Search NASA⌕ Search

SEARCH · Search NASA

Results for “Machine Learning Algorithms”

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.

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

At least 235 records · Page 13
All sourcesNASA NTRSDOE OSTI
NASA NTRS2007

Automated Recognition of 3D Features in GPIR Images

A method of automated recognition of three-dimensional (3D) features in images generated by ground-penetrating imaging radar (GPIR) is undergoing development. GPIR 3D images can be analyzed to detect and identify such subsurface features as pipes and other utility conduits. Until now, much of the analysis of GPIR images has been performed manually by expert operators who must visually identify and track each feature. The present method is intended to satisfy a need for more efficient and accurate analysis by means of algorithms that can automatically identify and track subsurface features, with minimal supervision by human operators. In this method, data from multiple sources (for example, data on different features extracted by different algorithms) are fused together for identifying subsurface objects. The algorithms of this method can be classified in several different ways. In one classification, the algorithms fall into three classes: (1) image-processing algorithms, (2) feature- extraction algorithms, and (3) a multiaxis data-fusion/pattern-recognition algorithm that includes a combination of machine-learning, pattern-recognition, and object-linking algorithms. The image-processing class includes preprocessing algorithms for reducing noise and enhancing target features for pattern recognition. The feature-extraction algorithms operate on preprocessed data to extract such specific features in images as two-dimensional (2D) slices of a pipe. Then the multiaxis data-fusion/ pattern-recognition algorithm identifies, classifies, and reconstructs 3D objects from the extracted features. In this process, multiple 2D features extracted by use of different algorithms and representing views along different directions are used to identify and reconstruct 3D objects. In object linking, which is an essential part of this process, features identified in successive 2D slices and located within a threshold radius of identical features in adjacent slices are linked in a directed-graph data structure. Relative to past approaches, this multiaxis approach offers the advantages of more reliable detections, better discrimination of objects, and provision of redundant information, which can be helpful in filling gaps in feature recognition by one of the component algorithms. The image-processing class also includes postprocessing algorithms that enhance identified features to prepare them for further scrutiny by human analysts (see figure). Enhancement of images as a postprocessing step is a significant departure from traditional practice, in which enhancement of images is a preprocessing step.

Park, Han↗
NASA NTRSDate not supplied

Application of Machine Learning Techniques in Calibration and Data Reduction of Multi-Hole Probes

This work presents procedures for implementing machine learning methods into existing algorithms for multi-hole probe calibration and data reduction. It demonstrates that using artificial neural networks (ANNs) can decrease the amount of calibration data needed to achieve a specific calibration uncertainty by over 50%, while also significantly reducing data reduction times. Instead of surface fitting methods, ANNs are employed. Initially, directional calibration coefficients related to flow angles are computed based on pressure measurements, and then these flow angles serve as input parameters for subsequent ANNs to iteratively define Mach number, static pressure, and total pressure. In an alternative approach, new calibration coefficients directly relate pressure measurements from the five-hole probe to the quantities of interest, thereby eliminating the need for iterative algorithms used in conventional surface fitting methods. This method offers several advantages: an average increase of less than 1%in calibration uncertainty for flow angles and a significant reduction in data reduction times to a few seconds on average. Additionally, the methodology is confirmed to avoid both over- and under-fitting.

Machine Learning↗
NASA NTRS1996

Development of Collaborative Research Initiatives to Advance the Aerospace Sciences-via the Communications, Electronics, Information Systems Focus Group

