Browse Topic: Neural networks

Items (421)
test
The paper discusses the design and high-fidelity flight dynamics modeling of a 13-lb lift-plus-cruise unmanned aerial vehicle (UAV) using Rotorcraft Comprehensive Analysis System (RCAS) in order to (1) better understand its physics of flight during a wide range of maneuvers, and (2) provide insight into the fidelity needed to achieve quantitative accuracy when compared to flight test data. Wind tunnel tests of the full aircraft were performed at a 65% scale to provide lookup tables for the flight dynamics model. Flight test data was collected while providing high control inputs to excite a variety of dynamic states in hovering and cruising modes to systematically validate the physics model. Near quantitative agreement was observed between the model predictions and test data during hover; however, the predictions began to disagree at higher forward cruising speeds. To address the discrepancy between the prediction and experiment, the flight dynamics model was improved by learning a correction from flight test data using a neural network. This hybrid physics plus data-driven approach reduced the error between the physics model and experiment by 74% and only needing 12 minutes of flight data for training. This hybrid methodology presents an alternate approach to high fidelity modeling which only needs a relatively small amount of flight test data.
Stewart, Reuben-WayneBrown, CaydenBenedict, Moble
This study aims to develop Control Equivalent Gust Input (CEGI) and Rotor Control Equivalent Gust Input (RCEGI) profiles that accurately reproduce vehicle response to deterministic gusts. This involves creating an inverse model using adaptive neural networks in order to map vehicle response to pilot and rotor control inputs. The accuracy of the CEGI and RCEGI models are then quantified using the Time Domain Integrated Cost Function (Ref. 1) to determine trends within the CEGI and RCEGI models for gusts of varying shape, magnitude, and duration as well as at varying flight conditions. Analysis using the cost function shows that the CEGI and RCEGI models follow similar trends. Both models are more accurate for gusts of short duration and small amplitude, and both models are more accurate for sinusoidal gusts than top hat gusts.
Sinha, TanayaPrasad, J.V.R.
This paper presents updates to The Rotorcraft Optimization Tools (RCOTools) package to streamline iterative rotorcraft comprehensive design. The work is presented in three parts. Part I. a brief introduction to our simplified API is shown, in addition to a new mission profile dashboard. Part II. demonstrates high-throughput using the embarrassingly parallel paradigm to produce large-scale datasets structured by simple design of experiments (DOE) as shown by our discussion on urban air mobility (UAM) emission minimization. Such datasets provide a necessary component for rapid database generation and supervised machine learning. Part III. the API is used to couple rotor performance and sizing optimization. A simple technique for ultra-fast hover calibration is given, as well as possible applications for neural network modeling in comprehensive design. These enhancements accelerate design workflows and enable data-driven approaches for next-generation urban air mobility and planetary rotorcraft concepts.
Pereyra, CarlosKung, Esther
Establish a comprehensive taxonomy of Artificial Intelligence in aviation
G-34, Artificial Intelligence in Aviation
This paper discusses the development of a quantitatively-accurate non-linear hybrid flight dynamics model of a hover-capable Air-Launched Tailsitter Unmanned Aerial System (ALUAS) in order to 1) understand its dynamics during complicated maneuvers, and 2) provide a high-fidelity framework to develop novel control laws. Wind tunnel tests were conducted on a 1:1 scale model of the full aircraft to measure the airloads, which were used in the simulation as a lookup table. Flight tests of the ALUAS were performed in hover, transition, and cruise to collect a large amount of unique state measurements by providing large excitations to induce highly transient motion. The flight dynamics predictions using Rotorcraft Comprehensive Analysis System (RCAS) software were then compared with experimental flight test data. To correct any discrepancies in the RCAS physics-based predictions, a correction was learned from the experimental measurements, making use of the large amount of collected flight test data. Using a neural network to learn this correction, the end result was a quantitatively accurate neural network assisted flight dynamics model. The accuracy of current simulations in complex flight states successfully demonstrates the applicability of the proposed methodology for correcting the dynamics model of novel out-of-the-box aircraft configurations.
Stewart, Reuben-WayneDooher, JackBenedict, Moble
Blade–wake interaction (BWI) is a significant source of broadband noise and is often dominant in rotors with high blade counts. Accurately capturing the resulting unsteady blade loading is computationally expensive and, therefore, drives the cost of BWI noise calculation. To address this challenge, a low-fidelity BWI noise prediction tool was developed using aerodynamic data from the blade element momentum theory (BEMT) and the lattice Boltzmann method (LBM) for a series of rotor configurations with medium to high solidity. Starting from a six-bladed baseline rotor, 13 additional configurations were generated by varying blade twist, taper, root collective, solidity, and blade count. The relationship between vortex miss distance and blade loading unsteadiness was quantified to construct a semi-empirical BWI noise model. The model predicted BWI noise with a root mean square error of 3.9 dBA and a mean absolute percentage error of 1%. It was subsequently integrated into a BEMT framework to produce aerodynamic and acoustic data for training a tandem neural network (TNN) that was employed to optimize two rotor geometries. The optimized designs achieved up to a 7% reduction in BWI noise and a 7% improvement in performance. Additional geometric modifications—including blade tip anhedral, forward sweep, and a mixed configuration—were also assessed using LBM, each demonstrating notable noise reduction.
