Multiaxial vibration data for blade fault diagnosis in multirotor unmanned aerial vehicles

Abstract

This dataset presents multiaxial vibration signals collected from a multirotor unmanned aerial vehicle (UAV) operating in hover mode for the purpose of blade fault diagnosis. Vibration measurements were recorded at the geometric center of the UAV, where the centerlines of the four rotor arms intersect, using a triaxial accelerometer. The dataset captures variations across the X, Y, and Z axes under different blade fault conditions, including healthy, minor imbalance, severe imbalance, and screw loosening scenarios. Each flight scenario was repeated under controlled conditions to ensure consistency and high-quality labeling. The resulting soft-labeled dataset includes time-domain signals from numerous test flights and has been used in multiple prior studies involving classical and deep learning-based fault classification techniques. This curated data collection provides a valuable resource for researchers in UAV health monitoring, vibration analysis, and machine learning-based fault diagnosis. The dataset is particularly useful for the development and benchmarking of signal processing pipelines and classification models aimed at identifying blade-level faults in multirotor UAV systems.

Background & Summary

Unmanned aerial vehicles (UAVs) are increasingly integrated into modern applications across diverse fields, including logistics¹, surveillance², research³, and recreation⁴. As UAV deployment scales, the importance of real-time fault diagnosis systems grows in parallel to ensure operational safety in work efficiency and general reliability^5,6,7,8. Among the various components vulnerable to malfunction, propeller blades are particularly critical; their imbalances or structural anomalies can cause flight instability or failure. Vibration-based monitoring, especially when paired with Artificial Intelligence (AI) intelligent algorithms, offers a promising approach to identify such faults preemptively. With the advancement of onboard sensing, data acquisition, and AI-enabled processing, the development and public availability of curated vibration datasets is essential to accelerate research and standardize evaluation in UAV health diagnostics.

The dataset described in this paper has been extensively utilized in previously published works. In the first study, a structured protocol for UAV fault diagnosis was developed, where the dataset was used to demonstrate the workflow for collecting, preprocessing, and analyzing vibration signals through machine learning (ML)⁹. In the second study, multi-resolution transform features were extracted and fed into a deep neural network for fault classification¹⁰. In the third, an unbalance classification method was implemented using an optimized stochastic gradient descent logistic regression model¹¹. In the fourth study, feature selection using the ReliefF method was applied prior to fault classification using variant ML models, namely SVM and kNN models¹². In the fifth, signal filtering was compared with multiresolution analysis methods to assess their performance in fault detection¹³. In the sixth, finite element analysis was conducted alongside experimental validation using the dataset to study the impact of consumed propellers on UAV airworthiness¹⁴. In the seventh study, frequency-domain vibration signal analysis was applied on the dataset’s time-domain signals to classify unbalance faults¹⁵. Finally, in the eighth study, discrete wavelet and fast Fourier transforms were combined to enhance signal decomposition and improve fault identification¹⁶. The diverse applications and methodologies employed across these studies are summarized in Table 1.

Table 1 Summary of prior studies using the dataset.

This repeated application across eight peer-reviewed studies highlights the dataset’s robustness, consistency, and value across different domains. This ranged from purely analytical signal transformations to advanced AI-based deep learning classifiers. The dataset is particularly effective in capturing and distinguishing multiaxial vibration responses of multirotor UAVs under various propeller fault conditions which makes it a versatile and reusable resource for researchers, educators, and developers of embedded health monitoring systems. Its availability encourages continued experimentation, benchmarking, and innovation in UAV fault diagnosis. In prior studies utilizing this dataset, both time-domain statistical features (e.g., RMS, standard deviation) and frequency-domain features derived from FFT or wavelet transforms have been used for fault classification. The dataset’s proven utility across multiple published studies ensures it can serve as a foundation for future advancements in the field.

Recent advancements in UAV diagnostics have highlighted the benefits of incorporating richer sensor modalities, real-time processing, and embedded implementations. For instance, Kim et al. demonstrated effective fault localization and severity estimation in multicopters using accelerometer and gyroscope data with PCA and ANN/SVM models¹⁷. Yaman et al. proposed an embedded SVM-based fault detection method optimized for low-cost onboard implementation¹⁸, while Baldini et al. employed vibration analysis to enable real-time propeller fault detection under operational conditions¹⁹. In a more hardware-centric approach, Sadhu et al. implemented onboard deep learning fault classification using FPGA-based systems for real-time diagnosis²⁰. Additionally, the predictive scope of UAV reliability has been extended to battery performance, as shown by Al-Haddad et al., who utilized deep neural networks for forecasting energy consumption and efficiency degradation trends in UAV power systems²¹. Collectively, these works underscore the relevance of curated datasets like ours, which provide high-quality, reusable vibration data that can serve as a diagnostic benchmark, simulation input, or model training foundation within a rapidly evolving UAV research landscape.

