In this research, a theoretical calculation method is developed, and the water total collection coefficient and convective heat transfer coefficient on the skin surface were obtained using CFD calculations and rapid neural network prediction methods. The consistency between IPSI and ice protection system performance is verified through an aircraft wing ice protection wind tunnel test. The results show that the lower the slat skin temperature in the ice protection area, the higher the IPSI value, which means the more severe the ice protection system performance. The ice protection performance severity analysis method is carried out for the cases in the CCAR/FAR/CS-25 continuous maximum(CM) envelope of an aircraft, and the IPSI values of each case at some altitudes are calculated, as well as the distribution of the ice protection system performance severity cases in the CM altitude-temperature envelope. The ice protection system performance severity cases are not the highest or lowest atmospheric static temperature in the CM envelope, in which atmospheric static temperature ranges from -8°C to -23°C from low altitude (1500 ft) to high altitude (22000 ft). When the flight speed is constant, and the atmospheric static temperature is gradually reduced, the IPSI value decreases. When the atmospheric static temperature is constant, and the flight speed increases gradually, the IPSI value increases. The influence of water evaporation heat transfer on ice protection performance severity is greater than that of convection heat transfer. The total air temperature for the most severe condition is always 0°C at every altitude. The IPSI calculation model can be used to quickly predict the severity of ice protection system performance in other ice protection areas of aircraft, even wind blades.
Hu, Weixue
Time series forecasting is a practical cornerstone for decision-making in diverse fields including finance, energy management and meteorology, yet conventional Transformer-based models are plagued by high computational complexity inherent in their attention mechanism, as well as poor efficiency in capturing long-range dependencies when processing long-sequence time series data. Meanwhile, the recently proposed Mamba model built on a linear-complexity state space model has achieved strong performance in long-sequence modeling tasks, but the vanilla Mamba model still has obvious limitations in time series forecasting scenarios, including insufficient discriminability of channel features, unsatisfactory prediction accuracy, and information loss during feature transmission in the encoder layer. To address these issues, this paper proposes An Enhanced Mamba Integrating Dynamic Compression Ratio Channel Attention and Progressive Projection for Multivariate Time Series Forecasting, which incorporates three core improvement strategies: a lightweight dynamic compression ratio channel attention, a residual-optimized encoder, and a progressive prediction head. These strategies realize adaptive weighting of channel features, lossless information transmission in the encoder and gradual refinement of forecasting results respectively. Extensive experiments on multiple public time series datasets show that compared with the vanilla Mamba model and mainstream Transformer-based forecasting methods, the proposed model achieves significant improvements in prediction accuracy, along with stronger robustness and higher practical application value.
Zhu, Yongdong, Zou, Lei, Li, Na, Wang, Chenxi
To address the complexity of the influence of expansion-section geometry on the thrust performance of solid rocket motor nozzles and the low efficiency of numerical-simulation-based optimization, this study proposes a deep-learning-based method for rapid nozzle thrust prediction. With a representative 70-lb BATES nozzle as the research object, the expansion-section contour is parameterized using cubic spline functions. Combined with Latin hypercube sampling and two-dimensional inviscid numerical simulations, a nozzle geometry– thrust dataset covering a wide range of contour characteristics is constructed. On this basis, surrogate models based on a fully connected neural network, a one-dimensional convolutional neural network, and a two-dimensional convolutional neural network are developed, and the applicability of different network architectures to the nozzle thrust prediction task is systematically compared. The results demonstrate that convolutional neural networks outperform fully connected networks in terms of geometric feature representation capability and prediction stability. In particular, by explicitly modeling the spatial structural features of the nozzle contour, the two-dimensional convolutional model achieves a coefficient of determination close to unity on the test set, with an average relative error of 0.03%. Further hyperparameter sensitivity analysis clarifies the effects of batch size, number of training epochs, and learning rate on model convergence behavior and generalization performance. This work provides a feasible technical support for rapid thrust evaluation and efficient optimization-oriented nozzle design.
Yuan, Yuan, Wu, Wei-tao, Hua, Yue
In high-end aerospace mission scenarios such as distributed sensing and tactical reconnaissance, path planning for multi-UAV cooperative formation flight has become an indispensable core component of mission systems. However, a key unresolved technical bottleneck remains: how to achieve an optimal balance between strictly maintaining the UAV formation’s preset geometric configuration and realizing collision-free, flexible flight in dynamically changing complex environments. Multi-Agent Reinforcement Learning (MARL) can theoretically solve this coupled optimization problem, but the control instructions generated by the existing mainstream MARL methods often have high-frequency fluctuation issues. Direct deployment of these fluctuating commands on physical UAV hardware will increase alternating stress on mechanical structures, raise the risk of motor overheating, and significantly boost battery energy consumption. These problems directly restrict the engineering application of such methods. To address the above challenges, this paper proposes a robust control framework that accounts for UAV physical characteristics to solve the model mismatch between simulation environments and real flight scenarios during carrier-based UAV deployment. Firstly, to improve the learning efficiency of the algorithm and avoid the cold start problem in the initial training stage, this paper introduces the Adaptive Curriculum Learning (ACL) method. Unlike the fixed-difficulty training mode, ACL can dynamically adjust the environmental complexity based on the real-time task performance of the UAVs. It first allows agents to master basic formation maintenance skills, before training is gradually transitioned to more complex tasks such as navigation in intricate environments. Additionally, considering the computational limitations of the onboard embedded processors of UAVs, this paper only integrates the Graph Attention Network (GAT) into the centralized Critic network of the centralized training and decentralized execution (CTDE) framework. The core advantage of this design is that the system no longer solely relies on distance as the judgment basis but assesses the influence of adjacent UAVs based on the collision risk level. This not only improves the accuracy of value estimation, but also does not compromise the response speed of local control instructions- a feature that is crucial for the real-time requirements of UAV flight control. The proposed method was tested and validated in a high-fidelity simulation environment built in strict compliance with UAV kinematic constraints. The experimental results fully demonstrate its superiority: in dense obstacle areas, the success rate of the task reached 80.4%; compared with the aggressive baseline method, the jerk was reduced by 10.9%. Most importantly, the flight trajectories generated by this framework are smooth and adaptable to actual flight conditions, and the corresponding control commands can be directly deployed on real flight controllers without additional parameter adjustments. This result demonstrates that the proposed framework achieves substantial technical progress in enhancing the autonomous response capabilities of UAV clusters to complex scenarios.
