1. Introduction
In recent years, advancements in aerial research have led to various applications that are now part of everyday human life. These applications involve acquiring real-world images in dangerous locations such as active volcanoes, earthquake-prone regions, and wildfire-affected areas, as well as tracking industrial activities and national borders and predicting atmospheric conditions. In reality, these operations can be performed by aircraft and drones, with the twin rotor 
 system (
) being the most commonly used [
1,
2]. It has encouraged researchers in the control field by serving as a tool for many experiments and enabling real-time adjustments to aeronautical vehicles [
2].
In most cases, the 
 control system is required to accurately measure two angles: pitch and yaw. These are frequently controlled by calculating two ideal 
 voltages, where each one supplies power to an individual 
 motor [
3]. As a result, the system possesses two degrees of freedom, allowing both horizontal and vertical movements. Therefore, the given performance is often restricted by the adequate selection of two main keys [
4]. The first key is the determination of a modeling strategy that produces an accurate multi-input, multi-output model (
). The second key is the correct choice of a controller synthesis technique.
On this topic, a novel evolutionary proportional–integral–derivative (
) controller design for 
 using a simplified genetic algorithm was presented [
5]. The proposed controller achieves a good response and excellent performance on the 
  system. In [
6], the behavior dynamics of the 
 system was decoupled into two single-input, single-output (
) systems. Each unit is controlled separately using time-optimal, robust 
 regulators. The cross-coupling impact is regarded as a disturbance or the change of system parameters. In this approach, each 
 regulator guarantees an optimal Nyquist plot (
) margin against these uncertainties. Later, various traditional control methods and advanced control techniques were applied [
7] to the 
  system. These techniques ensure that the 
 beam moves rapidly and precisely, enabling it to track a trajectory and reach specific positions. A fuzzy 
 control scheme was proposed in [
8] for 
, in which a new control structure allows the 
 to achieve a target position more efficiently than previously. All the gain values of the 
 regulator are obtained by the real coded genetic algorithm (
) with the system performance index as a fitness function. A feedback model predictive control strategy was proposed in [
9] for the nonlinear 
, in which the control scheme makes the 
 beam move rapidly and precisely to the target positions (pitch and yaw angles). In addition, its structure was tested under disturbance, where an external force was manually applied to the system to assess its performance under significant disturbance conditions. It is evident that the yaw and pitch angles shift from their intended positions due to the applied external force. However, the control system successfully restores stability, ensuring the beam returns to its desired angles. In [
10], a decoupled compensator was designed for a physical 
 by decoupling its actual behavior using a minimal open-loop decoupler. The resulting decoupled 
 units are adjusted using two independent two degrees of freedom (
) controllers, ensuring reference tracking dynamics while maintaining closed-loop robustness and mitigating the cross-coupling effect.
In a recent study, Ref. [
11] introduced a linear quadratic regulator (
) with an integral controller, resulting in a novel robust 
 controller. The effectiveness of this controller was evaluated against several existing controllers from the literature, and its performance was validated through experimental testing. Later, Ref. [
12] developed a robust controller by integrating a sliding mode control (
) approach with a backstepping controller. The backstepping control scheme eliminates the chattering and oscillation issues typically associated with (
), and the resulting controller is capable of providing excellent reference tracking dynamics for both the pitch and yaw angles, even when accounting for parametric model uncertainties. A control scheme established on active disturbance rejection control and input shaping was proposed in [
13] for the 
. The control objectives involved accurately following the desired trajectories and ensuring disturbance rejection in both the horizontal and vertical planes. Additionally, the composite control demonstrated robustness against changes in system parameters, including variations in the natural frequency and damping ratio. In [
14], two robust 
 fractional-order regulators (
) were proposed, in which the control of the pitch angle is completely independent from the control of the yaw angle. A fractional order integral–proportional derivative (
) controller was simulated and trialed to control each rotor of the 
 in [
15]. The approach achieved enhanced disturbance rejection compared to the standard proportional integral derivative (
) controller and proportional derivative (
) controller. The controller parameters were optimized using the MATLAB
® function fmincom. A robust 
 controller was proposed in [
16] for the 
, which had been previously separated into two 
 systems. This controller contains a fractional-order precompensator to ensure accurate reference path tracking behavior and a fractional post-compensator to enhance the closed-loop robustness.
