In this study, we address the critical need for airworthiness verification and safety monitoring of civilian fixed-wing drones by proposing a flight maneuver recognition method based on Support Vector Machines (SVM). The increasing deployment of fixed-wing drones in various sectors demands robust frameworks to ensure they operate within certified flight envelopes. Our work systematically extracts maneuver recognition requirements from existing UAV-specific conditions, defines maneuver categories and their trajectory characteristics, and constructs a six-dimensional feature pattern. By optimizing the SVM model with Bayesian optimization, we achieve high accuracy and efficiency, meeting the stringent demands of airworthiness validation.
Introduction
The combination of flight maneuver types constitutes the operational profile of fixed-wing drones, and accurate recognition of these maneuvers is a prerequisite for airworthiness verification and safety monitoring. Existing research in flight maneuver recognition primarily falls into traditional methods and deep learning approaches. Traditional methods, such as parameter analysis and pattern matching, have been applied successfully but often require extensive domain knowledge. Deep learning methods, including LSTM and CNN, offer superior temporal modeling but demand large labeled datasets and significant computational resources, which are often unavailable in the airworthiness domain where data is limited. For fixed-wing drones, the small-sample nature of certification data (typically on the order of 10³ samples) makes SVM particularly attractive due to its strong generalization capability under limited data conditions. Moreover, the decision boundary of SVM can be physically interpreted as flight envelope parameters, aligning well with the requirements of airworthiness authorities. Therefore, we select SVM as the core algorithm for this study, systematically analyzing UAV-specific conditions to establish maneuver features and constructing a robust recognition model for fixed-wing drones.
Airworthiness-Oriented Flight Maneuver Recognition Framework
Currently, airworthiness requirements for UAVs are primarily defined through specific conditions following a “one model, one standard” principle. Through comprehensive analysis of several type-certificated UAV special conditions, we extracted key requirements covering initial airworthiness (flight control requirements, flight envelope design, performance design) and continued airworthiness (flight performance monitoring, flight envelope management). The recognition of flight maneuvers for fixed-wing drones plays a central role in these requirements: it supports flight envelope design, performance verification, and safety boundary setting during initial certification, and enables real-time monitoring, anomaly warning, and envelope optimization during continued airworthiness.
The basic flight maneuvers of fixed-wing drones can be categorized into five types: climb, level flight, turn, roll, and descent. These maneuvers and their combinations constitute the mission flight profile. Table 1 summarizes the maneuver categories, their key metric features, and the corresponding airworthiness verification objectives. Each maneuver must satisfy specific constraints to ensure compliance. For example, for climb maneuvers, the vertical speed must not exceed design limits, the pitch angle should be between 5° and 15°, and the roll angle should be near zero. For level flight, altitude deviation must be within ±5 m, airspeed deviation within ±5% of the target, and both roll and pitch angles near zero.
| Maneuver | Key Metrics | Airworthiness Verification Objective |
|---|---|---|
| Climb | Altitude, vertical speed, airspeed, heading, pitch, roll | Ensure maximum climb rate and envelope compliance |
| Level flight | Altitude, airspeed, heading, roll, pitch | Verify cruise stability |
| Turn | Altitude, airspeed, heading, roll | Validate coordinated turn bank angle limit (≤30°) |
| Roll | Altitude, airspeed, heading, roll, pitch | Test maneuver load factor limits |
| Descent | Altitude, vertical speed, airspeed, heading, pitch, roll | Confirm approach trajectory stability and landing configuration maintenance |
Based on the analysis of special conditions and the features defined in Table 1, we constructed a systematic airworthiness-oriented flight maneuver recognition framework for fixed-wing drones. The framework involves real-time data acquisition from onboard sensors (altitude, airspeed, heading, pitch, roll, etc.), feature extraction according to Table 1, data cleaning and normalization, SVM-based classification, real-time comparison with airworthiness clauses, safety analysis, and feedback loop for flight control optimization. This closed-loop system ensures that fixed-wing drones operate within certified limits.
SVM-Based Flight Maneuver Recognition Model for Fixed-wing Drones
Sample Set Construction
We obtained experimental data from the public PX4 log dataset available on the Flight Review data-sharing platform. All flight logs contain pre-labeled maneuver scoring tags and were sampled at 100 Hz. From the raw data, we extracted six key flight parameters: altitude (m), vertical speed (m/s), airspeed (m/s), heading angle (rad), pitch angle (rad), and roll angle (rad). These parameters comprehensively describe the motion state of fixed-wing drones. Considering the high correlation between turn and roll maneuvers in actual flight, we merged the five basic maneuvers into four representative categories: climb (label 1), level flight (label 2), turn (label 3), and descent (label 4). In the data preprocessing stage, we applied the Bollinger Bands algorithm to clean outliers. For each sliding window, we computed the moving average (MA) and standard deviation (σ), then defined the upper and lower bounds as:
$$ \text{UB}_t = \text{MA}_t + 2\sigma_t $$
$$ \text{LB}_t = \text{MA}_t – 2\sigma_t $$
Data points falling outside these bounds were considered anomalous and removed. For example, Figure 3 in the original paper shows the effect of applying Bollinger Bands on the vertical speed data for a climb maneuver; we observed a significant reduction in spurious spikes. After removing stationary-state data and abnormal fluctuations, we obtained 15,557 valid six-dimensional feature vectors. Finally, we performed min-max normalization to eliminate dimensional differences:
$$ x_{\text{norm}} = \frac{x – x_{\text{min}}}{x_{\text{max}} – x_{\text{min}}} $$
Kernel Function Selection
The choice of kernel function directly affects the performance of SVM for fixed-wing drones maneuver recognition. We tested four kernel functions: linear, polynomial, RBF, and sigmoid. The experimental results on our dataset are summarized in Table 2.
