Self-attention enhanced GraphSAGE for UAV fault diagnosis using vibration signals

Abstract

Traditional data-driven fault diagnosis methods for unmanned aerial vehicles (UAVs) often struggle to in effectively extract fault features from complex vibration signals. To address this challenge, this paper proposes an enhanced fault diagnosis framework based on GraphSAGE, incorporating a self-attention mechanism (GraphSAGE-SA). Initially, a customized vibration data acquisition system was developed using MPU6050 sensors to capture vibration data across various operational conditions. Subsequently, the K-Nearest Neighbors (KNN) algorithm was employed to convert one-dimensional vibration signals into graph-structured data, revealing explicit topological relationships. The proposed GraphSAGE-SA model aggregates information from both nodes and neighbors via a hierarchical sampling process. The integrated self-attention mechanism then adaptively adjusts the importance of different neighbors, facilitating more precise feature extraction. Experimental validation conducted on a quadrotor UAV platform demonstrated that the proposed method achieves an impressive fault classification accuracy of 98%, surpassing conventional GraphSAGE variants by 3–12% and other graph-based approaches. This superior performance is attributed to the model’s ability to capture local structural patterns and global dependencies simultaneously, enabled by the attention mechanism. This research presents a novel and effective solution for intelligent fault diagnosis in UAVs and other complex mechanical systems. In real-world applications, timely and accurate fault diagnosis is essential for ensuring UAV operational safety, mission reliability, and preventing catastrophic failures.

Introduction

Unmanned Aerial Vehicles (UAVs) have become essential assets across a wide spectrum of domains, including surveillance, disaster response, precision agriculture, and logistics, due to their operational flexibility, economic viability, and capability to harsh environments^1–3. Nevertheless, as UAV architectures grow increasingly complex and are frequently deployed in demanding conditions, they are prone to a range of faults, such as motor anomalies and structural impairments⁴. Undetected or misdiagnosed faults can escalate rapidly, potentially resulting in critical system failures that compromise mission integrity and operational safety⁵. Consequently, there is an urgent and growing need for the development of robust and efficient fault diagnosis strategies tailored to the unique operational characteristics of UAV systems.

In the past, fault diagnosis for UAVs primarily depended on visual inspections, a method that was time-consuming, labor-intensive, and often prone to inaccuracy. Certain faults were difficult to detect and could not be identified in a timely manner⁶. As a result, the rapid and efficient detection of faults has remained a central focus of research. Modern UAV fault diagnosis has typically evolved through the following stages:

Fault diagnosis in rotating machinery has been a key area of research for many scholars. Existing fault diagnosis techniques can be broadly classified into several categories: signal-based, model-based, machine learning-based, and deep learning-based methods. Signal-based methods identify faults by applying various data analysis techniques, such as time-domain analysis⁷, frequency-domain analysis⁸, and time–frequency analysis⁹, to detect signals that deviate from normal operating conditions. Model-based fault diagnosis involve constructing an input–output model for UAVs and detecting faults by comparing the actual output with the estimated output. Asadi et al. applied model-based approaches, such as Kalman filter observers, to detect rotor faults in quadrotor UAVs¹⁰. The main challenge of this approach lies in the need for an accurate mathematical model, which is often difficult to obtain.

Machine learning-based fault diagnosis methods leverage shallow models for feature extraction and employ classifiers for fault identification. The primary methods include Support Vector Machines (SVM), Decision Trees, and K-Nearest Neighbors (KNN). Baskaya et al. and Bondyra et al. proposed an SVM model for detecting propeller damage in multi-rotor UAVs based on vibration signals collected by the onboard Inertial Measurement Unit (IMU) sensor^11,12. Lai et al. developed a KNN approach for detecting motor damage, achieving an accuracy rate of 90.73%, even under mixed fault conditions¹³. Four distinct damage modes were investigated, including motor degradation, friction-induced failures, backlash, and external perturbation effects. These modes were detected using various machine learning methods, such as BayesNet, Naive Bayes, Decision Trees, induction rules, KNN, logistic regression, and Multilayer Perceptron. The results revealed that BayesNet exhibited the best performance, with an accuracy close to 90%¹⁴. Despite the successes of machine learning-based fault diagnosis techniques, these algorithms face challenges in adaptively extracting features and require significant expertise in feature selection, ultimately limiting diagnostic accuracy.