Jayasundara, DilharaGomez, PhillipRandall, Ian
This paper discusses the development of a flight dynamics model (or digital twin) of a compact and re-configurable coaxial-propeller-based micro air vehicle (MAV) in hover, edgewise, and maneuvering flight using a hybrid physics-based plus data-driven approach. The MAV has a mass of 366 grams (0.81 lb), and features a 52 mm (2.05 in) diameter cylindrical fuselage, foldable propellers, and a two-axis gimbal thrust vectoring mechanism for pitch and roll control. The aircraft has been successfully launched from a pneumatic cannon and has achieved stable and controlled flight. A physics-based flight dynamics model of this novel MAV has been developed using Rotorcraft Comprehensive Analysis System (RCAS). RCAS is able to predict the translational dynamics near hover reasonably well; however, the accuracy decreases for rotational dynamics in edgewise flight resulting in significant differences between predicted dynamics and flight test data, known as residual dynamics. The current hybrid model utilizes the residual dynamics via a data-driven approach to correct the physics-based model. Using the measured vehicle states and control inputs, a deep neural network (DNNs) was trained to learn the residual forces and torques. The resulting hybrid model reduced prediction errors by 55% on average compared to the RCAS model based on pure physics.
Nyancho, MiracleColeman, DavidBenedict, MobleStewart, Reuben-Wayne
The advent of electric propulsion technology has led to a paradigm shift in aircraft design over the past few decades. This shift has expanded the possibilities for design and optimization processes more than at any previous time. To support these advancements, efficient flight dynamics simulation models that can be employed in iterative optimization and design processes are essential. Among the modules of a typical flight dynamics framework—namely, control, flight dynamics, and aerodynamics—the aerodynamics module, which includes the rotor performance model, generally demands the most computational effort, thereby limiting simulation efficiency. In this study, a novel machine learning (ML)-assisted flight dynamics framework is developed, incorporating a Neural Network Blade Element Theory (NN-BET) model as the rotor performance module. The results show a 7- to 8-fold reduction in computational time compared to fast, physics-based frameworks utilizing efficient Blade Element Momentum Theory (BEMT) models, without compromising predictive accuracy. Furthermore, the modular architecture of the framework allows for easy adaptation to a wide range of practical applications by replacing modules with functionally equivalent alternatives. The demonstrated accuracy and computational efficiency of the proposed flight dynamics framework make it a highly promising candidate for optimization and design applications.
Hashem Dabaghian, PedramHalder, Atanu
This study investigates the application of neural network architectures to predict control inputs required to replicate rotorcraft responses under vertical gust disturbances. Two modeling approaches are developed: the Control Equivalent Gust Input (CEGI) model, using body-axis inputs and the Rotor Control Equivalent Gust Input (RCEGI) model using rotor-specific inputs. Initial models employed single-input single-output (SISO) LSTM networks, which demonstrated limitations in capturing transient behavior and exhibited delay in predicted control inputs. By incorporating multiple vehicle response features and increasing the number of hidden neurons, multiple-input single-output (MISO) architectures significantly improved accuracy and reduced Root Mean Square Error (RMSE). Further enhancement was achieved by implementing bidirectional LSTM (BiLSTM) layers, which reduced both delay and transient error. Comparisons with inverted linear time-invariant (LTI) approximations showed that neural networks provided superior performance, particularly in modeling nonlinear dynamics. The results highlight the potential of deep learning approaches to improve the accuracy of control input mapping and inform real-time control strategies in unsteady flight environments.
Sinha, TanayaHayajnh, MahmoudPrasad, J. V. R.
As part of maintenance improvement on helicopters, Airbus Helicopters has made available a proactive analysis service based on Health and Usage Monitoring System data generated during the flight. The present paper describes the new approach used to detect and classify any changes in time series behavior thanks to A.I. (Artificial Intelligence) especially computer vision. This new approach is more efficient and relevant than the classical approach based one statistical law [Ref 1]; in fact, it is acting, as the human eye, which is able to identify easily any abrupt change on the time series, and classifies it, whether Machine learning or Deep Neural Networks both have shown excellent results in term of classification accuracy. First part of this paper highlights how the learning data were prepared, then the second and the third parts give more details about how the time series are transformed into image presentation and how the different Artificial Intelligence models were selected and feed, ultimately the obtained results.
