The extensive adoption of civil fixed-wing UAVs in various operational scenarios has created an urgent demand for robust airworthiness verification and real-time safety monitoring. A fundamental prerequisite for meeting these requirements is the accurate and efficient recognition of flight maneuvers from trajectory data. In our study, I propose a systematic method for recognizing flight maneuvers of civil fixed-wing UAVs, specifically tailored for airworthiness validation. The approach begins by extracting maneuver recognition requirements from existing UAV special conditions, defining maneuver categories and their associated trajectory features. A feature pattern composed of six indicators—indicated airspeed, pitch angle, roll angle, yaw angle, altitude, and altitude change rate—is then constructed to characterize each maneuver. Finally, a recognition model based on Support Vector Machine (SVM) is built and optimized through Bayesian optimization, achieving high accuracy and practical efficiency. The experimental results demonstrate that the proposed method satisfies the stringent demands of fixed-wing UAV airworthiness verification and safety monitoring.
Introduction
Flight maneuver recognition is a critical capability for evaluating flight performance, protecting flight envelopes, and ensuring structural integrity of fixed-wing UAVs. In the context of civil UAV airworthiness, each maneuver type must be identified to verify that the aircraft operates within its certified limits. Traditional approaches have relied on parametric analysis, pattern matching, and deep learning methods such as LSTM and CNN. However, these methods often require large annotated datasets (104 level) that are typically unavailable in airworthiness contexts, suffer from limited interpretability, and demand high computational resources. In contrast, Support Vector Machine (SVM) offers robust performance in small-sample scenarios, and its decision boundary can be physically mapped to flight envelope parameters, making it well-suited for airworthiness verification. Therefore, I select SVM as the core classifier, systematically derive maneuver features from UAV special conditions, and construct a fixed-wing UAV flight maneuver recognition model.
Airworthiness-Oriented Flight Maneuver Recognition Framework
Current airworthiness requirements for civil fixed-wing UAVs are primarily based on special conditions following the “one model, one standard” principle. By analyzing multiple type-certified UAV special conditions, I extract key information regarding flight control requirements (flight mode selection, maneuver limitations), flight envelope design (limiting and normal flight envelopes), and performance design parameters (altitude performance, maneuverability, stability and controllability). For continued airworthiness, flight performance monitoring (real-time monitoring, remote intervention, maintenance planning) and flight envelope management (flight safety, control, trajectory) are emphasized. I find that flight maneuver recognition plays a central role in all these aspects: in initial airworthiness, it supports flight envelope design, performance verification, and safety boundary setting; in continued airworthiness, it enables real-time status monitoring, anomaly warning, envelope optimization, and maintenance support.
The basic flight maneuvers of a fixed-wing UAV can be categorized into five fundamental types: climb, level flight, turn, roll, and descent. These five maneuvers and their combinations constitute the mission flight profile and are the core targets of maneuver recognition. The relationships between flight states and airworthiness requirements are analyzed, and the metric features for each maneuver along with their airworthiness verification objectives are summarized in Table 1.
| Maneuver | Airworthiness Verification Objective | Key Metric Features | Certification Reference |
|---|---|---|---|
| Climb | Ensure maximum climb rate and envelope compliance | Altitude, altitude change rate, airspeed, yaw angle, pitch angle, roll angle | SC-21-003 TP.B011 |
| Level Flight | Verify cruise stability | Altitude, airspeed, yaw angle, roll angle, pitch angle | High-risk UAV trial standard HY.1353 |
| Turn | Validate coordinated turn bank angle limits (≤30°) | Altitude, airspeed, yaw angle, roll angle | High-risk UAV trial standard HY.1975 |
| Roll | Test maneuver load limits | Altitude, airspeed, yaw angle, roll angle, pitch angle | SC-92-001 WL.143 |
| Descent | Confirm approach trajectory stability and landing configuration retention | Altitude, altitude change rate, airspeed, yaw angle, pitch angle, roll angle | SC-92-001 WL.75 |
To ensure that fixed-wing UAV maneuvers comply with airworthiness regulations and safety requirements, the recognition system must enforce strict constraints on key parameters for each maneuver type:
- For climb, the core is to verify the maximum climb rate and envelope compliance. The system must monitor the vertical rate not exceeding design limits, while maintaining pitch angle between 5° and 15° and roll angle near zero to ensure wing balance.