Huang, Xingyu, Zhang, Jiyu, Jiang, Yanshu, Fan, Yuhang
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
The development of the low-altitude economy presents abundant opportunities for data acquisition in Intelligent Transportation Systems (ITS). High-density interactions among heterogeneous traffic participants at complex urban intersections pose significant challenges for autonomous vehicle trajectory prediction. Existing Graph Neural Network (GNN) models often struggle to adequately capture interaction variances and global spatial dependencies among these heterogeneous agents. To address this, we propose STGANet, a Spatial-Temporal Graph Attention Network integrating Transformer-based multi-head attention with graph convolutions. STGANet employs a local-global feature extraction strategy: a dynamic multi-group GCN first extracts multi-semantic local spatial information, while a temporal attention mechanism captures long-term sequential dependencies. A feature fusion gate then adaptively integrates these spatio-temporal features. Subsequently, a global spatial multi-head attention module is applied to thoroughly mine the overarching heterogeneous interaction dynamics. Evaluated on the real-world ApolloScape trajectory dataset, STGANet demonstrates superior predictive performance in heterogeneous traffic flows. Compared to the baseline GRIP++ model, it reduces the Weighted Sum Average Displacement Error (WSADE) and Weighted Sum Final Displacement Error (WSFDE) by approximately 9% and 19%. Furthermore, with an average inference time of 12.8 ms, STGANet achieves high precision while strictly satisfying the real-time latency constraint (< 20 ms) required for practical autonomous driving applications.
Ou, Jingyi, Ou, Sijia, Zhang, Zhili, Zhu, Wenbo, Yu, Wei, Xie, Kan
In the modeling process, existing auto parts demand forecasting methods often fail to clearly distinguish structural differences at the demand-frequency level, which results in limited adaptability in real scenarios characterized by sparse demand and a complex category structure. In response to the above problems, this paper proposes a category-aware Transformer framework for auto parts demand forecasting. To explicitly represent demand differences across frequency-based accessory groups and to reduce training bias caused by long-tail sample imbalance, the framework first constructs frequency categories based on statistical demand characteristics and incorporates them into the model as embedding-based structural priors. It then designs a category-perception attention mechanism within the Transformer to guide differentiated modeling of category-related temporal dependencies. At the same time, in order to improve the expression ability of sparse and nonlinear disturbances, this paper introduces a nonlinear mapping module at the prediction end and uses a weighted loss function based on sample distribution to constrain the training target. Experimental results on three real auto-parts demand datasets from Dongfeng Automobile Co., Ltd. show that the proposed method outperforms five benchmark models in terms of RMSE and MAE, with especially strong performance in predicting intermediate-frequency and long-tail parts.
Yi, Yuxuan, Wang, Xiaodong, Yang, Wei, Cai, Yonghua
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
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
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
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
The rapid growth of the low-altitude economy has accelerated the deployment of unmanned aerial vehicle (UAV) technology in a wide range of applications, including power infrastructure inspection and disaster relief operations. These scenarios impose stringent real-time processing demands on onboard vision systems. However, aerial images are characterized by a high density of small and distant targets, which significantly weakens feature representation quality, and current state-of-the-art detection models carry substantial computational overhead that conflicts with the limited hardware resources available on embedded UAV platforms. This paper addresses both challenges simultaneously by proposing a novel lightweight, end-to-end detection framework designated the High-frequency Enhancement Retentive Detection Transformer (HER-DETR). The framework introduces three coordinated technical contributions. First, a lightweight backbone network termed the Cross Stage Partial High-frequency Enhancement Network is constructed. This backbone integrates a cross-stage high-frequency residual enhancement mechanism to alleviate feature attenuation commonly observed in small-target scenarios. Second, a Depthwise Intrascale High-Low Feature Interaction block is designed for intra-scale feature interaction at the highest semantic level, strengthening fine-grained representation through a dual-frequency decoupling strategy. Third, a Retention Feature Fusion module is developed as the core cross-layer feature fusion component, enabling bidirectional multi-scale feature integration via a lightweight sparse attention mechanism. Comprehensive experiments on the VisDrone2019 and HIT-UAV benchmark datasets confirm that HER-DETR achieves state-of-the-art performance. Compared with the RT-DETR-r18 baseline, the proposed framework improves mean average precision at a 0.50 intersection-over-union threshold by 2.0 percentage points on VisDrone2019 and by 1.6 percentage points on HIT-UAV, while simultaneously reducing model parameter count and computational floating-point operations by 41.6% and 35.3%, respectively. These results validate the framework’s capability to balance detection accuracy and computational efficiency for real-time UAV deployment.
Jiao, Minghai, Liang, Yinghan, Zhang, Ende, Kan, Shuang
Detecting surface defects on aircraft is critical to safety in the aerospace field. The current detection methods suffer from low efficiency, high false positive rate and poor adaptability to complex environment. Although the existing target detection algorithms have achieved notable performance in this field, at present, many models in this field still face considerable challenges. For example, due to the constraints of hardware platforms, machines that can be deployed on edge devices such as UAVs and robots need lightweight and easy-to-deploy models, but lightweight designs often a certain degree of accuracy degradation. At the same time, most of the surface defects of the aircraft are small targets, which impose higher demands on the model. In order to solve these limitations, this paper proposes an improved YOLO model ME-YOLO (multi directional edge fusion YOLO) based on edge enhancement and multi-scale fusion. Firstly, the EdgePyramid module is designed, which uses the rich edge responses extracted by the Multi Orientation Attention Sobel (MOAS) module and through the three-layer cascade pool structure to construct the three-scale edge features aligned with the backbone network. Then, the EdgeEnhance module is introduced to adaptively fuse the conventional convolution features and edge features at the corresponding resolution, effectively enhancing the sensitivity of the model to the corrosion boundary. Experimental results show that, compared with the baseline YOLOv11, ME-YOLO improves mAP@0.5 by 2.7% and simultaneously increases Recall by 2.3%, which helps reduce missed detections of corrosion targets. Although the introduction of multi-directional edge priors brings a slight increase in parameters (+3.66M) and a minor decrease in inference speed (--8.6%), the model still maintains a real-time detection capability of 97.54 FPS, meeting the deployment requirements for edge devices such as UAVs. This strategy of trading a modest computational overhead for improved detection rates in safety-critical tasks offers a potentially practical and robust alternative for intelligent operation and maintenance of aviation equipment.