Boutaleb, AbdelhafidDiaz, Alexandre
The rotorcraft community faces significantly higher accident rates compared to fixed-wing commercial aircraft, underscoring the critical need for enhanced safety measures. While Helicopter Flight Data Monitoring programs hold promise in improving safety, their widespread adoption remains limited, partly due to challenges associated with the acquisition and analysis of flight data. This paper proposes a Deep Learning (DL) solution to address safety concerns within the rotorcraft community by efficiently acquiring and analyzing flight data for a more automated and comprehensive safety assessment. Specifically, we leverage data obtained with cost-effective off-the-shelf cameras, and process it through Convolutional Neural Networks for automated detection and classification of gauges from several helicopters' cockpits. Our DL pipeline integrates a classifier for helicopter identification, an object detector for cockpit gauges detection and classification, and a network to infer the reading of each detected gauge. The contribution of this work is two-fold: (1) enhance rotorcraft safety by developing a DL framework capable of detecting, classifying, and inferring gauge readings for different helicopter types, and (2) boost research in the field by constructing a curated dataset valuable for aviation and machine learning communities.
Khelifi, AmineJohnson, Charles C.Thompson, LaceyBouaynaya, Nidhal C.Carannante, GiuseppinaTrabelsi, Mohamed Ali
This study investigates the use of machine learning (ML) models to estimate the gross weight (GW), the longitudinal position of the center of gravity (CGx), and 1/rev cyclic flapping angles (Δ1c and Δ1s) of a compound helicopter with three redundant controls - main rotor RPM, collective propeller thrust, and stabilator angle. Neural Network (NN), Gaussian Process for Regression (GPR), and Support Vector Machine (SVM) algorithms are employed to develop estimation models using supervised training. The airspeed, redundant controls, main rotor controls, aircraft attitudes, and main rotor torque are selected as input variables (predictors) to the models due to their accessibility through the aircraft Health and Usage Monitoring System (HUMS). The dataset is split into low-speed and high-speed regimes to compare the prediction accuracy and training cost of separate regime models against a combined full-regime model. Separate airspeed regime GPR models showed superior performance in GW estimation, with higher accuracy and cost-effectiveness compared to a single full-regime model. For CG estimation, GPR again outperformed NN and SVM, although the maximum outlier errors increase significantly if a 95% confidence interval is considered. Finally, for 1/rev cyclic flapping angle predictions, SVM estimations, though not superior to GPR or NN, were acceptable and had a significantly lower computational cost. The study also examined the importance of predictors, highlighting that, on average, certain predictors like rotor RPM and rotor torque are less influential, but their removal degraded performance and had no cost benefit.
Halder, AnubhavMakkar, GauravGandhi, Farhan
This paper investigates the feasibility of using machine learning to predict whirl flutter bifurcation diagrams. The machine learning techniques selected for the study are XGBoost and the long short-term memory neural network. These techniques are selected for their suitability for sequential and nonlinear data. The techniques are investigated for a propeller-nacelle test case with polynomial structural nonlinearities resulting in supercritical or subcritical whirl limit-cycle oscillations. The techniques are trained to learn the bifurcation diagram for the amplitude variation of pitch angle limit-cycle oscillations of the propeller-nacelle system as a function of the forward speed for various levels of cubic structural nonlinearity. Bifurcation diagram learning and testing data are generated using the bifurcation forecasting method. XGBoost is computationally faster to train but less accurate for low amounts of learning data, especially for the most weakly and strongly nonlinear cases. The long short-term memory neural network is more computationally expensive to train but shows a less scattered error pattern for sparse learning data. However, it is sensitive to the amplitude resolution of the bifurcation diagrams. The approach to sample the cubic nonlinearity range does not significantly impact the results once the techniques have a sufficient amount of data to learn from. The data requirements observed in the study suggest that, for these techniques, direct learning of bifurcation diagrams may not scale beyond a handful of input parameters.
Gatlin, MaiaRiso, Cristina
In the field of aerodynamics, there is a growing need for rapid load prediction in engineering applications. Surrogate modeling offers a promising solution, providing faster results compared to high-fidelity computational models. This study focuses on a Machine Learning (ML) framework tailored for surrogate modeling, specifically for integrated aerodynamic load predictions in aircraft design. Central to this framework is a Deep Neural Network (DNN) component capable of handling both steady-state and fluctuating aerodynamics. A key challenge for surrogate models lies in maintaining prediction accuracy, especially in scenarios involving nonlinear flow phenomena like flow separation and transonic shifts. To address these challenges, we introduce a two-step physics-state predictor that integrates an intermediate Convolutional Neural Network (CNN) component. This approach enhances the surrogate model's capability to accurately represent dynamic separated flows and other nonlinear patterns without relying on unrealistic user inputs. Results are presented for NACA0015 dynamic stall predictions for two different physics-state inputs.