The adaptive explicit nonlinear model predictive control (
) method introduced in [
17] was applied to the 
 considering a strongly coupled nonlinear 
 system with uncertain parameters. This approach provides improved control outcomes, including enhanced tracking performance, a smoother control law, and increased energy efficiency. More recently, a combination of proportional integral and proportional-derivative (
) controllers was proposed [
18] for 
, to control the pitch and yaw angles.
Flight systems often experience delays, which can result from various factors, including high-order rotor response, control actuator dynamics, filtering processes, and computational delays [
19]. Similarly, for 
 flight systems that contain the time delay produced by the collection of sensor data [
20] or that resulting from the cascading of multiple dynamic components used in the complex system [
21], the design of the controller for the flight system becomes a big challenge. All of the above are grouped in what is denoted hardware-induced delays 
. Moreover, an additional delay is produced when the 
 is operated in teleoperated mode. Recent studies in the literature propose controller designs that tackle time delays in flight control, utilizing advanced adaptive control theory, such as integrating the Smith predictor with adaptive controllers [
20], adaptive compensators based on observers [
22], 
 controllers, adaptive podcast controllers [
23], and robust adaptive controllers [
24]. In these investigations, the degradation of the flying quality due to the time delay of the systems is considered, and the difficulty in controlling them is apparent.
Persistent disturbances such as wind introduce errors in the position and orientation of 
 [
13]. Controllers designed not only for positioning and orientation but also for disturbance rejection tend to be more complex to tune [
25], and the inherent time delay of the 
 often leads to control failures [
21,
26]. To address these challenges, we aim to design a robust controller capable of mitigating the effects of disturbances and compensating for time delays.
The Smith predictor control scheme 
 is effective at eliminating the effect of time delay in a closed-loop system [
27]. However, its performance degrades when it is required to reject persistent disturbances. Thus, the main contribution of this paper is a modified Smith predictor controller specifically tailored to our 
, which is effectively represented by a high-order dynamic model with time delay. The proposed controller is designed to accurately track the pitch and yaw angles in the 
, while compensating for delay and rapidly rejecting disturbances.
Few studies have proposed 
 extensions of the 
. Among them, most have focused on addressing robustness issues related to varying process parameters (e.g., [
28]), while only a few have aimed to improve disturbance rejection. In most cases, the 
  has been combined with a decoupler (e.g., [
29,
30]), allowing the 
 control problem to be decomposed into several 
  control problems.
Most modifications to the  aimed at enhancing disturbance rejection have been developed for  systems. Consistent with the approach mentioned above, extensions of the  to  systems typically rely on combining these -based modifications with process decouplers. As a result, the disturbance rejection capabilities of the modified   are inherited by the corresponding  control system.
In previous work [
31,
32,
33], we developed a modified 
 specifically designed to mitigate the effects of step disturbances in 
 systems. In this paper, we propose extending this control strategy to the 
 case to address the control requirements of our 
.
The most relevant disturbance rejection methods for 
 processes with time delay—including the filtered SP modification, internal model control (
), and feedforward strategies—were compared in [
31] with our proposed 
 structure in the context of a second-order plant with time delay. Our 
 modification was further compared with several other 
 variants (including a sliding mode controller) in [
32] for 
 integrating processes with time delay. Additionally, it was benchmarked against the standard 
 in the case of a first-order process with time delay in [
33] In all these comparisons, our modified 
 demonstrated superior disturbance rejection performance. Therefore, we anticipate that the 
 extension of our modified 
 will also outperform 
 schemes based on other 
 disturbance rejection strategies combined with decouplers.
The remainder of the paper is organized as follows. 
Section 2 is dedicated to the dynamic model of the 
. Our proposed control scheme is presented in 
Section 3. 
Section 4 provides simulated results of a 
, while 
Section 5 presents the experimental results. Finally, 
Section 6 presents the conclusions.
