Browse Topic: Neural networks
Terminal guidance is critical for ensuring strike precision in the final phase of flight. However, traditional methods, such as proportional navigation and optimal guidance laws, face significant challenges regarding real-time performance and adaptability to dynamic targets. To address these issues, neural networks offer a promising solution by enabling adaptive adjustments to guidance parameters, thereby improving performance under various constraints.
Efficient optimization of aerodynamic shapes is a critical challenge in aircraft design. Traditional CFD-based optimization workflows suffer from high computational costs and low efficiency, which severely restricts their practical engineering application. In this paper, a novel aerodynamic optimization method based on a hierarchical neural network with adaptive activation functions is proposed. The network adopts learnable B-spline activation functions and is hierarchically constructed in accordance with the sharing status of B-spline control points. After being trained to achieve fast and accurate prediction of aerodynamic performance, the network can effectively replace the traditional CFD module in the optimization loop. The primary advantage of the proposed method is that it significantly reduces the computational cost during the optimization process while ensuring that the prediction accuracy is not compromised. This work thereby presents a novel strategy and technical framework for streamlining the design process of hypersonic vehicles.
Software-defined, highly customizable vehicle architectures drastically increase the number of hardware–software constellations that must be validated, especially under safety and timing constraints. Traditional unit and integration testing, as well as current regression and combinatorial methods, cannot practically cover this configuration space or reliably capture emergent effects arising from complex interactions, such as bandwidth contention and non-linear latency behavior. This work presents a proof-of-concept for predictive, situational validation of self-describing hardware and software components within realistic automotive E/E architectures. Proposing a novel Machine Learning- (ML) based method for early systemic feasibility prediction of automotive configurations using Graph Neural Networks (GNNs). Specifically, the subclass Graph Isomorphism Networks (GINs) is applied to predict the compatibility of a randomly composed configuration of software and hardware components, assessing both structural compatibility and functional stability. The trained models achieve recall and accuracy above 90%, even when detailed behavioral metadata is hidden during training, indicating that systemic incompatibilities are learnable from topological features alone. Results were achieved from training on a realistic, synthetic dataset representing less than 10e−27% of all possible permutations without finetuning or further parameter optimization. It demonstrates the potential of GIN-based graph learning to enable early, automated feasibility assessment, substantially reducing testing time and development effort for modular, personalized, and update-capable vehicle architectures.
In this study, we propose a methodology for predicting the acoustic modes and natural frequencies of a sedan using artificial intelligence and demonstrate the feasibility of controlling its acoustic characteristics by modifying the hole distribution of the package tray. In typical sedan structures, the cabin cavity and trunk cavity are acoustically coupled through holes in the package tray. The distribution of these holes significantly affects the natural acoustic modes and frequencies of the vehicle. However, once the exterior shape of the vehicle is finalized during the design stage, options for structural modifications to mitigate noise issues caused by these modes become extremely limited. To address this challenge efficiently, we develop a deep learning-based neural network model trained on data derived from a simplified acoustic analysis model of a sedan that includes a package tray. Finite element analysis is performed to generate acoustic modes and natural frequencies, which serve as training data, for various hole distributions. The trained model is then used to predict acoustic natural modes and natural frequencies from unseen input images representing different hole configurations in the package tray. These predictions are made in a fraction of the time required for traditional simulation methods, thereby validating the model’s effectiveness. Furthermore, we demonstrate that the latent variables embedded in the trained model can be manipulated to control the acoustic modes and natural frequencies of the sedan. This indicates the potential for artificial intelligence-driven acoustic design optimization in early-stage vehicle development, offering both time efficiency and design flexibility without physical prototyping or extensive simulations.
