SEARCH · Engineering Papers

Results for “Neural Networks”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 records

All sources NASA NTRS DOE OSTI

NASA NTRS2003

Modeling Aircraft Wing Loads from Flight Data Using Neural Networks

Neural networks were used to model wing bending-moment loads, torsion loads, and control surface hinge-moments of the Active Aeroelastic Wing (AAW) aircraft. Accurate loads models are required for the development of control laws designed to increase roll performance through wing twist while not exceeding load limits. Inputs to the model include aircraft rates, accelerations, and control surface positions. Neural networks were chosen to model aircraft loads because they can account for uncharacterized nonlinear effects while retaining the capability to generalize. The accuracy of the neural network models was improved by first developing linear loads models to use as starting points for network training. Neural networks were then trained with flight data for rolls, loaded reversals, wind-up-turns, and individual control surface doublets for load excitation. Generalization was improved by using gain weighting and early stopping. Results are presented for neural network loads models of four wing loads and four control surface hinge moments at Mach 0.90 and an altitude of 15,000 ft. An average model prediction error reduction of 18.6 percent was calculated for the neural network models when compared to the linear models. This paper documents the input data conditioning, input parameter selection, structure, training, and validation of the neural network models.

Allen, Michael J.↗

NASA NTRS1990

Learning and diagnosing faults using neural networks

Neural networks have been employed for learning fault behavior from rocket engine simulator parameters and for diagnosing faults on the basis of the learned behavior. Two problems in applying neural networks to learning and diagnosing faults are (1) the complexity of the sensor data to fault mapping to be modeled by the neural network, which implies difficult and lengthy training procedures; and (2) the lack of sufficient training data to adequately represent the very large number of different types of faults which might occur. Methods are derived and tested in an architecture which addresses these two problems. First, the sensor data to fault mapping is decomposed into three simpler mappings which perform sensor data compression, hypothesis generation, and sensor fusion. Efficient training is performed for each mapping separately. Secondly, the neural network which performs sensor fusion is structured to detect new unknown faults for which training examples were not presented during training. These methods were tested on a task of fault diagnosis by employing rocket engine simulator data. Results indicate that the decomposed neural network architecture can be trained efficiently, can identify faults for which it has been trained, and can detect the occurrence of faults for which it has not been trained.

Whitehead, Bruce A.↗

NASA NTRS2002

Modeling of UH-60A Hub Accelerations with Neural Networks

Neural network relationships between the full-scale, flight test hub accelerations and the corresponding three N/rev pilot floor vibration components (vertical, lateral, and longitudinal) are studied. The present quantitative effort on the UH-60A Black Hawk hub accelerations considers the lateral and longitudinal vibrations. An earlier study had considered the vertical vibration. The NASA/Army UH-60A Airloads Program flight test database is used. A physics based "maneuver-effect-factor (MEF)", derived using the roll-angle and the pitch-rate, is used. Fundamentally, the lateral vibration data show high vibration levels (up to 0.3 g's) at low airspeeds (for example, during landing flares) and at high airspeeds (for example, during turns). The results show that the advance ratio and the gross weight together can predict the vertical and the longitudinal vibration. However, the advance ratio and the gross weight together cannot predict the lateral vibration. The hub accelerations and the advance ratio can be used to satisfactorily predict the vertical, lateral, and longitudinal vibration. The present study shows that neural network based representations of all three UH-60A pilot floor vibration components (vertical, lateral, and longitudinal) can be obtained using the hub accelerations along with the gross weight and the advance ratio. The hub accelerations are clearly a factor in determining the pilot vibration. The present conclusions potentially allow for the identification of neural network relationships between the experimental hub accelerations obtained from wind tunnel testing and the experimental pilot vibration data obtained from flight testing. A successful establishment of the above neural network based link between the wind tunnel hub accelerations and the flight test vibration data can increase the value of wind tunnel testing.

UH-60A AIRCRAFT↗

NASA NTRS1991

Terminal Attractors In Neural Networks

Neural networks made to "learn" faster. Report presents theoretical study of terminal attractors in neural networks. Includes systematic analysis of applications to activation dynamics of neural networks.

