In the rapidly evolving landscape of unmanned aerial systems, the integration of advanced safety mechanisms has become paramount for ensuring reliable operation under adverse conditions. Our research focuses on addressing a critical challenge in drone technology: the degradation of safe maneuvering capabilities when unmanned aerial vehicles (UAVs) are subjected to fault risks. This work presents a comprehensive framework that combines fault risk quantification, machine learning, and adaptive control to enhance the resilience of quadrotor UAVs against actuator failures.
The proliferation of drone technology across civilian and commercial domains—including disaster response, environmental monitoring, logistics, and urban air mobility—has underscored the necessity for robust safety control strategies. Traditional control methods often rely on passive robustness to handle uncertainties, but they lack explicit mechanisms to quantify and adapt to fault risks in real time. Our approach bridges this gap by introducing a risk-aware control paradigm that explicitly models, learns, and compensates for fault-induced uncertainties.
| Challenge | Description | Impact on Drone Technology |
|---|---|---|
| Actuator Faults | Loss of effectiveness in motors due to wear or damage | Degraded maneuverability and trajectory tracking |
| Model Uncertainty | Incomplete knowledge of aerodynamic and dynamic parameters | Reduced control accuracy and stability margins |
| Environmental Disturbances | Wind gusts, turbulence, and other external factors | Increased positional uncertainty and safety risks |
| Multi-source Coupling | Interaction between faults, disturbances, and model errors | Complex nonlinear behavior threatening flight safety |
Problem Formulation and System Modeling
To establish a rigorous foundation for our safety control framework, we first develop a comprehensive mathematical model of the quadrotor UAV dynamics and actuator fault characteristics. The drone technology under consideration operates in a three-dimensional environment with six degrees of freedom, comprising translational and rotational motions.
We define the world coordinate frame FW and the body-fixed frame FB to describe the UAV’s motion. The position vector P = [px, py, pz]T ∈ ℝ3 and the velocity vector V = [vx, vy, vz]T ∈ ℝ3 represent the translational dynamics in the world frame. The Euler angles Ω = [ϕ, θ, ψ]T ∈ ℝ3 describe the orientation, where ϕ, θ, and ψ denote the roll, pitch, and yaw angles, respectively.
The complete nonlinear dynamics of our quadrotor system are expressed as:
$$
\begin{cases}
\dot{\mathbf{P}} = \mathbf{V}, \\[4pt]
\dot{\mathbf{V}} = \frac{1}{m} \mathbf{R}_{WB} F_m \mathbf{e}_3 – g \mathbf{e}_3, \\[4pt]
\dot{\mathbf{\Omega}} = \mathbf{R}_0 \boldsymbol{\omega}, \\[4pt]
\dot{\boldsymbol{\omega}} = \mathbf{f}(\boldsymbol{\omega}) + \mathbf{J}^{-1} \boldsymbol{\tau},
\end{cases}
$$
where m ∈ ℝ denotes the mass of the UAV, g ∈ ℝ is the gravitational acceleration, RWB ∈ ℝ3×3 is the rotation matrix from the body frame to the world frame, R0 ∈ ℝ3×3 is the attitude mapping matrix, ω = [p, q, r]T ∈ ℝ3 represents the angular velocity in the body frame, J = diag{Jx, Jy, Jz} ∈ ℝ3×3 is the inertia matrix, and τ = [τϕ, τθ, τψ]T ∈ ℝ3 is the total torque vector.
The total thrust Fm = ∑r=14 fr is generated by four brushless motors, where fr ∈ ℝ represents the lift force from the r-th motor. The control input vector u = [Fm, τϕ, τθ, τψ]T ∈ ℝ4 is related to the individual motor forces through a known mapping matrix Mu ∈ ℝ4×4.
Table 2: Physical Parameters of the Quadrotor UAV Used in Our Experiments
| Parameter | Symbol | Value | Unit |
|---|---|---|---|
| Mass | m | 0.62 | kg |
| Arm length | l | 0.23 | m |
| Moment coefficient | CM | 0.01 | m |
| Roll inertia | Jx | 1.5 × 10-3 | kg·m2 |
| Pitch inertia | Jy | 1.5 × 10-3 | kg·m2 |
| Yaw inertia | Jz | 2.0 × 10-3 | kg·m2 |
Actuator faults in drone technology often manifest as a reduction in motor efficiency due to factors such as bearing wear, voltage drops, or propeller damage. We model these faults using a multiplicative efficiency loss factor:
$$
\mathbf{F}_r^* = \boldsymbol{\Gamma} \mathbf{F}_r, \quad \boldsymbol{\Gamma} = \operatorname{diag}(\gamma_1, \gamma_2, \gamma_3, \gamma_4),
$$
where γi ∈ [0, 1] represents the capability coefficient of the i-th motor. A value of γi = 1 indicates healthy operation, while γi = 0 represents complete failure. This multiplicative model captures a wide range of fault scenarios encountered in practical drone technology, including partial and total motor failures.