Abras, JenniferHariharan, Nathan
Tailsitter configurations that operate in both fixed and rotary wing flight modes are typically capable of generating large control forces and moments, making them inherently capable of rapid transitions and aggressive maneuvers. However, harnessing these capabilities requires feedback control strategies that can effectively estimate the non-linear aerodynamics loads involved to successfully exploit them. This paper describes initial steps in combining an onboard flow sensing strategy with a data-driven approach to estimating inflight air loads. A neural network is trained to use measurements from a multi-hole probe to predict the output from a set of pressure sensors embedded in a wing section undergoing a series of pitch motions in a wind tunnel. We hypothesize that this limited context of emulating a sensor network represents a focused and compartmentalized approach to applying emerging data-driven techniques to challenging aeronautical problems. We compare estimation results from a set of neural networks with varying input configurations to assess the feasibility of our approach and the significance of different sensing modalities on overall performance. Current results show that a gated recurrent network (GRU) trained with unsteady pressure measurements was able to predict the chordwise pressure distribution on a pitching NACA 2412 airfoil using probe measurements, reproducing the transient and non-linear effects observed in our dataset.
Yeo, DerrickFloros, MatthewReddinger, Jean-PaulGerdes, JohnShrestha, Elena
In this work, a unified framework integrating global and local SHM methods for structural health monitoring (SHM) of rotorcraft structures is proposed. This framework integrates both "local" ultrasonic-guided wave-based and "global" vibration-based SHM schemes for tackling damage detection, identification, and quantification under uncertainty. The local SHM is completed by training a variation of variational auto-encoder (MMD-VAE) along with feed-forward neural networks (FFNN). The compressed latent space vector obtained during the training process is applied to achieve both signal reconstruction and state prediction. In terms of the global model, functionally pooled auto-regressive models with exogenous excitation (VFP-ARX) models are applied including to capture low-frequency vibrations. The complete experimental evaluation and assessment of the proposed framework are presented for an Airbus H125 helicopter blade under both low-frequency vibrations and ultrasonic guided waves for SHM.
Fan, YimingKopsaftopoulos, FotisForrester, DavidZhou, Peiyuan
Rotor blade optimization presents a multifaceted challenge as traditional design methodologies rely on computationally exhaustive high-fidelity computational fluid dynamics (CFD). Conversely, low-fidelity techniques such as potential flow based codes are inaccurate, especially in the regions of flow separation. This paper proposes leveraging artificial neural networks (ANNs) to predict the performance polar of a given airfoil geometry, and to facilitate the inverse design of airfoil, a modified form of ANNs (known as Tandem Neural Networks (T-NNs)) is implemented. The airfoil inverse design is a multi-point optimization problem (at multiple angles of attack) and therefore, the T-NNs are trained on the vectors of performance polar instead of individual angles of attack. The paper also delves into a comprehensive analysis of data wrangling, airfoil parametrization and design of experiments to cover a wide range of rotorcraft airfoils. A novel way of including practical design constraints for airfoil geometry is also included. Finally, this work demonstrates the application of the proposed methodology for airfoil inverse design, statistical analysis for generating a family of airfoils and optimization of HART-II rotor using T-NNs and Genetic Algorithm (GA).
Anand, ApurvaBaeder, James DMarepally, Koushik
Abstract In recent years, demands of flat wipers have rapidly increased in the vehicle industry due to their simpler structure compared to the conventional wipers. Procedures for evaluating the appropriate metallic flexor geometry, which is one of the major components of the flat wiper, were proposed in the authors’ previous study. However, the computational cost of the aforementioned procedures seems to be unaffordable to the industry. The discrete Winkler model regarding the flexor as the Euler–Bernoulli beam is established as the mathematical model in this study to simulate a flexor compressed against a surface at various wiping angles. The deflection of the beam is solved using a finite difference method, and the calculated contact pressure distributions agree fairly with those based on the corresponding finite element model. Flexor designs are paired with various windshield surfaces to accumulate a sufficiently large simulation database based on the mathematical model. An artificial neural network (ANN) approach is developed to predict contact pressure distributions of the flexor much faster than the mathematical model. Geometry of the curved surface is represented by a shape code obtained via a principal component analysis (PCA) and used in the ANN model. The ANN algorithm is also applied to efficiently evaluate the wiping patterns according to the simulated contact pressure distributions. These patterns are then classified by using a convolutional neural network (CNN) to identify several suitable flexor designs for the specific windshield. The flat wiper suggested by the current procedures is experimentally validated to justify its qualified wiping performances.
