In this paper, we present a comprehensive study on adaptive safety control for China UAVs (unmanned aerial vehicles) operating under fault risk conditions. The rapid development of China UAV technology has enabled widespread applications in disaster relief, environmental monitoring, logistics, and urban air mobility. However, the safety and reliability of China UAVs face significant challenges from various risk factors, including actuator faults, external disturbances, and model uncertainties. To address these challenges, we propose a novel framework that integrates fault risk quantification, lightweight neural network learning, and adaptive risk-tendency control compensation to enhance the safety and performance of China UAVs under fault conditions.
1. Introduction
China UAVs have emerged as a transformative technology in both civilian and commercial sectors. The growing low-altitude economy in China has accelerated the deployment of UAVs in complex urban environments, where they must operate safely despite potential actuator failures and environmental uncertainties. The safety control of China UAVs under fault conditions has become a critical research priority, as traditional control methods often lack the ability to explicitly quantify and adapt to fault risks in real time.
Our research focuses on developing a risk-aware control framework that explicitly models the uncertainty in fault estimation and adapts the control strategy accordingly. Unlike conventional fault-tolerant control approaches that rely on deterministic fault estimates, our method captures the probabilistic nature of fault risks and provides a systematic way to adjust control actions based on the severity of potential safety losses. This is particularly important for China UAVs operating in safety-critical missions where the consequences of control failures can be severe.
The main contributions of this work are threefold. First, we develop a fixed-time fault observer that provides rapid and accurate estimation of actuator faults, enabling timely detection of fault risks. Second, we introduce the conditional value-at-risk (CVaR) as a risk metric to quantify the tail risk of position uncertainty caused by actuator faults, which allows for explicit risk perception and effective capture of extreme events. Third, we design a lightweight backpropagation neural network to learn the mapping from fault risk to position uncertainty, and further develop an adaptive risk-tendency control compensation strategy that adjusts the control policy smoothly and continuously. Experimental results demonstrate that our method significantly improves the response speed and trajectory tracking accuracy of China UAVs under various fault conditions compared to conventional integral sliding mode control methods.
2. Problem Formulation
2.1 Dynamics Modeling of China UAVs
We consider a quadrotor China UAV with nonlinear dynamics described by the following equations:
$$
\begin{cases}
\dot{P} = V, \\
\dot{V} = \frac{1}{m}R_{WB}F_me_3 – ge_3, \\
\dot{\Omega} = R_0\omega, \\
\dot{\omega} = f(\omega) + J^{-1}\tau,
\end{cases}
$$
where $P = [p_x, p_y, p_z]^\top \in \mathbb{R}^3$ denotes the position in the world frame, $V = [v_x, v_y, v_z]^\top \in \mathbb{R}^3$ the linear velocity, $m$ the mass of the China UAV, $F_m$ the total thrust generated by the four motors, $g$ the gravitational acceleration, $\omega = [p, q, r]^\top \in \mathbb{R}^3$ the angular velocities, $J$ the inertia matrix, and $\tau$ the total torque. The rotation matrix $R_{WB}$ and the mapping matrix $R_0$ are defined based on the Euler angles $\Omega = [\phi, \theta, \psi]^\top$.
The control input $u = [F_m, \tau_\phi, \tau_\theta, \tau_\psi]^\top \in \mathbb{R}^4$ relates to the motor forces $F_r = [f_1, f_2, f_3, f_4]^\top \in \mathbb{R}^4$ through the mapping matrix $M_u$:
$$
u = M_u F_r = \begin{bmatrix}
1 & 1 & 1 & 1 \\
-\frac{l_\phi}{2} & \frac{l_\phi}{2} & \frac{l_\phi}{2} & -\frac{l_\phi}{2} \\
\frac{l_\theta}{2} & -\frac{l_\theta}{2} & \frac{l_\theta}{2} & -\frac{l_\theta}{2} \\
c_M & c_M & -c_M & -c_M
\end{bmatrix}
\begin{bmatrix}
f_1 \\ f_2 \\ f_3 \\ f_4
\end{bmatrix}.
