Abstract

The fault diagnosis of general aviation aircraft presents significant challenges due to the limited availability of real fault data, the variety of fault types, and the often weak fault signatures that can complicate detection. This paper introduces an innovative framework for intelligent fault diagnosis that leverages a multi-fidelity digital twin approach. The proposed system integrates four key modules: a high-fidelity flight dynamics simulation, fault injection driven by Failure Mode and Effects Analysis (FMEA), multi-fidelity residual feature extraction, and a large language model (LLM) for generating interpretable diagnostic reports.

Core Methodology

The foundation of the proposed framework is a digital twin constructed using the JSBSim six-degree-of-freedom (6-DoF) flight dynamics engine. This engine generates comprehensive engine health monitoring data across 23 channels through semi-empirical sensor synthesis equations, which simulate the behavior of aircraft systems under various conditions.

To address the diverse fault types, a three-layer fault injection engine is developed based on FMEA. This engine models the physical causal propagation of 19 distinct engine fault types, allowing for a systematic exploration of how faults can affect aircraft performance. The integration of FMEA ensures that the fault injection process is grounded in established knowledge of potential failure modes and their effects.

The multi-fidelity residual computation framework is a critical innovation of this research. It employs paired-mirror residuals and a Gated Recurrent Unit (GRU) surrogate model for residual prediction. The high-fidelity path utilizes nominal mirror trajectories with identical initial conditions to extract clean fault deviation signals, while the low-fidelity path enables real-time residual computation through the GRU model, which predicts outcomes based on historical data. This dual-path approach allows for a more robust analysis of fault signatures, enhancing the system’s diagnostic capabilities.

For the classification of faults, a 1D Convolutional Neural Network (CNN) is employed, which performs end-to-end diagnosis across 20 different fault classes. The integration of a large language model for report generation is particularly noteworthy. This LLM diagnostic report engine synthesizes the classification results, residual evidence, and domain-specific causal knowledge to produce natural language reports that are interpretable and actionable for engineers and maintenance personnel.

Experimental results demonstrate the efficacy of the proposed framework. The paired-mirror residual scheme achieved an impressive Macro-F1 score of 96.2% on the 20-class fault diagnosis task, indicating high accuracy in identifying faults. Additionally, the GRU surrogate scheme provided a significant inference acceleration of 4.3 times, with only a minimal performance cost of 0.6%. This highlights the efficiency of the proposed methods in real-time applications.

Furthermore, a comparative analysis across 24 different schemes revealed a crucial insight: the quality of residual features plays a more significant role in diagnostic performance than the architecture of the classifier itself, with a contribution factor of approximately five times. This establishes a new design principle within the field, emphasizing the importance of prioritizing residual quality in the development of diagnostic systems.

Why this matters for the future

The implications of this research are profound for the future of aviation safety and maintenance. As the aviation industry increasingly relies on advanced technologies, the ability to diagnose faults accurately and efficiently will be paramount. The integration of digital twin technology with machine learning and FMEA knowledge represents a significant step forward in predictive maintenance practices. By improving fault diagnosis, this framework can lead to enhanced aircraft reliability, reduced downtime, and ultimately safer flight operations.

Moreover, the approach outlined in this paper can be adapted and applied to other domains beyond aviation, such as automotive engineering, industrial machinery, and even healthcare systems. The principles of multi-fidelity modeling and the emphasis on residual quality can inform the development of diagnostic tools across various industries, paving the way for smarter, more resilient systems.

Conclusion

In conclusion, this paper presents a comprehensive and innovative framework for fault diagnosis in general aviation aircraft, addressing critical challenges through the use of multi-fidelity digital twins and enhanced knowledge integration. The successful implementation of this framework not only demonstrates high diagnostic accuracy and efficiency but also establishes foundational principles that can guide future research and applications in fault diagnosis across multiple sectors. As we move forward, the insights gained from this study will be instrumental in shaping the next generation of intelligent diagnostic systems, ultimately contributing to safer and more reliable operations in aviation and beyond.