  3. Modified Smith Predictor Control Scheme
A modified 
 for 
 is depicted in 
Figure 2. In this figure, the original 
 scheme is illustrated by the black lines and boxes. In this scheme, 
 is the actual dynamic model of the 
, as defined in the Equation (
6), while 
 represents a nominal model of 
, and 
 represents a nominal time delay model of 
. 
 is the reference vector of the system, where 
 is the reference for the vertical plane (pitch angle), and 
 for the horizontal plane (yaw angle). 
 = 
 performs the control function and 
 is a decoupler. 
 is the control signal vector, where 
 is the control signal for the vertical plane, and 
 for the horizontal plane. 
 is the fictitious control signal vector applied to the decoupled system. 
 = 
 is the disturbance controller. 
 = 
 is the external disturbance applied to the 
. 
 is the response of the system vector, where 
 is the yaw angle in the vertical plane, and 
 in the horizontal plane.
We denote the decoupled system (removing the delay) as  and its nominal value as .
  3.1. Design of the Main Controller
The main controller is composed of a regulator  and a decoupler . The design procedures for these two components are outlined next.
- Design of the Decoupler  - The decoupler splits the control structure into two subsystems: one controlling the vertical plane and other controlling the horizontal plane. It eliminates the interactions between  -  and  -  and between  -  and  - , allowing each output to be independently regulated by its corresponding input. A simplified decoupling structure, aligned with this tuning approach [ 34- ] is employed: - 
            where  -  are the elements of the nominal transfer function matrix  - . Then, we define - 
            which, in general, is not a diagonal matrix. In the case of the nominal process, we have that - 
            where  -  and  - . Taking into account the commutative property of the product of diagonal matrices, we have that - 
            where  -  is the nominal time delay. 
- Design of the Regulator  -  controllers are designed in the diagonal elements of  - : - 
            using the nominal process, i.e,  -  and  - . 
- Frequency domain techniques are often used to design these controllers. In particular, they are designed using the phase margin  -  and the gain margin  - , according to [ 7- ]: 
- The parameters of the controllers are defined as follows: - 
            where: 
- The specifications of the  controller for the vertical plane are as follows: - Phase margin: . - Phase crossover frequency:  rad/s. - Gain crossover frequency:  rad/s. - The specifications of the  controller for the horizontal plane are as follows: - Phase margin: . - Phase crossover frequency:  rad/s. - Gain crossover frequency:  rad/s. - These specifications for both the vertical and horizontal plane control of the  -  were selected to balance stability, robustness, and response speed for the vertical plane control of the  - . A phase margin of  -  ensures robustness against model uncertainties while avoiding excessive oscillations. The phase crossover frequency ( - ) corresponds to a settling time of approximately 26 s, ensuring smooth and stable response. The gain crossover frequency ( - ) provides sufficient bandwidth for disturbance rejection while maintaining robustness against high-frequency noise and unmodeled dynamics. Then, the parameters of the controller  -  are the following: 
  3.2. Standard Smith Predictor
The response of the 
 using a standard 
 is [
32,
35]:
        where 
 represents the relationship between the response output 
 and the desired reference input 
, and 
 denotes represents the influence of the disturbance 
 on 
. These are provided, respectively, by:
In a nominal situation, where 
 = 
 and 
 = 
, the previous equations become:
        where
  3.3. Modified Smith Predictor
The innovative structure, based on a modified Smith predictor for controlling the 
, introduces a compensator 
, whose output is inserted between the 
 and the decoupler 
. Placing H(s) at this point allows it to operate on the decoupled control signals, effectively mitigating external disturbances and modeling uncertainties in both the vertical and horizontal channels. This configuration ensures that disturbance compensation is performed before the cross-coupling dynamics, handled by D(s), are reintroduced, thereby improving the overall robustness and performance of the 
 control system. This structural modification is represented by the blue lines in 
Figure 2. The compensator is integrated with the original 
 architecture to generate a modified control signal. It will be shown that, for the nominal process, this block affects only the disturbance response while leaving the set-point tracking performance unchanged. Therefore, it can be used reduce the impact of disturbances on the system. The proposed control system, hereafter referred as 
, was initially designed for a heating furnace modeled as a 
 second-order process with a time delay [
31]. It was later adapted to control a marginally stable 
 motor (i.e., with a pole at the origin) [
32]. In the present work, we have redesigned the control system to handle a high-order 
 system along with time delays, which characterizes the 
. Additionally, we examined the benefits of this new 
 structure over the standard one in mitigating the effects of wind disturbances on the 
. The transfer function matrices of the 
 with the 
 control scheme are expressed as:
In the nominal situation, the above equations become: 
        where
Transfer function (31) demonstrates that the set-point tracking performance is influenced solely by . In contrast, transfer function (32), (33) is affected by both  and . Therefore, a careful design of  can mitigate the impact of the disturbance  on the output , which would otherwise be present in the standard  configuration.