Zak, Michail A.↗

NASA NTRS1993

Program Helps Simulate Neural Networks

Neural Network Environment on Transputer System (NNETS) computer program provides users high degree of flexibility in creating and manipulating wide variety of neural-network topologies at processing speeds not found in conventional computing environments. Supports back-propagation and back-propagation-related algorithms. Back-propagation algorithm used is implementation of Rumelhart's generalized delta rule. NNETS developed on INMOS Transputer(R). Predefines back-propagation network, Jordan network, and reinforcement network to assist users in learning and defining own networks. Also enables users to configure other neural-network paradigms from NNETS basic architecture. Small portion of software written in OCCAM(R) language.

Villarreal, James↗

NASA NTRS1991

Radar signal categorization using a neural network

Neural networks were used to analyze a complex simulated radar environment which contains noisy radar pulses generated by many different emitters. The neural network used is an energy minimizing network (the BSB model) which forms energy minima - attractors in the network dynamical system - based on learned input data. The system first determines how many emitters are present (the deinterleaving problem). Pulses from individual simulated emitters give rise to separate stable attractors in the network. Once individual emitters are characterized, it is possible to make tentative identifications of them based on their observed parameters. As a test of this idea, a neural network was used to form a small data base that potentially could make emitter identifications.

Anderson, James A.↗

DOE OSTI2026

Input specific neural networks

Neural networks have emerged as powerful tools for mapping between inputs and outputs. However, their black-box nature limits the ability to encode or impose specific structural relationships between inputs and outputs. Many scientific and engineering problems, such as constitutive modeling in solid mechanics, require networks that can enforce convexity, monotonicity, or other structural constraints to ensure physical consistency. Here, we introduce the Input Specific Neural Network (ISNN), a new architecture that enables multiple, distinct constraints to be imposed on different input subsets for scalar-valued outputs. This framework unifies convex, monotone–convex, monotone, and arbitrary mappings within a single network for the first time. Two ISNN architectures with analytical first- and second-order derivatives are developed. We demonstrate the performance on synthetic toy problems, inverse problems in isotropic hyperelasticity, and finite element simulations. ISNNs achieve improved extrapolation behavior, require fewer invariant inputs than standard input convex networks for polyconvex potentials, and enable significant computational savings via manual differentiation. We also show how ISNNs can be used to learn structural relationships between inputs and outputs via a binary gating mechanism. Particularly, ISNNs are employed to model a homogenized anisotropic free energy potential in a decoupled multiscale setting. The network learns whether or not the potential should be modeled as polyconvex and retains only the relevant layers while using the minimum number of inputs. ISNNs provide a flexible foundation for embedding structural priors into neural networks, enhancing both interpretability and stability. They are broadly applicable across computational mechanics and other scientific domains requiring constrained functional relationships.

Jadoon, Asghar A. [Univ. of Texas, Austin, TX (Uni↗

NASA NTRS1999

Classification of ion mobility spectra by functional groups using neural networks

Neural networks were trained using whole ion mobility spectra from a standardized database of 3137 spectra for 204 chemicals at various concentrations. Performance of the network was measured by the success of classification into ten chemical classes. Eleven stages for evaluation of spectra and of spectral pre-processing were employed and minimums established for response thresholds and spectral purity. After optimization of the database, network, and pre-processing routines, the fraction of successful classifications by functional group was 0.91 throughout a range of concentrations. Network classification relied on a combination of features, including drift times, number of peaks, relative intensities, and other factors apparently including peak shape. The network was opportunistic, exploiting different features within different chemical classes. Application of neural networks in a two-tier design where chemicals were first identified by class and then individually eliminated all but one false positive out of 161 test spectra. These findings establish that ion mobility spectra, even with low resolution instrumentation, contain sufficient detail to permit the development of automated identification systems.

Non-NASA Center↗

NASA NTRS1990

Ground states of partially connected binary neural networks

Neural networks defined by outer products of vectors over (-1, 0, 1) are considered. Patterns over (-1, 0, 1) define by their outer products partially connected neural networks consisting of internally strongly connected, externally weakly connected subnetworks. Subpatterns over (-1, 1) define subnetworks, and their combinations that agree in the common bits define permissible words. It is shown that the permissible words are locally stable states of the network, provided that each of the subnetworks stores mutually orthogonal subwords, or, at most, two subwords. It is also shown that when each of the subnetworks stores two mutually orthogonal binary subwords at most, the permissible words, defined as the combinations of the subwords (one corresponding to each subnetwork), that agree in their common bits are the unique ground states of the associated energy function.