| Kernel Function | Classification Accuracy (%) | Prediction Time (s) |
|---|---|---|
| Linear | 71.80 | 1.19 |
| Polynomial | 72.51 | 1.16 |
| RBF | 86.59 | 4.98 |
| Sigmoid | 72.85 | 1.61 |
The RBF kernel achieved the highest accuracy of 86.59%, outperforming the linear kernel by 14.79%, polynomial by 14.08%, and sigmoid by 13.74%. Although its prediction time is slightly longer, the accuracy gain is critical for airworthiness applications. Previous research indicates that SVM with RBF kernel automatically determines thresholds and central weights, reducing the upper bound of expected test error. Therefore, we selected the RBF kernel for our fixed-wing drones maneuver recognition model:
$$ K(\mathbf{x}_i, \mathbf{x}_j) = \exp\left(-\gamma \|\mathbf{x}_i – \mathbf{x}_j\|^2\right) $$
where γ is the kernel width parameter to be optimized.
Model Optimization via Bayesian Optimization Algorithm
The Bayesian Optimization Algorithm (BOA) is a systematic hyperparameter tuning method particularly suitable for machine learning models with limited evaluation budgets. For our SVM model, BOA constructs a probabilistic surrogate model using Gaussian Processes, dynamically updates the prior distribution based on historical parameter-performance pairs, and intelligently selects the next evaluation point using an acquisition function to efficiently locate the global optimum. The parameter tuning process for SVM is illustrated conceptually in Figure 4 of the original paper. We set the search ranges for the penalty parameter C and kernel parameter γ as [1, 100] and [0.1, 1000] respectively, and performed 50 iterations with 10-fold cross-validation. The BOA convergence curve shows that the algorithm stabilizes after only 15 iterations, demonstrating fast convergence. The optimal parameters found were:
$$ C_{\text{opt}} = 32.28, \quad \gamma_{\text{opt}} = 2.15 $$
Using these optimized parameters, the SVM model achieved a significant improvement in classification accuracy on the test set—from 86.6% (default parameters) to 98.3%, representing a 12% increase. This demonstrates the effectiveness of Bayesian optimization for tuning the SVM model tailored to fixed-wing drones flight maneuver recognition.
Reliability Analysis of the Flight Maneuver Recognition Model
We evaluated the optimized BO-SVM model on the held-out test set. The prediction runtime was 0.91 seconds, well within real-time monitoring requirements. The overall prediction accuracy reached 98.29%, as shown in the confusion matrix (Figure 6 in the original paper). The matrix indicates that the recognition accuracy for each maneuver class exceeded 97%, with the highest accuracy for level flight and descent maneuvers. Specifically, the precision, recall, and F1-score for each class are listed in Table 3.
| Model | C | γ | Accuracy (%) | Macro Precision (%) | Macro Recall (%) | Macro F1 (%) |
|---|---|---|---|---|---|---|
| BO-SVM | 32.28 | 2.15 | 98.3 | 98.2 | 98.5 | 98.4 |
| SVM (default) | 1 | 0.25 | 86.6 | 89.0 | 83.1 | 84.7 |
The optimized BO-SVM model demonstrates substantial improvements: recall increased from 83.1% to 98.5%, indicating a dramatic reduction in missed detections of dangerous maneuvers. This is critical for safety monitoring of fixed-wing drones. Furthermore, we compared BO-SVM against deep learning models (LSTM and CNN-BiLSTM) under the same dataset and hardware environment (Intel i9-13980HX). Table 4 summarizes the results.
| Evaluation Dimension | BO-SVM | CNN-BiLSTM | LSTM |
|---|---|---|---|
| Accuracy (%) | 98.29 | 93.91 | 76.62 |
| Training time (s) | 32 | 312 | 228 |
| Prediction time (s) | 0.91 | 15.2 | 12.7 |
| Interpretability | Support vector visualization | Weak (black-box) | Not interpretable |
| Data required for convergence | ~3,000 | ~12,000 | ~10,500 |
BO-SVM outperforms both deep learning models by a significant margin in accuracy, while requiring only 32 seconds for training (9.7 times faster than CNN-BiLSTM) and 0.91 seconds for prediction (16.7 times faster). Moreover, BO-SVM provides interpretability through visualization of support vectors, which is highly desirable for airworthiness certification where decisions must be justified. The memory footprint of BO-SVM is only 48 MB, compared to 620 MB for CNN-BiLSTM, reducing hardware cost and energy consumption for onboard deployment of fixed-wing drones.
Conclusion
In this work, we proposed a flight maneuver recognition method based on Support Vector Machines tailored for the airworthiness verification and safety monitoring of civilian fixed-wing drones. By analyzing UAV-specific conditions, we identified key requirements and defined a six-dimensional feature pattern consisting of indicated airspeed, pitch angle, roll angle, yaw angle, altitude, and altitude change rate. The SVM model was optimized using Bayesian optimization, achieving a test accuracy of 98.29% with a prediction time of 0.91 seconds. Comparative experiments demonstrated that our BO-SVM model significantly outperforms both default SVM and deep learning approaches (LSTM, CNN-BiLSTM) in terms of accuracy, training efficiency, inference speed, interpretability, and resource consumption. The proposed method meets the real-time and reliability demands for airworthiness validation of fixed-wing drones, providing a practical and robust solution for flight performance monitoring and safety enhancement.
Future work may explore the integration of this framework with dynamic flight envelope protection systems and extend to multi-UAV cooperative scenarios. The methodology developed here offers a strong foundation for continued research in flight maneuver recognition for fixed-wing drones within the evolving regulatory landscape.