The primary goal of the Adaptive Vision Laboratory Research project was to develop advanced computer vision systems for automatic target recognition. The approach used in this effort combined several machine learning paradigms including evolutionary learning algorithms, neural networks, and adaptive clustering techniques to develop the E-MOR.PH system. This system is capable of generating pattern recognition systems to solve a wide variety of complex recognition tasks. A series of simulation experiments were conducted using E-MORPH to solve problems in OCR, military target recognition, industrial inspection, and medical image analysis. The bulk of the funds provided through this grant were used to purchase computer hardware and software to support these computationally intensive simulations. The payoff from this effort is the reduced need for human involvement in the design and implementation of recognition systems. We have shown that the techniques used in E-MORPH are generic and readily transition to other problem domains. Specifically, E-MORPH is multi-phase evolutionary leaming system that evolves cooperative sets of features detectors and combines their response using an adaptive classifier to form a complete pattern recognition system. The system can operate on binary or grayscale images. In our most recent experiments, we used multi-resolution images that are formed by applying a Gabor wavelet transform to a set of grayscale input images. To begin the leaming process, candidate chips are extracted from the multi-resolution images to form a training set and a test set. A population of detector sets is randomly initialized to start the evolutionary process. Using a combination of evolutionary programming and genetic algorithms, the feature detectors are enhanced to solve a recognition problem. The design of E-MORPH and recognition results for a complex problem in medical image analysis are described at the end of this report. The specific task involves the identification of vertebrae in x-ray images of human spinal columns. This problem is extremely challenging because the individual vertebra exhibit variation in shape, scale, orientation, and contrast. E-MORPH generated several accurate recognition systems to solve this task. This dual use of this ATR technology clearly demonstrates the flexibility and power of our approach.

Knasel, T. Michael↗
NASA NTRS2005

Experiments on Supervised Learning Algorithms for Text Categorization

Modern information society is facing the challenge of handling massive volume of online documents, news, intelligence reports, and so on. How to use the information accurately and in a timely manner becomes a major concern in many areas. While the general information may also include images and voice, we focus on the categorization of text data in this paper. We provide a brief overview of the information processing flow for text categorization, and discuss two supervised learning algorithms, viz., support vector machines (SVM) and partial least squares (PLS), which have been successfully applied in other domains, e.g., fault diagnosis [9]. While SVM has been well explored for binary classification and was reported as an efficient algorithm for text categorization, PLS has not yet been applied to text categorization. Our experiments are conducted on three data sets: Reuter's- 21578 dataset about corporate mergers and data acquisitions (ACQ), WebKB and the 20-Newsgroups. Results show that the performance of PLS is comparable to SVM in text categorization. A major drawback of SVM for multi-class categorization is that it requires a voting scheme based on the results of pair-wise classification. PLS does not have this drawback and could be a better candidate for multi-class text categorization.

Namburu, Setu Madhavi↗
NASA NTRSDate not supplied

The Added Value of SMAP Soil Moisture in Crop Yield Forecasting Over Argentina

Argentina is one of the major producers and exporter of soybeans, corn, and wheat to the world market; therefore, the accurate and timely forecasting of those crops yield is crucial to national crop management and global food security. Previous studies have mainly focused on developing forecasting models for a specific crop type and location using a single source of data (e.g., vegetation indices), thus providing little insight into the forecasting models' performance on different crop types and regions. Besides, these models are based on traditional statistical regression algorithms, while more advanced machine learning approaches have not been explored. This study investigated the estimation of crop yields of three major crops (corn, soybean, and winter wheat) using Multiple Linear Regression (MLR) and Support Vector Machine (SVM), over major growing provinces in Argentina. Our models were trained and evaluated on data from 2015 to 2020, where three remote sensing products (Normalized difference vegetation index (NDVI), SMAP soil moisture, and MODIS evapotranspiration) were used as predictors. Our results indicated that accurate crop yield forecasts using the developed regression models could be made one to two months before harvest. The MLR and SVM model performance varied among different crop types, where soybean and corn exhibited better predictability compare to the wheat. In most cases, the SVM outperformed the multiple linear regression model due to its ability to capture the nonlinear and complex features of the crop-production process. The forecasted model that combines data from multiple sources outperformed single-source satellite data. The highest accuracy was obtained when the three data sources were all considered in the model development. Results also indicated that the inclusion of SMAP soil moisture improved crop yield forecasting in most provinces, and the most significant improvements occurred in the drier region.