Baram, Yoram↗

NASA NTRSDate not supplied

On the Probabilistic Analysis of Neural Networks

Neural networks are powerful tools for automated decision-making,seeing increased application in safety-critical domains, such as autonomous driving. Due to their black-box nature and large scale,reasoning about their behavior is challenging. Statistical analysis is often used to infer probabilistic properties of a network, such as its robustness to noise and inaccurate inputs. While scalable, statistical methods can only provide probabilistic guarantees on the quality of their results and may underestimate the impact of low probability inputs leading to undesired behavior of the network.We investigate here the use of symbolic analysis and constraint solution space quantification to precisely quantify probabilistic properties in neural networks. We demonstrate the potential of the proposed technique in a case study involving the analysis of ACAS-Xu, a collision avoidance system for unmanned aircraft control.

neural nets↗

DOE OSTI2025

Accelerating Hamiltonian Monte Carlo for Bayesian inference in neural networks and neural operators

Hamiltonian Monte Carlo (HMC) is a powerful and accurate method to sample from the posterior distribution in Bayesian inference. However, HMC techniques are computationally demanding for Bayesian neural networks due to the high dimensionality of the network’s parameter space and the non-convexity of their posterior distributions. Therefore, various approximation techniques, such as variational inference (VI) or stochastic gradient MCMC, are often employed to infer the posterior distribution of the network parameters. Such approximations introduce inaccuracies in the inferred distributions, resulting in unreliable uncertainty estimates. In this work, we propose a hybrid approach that combines inexpensive VI and accurate HMC methods to efficiently and accurately quantify uncertainties in neural networks and neural operators. The proposed approach leverages an initial VI training on the full network. We examine the influence of individual parameters on the prediction uncertainty, which shows that a large proportion of the parameters do not contribute substantially to uncertainty in the network predictions. This information is then used to significantly reduce the dimension of the parameter space, and HMC is performed only for the subset of network parameters that strongly influence prediction uncertainties. This yields a framework for accelerating the full batch HMC for posterior inference in neural networks. We demonstrate the efficiency and accuracy of the proposed framework on deep neural networks and operator networks, showing that inference can be performed for large networks with tens to hundreds of thousands of parameters. Finally, we show that this method can effectively learn surrogates for complex physical systems by modeling the operator that maps from upstream conditions to wall-pressure data on a cone in hypersonic flow.

Bayesian inference↗

NASA NTRS1991

VLSI synthesis of digital application specific neural networks

Neural networks tend to fall into two general categories: (1) software simulations, or (2) custom hardware that must be trained. The scope of this project is the merger of these two classifications into a system whereby a software model of a network is trained to perform a specific task and the results used to synthesize a standard cell realization of the network using automated tools.

Beagles, Grant↗

NASA NTRS2006

Classification and Prediction of RF Coupling inside A-320 and A-319 Airplanes using Feed Forward Neural Networks

Neural Network Modeling is introduced in this paper to classify and predict Interference Path Loss measurements on Airbus 319 and 320 airplanes. Interference patterns inside the aircraft are classified and predicted based on the locations of the doors, windows, aircraft structures and the communication/navigation system-of-concern. Modeled results are compared with measured data and a plan is proposed to enhance the modeling for better prediction of electromagnetic coupling problems inside aircraft.

Jafri, Madiha↗

NASA NTRS1992

Resource constrained design of artificial neural networks using comparator neural network

We present a systematic design method executed under resource constraints for automating the design of artificial neural networks using the back error propagation algorithm. Our system aims at finding the best possible configuration for solving the given application with proper tradeoff between the training time and the network complexity. The design of such a system is hampered by three related problems. First, there are infinitely many possible network configurations, each may take an exceedingly long time to train; hence, it is impossible to enumerate and train all of them to completion within fixed time, space, and resource constraints. Second, expert knowledge on predicting good network configurations is heuristic in nature and is application dependent, rendering it difficult to characterize fully in the design process. A learning procedure that refines this knowledge based on examples on training neural networks for various applications is, therefore, essential. Third, the objective of the network to be designed is ill-defined, as it is based on a subjective tradeoff between the training time and the network cost. A design process that proposes alternate configurations under different cost-performance tradeoff is important. We have developed a Design System which schedules the available time, divided into quanta, for testing alternative network configurations. Its goal is to select/generate and test alternative network configurations in each quantum, and find the best network when time is expended. Since time is limited, a dynamic schedule that determines the network configuration to be tested in each quantum is developed. The schedule is based on relative comparison of predicted training times of alternative network configurations using comparator network paradigm. The comparator network has been trained to compare training times for a large variety of traces of TSSE-versus-time collected during back-propagation learning of various applications.