Methods

This section is to detail the adopted methodology, the used tools, and the experimental setup for the acquired dataset.

UAV model and tools

In this study, a DJI Mini 2 Combo quadcopter was used as the aerial platform for vibration data collection. This compact UAV offers reliable stability in hover mode and is suitable for replicable laboratory experiments. The key specifications of the UAV are listed in Table 2.

Table 2 Specifications of the DJI Mini 2 Combo UAV.

To capture the vibration signals, an ADXL335 triaxial analog accelerometer was mounted at the geometric center of the UAV—specifically at the point where the four motor arms intersect. This placement allows equal sensitivity to all axis-aligned vibrations resulting from blade imbalances. The ADXL335’s specifications are provided in Table 3.

Table 3 Specifications of the ADXL335 accelerometer.

The accelerometer was interfaced with a National Instruments DAQ-6008 USB data acquisition device, which relayed signals to the PC. Specifications of the DAQ system are shown in Table 4.

Table 4 Specifications of the NI USB-6008 Data Acquisition System.

Sensor calibration was carried out by subtracting the mean from each raw signal in a continuous loop within LabVIEW. This eliminated initial signal drift and ensured zero-centered vibration recordings. The calibration method ensured smooth acquisition and alignment across all three axes. The LabVIEW block diagram representing this calibration logic is illustrated in Fig. 1.

Fig. 1

LabVIEW block diagram used for real-time signal calibration.

It is noted that only external accelerometer data are included in this dataset, while onboard gyroscope and magnetometer readings were intentionally excluded to maintain a focused scope on time-domain vibration signals suitable for frequency-domain and time-frequency domain transformations.

Fault initiation and flight conditions

All vibration recordings were acquired while the UAV operated in hover mode at a fixed altitude of 1.2 meters. Two primary fault types were introduced to simulate propeller-level imbalances:

• Mass Removal Fault: A small section of the blade was trimmed to simulate mechanical wear or damage.

• Weight Unbalance Fault: Small adhesive sticky tape was affixed to one side of a blade to induce mass asymmetry.

Each type of fault was implemented independently on individual blades. For reference, blade numbering follows a clockwise pattern as viewed from above: Blade 1 (lower right), Blade 2 (upper right), Blade 3 (upper left), and Blade 4 (lower left). The fault configurations and blade numbering scheme are presented in Fig. 2.

Fig. 2

Blade numbering and fault application methods (mass removal and weight unbalance).

Each flight session lasted 7 minutes (420 seconds), during which approximately 400,000 vibration samples were collected per case. All experiments were conducted indoors in a controlled environment with no wind interference and stable room temperature, ensuring consistency and high-quality data across recordings.

This dataset is limited to a single UAV model operated at a fixed altitude under windless indoor conditions, which should be considered when evaluating its applicability to real-world scenarios. It is also acknowledged that the dataset is limited to hover-mode data collected in a controlled indoor setting, and future versions may incorporate dynamic flight conditions and environmental variations to enhance real-world applicability.

Control input data such as roll, pitch, and thrust commands were not recorded in this study, as the UAV was operated in a steady hover mode with minimal control variation; however, their integration will be explored in future datasets involving dynamic flight conditions.

Data acquisition and preprocessing

The full experimental workflow is summarized in Fig. 3, which shows the step-by-step process from UAV flight setup and sensor mounting to data recording, labeling, and storage.

Fig. 3

Graphical methodology outlining the data acquisition and preprocessing workflow.

The actual experimental setup including the UAV, DAQ system, sensors, and PC platform is visually represented in Fig. 4.

Fig. 4

Experimental setup showing the UAV, DAQ device, accelerometer, and PC platform.

Vibration signals were recorded directly in the time domain and stored as Excel (.xlsx) files on the PC platform. To ensure usability and clarity, each signal file was manually labeled by the experimenter, indicating the type and location of the fault. For each flight condition—healthy, mass removal on one blade, and weight unbalance on another—recordings were repeated five times, resulting in a comprehensive and reliable dataset spanning five healthy and five faulty scenarios.

Data Records

The dataset is available at Figshare²². The complete dataset consists of five Excel files, each corresponding to a specific flight condition of the UAV, including both healthy and faulty cases. These cases involve controlled modifications to the UAV’s propeller blades, such as mass removal (to simulate damage) and weight addition using adhesive tape (to induce imbalance). The dataset was collected under repeatable indoor hover flight conditions, using a centrally mounted ADXL335 triaxial accelerometer connected to a NI USB-6008 DAQ system, as described in Section 3. Each file contains approximately 400,000 rows of raw, high-resolution vibration signals recorded continuously over a 7-minute period. The files are named clearly to reflect the specific fault case they represent:

• Healthy

• Damaged Bottom Right Blade

• Damaged Top Right Blade

• Unbalanced Bottom Right Blade

• Unbalanced Top Right Blade

Each Excel file includes four columns, structured as follows:

• Column A: Time (in seconds)

• Column B: X-axis acceleration (m/s²)

• Column C: Y-axis acceleration (m/s²)

• Column D: Z-axis acceleration (m/s²)

Manual labeling was used to ensure maximum clarity, with filenames explicitly indicating the fault condition and blade position. All data are recorded in the time domain, suitable for transformation into frequency or time-frequency domains depending on user needs. These characteristics make the dataset especially useful for UAV diagnostics, machine learning classification tasks, and vibration signal analysis. Figure 5 details the data records.