- For level flight, the focus is cruise stability. Altitude fluctuations should be within ±5 m, airspeed deviation controlled within ±5% of the calibrated value, and roll and pitch angles near zero.
- For turn, the key is coordinated bank angle limitation (roll angle ≤30°). The yaw rate must dynamically match the roll angle to prevent sideslip, while altitude changes remain smooth.
- For roll, the system must strictly test maneuver loads, limiting the roll rate to a safe range and monitoring the load factor in real time to avoid structural overload. Short-term fluctuations in pitch angle and altitude should be controllable.
- For descent, the objective is to ensure approach trajectory stability and landing configuration retention. The altitude change rate should be stable in the negative range, airspeed maintained or slightly increased to keep lift controllable, pitch angle negative, and roll angle near zero.
Based on the analysis of UAV special conditions and the metric features defined in Table 1, I construct a systematic airworthiness-oriented flight maneuver recognition framework. The framework first collects core flight data (altitude, airspeed, yaw angle, pitch angle, roll angle) from onboard sensors in real time. Then, key feature parameters are extracted according to Table 1, followed by outlier cleaning and data normalization. Next, the SVM algorithm is used to classify the preprocessed feature data to determine the current flight maneuver type. The recognition result is then compared in real time with the corresponding airworthiness requirements to complete safety analysis and compliance verification. If potential hazards or non-compliant maneuvers are identified, the system immediately issues warnings and triggers corrective actions. Finally, based on the recognition results and compliance feedback, flight control strategies are dynamically optimized, forming a closed-loop improvement of flight performance and safety assurance.
SVM-Based Fixed-Wing UAV Flight Maneuver Recognition Model
Sample Set Construction
The experimental data originate from the public PX4 log dataset on the Flight Review platform. All flight logs include pre-annotated maneuver rating labels and are sampled at 100 Hz. From the raw data, I extract six key flight parameters: altitude (m), altitude change rate (m/s), airspeed (m/s), yaw angle (rad), pitch angle (rad), and roll angle (rad). These parameters comprehensively cover the major motion characteristics of a fixed-wing UAV. Considering the high correlation between turn and roll maneuvers in actual flight, I merge the five basic maneuvers into four more representative categories: climb (1), level flight (2), turn (3), and descent (4).
In the data preprocessing stage, I apply the Bollinger Bands algorithm for cleaning. By calculating a moving average and a range of ±2 standard deviations, abnormal data points caused by sensor errors or external interference are effectively identified and removed. For example, processing the altitude change rate data for the “climb” maneuver yields cleaned data. Then, normalization is performed to eliminate dimensional differences among parameters. After rigorous quality control—including removal of static state data and anomalous fluctuation points—I obtain 15,557 valid six-dimensional feature vector samples.
Kernel Function Selection
The choice of kernel function is closely related to the specific classification problem. I test four commonly used kernel functions—linear, polynomial, RBF, and sigmoid—on the fixed-wing UAV maneuver dataset to select the one with the highest classification accuracy. The experimental results 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 achieves the highest accuracy at 86.59%, outperforming the linear kernel by 14.79%, the polynomial kernel by 14.08%, and the sigmoid kernel by 13.74%. Additionally, previous research indicates that SVM with the RBF kernel automatically determines the threshold and center weights, thereby reducing the upper bound of expected test error. Therefore, I select the RBF kernel as the kernel function for the fixed-wing UAV recognition model and optimize its parameter γ.