With the advancement of computational resources and the proliferation of big data, deep learning-based fault diagnosis has gained considerable attention. These approaches typically utilize multi-layer neural networks capable of automatically learning and extracting relevant features from raw data inputs¹⁵. By leveraging hierarchical representations through hidden layers, deep learning methods significantly enhance the precision of fault identification tasks¹⁶. For instance, a convolutional neural network (CNN) was employed to detect two types of propeller damage in quadrotor UAVs using acoustic signals. The time–frequency representations of the sound data were used as inputs to the CNN, achieving an accuracy exceeding 90% in experimental validation¹⁷. In related work, sensor data from accelerometers, gyroscopes, and magnetometers, sampled at 100 Hz, were utilized to detect operational anomalies. CNN and long short-term memory networks (LSTMs) were adopted to perform fault classification, resulting in an accuracy of 88% based on empirical results¹⁸.

Traditional fault diagnosis techniques, including model-based methods and signal processing approaches, have been extensively applied in UAV systems. However, these methods often face challenges in dealing with the high-dimensional, nonlinear, and heterogeneous nature of UAV sensor data, particularly vibration data, which contains critical information regarding the health status of mechanical components¹⁹. Furthermore, the limited availability of labeled fault data further hampers the performance of data-driven methods, such as deep learning, which require large volumes of annotated data for effective training²⁰.

Graph Neural Networks (GNNs) present several distinct advantages over traditional deep learning methods, such as Convolutional Neural Networks (CNNs), especially in fault diagnosis applications. While CNNs are designed for grid-like data (e.g., images), they struggle with irregular or non-Euclidean structures. In contrast, GNNs are inherently suited for graph-structured data, making them ideal for modeling complex systems like sensor networks or industrial equipment. GNNs are particularly effective in capturing the intricate dependencies between components, which are critical for precise fault diagnosis²¹. Moreover, GNNs leverage message-passing mechanisms to aggregate both local and global information, offering a more holistic understanding of the system’s state. This is in contrast to CNNs, which typically focus on local features and may overlook broader contextual relationships²². Additionally, GNNs are highly adaptable to dynamic changes in system structure, providing robustness in real-world applications where data evolves over time. Their ability to process multi-modal data and provide interpretable insights further enhances their utility in fault diagnosis, making GNNs a superior choice for complex, interconnected systems compared to traditional deep learning approaches²³. For instance, Wang et al. proposed a fault diagnosis method for rotor-bearing systems based on FSASCA-VMD and GraphSAGE-SA, achieving over 98% diagnostic accuracy even under noisy conditions²⁴. The application of GNNs, particularly GraphSAGE, in UAV fault diagnosis is well-justified for several reasons. First, the multi-sensor system on a UAV naturally forms a graph-structured network where nodes (sensors) are spatially related. Faults often propagate through this physical structure, creating inherent topological dependencies. Unlike traditional deep learning models like CNNs, which process data in isolation, GNNs explicitly model these inter-sensor relationships. Second, the KNN-based graph construction transforms sequential vibration data into a graph, capturing the underlying manifold structure of the high-dimensional signal space. GraphSAGE is particularly well-suited to this task, as it is an inductive learning framework that enables the model to generate embeddings for new, unseen data samples without requiring a fixed graph structure, which is—an essential feature for real-time fault diagnosis.

Despite notable advancements in UAV fault diagnosis, three critical challenges remain unresolved: (1) many existing methods fail to adequately model the complex topological relationships among multi-source sensor data; (2) conventional approaches lack adaptive feature weighting mechanisms to handle varying fault patterns; and (3) the interpretability of diagnostic results is often insufficient for practical deployment.

In this paper, we propose a novel UAV fault diagnosis method based on Graph Neural Networks, specifically leveraging the GraphSAGE algorithm combined with a self-attention mechanism. Our approach begins by collecting vibration data from UAV sensors, which is then used to construct a graph representation of the system using the KNN algorithm. This graph captures the spatial and structural relationships among sensor nodes, enabling the effective modeling of fault propagation patterns. The GraphSAGE-SA algorithm is employed to learn node embeddings and perform fault classification, providing a scalable and inductive framework that generalizes well to unseen data.