$$
2.2 Actuator Fault Modeling
We consider multiplicative efficiency losses in the actuators of China UAVs, which can be caused by motor speed reduction, blade damage, or bearing wear. The actual force produced by each motor is given by:
$$
F_r^* = \Gamma F_r, \quad r = 1, 2, 3, 4,
$$
where $\Gamma = \text{diag}(\gamma_1, \gamma_2, \gamma_3, \gamma_4) \in \mathbb{R}^{4\times 4}$ is the fault matrix, and $\gamma_i \in [0, 1]$ represents the capability coefficient of the $i$-th motor. $\gamma_i = 0$ indicates complete failure, while $\gamma_i = 1$ indicates a healthy motor. This multiplicative fault model captures the gradual degradation of actuator performance, which is a common failure mode for China UAVs operating in harsh environments.
2.3 Fault Risk Definition
We define fault risk as the uncertainty in the flight state of China UAVs caused by component failures such as motor speed reduction and blade damage. This risk manifests as altitude drop, attitude oscillation, and trajectory deviation, which directly affect the safety of China UAVs during mission execution. The key challenge is to quantify this risk in a way that captures both the probability and severity of adverse events, enabling proactive control adaptation.
3. Methodology
3.1 Fixed-Time Fault Observer Design
We design a fixed-time fault observer to estimate the lumped fault term in the dynamics of China UAVs. The system state vector is defined as $x = [v_z, p, q, r]^\top \in \mathbb{R}^4$, and the state equation under actuator faults is:
$$
\dot{x} = f(x) + Bu + D,
$$
where $B = \text{diag}(\frac{C_\phi C_\theta}{m}, \frac{1}{J_x}, \frac{1}{J_y}, \frac{1}{J_z}) \in \mathbb{R}^{4\times 4}$ is the control coefficient matrix, and $D = BM_u(\Gamma – I_4)F_r \in \mathbb{R}^4$ is the lumped fault term. The fixed-time observer is designed as:
$$
\begin{cases}
\dot{z}_1 = k_1\Xi(\rho_1, \rho_2, \tilde{e}_d) + f(x) + z_2 + Bu, \\
\dot{z}_2 = k_2\Xi(2\rho_1 – 1, 2\rho_2 – 1, \tilde{e}_d),
\end{cases}
$$
where $z_1 = \hat{x}$ and $z_2 = \hat{D}$ are the estimates of the state and fault terms, $\tilde{e}_d = z_1 – x$ is the estimation error, and $\Xi(\rho_1, \rho_2, \tilde{e}_d) = |\tilde{e}_d|^{\rho_1}\text{sign}(\tilde{e}_d) + |\tilde{e}_d|^{\rho_2}\text{sign}(\tilde{e}_d)$ with $\rho_1 \in (0.5, 1)$ and $\rho_2 \in (1, 1.5)$. The motor capability coefficients are then obtained from:
$$
\gamma_r = a_r/f_r + 1, \quad r = 1, 2, 3, 4,
$$
where $a_r$ is the $r$-th element of $M_u^{-1}B^{-1}\hat{D}$.
The fixed-time observer ensures rapid convergence of the fault estimate, which is crucial for timely risk assessment in China UAVs. The convergence time is bounded by a constant independent of the initial condition, guaranteeing predictable performance even under severe fault conditions.
