Introduction
In the complex domain of aerospace engineering, NASA systems engineers are tasked with managing an immense volume of information, including thousands of requirements, variables, and schedules. This paper introduces SEVA (Systems Engineer’s Virtual Assistant), a specialized AI designed to navigate this information density. Unlike general-purpose assistants that rely on statistical probability, SEVA is built on the rigorous, logic-heavy foundations of Systems Engineering (SE), providing explainable reasoning and support for complex decision-making.
The Systems Engineering Challenge
Modern aerospace projects utilize Model-Based Systems Engineering (MBSE) tools to manage complexity. While powerful, these tools often have steep learning curves and lack the intuitive flexibility of natural language. Conversely, general AI assistants like Siri or Alexa lack the deep domain knowledge required for high-stakes engineering. Even large-scale systems like IBM Watson often prioritize statistical likelihood over hard logic. In engineering, where "probably correct" is insufficient, there is a critical need for a "workbench" assistant that can ingest technical documents, answer relational queries, and help engineers explore the consequences of design changes through trustworthy, explainable logic.
System Architecture and Methodology
SEVA’s architecture is modular and centers on a human-in-the-loop design, ensuring that the engineer remains the primary decision-maker. The system bridges the gap between natural language and formal knowledge bases through four primary components: an NLP Module, a Dynamic Knowledge Base, an Inference Engine, and an Experience Capturing unit.
Targeted Open Information Extraction (TOIE)
A core innovation of SEVA is its approach to information extraction. Standard Open Information Extraction (OIE) tools often struggle with the technical density of engineering documents, producing noisy or incomplete data. SEVA utilizes Targeted Open Information Extraction (TOIE), which leverages dependency parsing to decompose complex technical sentences into granular [subject-predicate-object] triples. This granularity is essential for populating a formal ontology capable of supporting logical reasoning.
Knowledge Base Construction
SEVA employs a hybrid ontology approach, combining a "Base" ontology of common-sense engineering rules with a "Custom" ontology of project-specific facts. This is implemented using OWL 2 (Web Ontology Language) and Description Logic. The construction process categorizes extracted information into three distinct "boxes":
- ABox (Assertion Box): Contains project-specific facts (e.g., "STI is an instrument").
- TBox (Terminology Box): Defines the conceptual hierarchy (e.g., "Mass Spectrometer is a type of Spectrometer").
- RBox (Role Box): Defines relationships and their properties, such as transitivity (e.g., "partOf").
Key Features and Insights
Temporal Awareness and Hypothetical Modeling
Engineering data is rarely static. SEVA time-tags every piece of information, allowing it to answer questions about past states and maintain comprehensive change logs. Furthermore, the system includes a "Hypothetical Mode" that allows engineers to "sandbox" the knowledge base. In this mode, users can temporarily modify facts—such as removing a component—to see how the change ripples through the system's logic without permanently altering project data.
Explainability over Probability
By utilizing Description Logic rather than purely statistical models, SEVA can provide the specific reasoning steps it took to reach an answer. This transparency is vital for building trust with engineers who must verify the logic behind system-generated responses.
Evaluation and Results
The researchers evaluated SEVA-TOIE against state-of-the-art extractors, including Stanford Open IE and ClausIE, using a corpus derived from NASA’s Systems Engineering Handbook. SEVA-TOIE consistently produced more "ontology-ready" results, successfully granularizing complex technical specifications that other extractors found noisy or incomplete. The system also demonstrated the ability to handle complex query types, including relational, recall, and hypothetical questions.
Limitations and Future Work
While SEVA represents a significant step forward, several challenges remain. The current system is limited to independent sentences; future iterations will need to incorporate coreference resolution to handle complex document structures. Additionally, the authors plan to connect SEVA to "Endowed Models"—external simulators for orbital mechanics or flight dynamics—to further enhance its analytical capabilities. Scaling the logic-based approach to millions of assertions remains a long-term goal, which the authors aim to achieve by maintaining a targeted, domain-specific focus.
Summary updated