Chu, Yi-TzuHuang, Ting-ChuanLiao, Kuo-Chi
ABSTRACT
Singh, RajneshSridharan, AnanthShalu, HrithwikGovindarajan, Bharath
In this work, we present a lightweight pipeline for robust behavioral cloning of a human driver using end-to-end imitation learning. The proposed pipeline was employed to train and deploy three distinct driving behavior models onto a simulated vehicle. The training phase comprised of data collection, balancing, augmentation, preprocessing, and training a neural network, following which the trained model was deployed onto the ego vehicle to predict steering commands based on the feed from an onboard camera. A novel coupled control law was formulated to generate longitudinal control commands on the go based on the predicted steering angle and other parameters such as the actual speed of the ego vehicle and the prescribed constraints for speed and steering. We analyzed the computational efficiency of the pipeline and evaluated the robustness of the trained models through exhaustive experimentation during the deployment phase. We also compared our approach against state-of-the-art implementation in order to comment on its validity.
Samak, Tanmay VilasSamak, Chinmay VilasKandhasamy, Sivanathan
ABSTRACT
Thurman, ChristopherZawodny,  Nikolas
A Semantic Slam System Based on Visual-Inertial Information and around View Images for Underground Parking Lot2021-01-00784/6/2021
As one of the most challenging driving tasks, parking is a common but particularly troublesome problem in large cities. Recently, an excellent solution-automated valet parking (AVP) has become a hot research topic, which allows the driver to leave the vehicle in a drop-off area, while the vehicle driving into the parking slot by itself. For AVP, the precise localization is an indispensable module. However, the global positioning system (GPS) cannot be used in the underground parking lot and the localization method based on lidar is too expensive. In response to solve this problem, we propose a simultaneous localization and mapping system with the semantic information of parking slots (PS-SLAM), which is based on visual-inertial and around view images. First, the calibration of multi-sensors is conducted to obtain their intrinsic and extrinsic parameters. In this way, the around view image and transformation matrices between sensors can be acquired. Then, the ORB-SLAM3 based on visual-inertial information is used to acquire the pose of the vehicle and sparse point cloud map. Next, the parking slot in the around view image is detected by the deep convolutional neural network (DCNN) model called VPS-Net. Finally, a parking-slot association method is devised to associate the detected parking slots with the point cloud map to generate a semantic map. The field experiments are conducted using a wire control chassis with 4 fisheye cameras, an inertial measurement unit (IMU), and a monocular camera. The results show that the proposed visual semantic SLAM system not only can achieve centimeter-level localization in the indoor parking lot but also generate a semantic map with parking slots.
LI, WeiLi, ChaohuiXiao, DongjieZhou, DongWang, TaoCao, Libo
No Cost Autonomous Vehicle Advancements in CARLA through ROS2021-01-01064/6/2021
Development of autonomous vehicle technology is expensive and perhaps more complicated than initially thought, as evidenced by the recent rollback of anticipated delivery dates from companies such as Tesla, Waymo, GM, and more. One of the most effective techniques to reduce research and development costs and speed up implementation is rigorous analysis through simulation. In this paper, we present multiple autonomous vehicle perception and control strategies that are rigorously investigated in the user friendly, free, and open-source simulation environment, CARLA. Overall, we successfully formulated potential solutions to the autonomous navigation problem and assessed their advantages and disadvantages in simulation at no cost. First, a lane finding method utilizing polynomial fitting and machine learning is proposed. Then, the waypoint navigation strategy is described, along with route planning. Object detection is then implemented using pre-trained convolutional neural networks. A classic PID control strategy and the Stanley Method were investigated for lateral and longitudinal control of the vehicle. Finally, each of these components are simultaneously applied in the simulation environment using the robot operating system (ROS). As a result, we have achieved successful self-driving simulation. The key takeaway is that the perception and control strategies proposed can be easily transitioned towards physical implementation, through the use of ROS. The overall conclusion is that the CARLA simulation environment is a reliable workbench to test innovative solutions that could become technology enablers for the autonomous vehicle industry.