Nazmus Shams Sazib↗
NASA NTRSDate not supplied

Modeling Atmospheric Science Knowledge from Research Publications

NASA Earth Science Data Centers contain enormous amounts of remote sensing digital data. It is often a significant challenge for users to find data suitable for their research topic in these vast archives. One of the approaches is the usage-driven dataset discovery, where users seek publications on projects similar to their intended study. For this approach to be effective, users need a clear connection between the underlying data in the publications and the study objectives; this is not often apparent to non-expert users. Tools and methodologies that can help facilitate and organize these connections are therefore valuable for creating improved knowledge mappings, which can be further used by search engines to suggest data or publications best tailored to a user’s specific research goal. As an illustration of these challenges, in this work we focus on the atmospheric chemistry processes related to Earth environmental impacts such as ozone depletion, aerosols, smog formation, acid rain, and radiative forcing. We further limit our study to publications that use data from the Microwave Limb Sounder (MLS) instrument flown on the Aura Earth Observing System. To create knowledge representations of science carried out in these publications, we use existing ontologies such as the Global Change Master Directory (GCMD) and Semantic Web for Earth and Environmental Terminology (SWEET). These ontologies together encompass term dictionaries that include measured variables, names of molecules or radicals, mission and instrument names, locations, action words, among many others. Based on these terms acknowledge graph database was populated with the terms retrieved from scientific publications that study atmospheric chemistry. These databases can be used to further enhance the automation of knowledge discovery and facilitate machine learning and artificial intelligence algorithms or applications. These tools and methods can also be extended to apply to content from other related Earth science domains.

Irina Gerasimov↗
NASA NTRS2021

The Large Footprint of Small-scale Artisanal Gold Mining in Ghana

Gold mining has played a significant role in Ghana's economy for centuries. Regulation of this industry has varied over time and while industrial mining is prevalent in the country, the expansion of artisanal mining, or Galamsey has escalated in recent years. Many of these artisanal mines are not only harmful to human health due to the use of Mercury (Hg) in the amalgamation process, but also leave a significant footprint on terrestrial ecosystems, degrading and destroying forested ecosystems in the region. In this study, the Landsat image archive available through Google Earth Engine was used to quantify the total footprint of vegetation loss due to artisanal goldmines in Ghana from 2005 to 2019 and understand how conversion of forested regions to mining has changed over a decadal period from 2007 to 2017. A combination of machine learning and change detection algorithms were used to calculate different land cover conversions and the timing of conversion annually. Within the study area of southwestern Ghana, our results indicate that approximately 47,000 ha (⨦2218 ha) of vegetation were converted to mining at an average rate of ~2600 ha yr−1. The results indicate that a high percentage(~50%) of this mining occurred between 2014 and 2017. Around 700 ha of this mining occurred within protected areas as mapped by the World Database of Protected Areas. In addition to deforestation, increased artisanal mining activity in recent years has the potential to affect human health, access to drinking water resources and food security. This work expands upon limited research into the spatial footprint of Galamseyin Ghana, complements mapping efforts by local geographers, and will support efforts by the government of Ghana to monitor deforestation caused by artisanal mining.

Abigail Barenblitt↗
NASA NTRSDate not supplied

Using Machine Learning to Infer Material Properties of Debris Fragments from X-ray Images in the DebriSat Project

The DebriSat project is a collaboration effort with the NASA Orbital Debris Program Office, the U.S. Space Force Space Systems Command Center, The Aerospace Corporation, and the University of Florida. To date, over 200,000 fragments from this ground-based, hypervelocity impact experiment have been collected, and processing is underway to determine their physical characteristics, such as material, shape, color, characteristic length, and average cross-sectional area. The x-ray process is primarily used to identify the location of the fragments and estimated size for extraction, so that these physical characteristics can be assessed. This paper proposes a machine learning-based approach to characterize materials from x-ray images of debris fragments embedded in soft-catch foam used in the DebriSat project. The novel methodology discussed in this paper will highlight the use of x-ray imagery data to characterize these fragments without extraction or a human-in-the-loop. Both supervised and unsupervised machine learning techniques are utilized with this approach to infer the physical parameters of the fragments embedded in the soft-catch foam panels used in the impact experiment based on x-ray images of the foam panels. Additionally, 3D reconstructions of the extracted fragments are created with images taken from two different angles using the structure from motion (SfM) method. The characteristic lengths and shape from the 3D reconstruction, alongside the physical characteristics of the debris, are used in the inference of the material type. To develop and test the approach, a dataset of x-ray images of debris fragments of varying sizes and materials is collected. Supervised learning methods such as convolutional neural networks (CNNs), support vector machines (SVM), decision trees, and random forest classifiers are used due to the high-dimensional feature spaces of the debris and nonlinear decision boundaries for material categorization. Given the limited pre-labeled data of embedded debris materials smaller than 10 mm, unsupervised machine learning techniques such as clustering algorithms and autoencoders are used, in addition to supervised learning methods. The clustering algorithms group similar fragments together based on their physical properties, and autoencoders reduce the dimensionality of the x ray images and extract relevant features. The performance of the proposed approach's is analyzed using a range of statistical methods, including confusion matrices, receiver operating characteristic curves, and precision-recall curves. The results are compared with those obtained using a baseline approach that relies on manual identification and classification of debris fragments. To evaluate the effectiveness of different machine learning methods, statistical tests such as t-tests, ANOVA, and cross-validation are performed, comparing the performance of CNNs, SVMs, clustering algorithms, and autoencoders. Additional analysis needs to be conducted to identify any sources of bias or variability that may affect the results, such as variations in imaging conditions or fragmentation patterns. Other topics explored are limitations, refinements, and the potential use of semi-supervised learning techniques, such as self-training to label unlabeled datasets and co-training using x-ray images taken from two different angles as two different models.