Model Optimization Using Bayesian Optimization Algorithm
Bayesian Optimization Algorithm (BOA) is a systematic hyperparameter tuning method particularly suitable for machine learning models. It constructs a probabilistic surrogate model via Gaussian processes, dynamically updates the prior distribution using historical parameter information, and intelligently selects the next evaluation point based on an acquisition function to efficiently locate the global optimum. For the fixed-wing UAV flight maneuver recognition task, I apply BOA to tune the SVM parameters, significantly improving convergence speed and recognition accuracy.
I use 10-fold cross-validation to build the BO-SVM model on the preprocessed data. The parameter ranges for the penalty factor C and the kernel parameter γ are set to [1, 100] and [0.1, 1000] respectively, with the number of iterations set to 50. The optimization process shows rapid convergence: although the maximum iteration count is 50, the algorithm begins to converge after only about 15 iterations, indicating efficient parameter search. The optimal parameter combination found is C = 32.28 and γ = 2.15. When testing the optimized model on the test set, the classification accuracy improves by 12% compared to the default parameter configuration, demonstrating a substantial enhancement in model performance.
Reliability Analysis of the Flight Maneuver Recognition Model
The optimized SVM model achieves a prediction accuracy of 98.29% on the test set, with a prediction runtime of only 0.91 s. The confusion matrix shows that the recognition accuracy for each maneuver category exceeds 97%, indicating excellent overall performance for the fixed-wing UAV flight maneuver recognition model. To comprehensively evaluate the model, I conduct comparative tests under the same dataset (PX4 flight logs) and hardware environment (Intel i9-13980HX). The results are presented in Tables 3 and 4.
| Model | Penalty Factor C | Kernel Parameter γ | Accuracy (%) | Macro Precision (%) | Macro Recall (%) | Macro F1 Score (%) |
|---|---|---|---|---|---|---|
| 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 |
From Table 3, the default SVM exhibits lower precision (86.6%), indicating a relatively high number of false predictions across categories. In contrast, BO-SVM achieves significantly higher precision (98.3%), demonstrating a drastic reduction in misclassifications. In terms of recall, the default SVM is inadequate, implying a notable risk of missed detections; BO-SVM, with a recall of 98.5%, effectively avoids missing dangerous maneuvers. The F1 score of BO-SVM reaches 98.4%, reflecting a superior balance between precision and recall.
| 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) | Non-interpretable |
| Data Required for Convergence | 3,000 | 12,000 | 10,500 |
| Memory Usage | 48 MB | 620 MB | ~500 MB |
As shown in Table 4, BO-SVM outperforms deep learning models in accuracy (98.29% vs 93.91% for CNN-BiLSTM and 76.62% for LSTM). Moreover, BO-SVM is 9.7 times faster in training and 16.7 times faster in prediction than the best deep learning competitor (CNN-BiLSTM), fully satisfying the real-time requirements of airworthiness monitoring. The memory footprint of BO-SVM is only 48 MB, compared to 620 MB for CNN-BiLSTM, significantly reducing hardware cost and energy consumption.
In summary, the BO-SVM fixed-wing UAV flight maneuver recognition model exhibits clear advantages in accuracy metrics, real-time performance, and deployment cost. In practical applications, it can reliably support safe flight and autonomous control decision-making for fixed-wing UAVs, particularly for real-time monitoring of critical maneuvers in airworthiness verification.
Conclusion
In this paper, I have addressed the flight maneuver recognition requirements for civil fixed-wing UAVs in airworthiness verification and safety monitoring by proposing a method based on SVM. Through systematic analysis of UAV special conditions, I identified key maneuver recognition requirements and constructed a six-dimensional feature pattern consisting of indicated airspeed, pitch angle, roll angle, yaw angle, altitude, and altitude change rate. The SVM-based recognition model was established and optimized using Bayesian optimization, resulting in a significant improvement in both accuracy and efficiency. Experimental results demonstrate that the optimized model achieves a test accuracy of 98.29%, exhibits high real-time performance and interpretability, and effectively meets the demands of fixed-wing UAV airworthiness verification and safety monitoring. The proposed framework provides a practical and reliable solution for ensuring the operational safety and regulatory compliance of civil fixed-wing UAVs.