Prescinotti Vivan, GabrielGoberville, NickAsher, ZacharyBrown, NicolasRojas, Johan
DA-IVE: MLP Based Data Association Method for Instantaneous Velocity Estimation Using Multi-Radar: An Experimental Validation Study2021-01-00924/6/2021
This paper describes a novel Multi-Layer Perceptrons (MLP) learning-based association algorithm that is used in conjunction with an Instantaneous Velocity Estimator (IVE) to estimate the velocity of a surrounding vehicle using multi-radar sensors. The IVE algorithm requires at least two targets to be able to provide a velocity estimate. The approach suggested in this paper performs three stages of filtering on a list of targets available for the association to a given track. The algorithm identifies the one pair of targets that will provide the best instantaneous velocity estimation from all possible pairs. The three stages of filtering described ahead are, I - Semantic gating, II - MLP scoring, and III - Algebraic scoring. The IVE algorithm performs linear regression on the pair of targets it is finally provided to come up with a velocity estimation. This research also describes a novel method of labeling radar targets for use in the training of the neural network in association stage II. A thorough analysis of the correlation between a radar target’s quality and attributes is performed and presented here. The performance of the proposed algorithm is evaluated using real-world data collected through the ZF Automated Driving prototype vehicle.
Shakibajahromi, BaharehKrishnan, Anirudh SarathyAti, DilipJabalameli, AmirhosseinKanzler, StevenShayestehmanesh, Saeed
A robust framework for fault detection and identification of rotor degradation in multicopters while effectively rejecting the effects of gusts is introduced. The rotor fault detection and identification methods employed in this study are based on excitation-response signals of the aircraft under ambient turbulence to distinguish between an aircraft response to gusts and rotor faults. A concise overview of the development of statistical time series model for healthy aircraft using the aircraft attitudes as the output and controller commands as the input is presented. This model is utilized to extract quality features for training a simple neural network to perform effective online rotor fault detection and identification in a hexacopter exceptional speed of making a decision and accuracy of fault classification. It is shown that using a statistical time series model assisted neural network employed for online monitoring is capable of rejecting gusts, sensitive to even 20% rotor degradation and achieves fault detection and identification in less than 2 s after the fault with an accuracy over 99%.
Dutta, AirinGandhi, FarhanKopsaftopoulos, FotisMcKay, Michael
This paper presents a data-driven approach towards time-optimal trajectory generation for Unmanned Aerial Vehicles (UAV's) using a machine-learned trajectory generation mechanism for point-to-point time-optimal trajectories on-the-fly. To train this machine-learned black box trajectory generator off-line, a model-based optimization problem is first constructed for point-to-point time-optimal trajectory generation, with physical constraints on inputs, states, and rates. The formulated optimization problem is then solved off-line for a range of initial and terminal flight states to generate point-to-point data-sets that consist of the optimal state and input trajectories. This information is compressed by parameterizing the input and state trajectories using a set of basis functions. This data is then used to train the neural network-based trajectory planner. The output of the neural network is the basis function coefficient sets for the state and input trajectories (and the total flight time) which can then be used to reconstruct the flight trajectory. Once the neural network is trained, the data-driven on-board trajectory generator is ready to be deployed on the UAV for on-board planning. This approach is demonstrated for two scenarios: (1) the input to the neural network being the initial and terminal flight states and (2) the input to the neural network being initial and terminal flight states as well as physical constraints. To validate the performance of the machine-learned black-box trajectory generator, the root mean squared error between the neural network generated trajectories and the trajectories obtained from solving the optimization problem directly is statistically evaluated. These trajectories are also tested for violation of path constraints (which are not included explicitly in the training or input to the black box planner) by evaluating the mean constraint violation for each path-constrained variable.
Lai, RunyuZhao, DiMishra, Sandipan
A deep-learning powered single-strained electronic skin sensor can capture human motion from a distance. The single strain sensor placed on the wrist decodes complex five-finger motions in real time with a virtual 3D hand that mirrors the original motions. The deep neural network boosted by rapid situation learning (RSL) ensures stable operation regardless of its position on the surface of the skin.
Analysis of LiDAR and Camera Data in Real-World Weather Conditions for Autonomous Vehicle Operations2020-01-00934/14/2020
Autonomous vehicle technology has the potential to improve the safety, efficiency, and cost of our current transportation system by removing human error. With sensors available today, it is possible for the development of these vehicles, however, there are still issues with autonomous vehicle operations in adverse weather conditions (e.g. snow-covered roads, heavy rain, fog, etc.) due to the degradation of sensor data quality and insufficiently robust software algorithms. Since autonomous vehicles rely entirely on sensor data to perceive their surrounding environment, this becomes a significant issue in the performance of the autonomous system. The purpose of this study is to collect sensor data under various weather conditions to understand the effects of weather on sensor data. The sensors used in this study were one camera and one LiDAR. These sensors were connected to an NVIDIA Drive Px2 which operated in a 2019 Kia Niro. Two custom scenarios (static and dynamic objects) were chosen to collect sensor data operating in four real-world weather conditions: fair, cloudy, rainy, and light snow. An algorithm developed herein was used to provide a method of quantifying the data for comparison against the other weather conditions. The results from these performance algorithms show that sensor data quality degrades by an average of 13.88% for static objects and 16.16% for dynamic objects while operating in these conditions, with operations in rain proving to have the most significant effect on sensor data degradation. From this study, it is hypothesized that advancements in data processing algorithms can improve the usability of this degraded data. In future work, we seek to explore fault-tolerant sensor fusion algorithms that can overcome the effects of adverse weather.
