1. Introduction
High-value near-space vehicles (NSVs), with specific applications such as near-space exploration, have great significance and are attracting growing interest in not only academia but also industry [
1,
2]. NSVs should be designed with a high power/weight ratio, a high level of intelligence, and excellent reliability, and their hardware systems should support intelligent control algorithms. As an essential NSV subsystem, the actuation system is applied to realize flight and attitude control; therefore, it is defined as a safety system [
3,
4]. To ensure flight safety and further improve reliability, a hydraulic actuator (HA) and an electro-hydrostatic actuator (EHA) are designed and integrated as a dissimilar redundant actuation system (DRAS); such systems are widely used in large civil aircraft to avoid common-mode failures [
5,
6,
7]. However, for NSVs with specific applications, an EHA and an electro-mechanical actuator (EMA) can be designed and integrated as a new type dissimilar redundant actuation system (NT-DRAS) with a higher power/weight ratio compared with former systems. An NT-DRAS can also provide fault-tolerant control (FTC) based on its dissimilar redundancy hardware [
8], leading to higher mission reliability. Therefore, NT-DRASs are optimized actuation system schemes for NSVs. Attitude-corresponding sensor failure can lead the NSV states to be undetectable, and the state feedback control law cannot be performed using invalid states. Therefore, both NT-DRAS and sensor failures are essential factors for NSV flight safety.
As an important piece of flight safety technology, the FTC technique can be used to deal with actuator failure problems [
9,
10]. In some research works, this problem is modeled as a reliable robust flight tracking control one, where only a linear matrix inequality (LMI)-based control gain needs to be solved [
11,
12]. This kind of FTC controller is passive, since there is no controller reconfiguration. The adaptive technique is usually used in this case to enhance the passive FTC controller’s level of control [
13]. Once the passive controller cannot cope with the fault conditions, for example, when multiple faults occur, a hybrid FTC mechanism including normal, passive, and reconfigurable controllers is more effective and can achieve the desired FTC performance, even under changing actuator fault conditions [
14]. However, these results do not consider sensor faults in a flight control system. A simultaneous fault condition with both actuator and sensor failures is more complex and severe, because the controllable actuation system channel is limited and some flight output states are unknown. This challenging issue has been studied in some research [
15,
16,
17].
To further improve the effectiveness of the FTC mechanism under possible simultaneous actuator and sensor failures, similar to the FTC types (passive FTC and active FTC), there are also passive and active methods that can be applied to estimate unknown or inaccurate dynamic states. Marcello et al. presented a “passive” FTC design method using neural networks [
15], where the fixed PFTC controller, based on pre-trained neural networks, can respond to inaccurate failure identification and sensor failures in the case of prior known faults. In a recent, neural-based strategy [
18], aircraft outputs were constrained. Active methods include observer-based ones; for example, Lin et al. presented an observer-based networked control scheme to cope with the sensor delay problem [
19]. To improve the output feedback controller performance, the authors integrated an observer link into the overall scheme to measure random sensor delays. Among the various observer-based methods, the fuzzy observer-based FTC scheme also has wide applications, such as nonlinear large-scale systems and so on [
20,
21].
Xiao et al. presented a sliding mode observer (SMO)-based method to reconfigure the FTC strategy to cope with actuator faults [
22], where an SMO is used to measure system uncertainty. Li et al. presented an observer-based attitude control strategy for rigid aircraft [
23], where the observer is also used to address model uncertainty. Yu et al. presented a sliding mode- and adaptive technique-based aircraft FTC strategy for simultaneous actuator and sensor fault conditions. The authors investigated the integrated design of SMO and FTC fight control, providing valuable results [
24]. Except for system-level applications such as flight control models or actuation systems, the observer-based method can also be used in power electronic converters. For example, Xie et al. studied the diagnosis and resilient control problem for multiple sensor faults in cascaded H-bridge multi-level converters by using a reduced-order observer [
25]. Xu et al. studied diagnosis problems for power switch, grid-side current sensor, and DC-link voltage sensor faults in single-phase three-level rectifiers [
26]. Besides the observer-based FTC method, space vector modulation (SVM) can also be used in fault diagnoses and control problems [
27]. These studies all focused on specific system failures. Compared with neural network- or SVM-based methods, the observer-based strategy is more widely used. However, this kind of scheme needs to be more flexible and able to cope with more complex and changing fault cases, even including cases where the numbers whereby degraded and/or failed actuators and sensors change over time [
28]. The most important function of a designed observer is to provide equivalent states for a real system without false alarms. To solve this problem, Doostmohammadian and Charalambous analyzed the observational equivalence problem and proposed a Q-redundant observer design method using sensor networks [
29]. The sensor-network-structure-based algorithm was able to recover sensor networks accurately with polynomial-order complexity.