Fig. 5

Structure and contents of the dataset files. Each Excel file includes vibration data recorded across the X, Y, and Z axes, sampled over 400,000 instances per flight condition.

Technical Validation

To ensure the reliability, quality, and scientific usability of the presented vibration dataset, two independent validation techniques were employed. These methods served to verify the clarity of the signal response, confirm accurate labeling, and assess noise consistency throughout the recordings. Both approaches provided robust support for the technical validity of the dataset in the context of blade fault diagnosis in multirotor UAVs.

Classification-based validation using additional experimental data

In the first validation approach, an external test set comprising newly collected vibration signals—generated exclusively under blade damage scenarios involving mass removal—was used. These unseen samples were tested against machine learning classifiers originally trained on the main dataset. The objective was to assess whether the trained models could reliably generalize and correctly identify the same type of faults based on real-world vibrations not present during training. The results were found to be highly consistent with the expected outputs, confirming both the clarity of the measured signals and the strength of the class separability. Notably, the classifier accurately identified all fault instances with minimal false predictions. The classification achieved an accuracy, precision, and recall of 0.91, with a specificity of 0.978 and an area under the ROC curve (AUC) of 0.987 which confirms the reliability of the trained model. These results are visually summarized in Fig. 6, where each classified sample is mapped and annotated.

Fig. 6

Classification results of unseen damaged blade instances using pre-trained ML models.

Frequency-domain comparison of blade conditions

To complement the time-domain and time-frequency validation methods, a frequency-domain analysis was conducted using the Fast Fourier Transform (FFT) on the X-axis vibration signals from all five flight conditions. This transformation enables visualization of the signal energy distribution across frequencies, highlighting the spectral differences caused by blade faults. As illustrated in Fig. 7, the healthy UAV condition presents a relatively uniform low-frequency signature, whereas the damaged and unbalanced blade scenarios exhibit noticeable peaks and variations, indicating fault-induced harmonic content. These spectral characteristics offer an additional diagnostic dimension for future researchers and model developers, supporting more informed preprocessing strategies and feature extraction methods. The figure serves as a visual reference to underscore the separability and richness of the dataset in the frequency domain.

Fig. 7

FFT plots of X-axis vibration signals for all UAV blade conditions, illustrating frequency-domain differences between healthy, damaged, and unbalanced states.

Multi-resolution analysis using discrete wavelet transform

The second validation technique relied on a six-level Discrete Wavelet Transform (DWT) to examine the signal’s frequency content and identify any latent noise or inconsistency in the recordings. The Daubechies (db4) wavelet was used for decomposition due to its suitability for vibration signal analysis and effective time-frequency localization. Healthy and damaged blade signals were decomposed across multiple levels to isolate high- and low-frequency components. This analysis enabled evaluation of noise presence, the clarity of signal transitions, and determination of the most suitable decomposition level for feature extraction. The quality of the decomposition was remarkable, and no significant noise distortions were observed across any level, demonstrating the integrity of the original measurements. Additionally, a fusion of selected decomposition levels was tested to simulate hybrid signal processing strategies, which yielded superior signal clarity and class distinctiveness. Calibration of the accelerometer sensor also played a critical role in achieving consistent and artifact-free recordings, ensuring the dataset is ready for further transformation into either frequency-domain or time-frequency domain representations with minimal preprocessing effort. Figure 8 illustrates a side-by-side wavelet decomposition of healthy and damaged blade signals, highlighting the smooth transitions and well-preserved waveform structure across all levels.

Fig. 8

Six-level wavelet decomposition of vibration signals for healthy vs. damaged blades.

Usage Notes

The dataset is compatible with a variety of ML platforms and can be easily processed using both traditional and deep learning techniques. The current study particularly recommends the use of Orange Data Mining²³, an open-source data visualization and analysis tool, for rapid prototyping of classification pipelines^24,25. Orange allows users to create workflow-based data analysis structures with minimal coding experience as it is especially suitable for researchers and educators working in topics of fault diagnosis, specifically UAV fault detection.

Figure 9 illustrates an example Orange workflow used for classification in a previously published study based on this dataset. The visual workflow provides a clear representation of how features were selected, processed, and classified and generally serves as a helpful reference for future reuse and experimentation.

Fig. 9

Orange data mining prediction structural workflow utilizing a simple neural network¹⁰.