Wah, Benjamin W.↗

DOE OSTI2025

network-pruner (Neural network pruning analysis) [SWR-25-113]

This repository implements an iterative magnitude pruning algorithm for pruning neural networks in PyTorch. The pruning method involves gradually removing less significant weights from the model to achieve a specified sparsity, followed by fine-tuning the pruned model to recover performance.

Griffin, Kevin [National Renewable Energy Laborato↗

NASA NTRS2004

Real-Time Adaptive Color Segmentation by Neural Networks

Artificial neural networks that would utilize the cascade error projection (CEP) algorithm have been proposed as means of autonomous, real-time, adaptive color segmentation of images that change with time. In the original intended application, such a neural network would be used to analyze digitized color video images of terrain on a remote planet as viewed from an uninhabited spacecraft approaching the planet. During descent toward the surface of the planet, information on the segmentation of the images into differently colored areas would be updated adaptively in real time to capture changes in contrast, brightness, and resolution, all in an effort to identify a safe and scientifically productive landing site and provide control feedback to steer the spacecraft toward that site. Potential terrestrial applications include monitoring images of crops to detect insect invasions and monitoring of buildings and other facilities to detect intruders. The CEP algorithm is reliable and is well suited to implementation in very-large-scale integrated (VLSI) circuitry. It was chosen over other neural-network learning algorithms because it is better suited to realtime learning: It provides a self-evolving neural-network structure, requires fewer iterations to converge and is more tolerant to low resolution (that is, fewer bits) in the quantization of neural-network synaptic weights. Consequently, a CEP neural network learns relatively quickly, and the circuitry needed to implement it is relatively simple. Like other neural networks, a CEP neural network includes an input layer, hidden units, and output units (see figure). As in other neural networks, a CEP network is presented with a succession of input training patterns, giving rise to a set of outputs that are compared with the desired outputs. Also as in other neural networks, the synaptic weights are updated iteratively in an effort to bring the outputs closer to target values. A distinctive feature of the CEP neural network and algorithm is that each update of synaptic weights takes place in conjunction with the addition of another hidden unit, which then remains in place as still other hidden units are added on subsequent iterations. For a given training pattern, the synaptic weight between (1) the inputs and the previously added hidden units and (2) the newly added hidden unit is updated by an amount proportional to the partial derivative of a quadratic error function with respect to the synaptic weight. The synaptic weight between the newly added hidden unit and each output unit is given by a more complex function that involves the errors between the outputs and their target values, the transfer functions (hyperbolic tangents) of the neural units, and the derivatives of the transfer functions.

Duong, Tuan A.↗

NASA NTRSDate not supplied

Understanding and Verifying Neural Networks

Deep Neural Networks (DNNs) have gained immense popularity in recent times and have widespread use in applications such as image classification, sentiment analysis, speech recognition and also in safety-critical applications such as autonomous driving. However, they suffer limitations such as lack of explainability and robustness which raise safety and security concerns in their usage. Further, the complex structure and large input spaces of DNNs act as an impediment to thorough verification and testing. The SafeDNN project at the Robust Software Engineering (RSE) group at NASA aims at exploring techniques to ensure that systems that use deep neural networks are safe, robust and interpretable. In this talk, I will be presenting our technique Prophecy that automatically infers formal properties of deep neural network models. The tool extracts patterns based on neuron activations as preconditions that imply certain desirable output properties of the model. I would be highlighting case studies that use Prophecy in obtaining explanations for network decisions, understanding correct and incorrect behavior, providing formal guarantees wrt safety and robustness, and debugging neural network models. We have applied the tool on image classification networks, neural network controllers providing turn advisories in unmanned aircrafts, regression models used for autonomous center-line tracking in aircrafts and neural network object detectors

Deep Neural Networks↗

NASA NTRS1991

A consensual neural network

A neural network architecture called a consensual neural network (CNN) is proposed for the classification of data from multiple sources. Its relation to hierarchical and ensemble neural networks is discussed. CNN is based on the statistical consensus theory and uses nonlinearly transformed input data. The input data are transformed several times, and the different transformed data are applied as if they were independent inputs. The independent inputs are classified using stage neural networks and outputs from the stage networks are then weighted and combined to make a decision. Experimental results based on remote-sensing data and geographic data are given.

Benediktsson, J. A.↗