Table 3: Representative Fault Scenarios Considered in This Study
| Fault Case | Motor 1 | Motor 2 | Motor 3 | Motor 4 | Severity Level |
|---|---|---|---|---|---|
| Mild faults | 0% | 20% | 20% | 0% | Low |
| Moderate faults | 0% | 20% | 40% | 0% | Medium |
| Severe faults | 30% | 20% | 40% | 10% | High |
Fixed-Time Fault Observer Design
To enable rapid and accurate fault estimation—a critical requirement for safety-critical drone technology—we design a fixed-time fault observer. The state vector is defined as x = [vz, p, q, r]T ∈ ℝ4, and the state equation under actuator faults is expressed as:
$$
\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}) + \mathbf{B} \mathbf{u} + \mathbf{D},
$$
where B = diag(CϕCθ/m, 1/Jx, 1/Jy, 1/Jz) ∈ ℝ4×4 is the control coefficient matrix, and D = B Mu (Γ − I4) Fr ∈ ℝ4 represents the lumped fault term. The nonlinear function f(x) captures the gyroscopic and coupling effects inherent in quadrotor dynamics.
Our fixed-time observer is designed as:
$$
\begin{cases}
\dot{\mathbf{z}}_1 = k_1 \boldsymbol{\Xi}(\rho_1, \rho_2, \tilde{\mathbf{e}}_d) + \mathbf{f}(\mathbf{x}) + \mathbf{z}_2 + \mathbf{B} \mathbf{u}, \\[4pt]
\dot{\mathbf{z}}_2 = k_2 \boldsymbol{\Xi}(2\rho_1 – 1, 2\rho_2 – 1, \tilde{\mathbf{e}}_d),
\end{cases}
$$
where z1 = x̂ and z2 = D̂ are the estimates of the state and lumped fault, respectively. The error term is Ẽd = z1 − x, and the function Ξ(·) incorporates both linear and nonlinear gains to achieve fixed-time convergence:
$$
\boldsymbol{\Xi}(\rho_1, \rho_2, \tilde{\mathbf{e}}_d) = |\tilde{\mathbf{e}}_d|^{\rho_1} \operatorname{sign}(\tilde{\mathbf{e}}_d) + |\tilde{\mathbf{e}}_d|^{\rho_2} \operatorname{sign}(\tilde{\mathbf{e}}_d),
$$
with ρ1 ∈ (0.5, 1) and ρ2 ∈ (1, 1.5) being observer gains. This design ensures that the fault estimation converges within a bounded time independent of initial conditions, which is crucial for timely fault compensation in drone technology.
The motor capability coefficients are recovered from the fault estimate through the relationship:
$$
\gamma_r = a_r / f_r + 1, \quad r = 1, 2, 3, 4,
$$
where ar are elements of the vector Mu−1 B−1 D̂. This inverse mapping allows us to translate the lumped fault estimate into physically interpretable motor health indicators, enabling condition-based maintenance and adaptive control.
Fault Risk Quantification Using Conditional Value-at-Risk
A key innovation in our approach is the use of conditional value-at-risk (CVaR) to quantify fault risks in drone technology. Unlike traditional fault detection methods that provide point estimates, CVaR captures the tail behavior of the loss distribution, making it particularly suitable for safety-critical applications where rare but severe events must be considered.
We define the safety loss function Lf(P, Pd) ∈ ℝ3 as the element-wise squared position error:
$$
\mathbf{L}_f(\mathbf{P}, \mathbf{P}_d) = \mathbf{E}_p \odot \mathbf{E}_p,
$$
where ⊙ denotes the Hadamard product and Ep = P − Pd ∈ ℝ3 is the position deviation. The i-th element of this loss function, Lfi = EpT ei eiT Ep, quantifies the squared error in the i-th spatial direction.