3.2 Fault Risk Quantification Using CVaR
To quantify the fault risk, we first define a safety loss function based on the position uncertainty of China UAVs:
$$
L_f(P, P_d) = E_p \odot E_p,
$$
where $E_p = P – P_d \in \mathbb{R}^3$ is the position deviation, $P$ is the actual position, and $P_d$ is the desired position. The conditional value-at-risk (CVaR) of the safety loss function at confidence level $\alpha_f$ is defined as:
$$
\text{CVaR}_{\alpha_f}^{D_i}[L_f^i] = \mathbb{E}_{D_i}\{L_f^i \in \mathbb{R} | L_f^i > \text{VaR}_{\alpha_f}^{D_i}[L_f^i]\},
$$
which can be expressed in the equivalent form:
$$
\text{CVaR}_{\alpha_f}^{D_i}[L_f^i] = \min_{z_f^i \in \mathbb{R}} \mathbb{E}_{D_i}\left[ z_f^i + \frac{(L_f^i – z_f^i)^+}{1 – \alpha_f} \right].
$$
The CVaR metric captures the conditional expectation of the safety loss exceeding the value-at-risk threshold, which is particularly suited for quantifying tail risks in China UAV operations. Unlike traditional risk metrics that only consider average performance, CVaR explicitly accounts for extreme events that may lead to catastrophic consequences. This is essential for safety-critical applications of China UAVs, where even rare failures can have severe impacts.
The following table summarizes the properties of various risk metrics and highlights the advantages of CVaR for fault risk quantification in China UAVs:
| Risk Metric | Monotonicity | Sub-additivity | Tail Risk Capture | Distribution Invariance | Applicability to China UAV Fault Risk |
|---|---|---|---|---|---|
| Expected Cost | Yes | Yes | No | Yes | Limited |
| Mean-Variance | No | No | Partial | No | Limited |
| Entropic Risk | Yes | No | Partial | Yes | Moderate |
| Value-at-Risk | Yes | No | No | Yes | Moderate |
| Conditional Value-at-Risk | Yes | Yes | Yes | Yes | Excellent |
As shown in the table, CVaR satisfies all desirable properties of a coherent risk measure, including monotonicity, sub-additivity, and the ability to capture tail risks. These properties make CVaR particularly well-suited for quantifying fault risks in China UAVs, where the distribution of position uncertainty can be complex and heavy-tailed due to the coupling of multiple risk factors.
3.3 Neural Network Risk Learning
To model the temporal evolution of fault risk and its impact on position uncertainty, we design a lightweight backpropagation neural network combined with a sliding time window. The input features at time step $k$ consist of the motor capability coefficient sequence $\gamma_i(k – t)$, the fault risk confidence sequence $\alpha_f(k – t)$, and the UAV velocity sequence $V(k – t)$, where $t = 0, 1, \ldots, T_w – 1$. The input vector and corresponding output are defined as:
$$
x_k = [\gamma_i(k – T_w + 1 : k), \alpha_f(k – T_w + 1 : k), V(k – T_w + 1 : k)]^\top \in \mathbb{R}^{8 \times T_w},
$$
$$
y_k = [\text{CVaR}_{L_f}^j(k – T_w + 1 : k)]^\top \in \mathbb{R}^{3 \times T_w}.
$$
The neural network $f_\beta$ is designed with a single hidden layer using the hyperbolic tangent activation function:
$$
\begin{cases}
h = \tanh(W_1 x + b_1) \in \mathbb{R}^H, \\
\hat{y}_k = W_2 h + b_2 \in \mathbb{R}^3,
\end{cases}
$$
where $W_1 \in \mathbb{R}^{H \times 8}$ and $W_2 \in \mathbb{R}^{3 \times H}$ are weight matrices, $b_1 \in \mathbb{R}^H$ and $b_2 \in \mathbb{R}^3$ are bias vectors, and $H = 30$ is the number of hidden neurons. The mean squared error loss function for training is:
$$
\mathcal{L} = \frac{1}{N} \sum_{k=1}^N \|\hat{y}_k – y_k\|_2^2.
