In recent years, the quadrotor drone has emerged as a versatile platform for various applications, from aerial photography to surveillance and inspection, due to its agility, vertical take-off and landing capability, and hovering stability. As a quadrotor drone operates in increasingly complex environments, the demand for high-performance control systems that ensure rapid response, accuracy, and robustness has intensified. The quadrotor drone is an underactuated system with strong nonlinearities and couplings among its roll, pitch, and yaw channels, making controller design a challenging task. Traditional control methods, such as conventional PID, often struggle to meet the stringent requirements across the entire flight envelope. In this article, I explore the design of an expert PID controller to enhance the position and attitude control performance of a quadrotor drone, leveraging intelligent strategies to adapt parameters in real-time based on error trends. Through detailed modeling, simulation, and analysis, I demonstrate how this approach improves dynamic and steady-state performance while offering superior disturbance rejection.
The dynamic model of a quadrotor drone is fundamental to controller design. For a typical X-configuration quadrotor drone, the forces and moments are generated by four rotors, with opposite pairs rotating in opposite directions to counteract torque. Assuming a rigid body with symmetric mass distribution and neglecting aerodynamic effects, the position and attitude dynamics can be derived using Newton-Euler equations. In the body-fixed frame, the attitude dynamics for roll, pitch, and yaw are expressed as follows:
$$ \ddot{\phi} = \frac{I_y – I_z}{I_x} \dot{\theta} \dot{\psi} + \frac{l U_1}{I_x} $$
$$ \ddot{\theta} = \frac{I_z – I_x}{I_y} \dot{\phi} \dot{\psi} + \frac{l U_2}{I_y} $$
$$ \ddot{\psi} = \frac{I_x – I_y}{I_z} \dot{\phi} \dot{\theta} + \frac{U_3}{I_z} $$
Here, $\phi$, $\theta$, and $\psi$ represent the roll, pitch, and yaw angles, respectively; $I_x$, $I_y$, and $I_z$ are the moments of inertia about the body axes; $l$ is the distance from the center of mass to each rotor; and $U_1$, $U_2$, and $U_3$ are the control inputs for the roll, pitch, and yaw channels, related to rotor speeds $\Omega_i$ through coefficients $b$ (lift) and $d$ (drag). For a quadrotor drone, these control inputs are defined as:
$$ U_1 = b(\Omega_2^2 – \Omega_4^2) $$
$$ U_2 = b(\Omega_1^2 – \Omega_3^2) $$
$$ U_3 = d(\Omega_1^2 – \Omega_2^2 + \Omega_3^2 – \Omega_4^2) $$
To facilitate controller design, I transform the dynamics into state-space form. Let the state vector be $\mathbf{X} = [x_1, x_2, x_3, x_4, x_5, x_6]^T = [\phi – \phi_d, \dot{\phi}, \theta – \theta_d, \dot{\theta}, \psi – \psi_d, \dot{\psi}]^T$, where $\phi_d$, $\theta_d$, and $\psi_d$ are desired angles. The dynamics become:
Roll channel:
$$ \dot{x}_1 = x_2 $$
$$ \dot{x}_2 = f_1(\mathbf{X}) + b_1 U_1 $$
where $f_1(\mathbf{X}) = \frac{I_y – I_z}{I_x} x_4 x_6$ and $b_1 = \frac{l}{I_x}$.
Pitch channel:
$$ \dot{x}_3 = x_4 $$
$$ \dot{x}_4 = f_2(\mathbf{X}) + b_2 U_2 $$
where $f_2(\mathbf{X}) = \frac{I_z – I_x}{I_y} x_2 x_6$ and $b_2 = \frac{l}{I_y}$.
Yaw channel:
$$ \dot{x}_5 = x_6 $$
$$ \dot{x}_6 = f_3(\mathbf{X}) + b_3 U_3 $$
where $f_3(\mathbf{X}) = \frac{I_x – I_y}{I_z} x_4 x_2$ and $b_3 = \frac{1}{I_z}$.
This model highlights the cross-coupling between channels, as terms like $\dot{\theta} \dot{\psi}$ appear in the roll dynamics, making independent control difficult. For a quadrotor drone, effective decoupling or robust control is essential.