The value-at-risk (VaR) of the safety loss at confidence level αf is defined as:
$$
\operatorname{VaR}_{\alpha_f}^{\mathcal{D}_i}[L_f^i] = \inf \left\{ z_f^i \in \mathbb{R} \mid \Pr(L_f^i > z_f^i) > \alpha_f \right\},
$$
where αf ∈ (0, 1) represents the risk tolerance level. The CVaR then captures the expected loss in the tail beyond the VaR threshold:
$$
\operatorname{CVaR}_{\alpha_f}^{\mathcal{D}_i}[L_f^i] = \mathbb{E}_{\mathcal{D}_i} \left\{ L_f^i \in \mathbb{R} \mid L_f^i > \operatorname{VaR}_{\alpha_f}^{\mathcal{D}_i}[L_f^i] \right\}.
$$
An equivalent optimization-based formulation facilitates computational implementation:
$$
\operatorname{CVaR}_{\alpha_f}^{\mathcal{D}_i}[L_f^i] = \min_{z_f^i \in \mathbb{R}} \mathbb{E}_{\mathcal{D}_i} \left[ z_f^i + \frac{(L_f^i – z_f^i)^+}{1 – \alpha_f} \right],
$$
where (·)+ = max{·, 0}. This formulation allows efficient gradient-based optimization and integration with learning frameworks.
Table 4: Properties of Risk Measures for Drone Technology Safety
| Property | Definition | CVaR | VaR | Variance |
|---|---|---|---|---|
| Monotonicity | X ≤ Y ⇒ ρ(X) ≤ ρ(Y) | ✓ | ✓ | ✗ |
| Translation invariance | ρ(X + c) = ρ(X) + c | ✓ | ✓ | ✗ |
| Positive homogeneity | ρ(λX) = λρ(X), λ ≥ 0 | ✓ | ✓ | ✓ |
| Sub-additivity | ρ(X + Y) ≤ ρ(X) + ρ(Y) | ✓ | ✗ | ✓ |
| Tail risk capture | Sensitivity to extreme events | ✓ | Partial | ✗ |
Risk Learning with Lightweight Neural Networks
To model the complex relationship between fault risks and positional uncertainty in drone technology, we employ a lightweight backpropagation neural network combined with a sliding time window. This architecture captures the temporal dependencies inherent in fault propagation while maintaining computational efficiency suitable for real-time embedded deployment.
The input to our learning framework at time step k consists of a feature vector constructed from the past Tw time steps:
$$
\mathbf{x}_k = [\gamma_i(k – T_w + 1 : k), \; \alpha_f(k – T_w + 1 : k), \; \mathbf{V}(k – T_w + 1 : k)]^\top \in \mathbb{R}^{8 \times T_w},
$$
where γi are the motor capability coefficients estimated by the fixed-time observer, αf is the confidence level reflecting risk tolerance, and V is the UAV velocity measured by the motion capture system. The corresponding output is the CVaR-based risk measure across the three spatial directions:
$$
\mathbf{y}_k = [\operatorname{CVaR}_{L_f}^j(k – T_w + 1 : k)]^\top \in \mathbb{R}^{3 \times T_w}, \quad j = 1, 2, 3.
$$
The neural network fβ with one hidden layer of H = 30 neurons and hyperbolic tangent activation functions is defined as:
$$
\begin{cases}
\mathbf{h} = \tanh(\mathbf{W}_1 \mathbf{x} + \mathbf{b}_1) \in \mathbb{R}^H, \\[4pt]
\hat{\mathbf{y}}_k = \mathbf{W}_2 \mathbf{h} + \mathbf{b}_2 \in \mathbb{R}^3,
\end{cases}
$$
where W1 ∈ ℝH×8 and W2 ∈ ℝ3×H are weight matrices, and b1 ∈ ℝH and b2 ∈ ℝ3 are bias vectors. The mean squared error loss function is employed for training:
$$
\mathcal{L} = \frac{1}{N} \sum_{k=1}^N \|\hat{\mathbf{y}}_k – \mathbf{y}_k\|_2^2,
$$
where N is the number of training samples. We optimize the network using the Levenberg-Marquardt algorithm, which offers fast convergence for small-to-medium scale regression problems. The choice of network architecture—a single hidden layer with 30 neurons—represents a deliberate trade-off between model capacity and computational cost, ensuring that the risk learning module can run on the STM32F7 microcontroller used in our flight experiments.