$$
The network architecture is designed to be lightweight to enable real-time inference on embedded platforms commonly used in China UAVs. The following table summarizes the key parameters of the neural network model:
| Parameter | Value | Description |
|---|---|---|
| Input dimension | 8 | Motor coefficients, confidence, velocity |
| Hidden layer size | 30 | Number of hidden neurons |
| Output dimension | 3 | CVaR for x, y, z directions |
| Activation function | tanh | Hyperbolic tangent |
| Training algorithm | Levenberg-Marquardt | Optimization method |
| Window size $T_w$ | 10 | Sliding window length |
| Training error target | $10^{-3}$ | Mean squared error |
| Max training epochs | 2000 | Maximum iterations |
This lightweight design achieves a balance between computational efficiency and prediction accuracy, making it suitable for real-time deployment on China UAVs with limited onboard computing resources. The trained model can predict the CVaR-based fault risk measure in real time, providing critical information for adaptive control compensation.
3.4 Adaptive Risk-Tendency Control
Based on the learned risk model, we design an adaptive risk-tendency control compensation strategy. The baseline control law $u_b$ is compensated by the estimated fault term:
$$
u = u_b – B^{-1}\hat{D}.
$$
To further enhance the adaptability to varying fault risks, we introduce a risk-tendency weight vector $w_r \in \mathbb{R}^4$ that maps the three-dimensional fault risk to the four control channels:
$$
w_r = R \cdot \text{CVaR}_f, \quad \text{CVaR}_f = \begin{bmatrix}
\text{CVaR}_{\alpha_f}^{D_1}[L_f^1] \\
\text{CVaR}_{\alpha_f}^{D_2}[L_f^2] \\
\text{CVaR}_{\alpha_f}^{D_3}[L_f^3]
\end{bmatrix} \in \mathbb{R}^3,
$$
where $R \in \mathbb{R}^{4 \times 3}$ is the risk allocation mapping matrix. The final safety control law is:
$$
u = u_b – \text{diag}(w_r) B^{-1} \hat{D}.
$$
This formulation allows the control system to adjust the compensation intensity for each control channel based on the directional risk level. Channels with higher risk receive stronger compensation, enabling a risk-adaptive response that improves the overall safety of China UAVs under fault conditions. The risk-tendency approach ensures that the control policy transitions smoothly between different risk levels without abrupt changes that could destabilize the system.
3.5 Stability Analysis
We provide a stability analysis of the closed-loop system under the proposed adaptive safety control law. Considering the Lyapunov function candidate:
$$
V = x^\top \varpi x,
$$
where $\varpi = \varpi^\top \succ 0$. The error dynamics of the system are derived as:
$$
\dot{e} = \bar{f} + B(u_b – \text{diag}(w_r) B^{-1} \hat{D}) – B u_d + \Delta D,
$$
where the residual term $\Delta D$ captures the effects of observer errors, weight variations, and network approximation errors:
$$
\Delta D = \text{diag}(w_r)(\hat{D} – D) + (\text{diag}(w_r) – \text{diag}(\bar{w})) B^{-1} D + D_{NN}.
$$
Under boundedness assumptions on the observer errors, network prediction errors, and weight variation rates, we establish that the closed-loop system is uniformly ultimately bounded (UUB):
$$
\|e(t)\| \leq \begin{cases}
\Theta(t, \|e(0)\|), & 0 \leq t < T^*, \\
c_1(\varepsilon + \delta_0 + \bar{w}\bar{D}), & t \geq T^*,
\end{cases}
$$
where $\Theta(\cdot)$ describes the initial convergence, $c_1$ depends on system gains, $\varepsilon$ is the network approximation error bound, $\delta_0$ is the observer residual, $\bar{w}$ bounds the weight variation rate, and $\bar{D}$ bounds the lumped fault term. This result guarantees that the tracking error remains within a bounded set whose size can be adjusted by tuning the control gains.