  3.4. The Disturbance Rejection Controller  
The 
 controller serves as the disturbance rejection controller, aimed at mitigating and absorbing the impact of persistent perturbations like the wind effects on the 
 system [
13]. The control scheme of our 
 has two subsystems. This is a special feature that allowed us to design a new controller that can remove the disturbance effect that impacts the system. The new controller is based on adding an inner loop that feedbacks the difference between the process output and the nominal model output. In our control scheme, we define 
 as: 
        where 
 is a diagonal transfer function matrix such that 
. A simple choice of 
 is
        where 
 and 
 are chosen as the minima integer values that make the elements of matrix 
 proper. Substituting (34) in (33) gives that
        which coincides with 
 when 
 is set.
Comparison of (27) and (36), shows that the difference is that the term  of the first equation becomes  in the second one. This implies that, if the final value theorem were applied to determine the steady state effect of the disturbance on the output, this term is constant in the first equation while it tends to zero in the second when . This suggests a better disturbance rejection feature of the  at low frequencies than that of the .
The effect of persistent disturbances on the output is reduced if the magnitudes of 
 or 
 are reduced at low frequencies. Then, we define the index
        which allows us to evaluate the disturbance reduction features of our 
 compared to the standard 
. Since 
 and 
 are diagonal, the index 
 is composed of two functions:
        and
        where 
 and 
 represent the 
i-row, 
l-column element of matrices 
 and 
, respectively.
Comparing expressions (26) and (32), the only difference is the presence of  in the former and  in the latter. If , then the effect of the disturbance on the output is greater when using  than when using . Therefore,  rejects the disturbance better than  at frequency . Since these matrices are diagonal, this inequality can be tested for each channel of the multivariable system by comparing the corresponding diagonal elements of the matrices.
Translated to the individual diagonal elements, the inequality is equivalent to checking whether the ratio  is less than 1. Conversely, if this ratio is greater than 1, then  rejects the disturbance better than  at frequency .
Consequently, if ,  provides better disturbance rejection than . Conversely, if ,  offers better disturbance rejection than .
After selecting 
, 
, and 
, the transfer function 
 is as follows:
The values of  and  were tuned through simulation studies, with the aim of enhancing disturbance rejection performance while maintaining system stability. The selected values provided the best trade-off between response speed and robustness to external disturbances in both the vertical and horizontal channels. If the values of  and  are modified, it can impact the frequency range over which the disturbance rejection is optimized: (1) increasing  values can widen the effective disturbance rejection frequency range but may risk stability or increase sensitivity to noise, or (2) decreasing  values reduces disturbance rejection effectiveness, potentially narrowing the frequency range, but can improve the stability of the system.
Figure 3 plots 
 and 
. The green line represents the threshold line at 
. This line serves as a reference: when 
, 
 provides better disturbance rejection than 
; and when 
, 
 performs better. The red curve corresponds to 
, which evaluates the robustness function for the pitch angle 
. The black curve corresponds to 
, which evaluates the robustness function for the yaw angle 
.
 The figure demonstrates that  rejects disturbance effects more effectively than  in the frequency range of  rad/s for the pitch angle , and in the range of  rad/s for the yaw angle .
  4. Simulation Results
To confirm the effectiveness of the proposed 
 controller, the linearized model of the feedback instrument 
 33-220 was implemented in Simulink (Matlab). The step responses of this control scheme are presented in this section and compared to those of the conventional 
 controller, emphasizing both set-point tracking and disturbance rejection. Both 
 control schemes were applied to the linearized model of the 
 given in Equation (
8). The amplitudes of the reference step inputs were set to 
 rad for the vertical plane and 
 rad for the horizontal plane. A step disturbance of 
 rad was introduced at 80 s to simulate an external perturbation.