Saik Anam Siam↗
NASA NTRS2023

Using Machine Learning to Infer Material Properties of Debris Fragments from X-ray Images in the DebriSat Project

The DebriSat project is a collaboration effort with the NASA Orbital Debris Program Office, the U.S. Space Force Space Systems Command Center, The Aerospace Corporation, and the University of Florida. To date, over 200,000 fragments from this ground-based, hypervelocity impact experiment have been collected, and processing is underway to determine their physical characteristics, such as material, shape, color, characteristic length, and average cross-sectional area. The x-ray process is primarily used to identify the location of the fragments and estimated size for extraction, so that these physical characteristics can be assessed. This paper proposes a machine learning-based approach to characterize materials from x-ray images of debris fragments embedded in soft-catch foam used in the DebriSat project. The novel methodology discussed in this paper will highlight the use of x-ray imagery data to characterize these fragments without extraction or a human-in-the-loop. Both supervised and unsupervised machine learning techniques are utilized with this approach to infer the physical parameters of the fragments embedded in the soft-catch foam panels used in the impact experiment based on x-ray images of the foam panels. Additionally, 3D reconstructions of the extracted fragments are created with images taken from two different angles using the structure from motion (SfM) method. The characteristic lengths and shape from the 3D reconstruction, alongside the physical characteristics of the debris, are used in the inference of the material type. To develop and test the approach, a dataset of x-ray images of debris fragments of varying sizes and materials is collected. Supervised learning methods such as convolutional neural networks (CNNs), support vector machines (SVM), decision trees, and random forest classifiers are used due to the high-dimensional feature spaces of the debris and nonlinear decision boundaries for material categorization. Given the limited pre-labeled data of embedded debris materials smaller than 10 mm, unsupervised machine learning techniques such as clustering algorithms and autoencoders are used, in addition to supervised learning methods. The clustering algorithms group similar fragments together based on their physical properties, and autoencoders reduce the dimensionality of the x ray images and extract relevant features. The performance of the proposed approach's is analyzed using a range of statistical methods, including confusion matrices, receiver operating characteristic curves, and precision-recall curves. The results are compared with those obtained using a baseline approach that relies on manual identification and classification of debris fragments. To evaluate the effectiveness of different machine learning methods, statistical tests such as t-tests, ANOVA, and cross-validation are performed, comparing the performance of CNNs, SVMs, clustering algorithms, and autoencoders. Additional analysis needs to be conducted to identify any sources of bias or variability that may affect the results, such as variations in imaging conditions or fragmentation patterns. Other topics explored are limitations, refinements, and the potential use of semi-supervised learning techniques, such as self-training to label unlabeled datasets and co-training using x-ray images taken from two different angles as two different models.