The following table summarizes the key parameters used in the stability analysis and their physical meanings:
| Parameter | Definition | Bound |
|---|---|---|
| $\varepsilon$ | Network prediction error | $\|\text{CVaR}_{NN} – \text{CVaR}\| \leq \varepsilon$ |
| $\delta_0$ | Observer residual | $\|\hat{D} – D\| \leq \delta_0$ |
| $\bar{w}$ | Weight variation rate bound | $\|\dot{w}_r\| \leq \bar{w}$ |
| $\bar{D}$ | Lumped fault bound | $|D| \leq \bar{D}$ |
| $L$ | Fault derivative bound | $|\dot{D}| \leq L$ |
4. Experimental Validation
4.1 Experimental Setup
Our experimental platform consists of a quadrotor China UAV equipped with an STM32F7 microprocessor for safety control, an STM32F4 microprocessor for sensor fusion, and an IMU for attitude estimation. The navigation information is provided by an indoor motion capture system (NOKOV) and the onboard IMU. The China UAV uses a 3S LiPo battery with a nominal voltage of 11.1 V and a capacity of 3300 mAh, achieving a thrust-to-weight ratio of 3.2. The test trajectory is a circular path defined as:
$$
[1.8 \sin(\frac{2\pi}{T} t), 1.8 \cos(\frac{2\pi}{T} t), 0.8] \text{ m},
$$
where $T$ is the period of the circular trajectory that determines the flight speed. The experiments are conducted indoors to ensure controlled conditions and repeatable results for validating the proposed method on China UAVs.
4.2 Fault Scenarios
We define three representative fault cases with varying severity levels. The following table lists the capability coefficients for each case:
| Case | Motor 1 | Motor 2 | Motor 3 | Motor 4 | Severity Level |
|---|---|---|---|---|---|
| Case 1 | 0% | 20% | 20% | 0% | Mild |
| Case 2 | 0% | 20% | 40% | 0% | Moderate |
| Case 3 | 30% | 20% | 40% | 10% | Severe |
These fault cases cover a range of scenarios that are representative of real-world actuator failures in China UAVs, including single-motor faults, multi-motor faults, and asymmetric fault distributions. The fault severity levels are chosen to test the robustness and adaptability of the proposed control method under different operational conditions.
4.3 Fault Observer Performance
The fixed-time fault observer estimates the motor capability coefficients with good accuracy. For the three fault cases, the observer converges within approximately 1 second after fault injection. The estimation performance is summarized in the following table:
| Case | Motor 1 | Motor 2 | Motor 3 | Motor 4 | Convergence Time (s) |
|---|---|---|---|---|---|
| Case 1 (True) | 1.00 | 0.80 | 0.80 | 1.00 | ~1.0 |
| Case 1 (Estimated) | 0.98 | 0.81 | 0.79 | 0.99 | ~1.0 |
| Case 2 (True) | 1.00 | 0.80 | 0.60 | 1.00 | ~1.0 |
| Case 2 (Estimated) | 0.97 | 0.79 | 0.61 | 0.98 | ~1.0 |
| Case 3 (True) | 0.70 | 0.80 | 0.60 | 0.90 | ~1.2 |
| Case 3 (Estimated) | 0.69 | 0.81 | 0.59 | 0.89 | ~1.2 |
To smooth the estimated coefficients and mitigate high-frequency fluctuations, we apply a sliding time average window with a step size of 10, considering that the attitude control loop operates at 500 Hz. This filtering approach effectively suppresses noise while maintaining a low latency that is acceptable for real-time control of China UAVs.
4.4 Neural Network Training and Prediction
The dataset for training the fault risk model is collected from actual flight experiments with fault levels ranging from 5% to 40% in 5% increments, covering various combinations of dual-motor and quad-motor failures. The training dataset is split into training (70%), validation (15%), and test (15%) sets. The network achieves a mean absolute error of $6.8 \times 10^{-4}$ and a root mean square error of $8.2 \times 10^{-4}$ on the test set, both below the target error of $10^{-3}$. The prediction performance on the test set demonstrates that the network can accurately learn the mapping from fault conditions to CVaR-based position uncertainty.