In the control structure implemented in the simulation, the main controller parameters from Equation (21) were used to command the 
 system. Additionally, the decoupler parameters from Equation (
9) were applied to separate the system into two subsystems: vertical and horizontal. Furthermore, the disturbance rejection controller parameters from Equation (41) were utilized to enhance the system’s performance.
The responses of the control schemes applied to the 
 system in both the vertical and horizontal planes are illustrated in 
Figure 4 and 
Figure 5. 
Figure 4b presents the responses of both the 
 and 
 controllers on the vertical plane, while both controllers exhibit similar behavior in response to a reference change, they differ in their transient response when preventing external disturbances. Similarly, 
Figure 5b shows the responses of the same control schemes on the horizontal plane. Although the controllers demonstrate comparable behavior in terms of reference tracking, they display noticeable differences in transient performance during disturbance rejection. The control signals applied to the vertical and horizontal planes are displayed in 
Figure 4a and 
Figure 5a, respectively.
The classical 
 lacks a dedicated component for disturbance rejection; although it can attenuate disturbances, its response is slower compared to the proposed scheme. In the proposed 
 approach, the controller 
 significantly improves disturbance rejection capabilities. The simulation results outlined in 
Figure 4b and 
Figure 5b indicate that 
 exhibits superior disturbance rejection performance, providing improved transient behavior, in contrast to the conventional 
 scheme.
To evaluate the performance of our control system under varying operating conditions, we conducted simulations with 
 variations in all 
 parameters, except for the time delay. The time delay was excluded because the 
 scheme is sensitive to mismatches in time delay. However, it can be estimated in real time, allowing the controller’s time delay parameter to be retuned dynamically (e.g., [
36]). 
Figure 4 and 
Figure 5 present the simulation results, which demonstrate that (1) both 
 and 
 exhibit a wide range of stability robustness (tolerating more than 
 parametric variation); (2) the time responses to step-point tracking and step disturbances of both control systems show relatively small variations despite the parameter changes; (3) 
 rejects disturbances significantly more effectively than 
; and (4) the time responses to step-point tracking and step disturbances of 
 are more robust than those of 
, as they remain more consistent under parameter variations. We note that the system responses deteriorate when parameter variations exceed 
, and the system becomes unstable for variations around 
.
Table 2 provides a comparison of the time-domain responses of the 
 control structures from the simulation, focusing on disturbance rejection, considering the effect of airflow generated by a fan ventilator. In this table, 
 refers to the Integral of the absolute error, defined as:
  denotes the Integral of the squared error, given by:
      and 
 represents the total variation in the control signal, expressed as:
The first two indexes reflect the system’s ability to reject disturbance effects, while the third quantifies the control effort required. In this context, it measures the variation in the control signal during the application of the airflow disturbance on the . This metric is associated with actuator wear (as well as that of other physical components), the magnitude of the control signal, and the potential for actuator saturation.
  5. Experimental Results
To examine the practical effectiveness of the proposed controller, it was implemented into the 
 whose parameters are specified in 
Table 1. The 
 laboratory unit is a feedback system consisting of two blocks: a feedback encoder block and a feedback 
 block. The 
 uses two encoders to transmit the measurements of (
) and (
) to the feedback encoder block. This block has three parameters: a sample time of 
 , and offsets for channel one and channel two. Channel one corresponds to the first output of the encoder (
), while channel two corresponds to the second output of the encoder (
). The value of the digital input provided to the 
 is transformed into an analog response by the feedback 
 block.
The proposed 
 control schemes were tested in real time by applying Simulink in Matlab on the actual 
 as shown in 
Figure 6. To demonstrate the performance of the proposed approach, experimental tests were conducted for the trajectory tracking problem of the 
. The experiments were conducted in a university laboratory under controlled conditions. The 
 setup, including the fan, data acquisition card, and computer, was placed on a stable table and isolated from external environmental disturbances. The main source of noise was the airflow generated by the fan, which was considered during testing.