The main contributions of this work are as follows:

1. We propose a novel self-attention enhanced GraphSAGE framework (GraphSAGE-SA) which, in contrast to standard GNNs and generic attention-GNN hybrids, is specifically designed for vibration-based fault diagnosis in UAVs. This framework enables adaptive and interpretable aggregation of neighboring information.

2. We introduce a KNN-based graph construction method that effectively models the topological relationships between UAV sensor nodes from 1D vibration signals, a step that has not been commonly addressed in previous GNN-based fault diagnosis studies.

3. Through extensive experimental evaluations, we demonstrate that our approach significantly outperforms both traditional deep learning models and standard GNN variants, achieving state-of-the-art accuracy on a real-world UAV vibration dataset.

The structure of this paper is as follows: “Theoretical background” section provides a review of related work on UAV fault diagnosis and graph-based learning techniques. “Methodology” section describes the proposed methodology, covering data collection, graph construction, and the GraphSAGE-based fault diagnosis framework. “Experiments” section presents the experimental setup, results, and discussions. Finally, “Conclusion and future work” section concludes the paper and outlines future research directions.

This study not only contributes to the advancement of UAV fault diagnosis but also demonstrates the potential of graph-based learning in improving the reliability and safety of complex systems. By addressing the limitations of traditional methods and harnessing the power of GNNs, our approach offers a robust and scalable solution for fault diagnosis in UAVs and other complex systems.

Theoretical background

Graph neural networks (GNNs)

Graph data serves as a fundamental representation method for capturing relationships, where nodes represent entities and edges denote the relationships between these entities. In graph data, the relational information between data points is as critical as the data itself. This type of graph data, based on non-Euclidean space, possesses a topological structure that mirrors the geometric arrangement of the data, making it highly effective for analyzing high-dimensional data²⁵.

In graph theory, a graph is typically represented by a set of nodes and edges, as shown in Eq. 1:

1

where represents the set of nodes in the graph, and denotes the feature vector associated with node. E refers to the set of edges that encode the relationships among nodes.

GNNs are a class of neural network architectures specifically developed to learn representations from graph-structured data.

The core concept behind GNNs is to iteratively update the representation of each node by aggregating the representations of its neighboring nodes along with its own representation²⁶. This mechanism facilitates the propagation of information throughout the graph, allowing each node to progressively incorporate information from its neighbors, thereby capturing the intricate structures and dependencies embedded within the graph. Through multiple iterations, GNNs can learn effective node representations, which can then be leveraged for a range of downstream tasks, such as node classification, graph classification, or link prediction. This process can be mathematically represented by Eq. 2

2

here are the weight matrix, which computed by the neural network, because we aim to assign different levels of importance to different nodes, thereby enabling more effective extraction of critical features. The function is responsible for collecting information from the neighboring nodes, while updates the representation of the current node by integrating its own features with the aggregated information from its neighbors.

GraphSAGE

As previously noted, traditional GNNs require access to the complete graph structure during training. Specifically, in each iteration, the representations of all nodes are synchronously updated. However, this paradigm becomes computationally prohibitive when dealing with large-scale graphs, as both memory consumption and processing time increase substantially, potentially rendering training infeasible. To address these scalability limitations, Hamilton et al. proposed an inductive framework known as GraphSAGE, which eliminates the need to compute embeddings for all nodes during training. Instead of operating over the full graph, GraphSAGE employs a neighborhood sampling strategy wherein a fixed-size subset of neighboring nodes is randomly selected, and their feature representations are aggregated to generate embeddings for target nodes, including those unseen during training. This inductive approach significantly improves training efficiency and model scalability, making it particularly suitable for large and evolving graphs commonly encountered in real-world applications²⁷. The principle of GraphSAGE can be depicted as follows:

For each node  in the graph, GraphSAGE samples a fixed-size subset of its neighboring nodes. This sampling is performed recursively over  hops, where  corresponds to the number of layers in the GraphSAGE architecture. To mitigate the computational complexity of processing the entire graph, GraphSAGE constrains the number of neighbors sampled at each layer. The sampled neighborhood of node  at the layer  is denoted as 

At each layer , GraphSAGE aggregates feature information from the sampled neighbors of node  using a designated aggregation function, denoted as AGGREGATE_k. This function can be selected from several alternatives, including:

• Mean aggregation Computes the element-wise mean of the features from the sampled neighbors.

