Leveraging language models semantic similarity capabilities to facilitate information reuse in system engineering

Tools

Darm, Paul and Marchetti, Francesco and Garcia, Gérald and Redondo, Paloma Maestro and Riccardi, Annalisa and Fernández, Alberto González (2023) Leveraging language models semantic similarity capabilities to facilitate information reuse in system engineering. In: 74th International Astronautical Congress, 2023-10-02 - 2023-10-06, Heydar Aliyev Centre. (https://iafastro.directory/iac/paper/id/79692/summ...)

[thumbnail of Darm-etal-IAC-2023-Leveraging-language-models-semantic-similarity-capabilities-to-facilitate-information-reuse]
Preview
 Text. Filename: Darm-etal-IAC-2023-Leveraging-language-models-semantic-similarity-capabilities-to-facilitate-information-reuse.pdf
Accepted Author Manuscript
License: Strathprints license 1.0
Download (479kB)| Preview

Abstract

Model-Based Systems Engineering (MBSE) is a powerful approach for designing complex engineering systems, which also generates valuable data after each conducted study. However, currently there are few to no approaches for reusing this information in a systematic way. In this paper, we propose using state-of-the-art Natural Language Processing (NLP) methods and a graph database to analyze data from past missions and facilitate the design process of new missions. In particular, we firstly develop techniques for analysing a database of past-mission requirements. This includes the ability to identify semantic similar requirements from past missions for a given new requirement. We also fine-tune a language model in order to analyse the logical traceability between two requirements. These methods are meant to enable engineers to more efficiently define the requirement space for a new spacecraft.Secondly, we develop methods to analyse the physical and functional architectures of past missions. Based on an input for a new design, a graph database of past-mission design can be queried for similar design choices and functionalities by again leveraging the abilities of semantic similarity and a specialised breadth-first-search algorithm. Finally, we show how both the requirement and design analyses could be combined in order to automatically verify if the provisions of a requirements are reflected in the physical architecture. For this analysis, a language model is used to extract core concepts from a requirement. Then, in a second step, the concepts from the requirement are mapped to nodes in the graph database. For the actual verification, a relevant extract of the graph together with the requirement are then used as input for a large language model, which is prompted to reason if the requirement is fulfilled or not. By leveraging NLP and graph search techniques, we believe that these approaches can lead to more efficient and effective design processes for complex engineering systems by reusing information from past designs. The proposed techniques have been developed and tested on real past-mission requirements and design architectures in collaboration with Thales Alenia Space, RHEA group, and the European Space Agency.

ORCID iDs

Darm, Paul, Marchetti, Francesco ORCID logoORCID: https://orcid.org/0000-0003-4552-0467, Garcia, Gérald, Redondo, Paloma Maestro, Riccardi, Annalisa ORCID logoORCID: https://orcid.org/0000-0001-5305-9450 and Fernández, Alberto González;
CORE (COnnecting REpositories)