To evaluate the robustness of the proposed 
 control technique for rejecting external disturbances, a disturbance was introduced to the 
 at 
 s. This disturbance consisted of an airstream generated by a Duracraft fan (240 V, 50 Hz, 50 W) placed approximately half a meter away from the system, directed at both the tail and main rotors. The airflow began approximately 80 s after the start of the experiment. The experiment results are shown in 
Figure 7 and 
Figure 8.
Figure 7b shows the performance of both controllers in controlling the vertical plane. It can be observed that while both controllers achieve satisfactory tracking of the reference signal, the 
 controller exhibits a superior transient response, particularly in the presence of external disturbances. The 
 controller reduces the overshoot and improves compared to the standard SP scheme, highlighting its robustness under real-world conditions. Similarly, 
Figure 8b presents experimental outcomes for horizontal plane control. Both control schemes demonstrate effective reference tracking; however, the 
 controller shows enhanced disturbance rejection capabilities. During disturbance events, the 
 maintains a smoother and faster recovery, whereas the standard SP controller displays larger deviations. 
Figure 7a and 
Figure 8a illustrate the experimental control signals corresponding to the vertical and horizontal planes, respectively. These experimental findings confirm the improved performance of the proposed 
 method in practical scenarios, validating the advantages previously observed in simulation.
 Table 3 presents a comparison of the time-domain responses of the 
 and 
 controllers based on experimental results when subjected to disturbances from the ventilator’s airflow. The maximum error 
 in this table is defined as 
, and the 
, 
 and 
 are also reanalyzed. It is observed that the 
 and 
 index values for the 
 controller are reduced compared to those of the 
 controller, although this improvement comes at the expense of an increase in control action. Moreover, the performance indices obtained from the experiments closely align with those from the simulations, validating the accuracy of our models and reinforcing the relevance of the comparative simulation analysis presented in the previous section.
   6. Conclusions
This paper focuses on the mitigation of the airflow disturbance effects on the , with particular attention to reducing the impact of a fan ventilator. Due to the inherent time delay in the  dynamics, a control scheme based on the Smith predictor has been employed. Furthermore, a novel modification of the Smith predictor has been developed to enhance the control performance on the .
This research makes several key contributions, outlined as follows: (1) a multi-variable control method based on the Smith predictor controller capable of compensating for time delays and mitigating the effects of external disturbances while exhibiting good robustness features; (2) a novel methodology that adapts the 
 structure-originally developed in [
31] for a second order plus time-delay system—to control higher-order time delay systems, specifically applied to the 
; (3) a comparative simulation analysis demonstrates that the proposed scheme achieves the same set-point tracking performance as the widely known 
 (for the nominal process) while greatly improving the attenuation of air disturbances caused by the fan; and (4) an experimental evaluation of this control scheme on a 
 prototype, confirming its superior performance compared to the standard 
. All comparisons were conducted fairly, using the same controller 
 in each case.
The 
 structure is highly sensitive to inaccuracies in the nominal time delay. This paper has not addressed the issue of varying time delays. However, this limitation can be mitigated by rapidly estimating the time delay in real time using algebraic identification techniques, which enables retuning of the internal model used in the 
. We refer to our recent work [
36], where an adaptive 
 was designed and implemented on a teleoperated flexible-link robot with time-varying delay, based on real-time delay estimation using such an algebraic identification technique.
If the hardware-induced delay were known, it could be incorporated into the nominal model’s time delay and, consequently, into the  structure. If not, an identification method—such as the aforementioned algebraic identification technique—could be employed to estimate the delay in real time and retune the Smith predictor accordingly.
Our  eliminates steady-state errors caused by impulse or step-like disturbances. Since we are using linearized models, this property is validated independently of the disturbance amplitude. Such amplitude may introduce a steady-state error and/or alter the response shape only if it is large enough to drive the system dynamics outside the valid region of the linearization, or if it causes actuator saturation. It is worth noting, however, that ramp or parabolic-like disturbances would still result in steady-state errors. Even in these cases, our  enhances the disturbance rejection capabilities compared to the original .
The main contribution of this work is the enhancement of flight control for unmanned aerial vehicles (s) of the helicopter type in the presence of strong air currents, thereby improving the system’s overall stability and performance. This advancement enables more precise and reliable operation, even under challenging environmental conditions.