LLMs-based Approach and Genetic Algorithm for Sequencing Aeronautical Maintenance Tasks with Precedence Constraints

Historically, Brazilian airlines have operated under tight cost structures, making operational efficiency a determining factor for their survival. According to the 2021 Annual Operational Safety Report (RASO) of ANAC, aircraft maintenance accounts for approximately 22.4% of total operational costs, in addition to immobilizing the aircraft during its execution. In this context, planning heavy maintenance events, which may involve hundreds of interdependent tasks, represents one of the major operational challenges in the sector.

Such complexity is reflected in practice: the maintenance task scheduling problem belongs to the NP-hard class and has been predominantly addressed in the literature through heuristic approximation algorithms. However, a significant portion of existing studies faces difficulties in integrating real-world constraints and data, especially given the frequent updates of technical documents such as the Maintenance Planning Document (MPD), which defines the set of maintenance tasks applicable to each aircraft model.

To address this issue, the present work proposes a pipeline based on large language model agents (LLMs) integrated with a Genetic Algorithm, capable of directly processing the MPD and producing an optimized maintenance sequence. The approach is structured into three stages: first, an agent reads the MPD and automatically extracts the maintenance tasks along with their respective execution times; second, another agent semantically infers dependencies among these tasks based on the structure of the document itself, identifying, for example, that the cleaning of an aircraft component should precede its lubrication. Once the data are structured, a Genetic Algorithm receives the tasks, dependencies, and the number of available mechanics as input, and produces a parallel distributed sequence that minimizes the total maintenance time.

Preliminary results suggest the feasibility of the proposed approach. When processing the MPD related to the ATR72-212A aircraft, the pipeline extracted 756 tasks and automatically constructed a map of constraints and dependencies. The Genetic Algorithm found a maintenance sequence with a total execution time only 13.8% above the theoretical minimum reference, or lower bound, obtaining these results in approximately three minutes. In addition to its satisfactory performance, the solution reduces the manual effort required to update tasks whenever a new version of the MPD is released, and may be extended to other aircraft models, contributing to more efficient, reliable, and cost-effective maintenance planning. Future work may include: (i) evaluating the dependency matrix using additional documents, such as the Aircraft Maintenance Manual (AMM); (ii) comparing the approach with established methods; and (iii) incorporating real data from aircraft maintenance centers to improve the models.