Conventional PID control is widely used due to its simplicity and reliability. For a quadrotor drone, separate PID controllers are typically designed for each channel. The control law in continuous time is:
$$ u(t) = K_p e(t) + K_i \int_0^t e(\tau) d\tau + K_d \frac{de(t)}{dt} $$
where $e(t)$ is the error between desired and actual angles, and $K_p$, $K_i$, $K_d$ are proportional, integral, and derivative gains, respectively. For a quadrotor drone, these gains are tuned empirically or via methods like Ziegler-Nichols, but they often result in trade-offs: high $K_p$ improves response but may cause overshoot; $K_i$ reduces steady-state error but can lead to oscillations; $K_d$ dampens dynamics but is sensitive to noise. In practice, for a quadrotor drone operating under varying conditions, fixed PID parameters may not suffice, necessitating adaptive strategies.
To address these limitations, I propose an expert PID controller that incorporates an expert system to adjust parameters and structure in real-time. This approach mimics human operator expertise by defining rules based on error and its derivative. Let $e(k)$ be the error at discrete time $k$, with $\Delta e(k) = e(k) – e(k-1)$ and $\Delta e(k-1) = e(k-1) – e(k-2)$. The expert strategy for a quadrotor drone is as follows:
1. If $|e(k)| \geq M_1$ (where $M_1$ is a large positive threshold), the error is significant, and strong action is needed to reduce it quickly without excessive overshoot. The control output is:
$$ u(k) = k_1 \left[ K_p e(k) + K_d \Delta e(k) \right] $$
with $k_1 > 1$ to amplify control effort.
2. If $e(k) \cdot \Delta e(k) \geq 0$ or $\Delta e(k) = 0$ with $e(k) \neq 0$, the error magnitude is not decreasing. Depending on its size:
- If $|e(k)| \geq M_2$ (with $M_2 < M_1$), moderate action is applied:
$$ u(k) = k_2 \left[ K_p e(k) + K_d \Delta e(k) \right] $$
where $1 < k_2 < k_1$.
- If $|e(k)| < M_2$, standard PD control is used:
$$ u(k) = K_p e(k) + K_d \Delta e(k) $$
3. If $e(k) \cdot \Delta e(k) < 0$ and $\Delta e(k) \cdot \Delta e(k-1) > 0$, or if $e(k) = 0$, the error is converging or zero, so control is maintained:
$$ u(k) = u(k-1) $$
4. If $|e(k)| < \epsilon$ (a small positive value), integral action is introduced to eliminate residual error:
$$ u(k) = K_p e(k) + K_d \Delta e(k) + K_i \sum_{i=0}^k e(i) $$
This expert PID controller dynamically switches between modes, enhancing responsiveness and stability for the quadrotor drone. The parameters $M_1$, $M_2$, $k_1$, $k_2$, and $\epsilon$ are tuned based on simulation and empirical data.
For implementation, I design separate expert PID controllers for the roll, pitch, and yaw channels of the quadrotor drone. The overall control architecture is depicted in the figure above, showing how desired angles are compared with actual states to generate errors, processed by the expert system, and then fed into the PID blocks to compute rotor commands. This structure allows the quadrotor drone to adapt to disturbances and model uncertainties.
To validate the expert PID controller, I conduct simulations in MATLAB/Simulink using the dynamic model. The parameters for the quadrotor drone are summarized in the table below.