Zhou, Xing, Li, Yuan
Acoustic emission (AE) signals generated by incipient aero-engine damage are usually weak, intermittent, and highly dependent on sensor position, which makes diagnosis based on a single channel unreliable. To address this issue, this paper develops a multi-sensor AE diagnosis framework that combines entropy-guided feature fusion with a convolutional neural network (CNN). Time-domain and frequency-domain descriptors are first extracted from AE waveforms collected by multiple sensors. The consistency of each feature type across channels is then evaluated using feature entropy. Features with lower entropy are considered more informative for damage discrimination and are therefore retained for fusion. Based on the normalized channel contributions of these selected features, an adaptive weighted fusion matrix is constructed and used as the input to the CNN. The method is validated on a self-built aero-engine structural test rig equipped with 12 AE sensors. Fifteen representative initial-damage cases are designed, including signal excitation, crack-like excitation simulated by pencil-lead break (PLB), and impact-induced micro-deformation cases. Experimental results show that the proposed framework achieves a best average classification accuracy of 96.80% and provides more balanced class-wise performance than conventional machine-learning and deep-learning baselines. Additional noise-injection tests further indicate that the method maintains good robustness over a wide signal-to-noise-ratio range. These results suggest that entropy-guided multi-sensor AE fusion is a practical solution for intelligent diagnosis of aero-engine initial damage.
Jin, Yucheng, Han, Cong, Zhang, Leixuan, Yang, Guoan
To address the joint estimation of state and parameters for fixed-wing aircraft in wind gradient environments, this paper proposes a high-precision identification method based on a physics-informed neural network. By integrating wind gradient models with flight dynamics equations, a multi-objective optimization framework combining data-driven approaches with physical constraints is established. Employing polynomial time embedding and a dual-branch network architecture, coupled with a dynamic training strategy, simultaneous estimation of flight state and wind gradient parameters is achieved. Experimental results demonstrate that this method achieves approximately 7% parameter estimation error while accurately reconstructing unobserved attitude information. Requiring only four-dimensional observation data, this approach provides an effective solution for wind gradient glide navigation in low-cost aircraft.
Wu, Ting, Bai, Zhuoyue, Gao, Xianzhong
Rolling bearings are key components of aero-engine rotor systems, and their degradation can trigger abnormal dynamics and safety risks. Unplanned bearing failures increase downtime and inspection cost; early detection enables condition-based maintenance and reduces conservative replacement. Early faults are weak, intermittent, and often masked by operating variability, so diagnosis based on a single sensing modality is usually incomplete. Vibration signals describe global dynamic behavior and resonance responses, whereas acoustic emission (AE) captures micro-damage processes and short impulsive events. Given the differences in bandwidth and noise sources, combining vibration and AE can provide a fuller description of bearing states. This study proposes an STFT–MFDCNN framework for bearing fault diagnosis using fused vibration–AE information. First, synchronized vibration and AE signals are converted into time– frequency images via the short-time Fourier transform (STFT), preserving spectral content and temporal non-stationarity in a unified representation. Second, a dual-branch multi-scale feature dense convolutional neural network (MFDCNN) is constructed. In each branch, parallel convolutions with different kernel sizes capture fault patterns across receptive fields, and dense concatenation connects all preceding feature maps to improve feature reuse and gradient flow. Third, global average pooling generates compact feature vectors for both branches, which are concatenated in a fusion head and classified by a Softmax layer. Experiments on a vibration–AE dataset built in this work show that the proposed method reaches 97.78% identification accuracy. Compared with dual-branch fusion baselines based on AlexNet, ResNet, and a shallow CNN, accuracy increases by 3.34%, 4.45%, and 5.00%, respectively. Compared with single- modality models (Only-AE and Only-Vib), accuracy improves by 14.71% and 6.77%.
Li, Zhichuang, Wang, Xingfeng, Zhang, Yanpu, Yang, Guoan
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
Thermal barrier coatings (TBCs) serve as the critical protective layer for gas turbine components. In complex service environments, the thermally grown oxide (TGO) layer formed between the top coat (TC) and the bond coat (BC), along with the evolution of its microscopic morphology, is a key factor determining the service life of TBCs. To address the complex morphology of TGO layers that challenges conventional segmentation methods, this paper proposes MSF-SA ResU-Net, a multi-scale feature fusion self-attention model based on the ResU-Net architecture, for the image segmentation of thermally grown oxide layers. Building upon residual connections, this model incorporates two key enhancement modules at the bottleneck layer: Multi-Scale Fusion (MSF) and Self-Attention (SA). During the feature processing stage, the MSF module effectively captures the multi-scale morphological features of the TGO layer through receptive fields of different sizes, solving the difficulty of extracting long-strip and irregular morphologies. Subsequently, the SA module recalibrates the fused features by calculating the global dependencies between them, enhancing the network's attention to TGO boundaries and micro-defects. Comprehensive experiments on an in-house dataset of 3,000 SEM images demonstrate that MSF-SA ResU-Net outperforms standard U-Net and contemporary segmentation architectures. Furthermore, due to the feature extraction capability of the ResNet backbone network, the model maintains high accuracy and exhibiting good convergence performance. This makes the architecture an effective solution for the microscopic analysis of thermal barrier coatings.
She, Weicen, Song, Yan
Planetary gears are key transmission components in spacecraft drive systems and are prone to failures under long-term operation in harsh environments. To address the problem that fault diagnosis and life prediction are treated separately in existing research, this paper proposes an integrated diagnosis-prediction method. First, order tracking and time synchronous averaging are used to preprocess vibration signals, eliminating speed fluctuations and noise interference. Then, fault features are extracted via VMD/EEMD decomposition, and the gear health condition is identified through a deep convolutional neural network. Finally, ensemble learning prediction models are established separately for different fault states to realize life prediction. Experimental results show that the VMD+CNN method achieves higher diagnostic accuracy than traditional methods for gears in healthy, cracked, spalled, and pitted conditions. The ensemble prediction models matched with different states can further reduce the life prediction error. This method provides effective technical support for condition-based maintenance of spacecraft planetary gears.