Saik Anam Siam↗
NASA NTRSDate not supplied

Hyperspectral Sounder Spectral Fingerprinting: Using Machine Learning Techniques to Enhance Model-Based Physical Inversion

Different retrieval algorithms have been developed to process top-of-atmosphere (TOA) spectral radiance data provided by hyperspectral infrared sounder missions. Those algorithms are either optimal estimation method (OEM) based schemes with radiative transfer calculation involved in the retrieval process, or machine learning based methods that allow ultra-efficient data procession but lack of radiometric consistency validation based on the directly measured information. Combining both approaches leverages their respective technical advantages, leading to more accurate results. This study introduces a hyperspectral sounder fingerprinting algorithm to explore this hybrid approach. This approach involves the use of a spectral information-based classification method to identify an reference geophysical state and the corresponding radiative kernel. This enables the efficient retrieval of geophysical variables of interest through a radiative kernel-based linear inversion procedure. The fingerprinting method has been applied to analyze a decade-long hyperspectral sounder data record.

Wan Wu↗
NASA NTRS2000

Reinforcement Learning in Distributed Domains: Beyond Team Games

Distributed search algorithms are crucial in dealing with large optimization problems, particularly when a centralized approach is not only impractical but infeasible. Many machine learning concepts have been applied to search algorithms in order to improve their effectiveness. In this article we present an algorithm that blends Reinforcement Learning (RL) and hill climbing directly, by using the RL signal to guide the exploration step of a hill climbing algorithm. We apply this algorithm to the domain of a constellations of communication satellites where the goal is to minimize the loss of importance weighted data. We introduce the concept of 'ghost' traffic, where correctly setting this traffic induces the satellites to act to optimize the world utility. Our results indicated that the bi-utility search introduced in this paper outperforms both traditional hill climbing algorithms and distributed RL approaches such as team games.

Wolpert, David H.↗
NASA NTRS2004

Improving Search Algorithms by Using Intelligent Coordinates

We consider algorithms that maximize a global function G in a distributed manner, using a different adaptive computational agent to set each variable of the underlying space. Each agent eta is self-interested; it sets its variable to maximize its own function g (sub eta). Three factors govern such a distributed algorithm's performance, related to exploration/exploitation, game theory, and machine learning. We demonstrate how to exploit alI three factors by modifying a search algorithm's exploration stage: rather than random exploration, each coordinate of the search space is now controlled by a separate machine-learning-based player engaged in a noncooperative game. Experiments demonstrate that this modification improves simulated annealing (SA) by up to an order of magnitude for bin packing and for a model of an economic process run over an underlying network. These experiments also reveal interesting small-world phenomena.

Wolpert, David H.↗
NASA NTRSDate not supplied

Application of Machine Learning Techniques in Calibration and Data Reduction of Multi-Hole Probes

This work presents procedures to implement machine learning methods in the existing algorithms for multi-hole probe calibrations and data reduction. It is shown here, that utilizing artificial neural networks (ANNs) can reduce the amount of calibration data that needs to be acquired in order to obtain a specific calibration uncertainty, by more than 50% while simultaneously reducing data reduction times significantly. ANNs were used instead of the surface fitting methods, where first, the directional calibration coefficients related to the flow angles are calculated based on the pressure measurements, and then the flow angles are used as a set of the input parameters for the following ANNs to define Mach number and static and total pressure iteratively. In a second approach, novel calibration coefficients were used to directly relate the pressure measurements from five-hole probe to the quantities of interest thus, eliminating the need for iterative algorithms used in the conventional surface fitting methods. The advantageous features of this method are an average increase of less than 1% in the calibration uncertainty for flow angles and significant reduction of the data reduction times (few seconds). In addition, we confirmed the methodology to avoid over-fitting and under-fitting.