The training process of the lightweight neural network is characterized by the following convergence metrics:
| Metric | Training Set | Validation Set | Test Set |
|---|---|---|---|
| Mean Absolute Error | $5.2 \times 10^{-4}$ | $6.1 \times 10^{-4}$ | $6.8 \times 10^{-4}$ |
| Root Mean Square Error | $6.4 \times 10^{-4}$ | $7.5 \times 10^{-4}$ | $8.2 \times 10^{-4}$ |
| 95% Confidence Interval | $[-7.1 \times 10^{-4}, 7.3 \times 10^{-4}]$ | $[-8.5 \times 10^{-4}, 8.6 \times 10^{-4}]$ | $[-9.3 \times 10^{-4}, 9.5 \times 10^{-4}]$ |
| R-squared | 0.992 | 0.988 | 0.985 |
The high R-squared values across all datasets indicate that the network explains most of the variance in the CVaR output, confirming its effectiveness for fault risk prediction in China UAVs. After training, we apply a bias correction operation using 30% of the training samples to further enhance the prediction accuracy and stability of the model.
4.5 Comparative Experimental Results
We compare our proposed method with the integral sliding mode control (ISMC) method based on fixed-time fault observers. The comparison metrics include response time and tracking error. The response time is defined as the duration from when the z-direction tracking error first exceeds ±10% of the desired trajectory after a sudden fault until it returns to within the ±10% range. The tracking error is defined as the root mean square error:
$$
\mu = \sqrt{\frac{1}{M} \sum_{i=1}^M \|P_i – P_{d,i}\|^2},
$$
and $\bar{\mu} = \frac{1}{M_T} \sum_{j=1}^{M_T} \mu_j$ with $M_T = 5$ repetitions. The comparative results are summarized in the following table:
| Case | Method | Response Time (s) | Tracking Error (m) | Improvement in Response Time | Improvement in Tracking Error |
|---|---|---|---|---|---|
| Mild Faults | Proposed | 1.21 | 0.0608 | 43.2% | 14.9% |
| Mild Faults | ISMC | 2.13 | 0.0714 | – | – |
| Moderate Faults | Proposed | 1.17 | 0.0662 | 46.6% | 17.9% |
| Moderate Faults | ISMC | 2.19 | 0.0806 | – | – |
| Severe Faults | Proposed | 1.22 | 0.0860 | 43.8% | 15.4% |
| Severe Faults | ISMC | 2.17 | 0.1016 | – | – |
The experimental results demonstrate that our proposed method achieves approximately 44.6% faster response time on average compared to the ISMC method, while also improving tracking accuracy across all fault severity levels. The improvement in tracking error ranges from 14.9% for mild faults to 17.9% for moderate faults, with severe faults showing a 15.4% improvement. These results confirm that the adaptive risk-tendency control strategy effectively leverages the learned fault risk information to enhance the safety and performance of China UAVs under fault conditions.
4.6 Generalization Test
To validate the generalization capability of our method, we inject a random fault scenario that was not included in the training set: Motor 1 loses 10% efficiency, Motor 2 loses 30%, Motor 3 loses 20%, and Motor 4 loses 20%. The three-dimensional trajectory tracking results under this unseen fault case show that our method consistently outperforms the ISMC method, with significantly reduced altitude drop in the z-direction and smaller tracking errors in the x and y directions. This demonstrates that the learned risk model generalizes well to unseen fault conditions, making the proposed method practical for real-world deployment where fault patterns are unpredictable.