Liu, Yanyan, Gao, Tongxin, Zhu, Yangsong
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
Due to the limitation of unstable wireless links and resource-constrained edge computing device, it is challenging to transmit the unmanned aerial vehicles (UAVs) captured images with large data volume in real-time. To address this issue, this paper proposes a lightweight hybrid compression method named Jpeg Plus Plus Network (JppNet) for efficient multimodal image transmission. Different from existing methods that mainly focus on RGB image compression for offline storage and often use large networks for coding and reconstruction, the proposed JppNet focuses more on edge deployment under limited computing resources. At the encoder side, JPEG is first used to perform high-ratio compression on the RGB, D, and IR channels. Then, a lightweight convolutional encoder network extracts supplementary binary features for progressive transmission. At the decoder side, a medium-scale reconstruction network fuses the supplementary code with the JPEG reconstruction to restore image details. To verify the effectiveness of the proposed method, this paper built a multimodal data acquisition prototype and collected 20,000 high-quality RGBD-IR image samples for training and testing. Experimental results show that JppNet combines the high coding efficiency of JPEG with the detail recovery ability of convolutional neural networks. With low computational cost, JppNet obtains much lower reconstruction errors than JPEG on the RGB, D, and IR channels, especially at very low compression ratios. This supports the efficient transmission of UAV captured multimodal images under limited communication conditions.
Pi, Haoran, Ao, Yile, Li, Zhi, Tang, Gang, Jiang, Bowu
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
Inter-satellite links (ISLs) in large LEO constellations vary quickly with time, and their rapid change makes future link-state prediction a practical concern. Predicting ISLs with reasonable accuracy benefits routing decisions, resource scheduling, and overall network stability. Existing methods are still limited in several ways. Rule-based and physics-based methods build on hand-crafted assumptions, which become unreliable once links flicker frequently or topologies grow heterogeneous. Static graph neural networks describe the relations within a single snapshot, but their formulation does not represent how ISLs evolve from one moment to the next. Many recent dynamic graph models stay close to interpolation, and they exploit physical and temporal satellite attributes only in part, which weakens both node representations and prediction accuracy. To address these gaps, we develop an extrapolation-oriented dynamic link prediction framework with three contributions: (1) we shift the task from interpolation toward future-state extrapolation, which makes proactive management possible; (2) we extract physics- and temporal-aware node features by combining orbital dynamics with historical ISL information; (3) we restrict prediction to a candidate edge set, so the computation stays manageable while the method remains scalable. In experiments, the proposed method outperforms existing baselines on accuracy, efficiency, and scalability, suggesting that it can serve as a useful tool for managing large-scale LEO networks in practice.
Zhao, Jiayi, Fei, Shihan, Bao, Junpeng
To address the low computational efficiency problem by numerical calculation and large cumulative error by long-term sequential iterative prediction during the process of predicting the satellite’s transient temperature fields, this paper proposes a Physics-Informed U-Net (PI-Unet) model. Specifically, this PI-Unet model uses multi-branch spatiotemporal features as its inputs and combines hard constraints with a physics-informed loss to ensure the physical consistency of the predicted temperature field. Furthermore, for the error accumulation problem in long-sequence autoregressive prediction models, this paper presents a model training method based on the concept of filtering. Under the guidance of this method, the PI-Unet model can develop a long-term sequence prediction capability while keeping the accuracy of spatial temperature field prediction, achieving high-precision spatiotemporal prediction. By actively injecting prediction errors through a scheduled sampling strategy, along with a multi-step rolling training strategy, the PI-Unet model not only learns the governing physical laws of the spatial distribution of the transient temperature field, but also acquires the filtering capability to mitigate the accumulation of the long-term sequence rolling prediction errors. The experimental results indicate that the PI-Unet model maintains the maximum absolute error of predicting spatial transient temperature field stably below 5.5K and the error of the rolling-predicted transient temperature field in one orbit period less than 6K. In addition, the time for a single-step prediction is reduced to 0.005s, and the time for a total-period prediction is less than 0.5s. Therefore, this research provides a strong and high-efficiency technical support for the real-time monitoring of the satellite’s on-orbit temperature fields.
Li, Wanze, Xiao, Fuhong, Xue, Xiaoguang
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
Satellite edge computing (SEC) enables low-latency services by leveraging Low Earth Orbit (LEO) satellites. However, emerging applications increasingly generate dependency-aware tasks with complex precedence constraints, leading to challenging multi-DAG (directed acyclic graph) offloading under dynamic satellite environments. To address this issue, we propose a global task graph (GTG)-based multi-agent offloading frame-work named GTG-MARL, for dependency-aware task execution in regional satellite-ground collaborative networks. Specifically, we construct a GTG by merging multiple DAGs to capture both intra-task dependencies and inter-task coupling, thereby achieving unified dependency-aware modeling. Then, a graph neural network (GNN) is lev-eraged to encode the GTG into dependency-aware subtask embeddings. To handle dependency-aware offloading in dy-namic environments, we develop a multi-agent reinforce-ment learning framework based on generalized proximal policy optimization, where each LEO satellite acts as an agent for decision-making while a centralized critic stabilizes offline learning to eliminate real-time on-orbit communication overhead. During the execution phase, only the light-weight trained actors are deployed on satellites for on-orbit inference. Extensive experiments demonstrate that the pro-posed method significantly reduces task completion delay compared to state-of-the-art baselines, validating the practi-cal efficiency and low synchronization cost of the proposed framework for regional SEC deployment.