For the NSV FTC problem, there is a considerable real-time requirement in performing the proposed algorithm; therefore, computational complexity should be avoided. Motivated by model-based monitoring and fault diagnosis methods, the NSV states can be estimated by using a similar concomitant system with the same original structure and input. In this paper, a concomitant observer-based multi-level FTC strategy is proposed. First, under normal operating conditions, a fixed-structure observer is designed to estimate all the NSV states, and in the case when there are no sensor faults, the state feedback control law will use the original NSV states rather than the estimated observer states. However, if sensor failures occur, leading to some unknown NSV states, they can be estimated by implementing the concomitant observer, which can estimate the sensor failures caused unknown NSV states. This approach allows a hybrid state-feedback-based FTC law to be reconfigured by using the remaining normal NSV states and the corresponding concomitant observer states. In the case of additional sensor failures, the FTC law can be generated again to cope with the new faults by changing the observer dimension and structure. Consequently, since the unknown states caused by the sensor failures can be regenerated by the concomitant-observer, a new FTC law can be used for different actuator fault conditions. The linear quadratic regulator (LQR) technique is used to adjust the FTC gain [
30].
Compared with previous works, the novelty of this paper is its use of a new FTC framework for NSV complex changing fault conditions. It contains three innovative elements: the NT-DRAS model in the NSV input link; a flexible concomitant observer for the real-time changing sensor fault conditions; and a multi-level FTC algorithm. The above elements are integrated as a schematism for NSVs. This has significant effects and advantages. Since the NT-DRAS is used in the NSV, in order to utilize the NT-DRAS dissimilar redundancy resources, different level FTC strategies (NT-DRAS level and flight control level) are performed to form a multi-level FTC mechanism. By using the proposed intelligent FTC mechanism described in this paper, NSVs can realize “never give up” FTC based on the fundamental channel reconfiguration when facing complex changing fault conditions, considering both NT-DRAS and sensor faults.
4. Case Studies and Simulations
To present the results well, the sequence of the subsection settings are first given a specification.
Section 4.1 presents the NSV model data and baseline control gain under normal conditions.
Section 4.2 presents the NSV fault scenario set, from Case 1 to Case 4, as the NSV fault severity worsens, which can be exhibited in the simulation results of both the NSV and actuation system responses, respectively.
Section 4.3 presents the LQR controller-based FTC results without considering the sensors faults, in which
Section 4.3.1 illustrates the LQR-based FTC mechanism,
Section 4.3.2 and
Section 4.3.3 illustrate different levels of FTC effectiveness, and
Section 4.3.4 analyzes a sequence of different level FTC strategies. To further cope with sensors faults, an observer should be designed; therefore,
Section 4.4 presents the proposed observer solved gain and the convergence results.
Section 4.5 presents the final and most important findings about the comprehensive actuator system and sensor system changing fault conditions.
Section 4.5.1 presents three different control strategies,
Section 4.5.2 illustrates the compared results of S-1 and S-3. Finally,
Section 4.5.3 illustrates the final compared results of S-2 and S-3 under different fault conditions, including a specific step signal tracking process and worsening comprehensive fault conditions, presenting the advantages of the proposed concomitant observer-based multi-level FTC algorithm.
4.1. System Model Matrices and Solved Baseline Control Gains
To perform simulation studies, the data are from a NASA GTM model [
34]. Under normal conditions, the model matrices are given as:
Under normal conditions, each NT-DRAS drives its corresponding control surface by using an EHA channel. All the working channels of the NSV are in good condition; therefore, the performance status matrix has the value
. The sensor system is also in a fault-free condition; therefore, the sensor system states matrix has the value
. By using the proposed algorithm in Theorem 1 and the LQR function in Matlab toolbox, the baseline control law can be solved as in Equation (23):
4.2. NSV Fault Scenario Set
The NSV fault case scenarios are provided in
Table 1. The complete failure and performance degradations are studied in [
14]. In Case 1, all the NT-DRASs and sensors are in fault-free condition. In cases-2–4, complex changing fault cases occur, since both the current NT-DRAS working channels and sensor system suffer from performance degradation and/or failures. The fault scenarios are set to gradually deteriorate, which can be seen in performance status matrices
and
.
Figure 6 shows the NSV roll action response under different fault cases, in which a step reference signal with 12° amplitude in roll angle is given as the NSV solved command.
As shown in the four cases in
Figure 6, the baseline controller under normal conditions is considered first. The original control gain,
is determined and applied to simulate the system’s dynamic responses. In Case 1, the given reference roll angle can be closely tracked with very small sideslip angle oscillation. In Cases 2–4 the two NSV output states diverge with increasing severity of the conditions, which indicates that the sensor failure is a fatal failure for the NSV and needs to be compensated for by effective FTC strategies. Meanwhile, the FTC controller should be designed to effectively reconfigure the comprehensive control gain under complex changing fault conditions.
Figure 7 illustrates the responses of all the NT-DRAS current channels under comprehensive fault conditions.
As shown in
Figure 7, under normal conditions (Case 1), all control surfaces can undergo normal deflections. However, in Cases 2–4, it can be seen that due to abnormal outputs from the NSV model attitude, the calculated input of the NT-DRAS current channels continuously fluctuate within the saturation angle range, or even in a saturated actuating state.