The following table summarizes the key physical parameters of the China UAV used in our experiments:
| Parameter | Value | Unit |
|---|---|---|
| Mass $m$ | 0.62 | kg |
| Arm length $l$ | 0.23 | m |
| Torque coefficient $c_M$ | 0.01 | m |
| Moment of inertia $J_x$ | $1.5 \times 10^{-3}$ | kg·m² |
| Moment of inertia $J_y$ | $1.5 \times 10^{-3}$ | kg·m² |
| Moment of inertia $J_z$ | $2.0 \times 10^{-3}$ | kg·m² |
| Battery voltage | 11.1 | V |
| Battery capacity | 3300 | mAh |
| Thrust-to-weight ratio | 3.2 | – |
The hyperparameters used in our experiments are listed below:
| Hyperparameter | Value | Hyperparameter | Value |
|---|---|---|---|
| $\alpha_f$ (confidence level) | 0.95 | $T$ (trajectory period) | [5, 12] s |
| $\rho_1$ (observer gain 1) | 0.68 | $\rho_2$ (observer gain 2) | 1.28 |
| $R$ (risk allocation matrix) | $\begin{bmatrix} 0.1 & 0.1 & 0.8 \\ 0.7 & 0.2 & 0.1 \\ 0.2 & 0.7 & 0.1 \\ 0.2 & 0.2 & 0.6 \end{bmatrix}$ | $\Sigma$ (covariance matrix) | $\begin{bmatrix} 0.0019 & -0.00180 & -0.0110 \\ -0.0018 & 0.0018 & 0.0105 \\ -0.0110 & 0.0105 & 0.0798 \end{bmatrix}$ |
5. Discussion
The experimental results demonstrate that the proposed fault risk learning-based safety control method effectively enhances the performance of China UAVs under actuator fault conditions. The integration of CVaR-based risk quantification provides a principled way to capture tail risks that are critical for safety-critical operations. The lightweight neural network enables real-time risk prediction without significantly increasing the computational burden on the embedded platform, which is a key practical consideration for China UAVs with limited onboard computing resources.
One notable observation is that the response time improvement is consistent across all fault severity levels, indicating that the adaptive risk-tendency compensation effectively accelerates the control response regardless of the fault magnitude. This is particularly important for China UAVs operating in dynamic environments where fault conditions can change rapidly and unpredictably. The ability to quickly adapt to changing risk levels ensures that the UAV maintains safe operation even under adverse conditions.
Another important aspect is the generalization capability demonstrated by the random fault injection test. The fact that our method performs well on unseen fault configurations suggests that the neural network has learned a meaningful representation of the underlying relationship between fault characteristics and position uncertainty, rather than simply memorizing the training data. This generalization property is essential for practical deployment, as real-world fault scenarios are diverse and cannot be fully covered during training.
6. Conclusion and Future Work
In this paper, we have presented a fault risk learning-based adaptive safety control method for China UAVs. The main contributions of our work are summarized as follows: (1) a fixed-time fault observer that provides rapid and accurate fault estimation; (2) a CVaR-based risk quantification approach that captures tail risks in position uncertainty; (3) a lightweight neural network that learns the mapping from fault conditions to risk measures; and (4) an adaptive risk-tendency control compensation strategy that adjusts control actions based on the learned risk. Experimental results on a quadrotor China UAV platform demonstrate that our method achieves approximately 44.6% faster response time and 16.1% better tracking accuracy on average compared to the integral sliding mode control method.
Future research directions include extending our method to address coupled multi-source risks in China UAVs, such as the combined effects of actuator faults, sensor noise, and environmental disturbances. We also plan to integrate onboard computers with multi-sensor systems including depth cameras to enhance the risk learning and control capabilities in dynamic and highly uncertain environments. Furthermore, we will investigate the application of our risk-aware control framework to cooperative multi-UAV missions where safety risks propagate through the network, requiring coordinated risk management strategies. These advancements will contribute to achieving higher levels of autonomous safety for China UAVs operating in complex and safety-critical scenarios.
The proposed framework establishes a foundation for risk-aware safety control that can be extended to other types of unmanned systems, including fixed-wing UAVs, ground robots, and autonomous vehicles. By explicitly quantifying and adapting to risks, our approach provides a systematic way to balance performance and safety, which is essential for the widespread adoption of autonomous systems in safety-critical applications.