Mamba-MSTN: A Multi-scale Adaptive State-aware Sequence Learning Framework for Fixed-wing UAV Fault Diagnosis

This paper proposes a novel deep learning framework called Mamba-MSTN, specifically designed to address the challenge of fault diagnosis for fixed-wing UAVs under low-sample conditions. The framework integrates 1D-RCNN, BiGRU, Mamba, and multi-head self-attention mechanisms, and achieves accurate modeling of complex flight data through a multi-scale temporal feature extraction module, aiming to improve the accuracy and efficiency of fault diagnosis in low-sample scenarios.

Fixed-wing UAVs are widely used in modern agriculture, aerial monitoring, and intelligent plant protection, but their safe operation faces severe challenges: scarcity of fault diagnosis data, insufficient labeled samples, and complex and variable flight environments. Traditional supervised learning performs poorly in low-sample scenarios, and existing deep learning models struggle to capture both local details and global dependencies simultaneously. Additionally, UAV fault signals have transient and multi-scale characteristics, and different fault types have similar representations, posing technical challenges.

To address the above challenges, researchers propose the Mamba-MSTN framework, whose core design concept is to achieve high-precision fault diagnosis under low-sample conditions through multi-level feature extraction and adaptive state-aware mechanisms. The framework's name comes from its key component, the Mamba state space model, which has high computational efficiency and strong modeling capabilities when processing long sequence data. Compared to Transformer, it maintains global awareness while reducing computational complexity.

Multi-scale Temporal Feature Extraction (MSTFE)

Uses a parallel multi-subspace strategy to capture dynamic characteristics of data from different time granularities, focusing on both transient changes and long-term trends, thus addressing the limitations of single-scale modeling.

Hybrid Architecture Design

Integrates the advantages of multiple neural networks:

• 1D-RCNN: Extracts local features from raw sensor data and captures subtle changes
• BiGRU: Models bidirectional temporal dependencies and enhances understanding of temporal dynamics
• Mamba Module: Enables efficient global state awareness and handles long-distance dependencies
• MHSA: Dynamically adjusts the importance of feature channels and achieves content-adaptive filtering

Adaptive State-aware Mechanism

Dynamically adjusts internal state representations based on the characteristics of input data, improving adaptability and generalization ability for different fault modes.

The research team validated the framework on real UAV flight data, and the results show that Mamba-MSTN significantly outperforms mainstream methods (traditional machine learning and existing deep learning models) under low-sample settings. Key performance indicators:

• Diagnostic accuracy: Reaches a high level with limited labeled samples
• Computational efficiency: Significantly reduces resource consumption compared to pure Transformer
• Generalization ability: Stable performance under different flight conditions and fault types
• Real-time performance: Meets the timeliness requirements for online fault monitoring

This research is of great significance to agricultural plant protection UAV operators (reducing maintenance costs and improving operational safety) and aerial monitoring tasks (avoiding equipment loss and task interruptions). The framework's modular design can be adapted to different fixed-wing UAV platforms, and its low-sample learning characteristics reduce reliance on large amounts of labeled data, making deployment more economically feasible.

The success of Mamba-MSTN provides new ideas for the field of sequence learning: combining state space models with traditional deep learning components can improve computational efficiency while maintaining expressive power, and is expected to be extended to scenarios such as equipment health monitoring and predictive maintenance. Future research directions: Improving the ability to identify extremely rare faults, exploring unsupervised/semi-supervised learning to reduce reliance on labels, and developing lightweight versions friendly to edge computing.