Geng, Zeao, Zhao, Lu, Jin, Huiying, Jiang, Ping, Cai, Hui, Zhou, Jian
Bearing signals measured under different loads rarely keep the same waveform characteristics. Their amplitudes, spectral components, and statistical properties may change at the same time, which makes a classifier trained on one load condition less reliable on another. This study investigates a cross-load diagnosis method that trains a multi-scale convolutional squeeze-excitation network (MCSECNN) with multi-kernel maximum mean discrepancy (MK-MMD) and a domain-adversarial neural network (DANN). The convolutional branch captures vibration patterns over several temporal ranges, and the SE unit adjusts the importance of the channels generated by the parallel kernels. MK-MMD penalizes the distance between load-dependent feature distributions, whereas DANN weakens the load identity retained in the features. Tests using CWRU bearing data show that the integrated model gives better accuracy than the basic MCSECNN and the two single-adaptation variants. Ablation results indicate that the statistical and adversarial constraints contribute to transfer performance in different ways.
Ming, Mei, Ai, Yumei, Li, Shengnan, Wan, Haozheng, Sang, Shentian, Zhang, Haifeng
As part of the dtec.bw MORE project, serial hybrid powertrains featuring alternative combustion regimes, specifically homogeneous reactivity-controlled compression ignition (hRCCI), are under investigation to maximize thermal efficiency while minimizing NOx and soot emissions. The proposed hRCCI concept utilizes dual-fuel stratification, employing early direct injection (DI) of octanol (high-reactivity fuel) alongside port fuel injection (PFI) of ethanol (low-reactivity fuel). While 0D/1D engine modeling is essential for developing predictive powertrain simulation frameworks, conventional models often lack the robustness required to capture the complexities of low-temperature combustion (LTC). This study addresses this limitation by developing a multi-zone “onion skin” quasi-dimensional model. Since for LTC, ignition and heat release rates are highly sensitive to thermal and chemical stratification, capturing these gradients is critical. 3D computational fluid dynamics (CFD) simulations are capable of capturing thermal and chemical stratifications to a high degree of accuracy, whereas 0D simulation models, in general, do not consider them. The primary contribution of this work is the translation of high-fidelity 3D CFD data into a computationally efficient quasi-dimensional environment. A simplified 3D CFD architecture was utilized to map the effects of various operating parameters on mixture distribution. Using these results, a regression learning model was trained to predict octanol and temperature stratification within the combustion chamber. Validation against simulation data demonstrates that this coupled machine learning and multi-zone approach provides a robust, predictive tool for evaluating advanced LTC concepts within larger system-level simulations.
Sundaram, Pravin Kumar, Grundl, Larissa Michaela, Trapp, Christian Thorsten, Tinschmann, Georg
Unmanned ground vehicles (UGVs) operating in unstructured environments must account not only for terrain traversability but also for terrain-induced loads that affect component durability. Existing path planning approaches primarily consider obstacle avoidance and mobility, while neglecting cumulative structural degradation due to repeated loading. High-fidelity physics-based simulations can capture these effects but are computationally prohibitive for real-time applications. This study investigates machine learning-based surrogate models for predicting vehicle component reaction forces from terrain height sequences generated using a controlled parametric terrain formulation. Both feed-forward and recurrent neural network architectures are evaluated, and ensemble-based probabilistic techniques are incorporated to quantify predictive uncertainty. Results show that the ensemble long short-term memory (Ens-LSTM) model achieves the lowest prediction error (mean absolute error of 0.621 kN) while maintaining narrow 95% prediction intervals (3.22–4.28 kN). A simpler ensemble feed-forward network (Ens-NN) achieves comparable accuracy (0.626 kN) with reduced model complexity. These results demonstrate that data-driven surrogate models can provide accurate and uncertainty-aware force predictions, enabling the integration of structural reliability considerations into fatigue-aware path planning for UGVs.
Chua, Yang Kang, Mundiwala, Mohammad, Wang, Xudong, Castanier, Matthew, Hu, Zhen, Hu, Chao
This paper presents a deep learning-based approach for online rotor temperature estimation in electrically excited synchronous motors (EESMs). Accurate rotor temperature estimation is critical for ensuring safe operation, improving performance, and enabling reliable thermal management of electric traction motors. Recurrent neural network (RNN) architectures, including gated recurrent unit (GRU) and long short-term memory (LSTM) networks, are investigated to develop a data-driven thermal virtual sensor capable of capturing the temporal dynamics of motor operation. Experimental data collected from a 190 kW EESM prototype are used to train and evaluate the proposed models. A systematic training, testing, and 10-fold cross-validation framework is employed to assess prediction accuracy and generalization capability. The results demonstrate that the GRU-based model achieves higher prediction accuracy than the LSTM model while maintaining comparable inference latency. The proposed approach provides an efficient and lightweight solution for real-time rotor temperature estimation suitable for embedded motor control applications.
Tatari, Farzaneh, Aligoudarzi, Mohsen Mirza
This paper describes ongoing research and development of an efficient optimization/search–based modeling and simulation framework for rapidly identifying low-performance scenarios in advanced autonomous systems. Ensuring predictable, safe behavior across complex, integrated systems remains a core operational test-and-evaluation challenge. Our goal is to balance rigorous validation with timely deployment. We are developing TEAAS (Test & Evaluation of Advanced Autonomous Systems), a scalable, faster-than-real-time framework designed to uncover critical failure scenarios efficiently. Key features include GPU-accelerated parallel simulation and learning, computational intelligence–based search of optimal parameters, uncertainty quantification for reproducibility, and real-time physics-accurate sensor models. We conducted simulation experiments to evaluate and demonstrate the framework performance for two black-box ground-vehicle autonomous systems. Key results were that adequate uncertainty quantification can be achieved with as few as 10 repeated runs per simulation scenario, sensor realism has a significant effect on failure rate, distinct differences between the two autonomies failure modes were identified, and our efficient optimization/search methods identify critical performance regions in a small fraction of the number of simulations required by a naïve Monte Carlo search.
Snarski, S., Menozzi, A., Persons, B., Lazar, D., Khan, N.