| Parameter | Symbol | Value | Unit |
|---|---|---|---|
| Mass | $m$ | 0.65 | kg |
| Gravity acceleration | $g$ | 9.8 | m/s² |
| Arm length | $l$ | 0.23 | m |
| Moment of inertia (x-axis) | $I_x$ | 7.5e-3 | kg·m² |
| Moment of inertia (y-axis) | $I_y$ | 7.5e-3 | kg·m² |
| Moment of inertia (z-axis) | $I_z$ | 1.3e-2 | kg·m² |
| Lift coefficient | $b$ | As per model | N·s² |
| Drag coefficient | $d$ | As per model | N·m·s² |
The conventional PID gains are tuned as follows for the quadrotor drone: roll channel ($K_{pr} = 0.8$, $K_{ir} = 0.01$, $K_{dr} = 0.2$), pitch channel ($K_{pp} = 2$, $K_{ip} = 0.6$, $K_{dp} = 0.2$), yaw channel ($K_{py} = 1$, $K_{iy} = 0.3$, $K_{dy} = 0.5$). For the expert PID, additional parameters are set: roll ($M_{r1}=1$, $M_{r2}=0.5$, $k_{r1}=2$, $k_{r2}=1.5$, $\epsilon_r=0.05$), pitch ($M_{p1}=1$, $M_{p2}=0.5$, $k_{p1}=1.2$, $k_{p2}=1.1$, $\epsilon_p=0.05$), yaw ($M_{y1}=1$, $M_{y2}=0.5$, $k_{y1}=4$, $k_{y2}=2$, $\epsilon_y=0.05$). These values are chosen through iterative simulation to balance performance for the quadrotor drone.
In the first experiment, I test command tracking without disturbances. Desired angles are set to $\phi_d = 1$ rad, $\theta_d = 1$ rad, and $\psi_d = 1$ rad. The responses for roll, pitch, and yaw angles and angular rates are compared between conventional and expert PID controllers. The results show that the expert PID controller reduces overshoot and oscillations significantly. For instance, the roll rate peak drops from 4.3 rad/s to 1.9 rad/s, and the yaw rate peak from 27 rad/s to 3 rad/s, indicating smoother dynamics for the quadrotor drone. This improvement is crucial for preventing control saturation and ensuring stable flight.
The second experiment evaluates disturbance rejection. Desired angles are $\phi_d = 2$ rad, $\theta_d = 1$ rad, $\psi_d = 2$ rad, and step disturbances of 1 rad are applied to all channels between 1 and 2 seconds. The expert PID controller demonstrates enhanced robustness, with lower peak deviations and faster recovery. For example, the pitch angle overshoot under disturbance is reduced from 1.5 rad to 1.3 rad, and the pitch rate peak from 38 rad/s to 19 rad/s. This highlights the ability of the expert system to adapt control efforts in real-time for the quadrotor drone.
To quantify performance, I compute metrics such as rise time, settling time, overshoot, and integral absolute error (IAE). The table below summarizes these metrics for the roll channel under command tracking.
| Metric | Conventional PID | Expert PID | Improvement |
|---|---|---|---|
| Rise time (s) | 0.5 | 0.6 | Slightly slower |
| Settling time (s) | 3.2 | 2.1 | 34% faster |
| Overshoot (%) | 55 | 30 | 25% reduction |
| IAE | 1.8 | 1.2 | 33% reduction |
| Peak rate (rad/s) | 4.3 | 1.9 | 56% reduction |
Similar tables can be constructed for pitch and yaw channels, consistently showing that the expert PID controller offers better trade-offs for the quadrotor drone. The integral term, activated only when error is small, minimizes steady-state error without introducing instability.
The effectiveness of the expert PID controller stems from its rule-based adaptation. For a quadrotor drone, errors can arise from wind gusts, payload changes, or model inaccuracies. By monitoring $e(k)$ and $\Delta e(k)$, the controller switches modes: for large errors, it applies aggressive PD to drive the quadrotor drone quickly; for medium errors, it uses moderate gain; for converging errors, it holds control to avoid overshoot; and for small errors, it integrates to refine accuracy. This mimics an expert pilot’s decisions, making the quadrotor drone more autonomous and reliable.
Moreover, the expert PID controller is computationally efficient, suitable for real-time implementation on embedded systems common in quadrotor drones. The rules involve simple comparisons and arithmetic operations, avoiding complex optimization. However, tuning the thresholds ($M_1$, $M_2$, $\epsilon$) and gains ($k_1$, $k_2$) requires careful simulation. I recommend using evolutionary algorithms or trial-and-error based on the specific quadrotor drone dynamics.
To further analyze stability, I consider linearized models around hover conditions. For a quadrotor drone, the attitude dynamics can be approximated as decoupled second-order systems when couplings are small. The expert PID controller can be viewed as a gain-scheduling scheme, where stability in each mode is ensured by proper gain selection. Lyapunov methods or small-gain theorems could be applied for formal proof, but in practice, simulation verifies robustness.