Machine Learning↗
NASA NTRS1999

Avoiding Braess' Paradox Through Collective Intelligence

In an Ideal Shortest Path Algorithm (ISPA), at each moment each router in a network sends all of its traffic down the path that will incur the lowest cost to that traffic. In the limit of an infinitesimally small amount of traffic for a particular router, its routing that traffic via an ISPA is optimal, as far as cost incurred by that traffic is concerned. We demonstrate though that in many cases, due to the side-effects of one router's actions on another routers performance, having routers use ISPA's is suboptimal as far as global aggregate cost is concerned, even when only used to route infinitesimally small amounts of traffic. As a particular example of this we present an instance of Braess' paradox for ISPA'S, in which adding new links to a network decreases overall throughput. We also demonstrate that load-balancing, in which the routing decisions are made to optimize the global cost incurred by all traffic currently being routed, is suboptimal as far as global cost averaged across time is concerned. This is also due to "side-effects", in this case of current routing decision on future traffic. The theory of COllective INtelligence (COIN) is concerned precisely with the issue of avoiding such deleterious side-effects. We present key concepts from that theory and use them to derive an idealized algorithm whose performance is better than that of the ISPA, even in the infinitesimal limit. We present experiments verifying this, and also showing that a machine-learning-based version of this COIN algorithm in which costs are only imprecisely estimated (a version potentially applicable in the real world) also outperforms the ISPA, despite having access to less information than does the ISPA. In particular, this COIN algorithm avoids Braess' paradox.

Wolpert , David H.↗
NASA NTRS2006

Measuring Constraint-Set Utility for Partitional Clustering Algorithms

Clustering with constraints is an active area of machine learning and data mining research. Previous empirical work has convincingly shown that adding constraints to clustering improves the performance of a variety of algorithms. However, in most of these experiments, results are averaged over different randomly chosen constraint sets from a given set of labels, thereby masking interesting properties of individual sets. We demonstrate that constraint sets vary significantly in how useful they are for constrained clustering; some constraint sets can actually decrease algorithm performance. We create two quantitative measures, informativeness and coherence, that can be used to identify useful constraint sets. We show that these measures can also help explain differences in performance for four particular constrained clustering algorithms.

constraints↗
NASA NTRS1987

The power of neural nets

Implementation of the Hopfield net which is used in the image processing type of applications where only partial information about the image may be available is discussed. The image classification type of algorithm of Hopfield and other learning algorithms, such as the Boltzmann machine and the back-propagation training algorithm, have many vital applications in space.

Ryan, J. P.↗
NASA NTRSDate not supplied

NASA's Moon to Mars Autonomous Habitat Status

NASA is developing a strategy for sending humans to the Mars vicinity, known broadly as the Moon to Mars (M2M) Campaign. A critical part of this campaign is the development of in-space and surface habitation systems capable of substantially extending human presence beyond Low Earth Orbit (LEO). Mars missions feature an in-space transit habitat capable of supporting crews of four on ~850-1200-day missions, including transit to and from Mars and time in Mars orbit. Surface and transit habitats are complex elements which must keep crewmembers healthy and productive in deep-space environments with limited resources, long rescue times in contingency situations, and communication delays; all within constrained mass, volume, and power budgets. These habitats provide crew both living and workspace as well as most of the resources needed to support crew life. For deep space habitats, automation needs to be employed due to latency and for significant amounts of time when the habitats are uncrewed. Automation of systems is possible in space applications, but there are limitations. Outside of the Earth’s (or any) magnetosphere, radiation environments are harsh to both the physical hardware and the software components. Radiation (charged particles and ionizing electromagnetic waves) degrades and damages the hardware and causes single event upsets (SEUs) in software. If the hardware is damaged, data can be lost, or control actions not made. For software, SEUs cause algorithms to result in different solutions, or incorrect commands to be sent out. This means that algorithms and hardware used for deep space systems are different than what is used on Earth. Radiation-tolerant hardware is generations behind the current state-of-the-art hardware. Recent NASA missions, such as James Webb Space Telescope, continue to rely on older technologies such as the RAD750 processor, and the most advanced processors are still single core and less than 1.5 GHz. There have been attempts to use higher performance processors, but these often take multiple mitigation steps to handle the radiation environments, which limits the processing power and/or throughput. Current techniques for radiation mitigation have been redundancies, voting, physical separation of hardware, encasing materials, under-clocking hardware, and more. Some radiation mitigation techniques do provide benefits such as having a redundant system to improve the probability that a system will be available when needed. Autonomous software systems will have fewer interactions with humans on deep space missions and therefore need to be able to handle more off-nominal conditions. Microgravity also complicates the autonomous aspects of the mission because autonomous systems are usually built from known deterministic states, but microgravity causes physical objects to shift and move changing the location an autonomous system placed the object. Not only does the software need to be reliable and deterministic, losing resources due to a software error is not only costly but detrimental to reputation. The combination of having lower performance hardware and having to be able to verify and deterministically run software and an ever-changing environment makes deep space autonomous systems more complicated. Multiple gaps have been identified including verification of autonomous software algorithms (including artificial intelligence and machine learning), higher performance processors (graphics and general purpose), high speed networks (onboard and transmissions), memory, power distribution, data security, and variations from these. These gaps need to be closed for more advanced systems to be deployed and reduce the size, weight, and power impacts on the habitats.