Compression-ignition engines operating with biodiesel blends often exhibit variability in fuel properties, such as density, viscosity, and cetane number, which can lead to systematic deviations in injected fuel mass when using conventional physics-based models. These deviations can reduce combustion efficiency and increase brake-specific fuel consumption (BSFC). This study proposes a lightweight neural network–based approach to compensate for structural errors in baseline injection models, using a single-layer perceptron trained on the relative error (delta) between actual and modeled injected mass. By normalizing engine and fuel parameters and introducing a small amount of measurement noise, the network learns to predict a corrective factor that adapts the injected mass to match the desired target under varying fuel conditions. Simulation results demonstrate that the neural correction significantly reduces systematic bias: in test cases with intentionally introduced structural error, the average injection deviation of −1.7% in the baseline model is reduced to approximately 0.002% after correction. Root-mean-square (RMS) error over training and validation datasets remains below 0.16%, indicating robust generalization. The proposed method offers a computationally efficient solution suitable for embedded engine control units, requiring minimal additional complexity while ensuring precise fuel delivery. By eliminating bias caused by fuel property variability, the approach has the potential to improve fuel economy, reduce emissions, and maintain consistent engine performance under a wide range of operating conditions. This framework provides a practical path for integrating adaptive, data-driven correction mechanisms in diesel engines operating with heterogeneous or variable biofuels.
Gutierrez, Marcos, Taco, Diana
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
This study presents a computational framework that integrates an air-standard thermodynamic engine model with artificial neural networks to predict the performance of spark-ignition (SI) engines operating with alternative fuels of reduced lower heating value (LHV). A deterministic thermodynamic simulator was developed in Excel, incorporating engine geometric parameters (compression ratio, bore, stroke, displacement), operating speed, and fuel properties, with particular emphasis on LHV as the dominant energetic descriptor. The model computes in-cylinder states, thermal efficiency, indicated mean effective pressure, and power output under idealized air-standard assumptions. To extend predictive capability beyond fixed-parameter analyses, a feedforward neural network was trained using datasets generated from systematic parametric sweeps of engine geometry, speed, and fuel LHV. The neural network captures nonlinear interactions between compression ratio, combustion energy release, and performance indicators, enabling rapid estimation of engine response when substituting conventional gasoline with lower-LHV alternative fuels. Results demonstrate that the hybrid thermodynamic–neural approach accurately predicts trends in efficiency degradation and power reduction associated with decreasing LHV, while identifying compensatory design adjustments, particularly through compression ratio optimization. The methodology provides a low-cost and computationally efficient tool for preliminary evaluation of alternative liquid fuels in SI engines without resorting to complex CFD or experimental campaigns. This work contributes a transparent, reproducible modeling strategy suitable for early-stage engine design studies and fuel screening, supporting sustainable fuel transitions in spark-ignition propulsion systems.
Gutierrez, Marcos, Taco, Diana
With continuous advancements in load-side resources such as distributed photovoltaic systems, electric vehicles, and virtual power plants, the low-carbon and sustainable development attributes of power systems have been significantly enhanced. Meanwhile, the coupling intensity between sustainable power systems and meteorological conditions has been further consolidated. Considerable impacts are exerted by weather variations, particularly extreme weather events, on the dispatching and operation of sustainable power systems. Accurate load forecasting is critical for enabling sustainable power systems operators to optimize power generation strategy, ensuring supply stability and resilience against extreme weather-induced disruptions. However, the intrinsic non-stationarity and volatility of extreme weather events present significant challenges to conventional forecasting approaches. Herein, we introduce a hybrid algorithm integrating Newton–Raphson-based optimizer (NRBO) with extreme gradient boosting (XGBoost) to enhance short-term load predictions under such conditions. The model uses optimally selected meteorological and load features as inputs, while NRBO systematically tunes XGBoost’s hyper-parameters to maximize performance. Evaluated on an Irish dataset, the proposed framework is quantitatively compared against five baseline models, including traditional decision trees and neural networks. The case studies show that the mean absolute percentage error (MAPE) of the proposed model is 2.57%, which is the lowest among these decision tree and neural network algorithms.
Wang, Yi, Zhou, Jian, Wu, Gang, Ma, Tiannan, Ma, Ruiguang, He, Chuan, Zhu, Huixian
Brake pad wear is a major and growing source of non-exhaust particulate emissions, projected to reach 1.3 million tons annually by 2030 and contributing up to roughly 55% by mass of non-exhaust traffic-related PM10 in urban environments, underscoring the need for improved durability and material optimization. This study investigates a three-stage eXtreme Gradient Boosting (XGBoost) ensemble paired with a residual Fully Connected Neural Network (FCNN) corrector to predict brake pad wear rate and support formulation optimization. Experiments used a simplified FMVSS 135 protocol on a Universal Mechanical Tester (UMT) simulating realistic braking across eight friction regimes. Wear rate was the sole machine-learning prediction target, while coefficient of friction (CoF) was retained as an input feature rather than a target. Despite a limited but high-quality 280-cycle dataset, regime-aware stratified splitting, sample reweighting, and hyperparameter optimization enabled robust generalization. The three-stage XGBoost ensemble with residual FCNN correction achieved a global held-out test R2 of 0.976 for wear rate prediction. A Taguchi L8 design of experiments defined the brake pad compositions, reducing experimental time and material consumption compared to conventional approaches. The framework demonstrated strong agreement between measurements and predictions for the dominant low-severity regime, while per-regime analysis identified the high-severity minority regimes as the priority for additional data collection, since within-regime R2 remains negative for every regime given current sample sizes. A sequence-aware mean absolute scaled error (MASE) analysis further shows that, despite the high global R2, none of the four pipeline stages currently outperforms a naive one-cycle persistence forecast on absolute error, a distinction reported here for transparency. The scalable architecture enables straightforward integration of additional material and process parameters, supporting iterative brake formulation development in industrial settings and, by reducing empirical testing requirements, sustainable brake material development with reduced replacement frequency and associated emissions.