• LSTM aggregation Employs a long short-term memory (LSTM) network to capture sequential dependencies during feature aggregation.

• Pooling aggregation Applies a neural network—typically a fully connected layer followed by a max-pooling operation—to extract and combine neighbor features.

The aggregated neighborhood features for node  at layer  are computed as:

where represents the feature vector of neighbor  from the previous layer.

Following aggregation, the node representation is updated by combining its own feature vector from the preceding layer with the aggregated neighborhood features. This is subsequently passed through a nonlinear transformation using an activation function (e.g., ReLU). The update rule is defined as:

where

•  is the learnable weight matrix at layer ,

• is the activation function,

•  denotes the concatenation operation.

Self-attention based aggregation approach

The impact of the same information varies across different classification outcomes, and the influence of different types of information on fault classification differs in magnitude. Therefore, it is essential to consider the varying weights of neighboring nodes when aggregating information, rather than simply averaging the information from all neighbors. The Transformer model, introduced by Vaswani in 2017, departs from traditional attention mechanisms by relying solely on attention mechanisms and excluding RNN-based models²⁸. The self-attention mechanism is a key component of the Transformer model, which is composed solely of self-attention layers and fully connected layers. This architecture enables the model to allocate greater weights to more significant information and facilitates the extraction of global features from time series data. The self-attention layer can replace the RNN layer due to the identical input and output dimensions. While both GraphSAGE and self-attention mechanisms have been studied independently, their integration in this work is specifically designed to address the challenges of UAV vibration signal analysis. Unlike standard attention-augmented GNNs, which often apply attention mechanisms in a generic manner, our GraphSAGE-SA model uses self-attention to adaptively assign weights to the importance of neighboring sensor nodes based on the vibrational context. This enables the model to focus on the most relevant fault propagation paths and ignore noisy or irrelevant connections, thereby enhancing both feature discriminability and model interpretability—a critical advantage for UAV maintenance and safety in real-world applications. The Transformer model employs three learnable weight matrices, , , and , to linearly transform the input sequence , resulting in the Key, Value, and Query matrices. The calculation formulas are as follows.

The output of the self-attention mechanism, denoted as , is computed using the following equation:

In this equation, represents the transpose of , and denotes the dimensionality of , which helps maintain the stability of gradient values during training.

For a given sequence, the self-attention mechanism calculates the dot product between the Query and Key, followed by normalization using the Softmax function to obtain the attention weights. These weights determine the importance of each element within the sequence. As a result, each entity is represented as a weighted sum of all entities in the sequence.

Methodology

Based upon the conventional GNN framework, the network architecture illustrated in Fig. 1 is proposed. The acquired one-dimensional signal is segmented into fixed-length intervals to generate multiple samples of equal length. Following the principles of KNN-based graph construction, the K nearest neighbors are identified for each node. The edge weights between nodes are determined using a combination of Gaussian and weighting functions.

Fig. 1.

Fault diagnosis framework. (a) Vibration Data Acquisition, (b) Data segmentation, (c) Graph Construction using KNN, (d) GNN Processing with GraphSAGE and Self-Attention, (e) Fault Classification.

The edge weight between nodes i and j is calculated using a Gaussian kernel function based on the Euclidean distance between their feature vectors:

where and are the feature vectors of nodes and , respectively, and is a scaling parameter that controls the width of the kernel function.

The GraphSAGE model is utilized for neighborhood sampling and feature aggregation. The architecture of the proposed model is depicted in Fig. 1. The GraphSAGE-SA model comprises two graph convolutional layers, with each layer sampling 10 neighbors per node. The dimensionality of the node embeddings is set to 64 in the first layer and 32 in the second layer. A ReLU activation function is applied following each graph convolutional layer.