Adaptive Risk-Tendency Control Compensation
The core of our safety control framework lies in the adaptive risk-tendency compensation strategy, which dynamically adjusts the control effort based on the learned CVaR-based risk measures. This approach enables continuous, chatter-free adaptation of the control policy in response to evolving fault conditions.
We first construct a risk-tendency weight vector wr ∈ ℝ4 that maps the three-dimensional risk measures to the four control channels:
$$
\mathbf{w}_r = \mathbf{R} \cdot \mathbf{CVaR}_f, \quad \mathbf{CVaR}_f =
\begin{bmatrix}
\operatorname{CVaR}_{\alpha_f}^{\mathcal{D}_1}[L_f^1] \\
\operatorname{CVaR}_{\alpha_f}^{\mathcal{D}_2}[L_f^2] \\
\operatorname{CVaR}_{\alpha_f}^{\mathcal{D}_3}[L_f^3]
\end{bmatrix} \in \mathbb{R}^3,
$$
where R ∈ ℝ4×3 is the risk allocation mapping matrix learned during training. This matrix establishes the relationship between spatial risk directions and control channel priorities, enabling direction-sensitive compensation.
The baseline control law ub, designed using a drag utilization method for high-maneuverability tracking, is augmented with the fault compensation term:
$$
\mathbf{u} = \mathbf{u}_b – \operatorname{diag}(\mathbf{w}_r) \mathbf{B}^{-1} \hat{\mathbf{D}}.
$$
Here, diag(wr) forms a diagonal matrix from the weight vector, allowing different compensation intensities across control channels. Channels associated with higher risk levels receive stronger compensation, effectively prioritizing safety in the most critical directions.
Table 5: Risk Allocation Mapping Matrix R
| Control Channel | X-direction risk | Y-direction risk | Z-direction risk |
|---|---|---|---|
| Thrust (Fm) | 0.1 | 0.1 | 0.8 |
| Roll torque (τϕ) | 0.7 | 0.2 | 0.1 |
| Pitch torque (τθ) | 0.2 | 0.7 | 0.1 |
| Yaw torque (τψ) | 0.2 | 0.2 | 0.6 |
Stability Analysis
We establish the uniform ultimate boundedness (UUB) of the closed-loop system under the proposed adaptive risk-tendency control law. Consider the Lyapunov function candidate:
$$
V = \mathbf{e}^\top \boldsymbol{\varpi} \mathbf{e}, \quad \boldsymbol{\varpi} = \boldsymbol{\varpi}^\top \succ 0,
$$
where e = x − xd is the tracking error. The error dynamics incorporate the observer estimation error, the neural network approximation error, and the weight variation term:
$$
\dot{\mathbf{e}} = \bar{\mathbf{f}} + \mathbf{B}(\mathbf{u}_b – \operatorname{diag}(\mathbf{w}_r)\mathbf{B}^{-1}\hat{\mathbf{D}}) – \mathbf{B}\mathbf{u}_d + \boldsymbol{\Delta}_D,
$$
with the residual term ΔD given by:
$$
\boldsymbol{\Delta}_D = \operatorname{diag}(\mathbf{w}_r)(\hat{\mathbf{D}} – \mathbf{D}) + (\operatorname{diag}(\mathbf{w}_r) – \operatorname{diag}(\bar{\mathbf{w}}))\mathbf{B}^{-1}\mathbf{D} + \boldsymbol{\Delta}_{\text{NN}}.
$$
Under the assumptions that the observer error is bounded by δ0, the neural network approximation error is bounded by ε, and the weight variation rate is bounded by w̄, we obtain:
$$
\|\boldsymbol{\Delta}_D\| \leq c_2(\varepsilon + \delta_0 + \bar{w} \bar{D}).
$$
Taking the time derivative of the Lyapunov function and applying Lipschitz and Young inequalities yields:
$$
\dot{V} \leq -\lambda_{\min}(\mathbf{Q}) \|\mathbf{e}\|^2 + k \|\mathbf{e}\| (\varepsilon + \delta_0 + \bar{w} \bar{D}),
$$
which can be further bounded as:
$$
\dot{V} \leq -d_1 \|\mathbf{e}\|^2 + d_2 (\varepsilon + \delta_0 + \bar{w} \bar{D})^2,
$$
where d1 > 0 is adjustable through control gains and d2 > 0 is a constant. This inequality implies that the tracking error converges to a bounded neighborhood of the origin:
$$
\limsup_{t \to \infty} \|\mathbf{e}(t)\| \leq \sqrt{\frac{d_2}{d_1}} (\varepsilon + \delta_0 + \bar{w} \bar{D}).
$$
The ultimate bound depends on the observer accuracy, the neural network approximation quality, and the smoothness of the risk weight adaptation. By increasing the control gains, we can reduce the bound until it reaches the lower limit imposed by the observer and learning errors.