Katakam, Abhishek, Eslamiat, Hossein, Kancharla, Sai Krishna, Filip, Peter
Recently, there has been a drastic shift in the industry towards wire architectures like steer-by-wire and brake-by-wire. For safe and accurate force control, diagnostics, and consistent performance over the operating envelope, accurate plant modeling of the Electro-Mechanical Brake (EMB) is important. Classical approaches involved linearized dynamic EMB models and the use of the characteristic stiffness curve for calibration at the operating points. These methods often perform poorly over regions where hysteresis, compliance, and friction are strongly nonlinear. Prior research on state or force estimation for EMB has focused on pad contact detection, thermal adaptation, and hysteresis-aware clamp force estimation. However, there are still accuracy gaps in practical applications during transients and under shifting friction regimes. In this work, a digital twin based on Physics-Informed Machine Learning is introduced, following the governing dynamics of the actuator-caliper assembly of EMB while learning (i) a physically significant parameter—system damping (Bsys) and (ii) a non-linear friction term constrained as a function of the actuator motion states and operating conditions. Non-linear friction is captured through gray-box friction formulation and learning unmodeled residual dynamics such as hysteresis and backlash. An EMB test stand is used to collect steps, ramps, holds/engagements, APRBS, and swept-sine excitations, with signals including time-aligned force command, motor torque/current, actuator position/velocity, and pad force measurement from a force sensor for model training. Results demonstrate a decrease in pad-force prediction error, along with non-linear and residual friction estimation. The resulting digital twin can enable sensor-less force estimation, friction compensation design, predictive analytics, and health monitoring through tracking parameter drift and friction signatures.
Rai, Prakhar, Gadhvi, Tirth
Teardown evaluation of chassis system components plays a critical role in benchmarking, failure analysis, and competitive product assessment. These inspections rely heavily on experienced engineers who interpret visual defect patterns, material conditions, wear signatures, and manufacturing variations. However, expert driven evaluation processes are often subjective, difficult to standardize, and challenging to scale across global engineering teams. This paper presents a structured AI-assisted expert evaluation framework developed to enhance consistency, preserve institutional knowledge, and enable continuous improvement in chassis component teardown analysis. The proposed system integrates convolutional neural network architectures, including ResNet18 and its variants, into a human in loop inspection workflow. AI models perform initial classification of component images (e.g., OK/not OK and defect subclasses) and provide associated confidence scores. These predictions are presented as decision support, while final authority remains with the evaluating expert. Experts can confirm or override AI outputs, annotate defect regions using bounding boxes, assign subclass categories, and provide structured technical comments. All expert interactions, including AI disagreements, are systematically recorded. Correction instances are analyzed to identify model limitations, ambiguous defect conditions, and data gaps. Expert validated evaluations are incorporated into the training dataset to enable iterative model refinement. This closed loop process supports progressive improvements in model robustness and classification accuracy across varying teardown conditions and component types. A centralized cloud-based repository maintains full traceability of inspections, including timestamps, AI confidence levels, expert modifications, and annotation metadata. This structured knowledge capture converts tacit engineering judgment into a persistent digital asset, supporting auditability, cross-site alignment, and accelerated onboarding of new engineers. The framework demonstrates how AI can be effectively deployed as an assistive technology in chassis teardown evaluation, improving repeatability, enhancing data driven benchmarking, and enabling scalable knowledge preservation without displacing expert authority.
Divakaruni, Saikiran, Vaibhav, Veer, Hansen, Scott, Hood, Trevor, Agrawal, Rahul
Heavy-haul railway development is as important a strategic direction for China’s railway sector as high-speed railway development. With the continuous expansion of heavy-haul railway operational mileage and the growing transportation demand, the complexity of operational adjustment in heavy-haul railways continues to escalate. The operational adjustment for heavy-haul trains subjected to speed restrictions following maintenance windows, aimed at maximizing the restoration of timetable regularity and guaranteeing freight volume, constitutes a critical and urgent research problem. Grounding on the operational principle of heavy-haul railways that prioritizes freight volume preservation over strict punctuality, this paper proposes a rescheduling model that incorporates the timetable disruption level. By adopting the order entropy theory, the model is formulated to minimize the total system disruption level and freight time cost, thereby framing the rescheduling problem as an order entropy optimization problem. The train disruption level is a metric that quantifies the perturbation intensity of a rescheduled timetable relative to the original operational plan within a given railway section. By initializing the entropy value of the pristine operational order to zero, any subsequent entropy variation directly facilitates a systematic investigation into the evolution of train operational order. Effective timetable adjustment strategies are subsequently derived, contingent upon the available buffer time and freight volume constraints. Guided by actual operational timetables and sectional line characteristics, this research designs a case study. The Gurobi solver is employed to obtain a rescheduled heavy-haul train timetable that successfully restores the original freight volume. The proposed methodology contributes to the enrichment of railway transportation organization theory and provides robust decision-making support for the strategic deployment of heavy-haul train services.
Fu, Lu, Lin, Li, Xu, Shichao, Ji, Guanggang, Li, Zheng
This paper addresses the issue of regenerative braking energy recovery in new energy vehicles and designs and optimizes a braking force distribution strategy. The strategy uses an ANFIS controller to dynamically optimize the proportion of front-axle regenerative braking force. The introduction of a pruning algorithm reduces computational complexity, thereby enabling a significant increase in mileage while maintaining stable driving performance. Co- simulations integrating Simulink and AVL Cruise, alongside Hardware-in-the-Loop (HiL) tests, the proof is that this strategy can still maintain excellent stability under different braking intensities. Moreover, it exhibits significantly higher energy recovery efficiency compared to benchmark strategies, while its effectiveness and real- time performance are successfully validated.
Lin, Hui, Zhao, Xuezhan, Tian, Jiahao
In order to conduct more in-depth research on the driving sight distance of curved tunnels in mountainous highways, a systematic theoretical calculation model of spatial sight distance of curved tunnels based on three-dimensional characteristics is established, and the spatial sight distance value of curved tunnels in mountainous highways is recommended in combination with the changes of driving behaviour under different spatial sight distances. Firstly, the concept of spatial sight distance of curved tunnels is proposed, the theoretical calculation model of spatial sight distance of curved tunnels is established, and the model is verified by a multi-scale neural network; Secondly, five UC-win/road simulation models of curved tunnel with different spatial sight distances are established, and the simulation experiments are carried out in combination with mp160 multi-channel physiological recorder and SMI etgtm eye tracker; Finally, the mathematical statistics method and SPSS software are used to analyse the operation behaviour, psychological behaviour and eye movement behaviour of drivers in curved tunnels with different spatial sight distances, and to verify the different effects of the critical value of spatial sight distance on driving behaviour in the theoretical calculation. Furthermore, by taking the spatial sight distance as the independent variable, the regression model is established with the average speed, trajectory offset, heart rate change rate, and pupil diameter change rate as the dependent variables. Based on the driver’s behaviour threshold, the recommended spatial sight distance of a curved tunnel is proposed. The results show that the recommended range of spatial sight distance of the curved tunnel of mountainous highway with a design speed of 80 km/h is 125 m to 140 m, the limit value is 110 m, and the appropriate value is 155 m. There is a critical value between two-dimensional sight distance and spatial sight distance, which has a significant impact on the change of driving behaviour in a curved tunnel.