Furthermore, a self-attention mechanism is incorporated to adaptively learn the attention weights between each node and its neighbors, enabling more accurate aggregation operations. This enhances the model’s ability to capture the diverse relationships between a node and its neighbors, thereby generating more discriminative node representations. The proposed model is applied to fault diagnosis of UAVs.

Experiments

Platform

To evaluate the effectiveness of the proposed fault diagnosis framework, experimental studies were conducted using a quadrotor unmanned aerial vehicle (QUAV) platform assembled at ShanDong University of Aeronautics, as illustrated in Fig. 2. Actual operational fault data were collected to assess the diagnostic performance of the proposed methodology, and the detailed configuration of the UAV is presented in Table 1.

Fig. 2.

UAV platform.

Table 1.

Configuration of UAV.

Parameters	Descriptions
Wheelbase	450 mm
Controller	Pixhawk2.4.8
Battery	4800 mAh (3S)
Weight	1200 g
Motor	2212
Propeller	1045

The experimental hardware environment comprises an Intel (R) Core (TM) i9-7800 CPU operating at 3.60 GHz and 16 GB of RAM. The software environment includes the required platform and operating system for executing the experiments. Model implementation was carried out using Python 3.9 on a Windows 10 Professional operating system.

Data acquisition system

This study presents the design and implementation of a vibration data acquisition system based on the STM32 microcontroller and the MPU6050 sensor. As shown in Fig. 3, the schematic and physical structure of the system are illustrated. The system enables real-time acquisition and storage of vibration data on an SD card, providing a reliable data source for subsequent analysis. The STM32 microcontroller employed in this study is the STM32F103C8T6 from STMicroelectronics, which is built on the ARM Cortex-M3 core and operates at a frequency of 72 MHz. It offers a comprehensive set of on-chip resources, including flash memory, SRAM, timers, and serial interfaces, sufficient to meet the functional requirements of the system²⁹. The MPU6050 is a six-axis motion processing sensor that integrates a three-axis gyroscope and a three-axis accelerometer, allowing for simultaneous measurement of angular velocity and acceleration. It communicates with the STM32 via the I²C interface and is characterized by its compact size, low power consumption, and high precision. The SD card storage module interacts with the STM32 through the SPI interface and is responsible for storing the acquired vibration data. The FAT32 file system is employed to manage the SD card, facilitating efficient data access and analysis.

Fig. 3.

Data acquisition system.

Data collection

The collected experimental data comprise four operating conditions: the healthy state, loosening of the screw connecting the arm base and the upper plate as shown in Fig. 4a, loosening of the motor base screw as shown in Fig. 4b, and blade damage as illustrated in Fig. 4b.

Fig. 4.

Fault state: (a) loosening of the arm screw; (b) loosening of the motor screw; (c) the damaged propeller.

The data corresponding to each fault condition were collected while the UAV was in an idle state, ensuring consistency of environmental conditions during acquisition. Screw loosening was introduced by rotating the screw approximately 1.5 turns from the fully tightened position, while propeller damage was simulated by removing approximately 15% of the blade tip area. Vibration signals were sampled at a frequency of 100 Hz. For each of the four conditions (healthy and three fault types), 1000 samples were obtained, with each sample comprising a signal segment containing 1024 data points. The entire dataset was randomly divided into training (70%), validation (20%), and testing (10%) subsets, with strict control to prevent data leakage across partitions.

It is worth noting that the present study focuses on structural and propeller-related faults as a critical initial step. Motor faults, recognized as another prevalent and significant failure mode in UAVs, will be addressed in future research. The inclusion of motor fault data under dynamic flight conditions is considered an essential direction for further validating the practical robustness of the proposed model.

Data preprocessing

The statistical features of vibration signals, such as mean, root mean square error (RMSE), and peak-peak value (PP), typically vary under different operating conditions³⁰. Among these, RMSE is commonly employed to detect structural damage by quantifying the deviation between healthy and faulty states, and has been extensively utilized in damage detection applications³¹. RMSE measures the discrepancy between predicted values and actual measurements, and is defined as follows:

where and represent the predicted and measured values, respectively (Fig. 5).

Fig. 5.

The comparison of RMSE. Figure 5 depicted the RMSE values of the three axes across the four states.