Experimental Validation
We conducted extensive flight experiments on a custom-built quadrotor platform to validate the effectiveness of our proposed method. The experimental setup consists of:
- Flight platform: Quadrotor UAV with STM32F7 microcontroller for safety control (100 Hz translational control, 500 Hz attitude control) and STM32F4 for sensor fusion.
- Navigation system: NOKOV motion capture system for high-precision position tracking, combined with onboard IMU for attitude estimation.
- Communication: Ultra-wideband (UWB) link providing 50 Hz telemetry between the UAV and ground station.
- Power system: 3S LiPo battery (11.1 V, 3300 mAh) with thrust-to-weight ratio of 3.2.
- Reference trajectory: Circular path defined as [1.8 sin(2πt/T), 1.8 cos(2πt/T), 0.8] m, with period T adjustable to vary flight speed.
The training dataset for the fault risk model was collected through systematic flight experiments covering fault severities from 5% to 40% at 5% intervals, with combinations of faults on motors 1-4. A total of 70% of the data was used for training, 15% for validation, and 15% for testing.
Table 6: Neural Network Training Performance
| Metric | Value |
|---|---|
| Mean absolute error (test set) | 6.8 × 10-4 |
| Root mean square error (test set) | 8.2 × 10-4 |
| 95% confidence interval | [−9.3 × 10-4, 9.5 × 10-4] |
| Training epochs | 2000 |
| Target error | 1.0 × 10-3 |
We compared our proposed adaptive risk-tendency control method against the integral sliding mode control (ISMC) approach with fixed-time fault observation. The comparison metrics include:
- Response time: Time from fault injection to recovery of z-axis tracking error within ±10% of the desired trajectory.
- Tracking error (TE): Root mean square error averaged over 5 repeated flight tests.
Table 7: Experimental Comparison Results
| Fault Case | Method | Response Time (s) | Tracking Error (m) |
|---|---|---|---|
| Mild faults | Proposed method | 1.21 | 0.0608 |
| Mild faults | ISMC method | 2.13 | 0.0714 |
| Moderate faults | Proposed method | 1.17 | 0.0662 |
| Moderate faults | ISMC method | 2.19 | 0.0806 |
| Severe faults | Proposed method | 1.22 | 0.0860 |
| Severe faults | ISMC method | 2.17 | 0.1016 |
The experimental results demonstrate a consistent improvement across all fault scenarios. Our proposed method achieves approximately 44.6% faster response time compared to the ISMC method, attributed to the proactive risk-aware compensation that anticipates fault effects rather than reacting after fault estimation. The tracking error improvement ranges from 14.8% in mild fault cases to 15.4% in severe fault cases, indicating that the risk-tendency adaptation effectively compensates for both small and large fault-induced disturbances.
To further validate the generalization capability of our learned risk model, we conducted additional experiments with a randomly injected fault configuration (10%, 30%, 20%, and 20% efficiency loss on motors 1-4 respectively)—a scenario not present in the training dataset. The three-dimensional trajectory tracking results showed that our method maintained superior performance compared to ISMC, with improved altitude hold (reduced drop) and tighter lateral tracking. This demonstrates that the neural network successfully learned the underlying relationship between fault risks and positional uncertainty, enabling robust performance on unseen fault combinations.
Table 8: Computational Resource Requirements
| Component | CPU Usage (%) | Memory (KB) | Execution Time (ms) |
|---|---|---|---|
| Fixed-time observer | 3.2 | 4.8 | 0.12 |
| Risk quantification (CVaR) | 1.1 | 2.1 | 0.05 |
| Neural network inference | 8.7 | 12.4 | 0.31 |
| Risk-tendency compensation | 0.8 | 1.2 | 0.03 |
| Baseline control | 5.4 | 6.7 | 0.18 |
| Total | 19.2 | 27.2 | 0.69 |
The lightweight design of our neural network (single hidden layer with 30 neurons) ensures that the entire risk learning and adaptive control pipeline consumes less than 20% of the STM32F7 CPU resources and 27.2 KB of memory, with a total execution time of 0.69 ms per control cycle. This leaves ample computational headroom for other safety-critical tasks, demonstrating the practical feasibility of our approach for resource-constrained embedded drone technology platforms.
