Year: 2026
Type: dissertation
Abstract: Many industrial products need to comply with regulations, standards, or other compliance requirements to ensure baseline security and safety. Manual compliance assessments against these requirements tend to be slow and cumbersome. Thus, automated compliance checking has been introduced to make compliance processes more efficient. Recent methods have applied artificial intelligence (AI) to this task. However, more complex AI models are often opaque in how they reach their decisions, reducing the traceability and transparency of the systems. This thesis investigates the use of large language models (LLMs) to support compliance assessments, with particular attention to transparency and traceability in the system design. The proposed system has two stages. In the first stage, an audit checklist is generated for each compliance requirement in an input standard or regulation. In the second stage, these checklists are used to assess the provided product evidence for compliance, and compliance for each requirement is determined using a deterministic decision rule. The system also generates suggestions on how to improve the presented evidence to increase compliance in the future. The system was evaluated using compliance data from two IoT devices and requirements from the EN IEC 62443-4-2:2019/prAA:2026 draft standard amendment [15]. The proposed system was compared against a non-checklist based approach, called the direct LLM baseline. An expert questionnaire was conducted to assess the usefulness and clarity of outputs and preference between the two approaches. Other performed evaluations examined output stability, evidence retrieval, and recommendation quality. The results indicate that the proposed approach produces stable outputs and is perceived as useful by experts. Overall, the evaluation suggests that the checklist-based approach is a feasible method for automated compliance checking and warrants further investigation.
Source: Aaltodoc (Aalto University)
Author: Ahola, Taina
Institution: Aalto University
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204017660
