Correcting and Optimizing AICODE-Generated Aeronautical Control Code via a Knowledge Graph-Based Mechanism

Artificial Intelligence Code Generation (AICODE) technology holds immense potential for enhancing software development efficiency. However, when applied to safety-critical domains like aeronautical control, the generated code often exhibits logical flaws and compliance risks due to its lack of domain knowledge. To address this issue, this paper proposes and implements a post-processing prototype mechanism based on a Knowledge Graph (KG), specifically designed for the automated correction and optimization of AICODE-generated aeronautical control code. The mechanism begins by constructing an Aeronautical Control Knowledge Graph (ACKG) for typical UAV attitude and navigation control tasks, which integrates principles of flight control, a subset of airworthiness-related standard clauses, and historical defects. This ACKG is integrated into a prototype system with static analysis and simulation verification tools. Upon receiving a piece of code generated by AICODE, the mechanism leverages the ACKG to perform semantic-level error detection and defect tracing, and then drives AICODE to generate targeted repair solutions. Subsequently, it iteratively optimizes the functionally correct code based on performance models and optimization strategies within the ACKG. Case studies on two typical control modules (a PID attitude loop and an EKF navigator) demonstrate that, under the given dataset and experimental setup, the prototype system increased the simulation pass rate of the initial code from 68.5% to 99.1%, achieved a defect detection rate of 96.4%, and improved the performance of key modules by an average of approximately 18.5%. It must be emphasized that the main novelty lies in applying this integrated framework to the specific domain of AICODE-generated flight control software, and results are based on limited-scale case studies. This work serves as a proof of concept, and the proposed mechanism should be viewed as an auxiliary tool for development and simulation, not a qualified tool for certification under standards like DO-178C. From a research perspective, this work validates the feasibility and application potential of a synergistic ‘Knowledge Graph [Formula: see text] Large Model’ approach to safeguarding the quality of AICODE-generated code.