Scott B. Tashakkor↗
NASA NTRSDate not supplied

Towards an Aviation Large Language Model by Fine-tuning and Evaluating Transformers

In the aviation domain, there are many applications for machine learning and artificial intelligence tools that utilize natural language. For example, there is a desire to know the commonalities in written safety reports such as voluntary post incidents reports or aerial wildfire operations reports to better understand the risks present. Another use-case is the possibility of extracting airspace procedures and constraints currently written in documents such as Letters of Agreement. These applications can benefit from the use of state-of-the-art natural language processing techniques when adapted to the language/phraseology specific to the aviation domain. This paper evaluates the viability of adaptation of NLP tools to the aviation domain by fine-tuning transformer based models using aviation data sets. In 2018, a novel language model based on neural units (also called transformers) was created and became known as “Bidirectional Encoder Representations from Transformers” or BERT. This architecture combined with large amounts of English training data and innovative semi-supervised training tasks set the standard for what would later emerge as Large Language Models. The performance of these models was further improved by hyperparameter tuning and refinement of the semi-supervised training task and resulted in “Robustly Optimized BERT Pre-training Approach through hyperparameter tuning” or RoBERTa models. These pre-trained Large Language Models proved to be useful for a wide variety of natural language processing tasks such as text classification and question answering through a process called fine-tuning. The transformer architecture with pre-trained weights served as the basis with the last few layers replaced with layers fine-tuned to perform a new task e.g., a layer that provides a label for the entire input text. This process of fine-tuning can also be used to adapt the models to new domains; e.g., BioBERT started with the pre-trained BERT model and was completed by additional fine-tuning and training on biomedical documents. Transformer-based architectures can also be used to create rich representations of text called embeddings which can serve as the input to other machine learning models. This allows simpler algorithms such as logistic regression to use context-rich representations of the text while still remaining quick to train and evaluate. In the world of aviation, there is a growing demand for natural language processing and understanding but the domain presents unique challenges. Due to the technical content (and specialized language) of most aviation documents, fine-tuning pre-trained Large Language Models to specific tasks has not met the benchmark on natural language processing tasks set by simpler models trained from scratch on the data. To address this deficiency, this paper evaluates the improvements from fine-tuning a Large Language Model on a large set of aviation documents using the original semi-supervised training tasks before performing specific natural language tasks. In fine-tuning, a domain-specific dataset is used on the original training task but with the pre-trained Large Language Model instead of starting from a random initialization. This approach allows the model to be adapted to the specific domain language without discarding the information gained from training on general English data. This paper utilized two major dataset types to train and assess the RoBERTa fine-tuning performance. The first are 7,057 Letters of Agreement which are Federal Aviation Administration (FAA) documents that formalize airspace operations across the national airspace system. They contain many examples of ‘aviation English’ using domain specific terminology and phrasing which serves as a representative basis to perform the semi-supervised fine-tuning. The second type is the 494 document classification labels to be used for evaluation. This down-stream evaluation aims to show the performance of the fine-tuned model, better understand how much data is needed for an effective fine-tuning, and how fine-tuning can be adapted for different applications in-the domain. After semi-supervised training, evaluation begins by encoding the documents for classification using the fine-tuned RoBERTa model. Then a logistic regression classifier is trained to label the document type and compared against our ground truth labels. This currently leads to a 82.8% accuracy on 10-fold cross validation showing improvement over baseline RoBERTa which achieved 81.0%. We plan to measure the improvements on additional tasks and it is expected that these improvements will lead to more robust models that can tackle the natural language processing challenges present in aviation datasets.

ATM↗