Discussion and Future Directions
Our adaptive safety control framework represents a significant advancement in fault-tolerant drone technology, addressing the critical gap between fault detection and risk-aware control adaptation. By explicitly quantifying fault risks using CVaR and learning their impact on positional uncertainty, we enable proactive rather than reactive safety responses. The experimental validation confirms that this approach delivers measurable improvements in both response speed and tracking accuracy across diverse fault scenarios.
Key contributions of our work include:
- Risk-aware fault quantification: We introduced CVaR as a risk measure for fault-induced uncertainties in drone technology, extending its application from collision risk and energy risk to the domain of actuator fault risks. The tail-risk capture property of CVaR is particularly valuable for safety-critical UAV operations where rare but severe fault events must be anticipated.
- Learning-based risk prediction: Our lightweight neural network architecture with sliding time window effectively captures the temporal evolution of fault risks and their influence on positional uncertainty. The network’s compact design (30 hidden neurons) enables real-time inference on embedded flight controllers without compromising prediction accuracy.
- Adaptive risk-tendency compensation: The direction-sensitive risk allocation mechanism translates multi-dimensional risk measures into channel-specific control compensation, allowing the UAV to prioritize safety in the most affected directions. The continuous adaptation ensures smooth control transitions without the chattering typically associated with switching-based methods.
Future research directions for advancing risk-aware drone technology include:
- Multi-source risk coupling: Investigating the interaction between actuator faults, environmental disturbances, and model uncertainties to develop comprehensive risk models that capture compound effects.
- Online learning and adaptation: Extending our offline-trained neural network framework to enable continuous online learning, allowing the risk model to adapt to evolving fault patterns and environmental conditions.
- Sensor integration: Incorporating onboard depth cameras, lidar, and other exteroceptive sensors to expand the risk perception capability beyond proprioceptive measurements, enabling awareness of external hazards.
- Distributed risk awareness: Extending the risk-aware control paradigm to multi-UAV systems, where collaborative risk assessment and coordinated safety responses can enhance overall mission resilience.
- Formal verification of safety bounds: Developing rigorous mathematical frameworks to certify the safety of risk-aware control policies under probabilistic fault models, providing guarantees for certification in safety-critical applications.
Conclusion
This paper presented a comprehensive framework for adaptive safety control of quadrotor UAVs under actuator fault risks, addressing a fundamental challenge in modern drone technology. By integrating fixed-time fault observation, CVaR-based risk quantification, lightweight neural network learning, and adaptive risk-tendency compensation, our approach enables explicit risk awareness and proactive safety adaptation that significantly outperforms conventional methods.
The experimental results validated that our method achieves approximately 44.6% faster response time and 14.8-15.4% improvement in tracking accuracy compared to integral sliding mode control, while maintaining lightweight computational requirements suitable for embedded flight controllers. The successful generalization to unseen fault configurations demonstrates the robustness and practical utility of the learned risk models.
As drone technology continues to expand into increasingly complex and safety-critical applications—including urban air mobility, autonomous delivery, and emergency response—the ability to quantify, learn, and adapt to fault risks will become essential. Our work provides a foundational framework for risk-aware autonomy that balances performance and safety, paving the way for more resilient and trustworthy unmanned aerial systems.
Table 9: Summary of Key Performance Improvements
| Metric | ISMC Method | Proposed Method | Improvement |
|---|---|---|---|
| Average response time (s) | 2.16 | 1.20 | 44.4% faster |
| Mild fault tracking error (m) | 0.0714 | 0.0608 | 14.8% better |
| Moderate fault tracking error (m) | 0.0806 | 0.0662 | 17.9% better |
| Severe fault tracking error (m) | 0.1016 | 0.0860 | 15.4% better |
| CPU utilization (%) | 15.8 | 19.2 | 3.4% higher (acceptable) |
| Memory footprint (KB) | 18.5 | 27.2 | 8.7 KB higher (acceptable) |
The slight increase in computational resource consumption is well within the capacity of modern flight controllers and represents a worthwhile investment for the substantial safety and performance gains achieved. Our ongoing work aims to further optimize the neural network architecture through techniques such as pruning and quantization, reducing the resource footprint while maintaining prediction accuracy.