Goberville, NickEl-Yabroudi, MohammadOmwanas, MarkRojas, JohanMeyer, RickAsher, ZacharyAbdel-Qader, Ikhlas
A Visible and Infrared Fusion Based Visual Odometry for Autonomous Vehicles2020-01-00994/14/2020
An accurate and timely positioning of the vehicle is required at all times for autonomous driving. The global navigation satellite system (GNSS), even when integrated with costly inertial measurement units (IMUs), would often fail to provide high-accuracy positioning due to GNSS-challenged environments such as urban canyons. As a result, visual odometry is proposed as an effective complimentary approach. Although it’s widely recognized that visual odometry should be developed based on both visible and infrared images to address issues such as frequent changes in ambient lightening conditions, the mechanism of visible-infrared fusion is often poorly designed. This study proposes a Generative Adversarial Network (GAN) based model comprises a generator, which aims to produce a fused image combining infrared intensities and visible gradients, and a discriminator whose target is to force the fused image to retain as many details that exist mostly in visible images as possible. Based on the fused image, the Features from Accelerated Segment Test (FAST) algorithm is adopted to extract feature points which are then traced with the Lucas-Kanade (LK) algorithm in subsequent images. Furthermore, to remove mismatched feature points, the Random Sample Consensus (RANSAC) method is employed to detect the outliers iteratively and to compute the essential matrix. Experiments are conducted utilizing the KAIST benchmark dataset. A significant improve in positioning accuracy is observed from experimental results, as compared to visual odometry built upon visible and infrared images only. The proposed visual odometry can provide high-accuracy positioning when the GNSS is challenged.
Zhou, YunfeiYin, Zhishuai
Autonomous Vehicles Camera Blinding Attack Detection Using Sequence Modelling and Predictive Analytics2020-01-07194/14/2020
Autonomous vehicles are waiting to address the global automotive mobility challenges through an intelligent smart transportation system, which includes advanced sensor-actuator configurations to control, navigate, and drive the vehicles. Multi-sensor data fusion from the key sensors such as camera, radar, and lidar is used to achieve the environmental perception for autonomous vehicles by capturing the various attributes of the environment. Cameras are the dominant sensors to achieve the perception by providing vision capability to vehicles. The direct interface of the cameras with the dynamic driving environment carries numerous attack surfaces on the camera. Blinding attacks on the cameras are one of the critical attacks with an intention to blind the cameras either fully or partially by projecting light into the cameras to hide the objects which results in failure in object detection. Here, the blinding attack detection approach is proposed which detects the blinding attacks on the camera in a dynamic driving environment by camera data predictive analytics. The proposed system predicts the future next frame of the video at each time instance and compares the received frame from the camera with the predicted frame at that instance to detect the blinding attacks. The incoming frames from the camera are sequentially modeled using a convolutional encoder-decoder neural network to predict the consecutive future frames, and the predicted frames are compared with the received camera frames to identify the similarity measure between the predicted and incoming camera frames of the same instance. Further, the approach detects the blinding on the camera, if the similarity measure calculated falls below a fixed threshold. The similarity measure which is inversely proportional to the amount of blinding is used to identify the blinding attacks. The predictive analytics of the sequentially modeled video frames with similarity measurement is used for the successful detection of blinding attacks.
D H, Sharath YadavAnsari, Asadullah
Joint Calibration of Dual LiDARs and Camera Using a Circular Chessboard2020-01-00984/14/2020
Environmental perception is a crucial subsystem in autonomous vehicles. In order to build safe and efficient traffic transportation, several researches have been proposed to build accurate, robust and real-time perception systems. Camera and LiDAR are widely equipped on autonomous self-driving cars and developed with many algorithms in recent years. The fusion system of camera and LiDAR provides state-of the-art methods for environmental perception due to the defects of single vehicular sensor. Extrinsic parameter calibration is able to align the coordinate systems of sensors and has been drawing enormous attention. However, differ from spatial alignment of two sensors’ data, joint calibration of multi-sensors (more than two sensors) should balance the degree of alignment between each two sensors. In this paper, we assemble a test platform which is made up of dual LiDARs and one monocular camera and use the same sensing hardware architecture as intelligent sweeper designed by our laboratory. Meanwhile, we propose the related joint calibration method using a circular chessboard. The center of circular chessboard is respectively detected in camera image to get pixel coordinates and in point cloud of LiDAR to get 3D coordinates. The calibration problem is then converted into a 3D-2D PnP matching problem and the center of the chessboard is set as corresponding points to construct the geometric constraints to get initial calibration values. Further, a proper global loss function is elaborately designed for Levenberg-Marquardt nonlinear optimization to obtain the final calibration parameters, and then the extrinsic parameters between any two sensors are estimated simultaneously. Experimental results show that the proposed method is suitable for the joint calibration of fusion system composed of LiDARs and camera, and the calibration results have high accuracy and stability.
