Large Language Models for Signal Processing Pipelines : Advancing Military Electronic Intelligence Workflows

Sammanfattning

Artificial intelligence (AI) has come to play a critical role in the modern defense landscape, where data has increasingly emerged as a strategic asset. Signal processing of radar emissions for detection, identification, and classification is no exception, but broader adaptability in the electronic intelligence (ELINT) domain has remained particularly unexplored by public research. Whereas the environment is typically challenged by complex and incompatible datasets, along with high data volumes and information overload, this thesis investigates whether and how large language models (LLMs) can be utilized to address these challenges. An AI-powered chatbot was proposed, capable of integrating structured and unstructured data, with a modular design intended to enhance transparency and explainability. The system was evaluated on a synthetic dataset, with architectural and design choices based on assumptions constrained by the sensitive nature of the domain. Results from unit testing showed a classification accuracy of 97.8% for prompt categorization, correct results returned by 93.3% of generated SQL queries, and significant time savings compared to manual querying. The system also enabled explainability and offered transparency through intermediate logging and decision justification. However, conversational tests exposed limitations such as error propagation, vulnerability to ambiguous inputs, and dependence on model capabilities and user proficiency. While many of these issues could be addressed, the findings revealed a strong correlation between system reliability and available resources, highlighting both opportunities and risks in operational contexts. Despite limitations due to simulated data and general assumptions, the prototype demonstrates feasibility and potential for broader adoption of LLMs in intelligence environments.