Evaluating rotor component clearances is a multidisciplinary process aimed at ensuring that no contact occurs between rotor parts during a rotorcraft's operational life. It begins with calculating relative distances between components across all possible displacements and deformations combinations using a rotor kinematic model, and ends with clearance verification through flight data analysis and simulation. This task requires coupling detailed rotor aeroelasticity with flight mechanics to predict deformation under load, which is computationally expensive and unsuitable for real-time use. This work proposes a machine learning–based alternative: a neural network to estimate rotor clearances from flight mechanics inputs, with a specific application demonstrated in a simulated tiltrotor emergency maneuver with a pilot in the loop. The trained model successfully captures nonlinear relationships between maneuver parameters and rotor structural response, providing accurate predictions with reduced computational cost. The proposed framework can support both preliminary and detailed rotorcraft design and offers potential for real-time load prediction in applications where conventional aeroelastic tools are computationally expensive.
The proliferation of Autonomous Aerial Vehicles (AAVs) necessitates robust solutions for dynamic obstacle avoidance, particularly against non-cooperative intruders whose trajectories are unpredictable. While traditional path-planning algorithms excel in static environments, they struggle with dynamic obstacles due to the inherent difficulty in accurately estimating and registering their real-time depth and velocity into a world model. This paper presents a novel two-stage vision-based framework that leverages deep learning for reactive avoidance of non-cooperative dynamic intruders. Our approach decouples the perception and decision-making processes: an object detection deep neural network first processes monocular camera images to detect and track the 2D pixel coordinates of intruders. This perceptual output is then fed into a deep reinforcement learning agent, which learns a mapping from the intruder's image-space location to a high-level avoidance maneuver. This leads to more efficient learning, as the RL agent focuses solely on the policy without the burden of learning visual features. The advantage of using RL lies in its ability to handle partially observable situations—because reliable depth or full 3-D position information is not always readily available from monocular imagery, the RL agent learns to act based on the observable visual cues. Simulation results confirm that our proposed framework provides an effective solution for vision-based, non-cooperative intruder avoidance.
This paper presents a spatio-temporal graph neural network (STGNN) centric approach to enable heterogeneous agents to collaborate and cooperate for different types of missions. The STGNN-centric approach and corresponding autonomy are encapsulated in the Advanced Graph-enabled Network Technology for Collaborative Autonomous Agents (AGENTCA) technology. Various decentralized and distributed control architectures are reported in the literature, but in some instances these approaches do not leverage the inherent graph network which can increase scalability to larger teams and algorithmic efficiency. Specifically, in this paper advances in artificial intelligence are leveraged to parameterize and encode optimal, or nearly optimal, swarm control techniques. For this work, the team focused on developing a diffusion-based STGNN swarm controller using imitation learning. An expert, centralized swarm control law was used to guide the STGNN during the learning process. The STGNN controller enables the swarm to follow a leader while avoiding static and dynamic obstacles and maintaining a desired separation distance from neighbors and obstacles. The approach is demonstrated in simulation with hundreds of agents and in flight tests with up to thirteen test vehicles.
This paper introduces a robust supervised machine learning framework for estimating helicopter gross weight during the takeoff phase. The methodology leverages high-fidelity datasets from Airbus's global in-service fleet to ensure a reliable training foundation. At the core of the approach is a long short-term memory recurrent neural network, supported by a patented data-curation pipeline designed to maintain high data integrity. To align with rigorous aviation safety standards, the study outlines a learning assurance process compliant with EASA guidelines, specifically addressing safety assessment objectives for machine learning. A central innovation is the characterization and monitoring of the model's operational design domain through multidimensional functional principal component analysis. By projecting high-dimensional, non-linear sensor data into a manageable tabular subspace, this approach enables the definition of safety envelopes using explainable and efficient classical methods. Validated against diverse real-world flight profiles, the framework demonstrates high predictive accuracy, marking a significant milestone toward deploying the model on airborne targets for safety-critical functions such as condition-based maintenance.