The growth rate (GR) refers to the ratio of the increase in a specific data index over a given period to the value of that index in the base period. The basic formula for GR is defined as follows:

The GR can more intuitively reflect the change of a certain feature.

It can be observed that damage induces more pronounced changes in the RMSE along the Y-axis compared to the other two axes, with the GR of RMSE reaching up to 166.6% (Table 2). Therefore, in the subsequent analysis, vibration signals in the Y-axis direction were selected to construct the graph structure as input to the model.

Table 2.

The GR of RMSE in different directions.

RMSE	Normal	Case 1	Case 2	Case 3	GR of RMSE (case 1) (%)	GR of RMSE (case 2) (%)	GR of RMSE (case 3) (%)
X-axis	0.005	0.008	0.006	0.007	60.00	20.00	40.00
Y-axis	0.0045	0.01	0.012	0.0095	122	166.67	111.11
Z-axis	0.0028	0.003	0.0029	0.0035	7.14	3.57	25.00

Results and analysis

Fault diagnosis based on GraphSAGE

The influence of K value selection

The selection of the parameter K in the KNN algorithm plays a critical role in the construction of the graph structure, which subsequently influences the performance of the GNN in fault diagnosis tasks. Experimental results indicate that varying the value of K leads to distinct graph topologies, thereby affecting the model’s capacity to capture underlying patterns in one-dimensional vibration signals. Specifically, a smaller K value tends to generate a sparse graph, which may omit important relationships between data points, resulting in reduced diagnostic accuracy. In contrast, a larger K value produces a denser graph, potentially introducing noise and redundant connections that could degrade model performance. The optimal K value achieves a balance between sparsity and connectivity, enabling the graph to effectively represent the intrinsic structure of the data while minimizing the influence of irrelevant connections. This balance is essential for achieving high diagnostic accuracy and robustness in GNN-based fault diagnosis systems. The optimal K value was determined via a grid search over K = {1, …, 9, 10}. The classification accuracy under differernt K values is presented in Fig. 6. Based on the validation results, K = 3 was selected for the final experiments, as it yielded the highest fault classification accuracy while effectively balancing graph connectivity and sparsity to capture the most relevant structural relationships in the vibration data.

Fig. 6.

The k-value affection for accuracy.

The choice of aggregation method

The choice of aggregation method in GraphSAGE is pivotal to the effectiveness of the GNN in fault diagnosis applications. Experimental results demonstrate that different aggregation strategies—such as mean, LSTM, and pooling—exhibit varying capabilities in capturing structural and feature-level information from the graph. The mean aggregator, while offering high computational efficiency, tends to oversmooth local variations, potentially suppressing critical diagnostic features in one-dimensional vibration signals. In contrast, the LSTM aggregator, equipped with sequential processing capabilities, excels at modeling complex dependencies and temporal patterns, albeit with increased computational cost. The pooling aggregator provides a compromise between efficiency and representational power, leveraging its selective aggregation mechanism to emphasize salient features. The optimal choice of aggregator depends on the characteristics of the dataset and the specific diagnostic requirements. Findings suggest that the LSTM aggregator generally yields superior performance when handling datasets with intricate temporal dependencies, whereas the pooling aggregator is better suited for scenarios demanding a trade-off between accuracy and computational overhead. These observations underscore the critical role of aggregator selection in enhancing the diagnostic capability of GraphSAGE for fault diagnosis tasks.

As illustrated in Fig. 7, the proposed GraphSAGE-SA model achieves a final classification accuracy of approximately 98%, outperforming the baseline GraphSAGE model with mean aggregation, which attains an accuracy of around 93%, by a notable margin of 5 percentage points. This result highlights the substantial enhancement in feature extraction capability introduced by the integrated self-attention mechanism. Traditional methods, including CNN and SVM, were also included in the comparative experiments.

Fig. 7.

The aggregation method affection.

To visually evaluate the classification performance of the models, the confusion matrix is widely recognized as an effective statistical and visualization tool in data mining and analysis³².