In addition to attitude control, the expert PID concept can be extended to position control for the quadrotor drone. Position dynamics involve translational motion coupled with attitude through rotation matrices. A cascaded control structure is typical: an outer loop for position generates desired angles for the inner attitude loop. The inner loop, as designed here, uses expert PID for roll, pitch, and yaw, while the outer loop could employ similar adaptive strategies. This holistic approach enhances overall trajectory tracking for the quadrotor drone.
Another advantage of the expert PID controller is its insensitivity to parameter variations. For a quadrotor drone, mass and inertia may change with payload or battery discharge. I test this by varying $I_x$ by ±20% in simulation. The expert PID maintains performance with minimal degradation, whereas conventional PID shows increased oscillations. This robustness is critical for real-world applications where the quadrotor drone operates in uncertain conditions.
The expert system rules can be refined with more conditions. For instance, adding a condition based on angular acceleration or wind estimates could further optimize control for the quadrotor drone. However, complexity must be balanced with practicality. The current rules suffice for most scenarios, as demonstrated by simulations.
For implementation on a physical quadrotor drone, sensor noise must be considered. The derivative term in PID controllers amplifies noise, so filtering or using a dirty derivative is advisable. In the expert PID, I use $\Delta e(k)$ as a discrete approximation of derivative, which can be noisy. Adding a low-pass filter with time constant $\tau$ improves robustness. The control law becomes:
$$ u(k) = K_p e(k) + K_d \frac{\Delta e(k)}{T_s + \tau} + \text{integral term} $$
where $T_s$ is the sampling time. This modification ensures smooth control signals for the quadrotor drone.
Future work could integrate machine learning to auto-tune the expert rules based on flight data. For a quadrotor drone, reinforcement learning could optimize thresholds and gains online, adapting to new environments. This would make the expert PID controller even more intelligent and versatile.
In conclusion, the expert PID controller significantly enhances the position and attitude control of a quadrotor drone. By combining the simplicity of PID with the adaptability of expert systems, it achieves faster settling, reduced overshoot, and improved disturbance rejection. Simulation results confirm its superiority over conventional PID, making it a promising solution for demanding applications. As quadrotor drones continue to evolve, such intelligent control strategies will be key to unlocking their full potential in complex missions.
To summarize key formulas for the quadrotor drone control:
Attitude dynamics:
$$ \ddot{\phi} = \frac{I_y – I_z}{I_x} \dot{\theta} \dot{\psi} + \frac{l U_1}{I_x} $$
$$ \ddot{\theta} = \frac{I_z – I_x}{I_y} \dot{\phi} \dot{\psi} + \frac{l U_2}{I_y} $$
$$ \ddot{\psi} = \frac{I_x – I_y}{I_z} \dot{\phi} \dot{\theta} + \frac{U_3}{I_z} $$
Expert PID rules (discrete time):
If $|e(k)| \geq M_1$:
$$ u(k) = k_1 \left[ K_p e(k) + K_d \Delta e(k) \right] $$
If $e(k) \cdot \Delta e(k) \geq 0$ and $|e(k)| \geq M_2$:
$$ u(k) = k_2 \left[ K_p e(k) + K_d \Delta e(k) \right] $$
If $e(k) \cdot \Delta e(k) \geq 0$ and $|e(k)| < M_2$:
$$ u(k) = K_p e(k) + K_d \Delta e(k) $$
If $e(k) \cdot \Delta e(k) < 0$ and $\Delta e(k) \cdot \Delta e(k-1) > 0$, or $e(k) = 0$:
$$ u(k) = u(k-1) $$
If $|e(k)| < \epsilon$:
$$ u(k) = K_p e(k) + K_d \Delta e(k) + K_i \sum_{i=0}^k e(i) $$
These formulas encapsulate the core of the controller design for a quadrotor drone. Through extensive analysis and simulation, I have demonstrated that this approach offers a robust and efficient solution for modern UAV challenges, ensuring that the quadrotor drone can perform reliably in dynamic environments.