This paper focuses on the implementation of a novel supervised Machine Learning model for estimating helicopter weight during takeoff, utilizing extensive datasets from Airbus's global in-service fleet. The study details a learning assurance process aligned with the EASA concept paper for machine learning application, and with the on-going Eurocae ED-324. We propose a set of Machine Learning Requirements, a Machine Learning Model Description, and its implementation for a long short-term memory recurrent neural network. Finally, we verify the requirements on the implementation. Demonstrated on legacy avionics computers, the implementation is suitable for the deployment of the developed Machine Learning Model weight estimator on airborne targets for critical functions such as on-board alerting.
This paper extends a previously developed adaptive pilot model framework for inner-loop roll-attitude tracking [1] to outer-loop position tracking tasks. Pilot Model identification is performed for two command signal types - a discrete step-like signal and a continuous Sum-of-Sines (SOS) signal - yielding distinct parameter signatures that reflect the different anticipatory and tracking demands of each signal type. An adaptive pilot model for the outer-loop position tracking task is formulated using a model-reference neural network (MRNN) architecture with a linearly parameterized neural network updated by a Lyapunov-stable adaptive law. Simulation results for both discrete and continuous tasks demonstrate that the adaptive pilot model remains stable and maintains position tracking performance under both a doubling and a halving of the nominal control sensitivity. Preliminary results are also presented for a multi-axis maritime task, extending the framework to simultaneous lateral and vertical position tracking.
Accurate prediction of equilibrium combustion products and thermodynamic properties is essential for optimizing engine performance, enhancing combustion efficiency, and reducing emissions in diesel-powered systems. Traditional methods for combustion modeling often involve solving complex chemical equilibrium equations or thermodynamic relations, which could be computationally expensive and time-consuming. In this study, we present a data-driven approach using a deep neural network (DNN) model to predict the equilibrium combustion products and key thermodynamic characteristics of diesel under varying thermodynamic conditions. The proposed DNN model is trained on a comprehensive dataset generated from equilibrium calculations. The inputs include pressure, temperature, and equivalence ratio, covering a relatively wide range to encompass diesel equilibrium combustion under various conditions. Outputs are equilibrium combustion products and thermodynamic properties, including enthalpy, internal energy, specific volume, entropy, specific heat, and mole fractions of ten primary combustion species. The model is built using TensorFlow/Keras with a three-layer feedforward architecture and optimized using the Adam optimizer. Input and output variables are normalized using standard scaling, and model performance is evaluated using MSE, MAE, and R2 metrics. Results show that the DNN achieves high accuracy across most outputs, with R2 values exceeding 0.99 for major thermodynamic variables, and provides fast inference suitable for real-time applications. This study shows that deep learning provides a viable and efficient alternative to traditional equilibrium solvers for combustion modeling, with potential applications in engine control, system-level simulations, and design optimization.
Autonomous vehicle navigation requires accurate prediction of driving path curvature to ensure smooth and safe trajectory planning. This paper presents a novel approach to curvature prediction using deep neural networks trained on GPS-derived ground truth data, rather than model predictions, providing a more accurate training signal that reflects actual vehicle motion. We develop a multi-modal neural network architecture with temporal GRU encoders that processes vision features, driver intent signals, historical curvature, and vehicle state parameters to predict curvature. A key innovation is the use of GPS-based actual curvature measurements computed from vehicle motion data (κ = ωz/v) as training supervision, enabling the model to learn from real-world driving patterns. The model is trained on 5,322 samples from real-world driving data collected on The University of Oklahoma’s Norman Campus using a Comma 3X device and a 2025 Nissan Leaf electric vehicle. Experimental results demonstrate high steering curvature prediction accuracy with a Pearson correlation coefficient of 0.805, Mean Absolute Error of 0.027654, and Root Mean Squared Error of 0.034402 on the validation set. The model achieves stable convergence within 10 epochs and maintains consistent performance across diverse driving scenarios, from straight highway segments to complex turning maneuvers. This work contributes to autonomous driving technology by demonstrating the effectiveness of GPS-supervised learning for curvature prediction, successfully deployed in OpenPilot’s production system with real-time inference at 5 Hz.
Items per page:
50
1 – 50 of 1486