Deng, ZhenwenXiong, LuYin, DongShan, Fengwu
Deep Optimization of Catalyst Layer Composition via Data-Driven Machine Learning Approach2020-01-08594/14/2020
Proton exchange membrane fuel cell (PEMFC) provides a promising future low carbon automotive powertrain solution. The catalyst layer (CL) is its core component which directly influences the output performance. PEMFC performance can be greatly improved by the effective optimization of CL composition. This work demonstrates a deep optimization of CL composition for improving the PEMFC performance, including the platinum (Pt) loading, Pt percentage of carbon-supported Pt and ionomer to carbon ratio of the anode and the cathode,. The simulation results by a PEMFC three-dimensional (3D) computation fluid dynamics (CFD) model coupled with the CL agglomerate model is used to train the artificial neural network (ANN) which can efficiently predict the current density under different CL composition. Squared correlation coefficient (R-square) and mean percentage error in the training set and validation set are 0.9867, 0.2635% and 0.9543, 1.1275%, respectively. It illustrates that the well-trained ANN has a comparable accuracy with the physical model. Then, the ANN is utilized as the fitness function in the genetic algorithm (GA) to search the optimal CL composition for maximizing the current density. For verification, the optimal solution of CL composition is returned to the physical model and the comparison between the ANN predicted current density and the physical model simulated current density is provided. The percentage error is only 2.418% which can illustrate the validity of this work.
Wang, BowenXie, BiaoXuan, JinGu, WenZhao, DezongJiao, Kui
Reference Test System for Machine Vision Used for ADAS Functions2020-01-00964/14/2020
Advanced Driver Assistance Systems (ADAS) like Lane Departure Warning (LDW) and Lane Keep Assist (LKA) have been available for several years now but has experienced low customer acceptance and market penetration. These deficiencies can be traced to the inability of many of the perception systems to consistently recognize lane markings and localize the vehicle with respect to the lane markings in the real-world with poor markings, changing weather conditions and occlusions. Currently, there is no available standard or benchmark to evaluate the quality of either the lane markings or the perception algorithms. This work seeks to establish a reference test system that could be used by transportation agencies to evaluate the quality of their markings to support ADAS functions that rely on pavement markings. The test system can also be used by designers as a benchmark for their proprietary systems. To support this development, an extensive video dataset was collected at different times of day and weather conditions on various roads in Central Texas. The videos were evaluated on different state-of-the art lane detection algorithms and their performance was ranked based on a set of metrics specifically developed for evaluating the effectiveness of the lane estimation system. The test scenarios are comprised of a set of roadways and environmental features, as well as the pavement marking presence and luminance variables. A systems approach is presented by correlating the algorithm performance data to the environmental factors, lane marking types, color, material, and the retroreflectivity of pavement markings.
Nayak, AbhishekRathinam, SivakumarPike, AdamGopalswamy, Swaminathan
Robust Sensor Fused Object Detection Using Convolutional Neural Networks for Autonomous Vehicles2020-01-01004/14/2020
Environmental perception is considered an essential module for autonomous driving and Advanced Driver Assistance System (ADAS). Recently, deep Convolutional Neural Networks (CNNs) have become the State-of-the-Art with many different architectures in various object detection problems. However, performances of existing CNNs have been dropping when detecting small objects at a large distance. To deploy any environmental perception system in real world applications, it is important that the system achieves high accuracy regardless of the size of the object, distance, and weather conditions. In this paper, a robust sensor fused object detection system is proposed by utilizing the advantages of both vision and automotive radar sensors. The proposed system consists of three major components: 1) the Coordinate Conversion module, 2) Multi level-Sensor Fusion Detection (MSFD) system, and 3) Temporal Correlation filtering module. The proposed MSFD system employs the principles of artificial intelligence beyond simple comparison of data variance of the sensors. And then, its performance is further improved by using the temporal correlation information with an adaptive threshold scheme. The proposed system is evaluated with the collected video data (6,854 image frames with 18,918 labeled objects). Based on the laboratory testing and in-vehicle validation, the proposed system demonstrates its high accuracy for detecting any size of objects in real-world data.
Park, JungmeJayachandran Raguraman, SriramAslam, AakifGotadki, Shruti
Items per page:
1 – 50 of 421