For each operational condition, the test set consists of 100 samples, and the confusion matrix of the diagnostic results is presented in Fig. 8. The conventional graph neural network based on GraphSAGE achieves a maximum diagnostic accuracy of 95% and a minimum of 86%. With the integration of the attention mechanism, the accuracy improves significantly, reaching a maximum of 100% and a minimum of 98%. The diagnostic approach proposed in this study demonstrates a clear advantage over traditional methods, exhibiting both the highest fault classification accuracy and superior stability. These results indicate that the incorporation of the attention mechanism enhances the extraction of fault-related features, thereby improving the overall performance of the model.

Fig. 8.

The confusion matrix of UAV fault diagnosis.

Robustness and stability analysis

The diagnostic reliability of GraphSAGE-SA is theoretically supported by the principles of graph smoothness and Lipschitz stability. Given the construction of a KNN graph with Gaussian kernel weighting, the graph Laplacian ensures that minor perturbations in node features induce only limited variations in node embeddings. In addition, the incorporation of a self-attention mechanism enhances model robustness by assigning greater weights to relevant neighbors, thereby effectively suppressing noisy or irrelevant connections. Formally, the attention-based aggregation satisfies a form of Lipschitz continuity in the feature space, which contributes to consistent fault identification under varying levels of noise. The experimental results, which consistently achieve accuracy exceeding 98% across all fault categories and data partitions, provide empirical validation of this theoretical stability.

Conclusion and future work

This study proposes a self-attention enhanced GraphSAGE framework (GraphSAGE-SA) for intelligent fault diagnosis of UAVs using vibration signals. The core of the approach lies in transforming one-dimensional vibration data into graph-structured representations via a KNN algorithm, thereby explicitly capturing the topological relationships among sensor nodes. By integrating a self-attention mechanism into the GraphSAGE architecture, the model performs adaptive, non-local aggregation of neighborhood information, enabling dynamic focus on the most relevant fault-related features while effectively suppressing noise.

The proposed framework was rigorously evaluated on a real-world quadrotor UAV platform under various structural and propeller fault conditions. Experimental results demonstrate that GraphSAGE-SA achieves a high fault classification accuracy of 98%, substantially outperforming conventional GraphSAGE variants (by approximately 5%), other graph-based models such as GCN and GAT, as well as traditional deep learning baselines including CNN and LSTM. These results confirm the effectiveness of combining graph-based structural learning with attention-driven feature refinement for complex vibration signal analysis, offering a novel and robust solution to the challenges of feature extraction and dependency modeling in UAV fault diagnosis.

Despite these promising results, this study has several limitations that outline directions for future research:

• (i)Fault and Operational Scenarios: The current validation was performed under a limited set of structural faults and in an idle state. Future work should incorporate critical motor faults (e.g., bearing wear, winding imbalances) and evaluate the model under dynamic flight conditions with real aerodynamic loads to assess its practical robustness and generalizability.
• (ii)Quantitative Fault Severity: While fault conditions were introduced, a more precise quantification of fault severity (e.g., exact torque values for loosening, progressive levels of propeller damage) is necessary to systematically evaluate the model’s sensitivity and its potential for prognostics.
• (iii)Platform Generalizability: The model was validated on a single UAV platform. Investigating its performance across necessary UAV models, sizes, and configurations is crucial to verify its broad applicability.

Accordingly, our future work will focus on three primary directions:

• (i)Expanding fault libraries and operational envelopes to include motor faults and in-flight data collection.
• (ii)Exploring model efficiency and deployment through model compression techniques (e.g., pruning, quantization) to enable real-time inference on resource-constrained UAV embedded systems.
• (iii)Enhancing generalizability through transfer learning and domain adaptation techniques, with the goal of developing a diagnosis system capable of quickly adapting to new UAV platforms or changing operational environments without extensive retraining.

In summary, this study not only proposes a high-accuracy fault diagnosis framework for UAVs but also establishes a graph-based paradigm that can be extended to the intelligent monitoring of other complex mechanical systems. By continuously addressing the identified limitations, the aim is to transition the proposed technology from laboratory validation to real-world deployment, thereby enhancing the safety, reliability, and autonomy of UAV operations.

Author contributions

Y.M. and H.W. conceived the experiment(s), Y.S. and B.L. conducted the experiment(s), L.C. and Y.M. analysed the results.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.