Tang, Xie, Zheng, LiWen, Lin, GuoJin, Gao, YanYang, Lan, FuAn
In recent years, China's urban rail transit sector has undergone rapid expansion, with passenger demand consistently increasing. Accurate passenger flow forecasting is essential for ensuring efficient and safe metro operations. This paper takes Nantong Metro Line 1 as a case study and applies an optimized forecasting approach that integrates a grey metabolism model with the Holt double-parameter exponential smoothing method. Based on an analysis of Automated Fare Collection (AFC) data from March 2023 to February 2024, passenger flow on this line demonstrates a clear linear growth trend, which aligns well with the assumptions of the grey metabolism model. The results indicate that the optimized grey metabolism model not only significantly enhances prediction accuracy but also greatly reduces the variance ratio, demonstrating high reliability in forecasting outcomes. This improved methodology provides a more robust tool for metro operators in planning services, managing capacity, and optimizing resource allocation.
Fan, Fan, Zhao, Zeheng, Zhang, Jin, Ma, Junhao, Qian, Beiyue
This study proposes a physics-informed graph convolutional reduced-order model, namely Phys-GCN, for high-fidelity and computationally efficient prediction of steady incompressible flow fields. In Phys-GCN, the incompressible Navier–Stokes equations are embedded into the loss function via residual constraints, such that the spatial feature extraction of graph convolutional networks is integrated with the physics-constrained learning strategy of physics-informed neural networks. This mixed design enables the model to capture complex nonlinear flow features while maintaining a clear level of physical interpretability. Benefiting from the node-edge encoding inherent to graph neural networks, Phys-GCN operates directly on unstructured CFD meshes to learn flow features from graph representations constructed using node attributes and adjacency relationships. In doing so, Phys-GCN dispenses with voxelization or SDF preprocessing and fully preserves the local geometric and topological characteristics of the flow domain. The proposed model is systematically evaluated on steady flows past circular and elliptical cylinders, where the predicted velocity and pressure fields are compared against reference CFD solutions in both interpolation and extrapolation scenarios. Results show that, for all physical quantities, the reconstructed steady flow fields achieve mean relative errors below 5%, exhibiting excellent agreement with the CFD benchmark solutions. After offline training, Phys-GCN achieves inference times that are several orders of magnitude faster than conventional CFD solvers, while maintaining comparable predictive accuracy. These findings demonstrate that Phys-GCN provides an accurate and efficient graph-based and physics-informed surrogate for steady flow-field reconstruction on non-uniform, unstructured meshes, thereby laying a solid foundation for future extensions to more complex three-dimensional and compressible flow configurations.
Xie, Haoran, Zhou, Hao, Yu, Changhao, Li, Qiang, Liu, Tianyu, Peng, Jiangzhou
The folding wing mechanism is widely used in aircraft design. Whether the folding wing surface can unfold smoothly determines whether the aircraft can fly normally. Therefore, studying the aerodynamic loads and structural deformations during the unfolding process of folded wing surfaces is very important. The motion process of a folded wing mechanism is a typical fluid-structure interaction (FSI) process. During deployment, the wing surface moves under the combined action of the actuator’s pull and the aerodynamic loads from the incoming flow, while the large deformation of the wing surface during its movement, in turn, affects the aerodynamic loads on the mechanism from the flow field. Considering the FSI effects during the unfolded motion process of the folded wing, simulation was conducted using the ALE algorithm in LS-DYNA to obtain the kinematic and dynamic parameters in the unfolded motion process, and also to get the aerodynamic torque on the wing under different angles and angular velocities. In practical engineering applications, the actuation force of the deployment mechanism can vary due to factors such as the amount and performance of the pyrotechnic material. Consequently, the final velocity and the whole motion process of the wing mechanism will also change. For the calculation of aerodynamic external loads under multiple operating conditions, using the ALE algorithm will consume a large amount of computational time and cost. Given the high computational cost and long computation time of finite element simulations, a BP neural network was established to calculate the aerodynamic loads on the wing surface under different actuation forces. This allows for a rapid assessment of whether significant deformation or damage will occur to the folding mechanism or nearby components during the deployment process.
Wei, Ting, Li, Naitian, Tong, Zongkai
Large-section tunnel construction using the mining method can significantly affect the operational safety of existing metro lines and ground stability, while their non-uniform settlement remains challenging to monitor comprehensively. In this study, the post-station section of a metro project in Chengdu was investigated to elucidate the vertical displacement and ground settlement behavior induced by a large-section tunnel undercrossing an existing metro line. A displacement reconstruction method integrating sparse-point monitoring with a radial basis function neural network (RBFNN) was developed to fit the full-field settlement distribution of both the existing line and the ground surface. A finite element model incorporating the existing shield tunnels, station structures, and the newly constructed mined tunnel was established, and the construction process was simulated. The numerical results indicated maximum settlements of 4.42 mm for the existing line and 4.37 mm for the ground surface, with fitting errors below 5.8% and 6.1%, respectively. Physical model tests further validated the approach, yielding maximum settlements of 0.86 mm and 1.01 mm for the existing line and ground surface, respectively, and an average fitting error below 3.5%. Both numerical and experimental findings confirmed that the induced displacements were within a controllable range and that the surrounding strata remained stable. The proposed method enables accurate and intuitive reconstruction of displacement distribution during under-crossing tunnel construction, reducing the number of required monitoring points while maintaining high fitting accuracy.
Wang, Rui, Chen, Jian, Cheng, Tao, Du, Lin, Li, Ruixiao
Items per page:
1 – 50 of 1616