4.3. Effectiveness Verification of LQR-Based FTC Controller for the NT-DRAS Faults
4.3.1. LQR Solved FTC Control Gain
When the NSV suffers from multiple NT-DRAS current working channels faults, controller switching would be performed to maintain a certain level of flight control performance. The original NT-DRAS fault condition is considered as Case 2, in which the corresponding lower rudder and inside aileron NT-DRAS channels have completely failed, and the performance of the corresponding spoiler NT-DRAS channel degrades to 60%. Under this condition, the LQR-based FTC gain can be solved using Equation (24).
Figure 8 shows a comparison of the results for the NSV control surface deflections under Case 1 and Case 2.
As indicated in
Figure 8, the NT-DRASs corresponding to the inside aileron and the lower rudder working channels have completely failed, resulting in no responses during the time range. The reconfigurable FTC gain matrix in Equation (24) shows that the control gains for the outside aileron and the upper rudder have increased, resulting in increased response magnitudes, ultimately compensating for failures caused by the failed NT-DRAS current channels. Although there is performance degradation in the corresponding spoiler NT-DRAS current channel, it still can achieve the desired response with the increased FTC gains.
4.3.2. NT-DRAS Level-Based FTC Effectiveness Simulation
Figure 9 shows the comparison results of the NT-DRAS level-based FTC effectiveness, so that the FTC improvement of the NT-DRAS channel switching can be presented.
As shown in
Figure 9, the switching strategy here involves mitigating the performance degradation or complete failure of the current EHA channel by switching to the EMA channel. The switching occurs in the 40th second. From the response of the roll and sideslip angles at the flight control level, it is evident that after the switching, the roll angle fits the normal condition curve completely. Before the channel switching of the NT-DRAS, the sideslip angle of the NSV model increases by approximately
during the roll motion compared with the normal condition. However, after the NT-DRAS switching in the 40th second, it fully restores to the quality of normal condition, i.e.,
, as the EMA channel completely replaces the previously failed or degraded EHA channel. Consequently, both the roll and sideslip responses can reach a normal condition level.
4.3.3. Flight Control Level-Based FTC Effectiveness Simulation
Figure 10 shows a comparison of the results of the flight control level-based FTC effectiveness (before and after the control gain adjustment using LQR), so that the FTC improvement of the LQR adjustment can be presented.
As shown in
Figure 10, by using the LQR controller-based flight control level FTC strategy, all the NT-DRASs do not switch channels. Instead, the control gains at the flight control level are globally adjusted through the LQR controller, essentially compensating for the required actions of the failed or degraded control surfaces by increasing the gains of other effective control surfaces. The switching occurs in the 40th second. Before switching, it can be observed from the response of the NSV model’s roll and sideslip angles that the response of the roll angle is basically the same as that of the normal condition. However, due to partial NT-DRAS channel failures, the sideslip angle fluctuates to about twice the normal level. After the gain adjustment through the LQR controller, the response of simulated sideslip angle in the subsequent 40s shows that although the sideslip angle does not completely overlap with the situation before the LQR controller gain adjustment, the amplitude of fluctuation is essentially suppressed to the level of the normal condition (
).
4.3.4. Summary of the Different Level-Based FTC Effectiveness Analysis
As can be seen from the simulation effect of the two-level switching strategies in
Figure 9 and
Figure 10, the NT-DRAS channel switching and the LQR gain adjustment are able to realize the desired FTC performance for the NSV model, whereby the NT-DRAS channel switching strategy at the actuation system level has the best FTC effect because it adopts a new channel to completely replace the faulty channel. The cost is that the backup margin resources are enabled. The LQR gain adjustment at the flight control level can basically achieve the same performance as that of the normal operating condition, but at the cost of a high level of controller gain changes, and the large gain may bring problems of energy consumption and the adaptability of the full operating conditions. In order to form a hierarchical FTC strategy, the solution given in this paper is based on the principle of safety combined with performance. During the first FTC process, the LQR global gain adjustment at the flight control level is carried out in the event of NT-DRAS channel failure or performance degradation. When this gain adjustment cannot meet the NSV safety and performance requirements, the faulty NT-DRAS current working channel is switched. This FTC measure is carried out at the expense of the residual resources. Finally, if all the NT-DRASs of the actuation system level have already carried out channel switching and no residual resources are left, then the LQR secondary global gain adjustment at the flight control level is used to perform FTC again, thus forming a hierarchical multi-level FTC strategy to cope with ever-increasing faults.
4.4. Verification of the Designed Observer Convergence
4.4.1. Solved Observer Gain Matrix
To generate the feedback control law when sensor system faults occur, the concomitant observer states are used to replace the sensors failures caused unknown NSV states. The concomitant observer gain can be solved by using the algorithm in Theorem 1. To improve the convergence efficiency, the adjustment coefficient was chosen as
. Finally, the used observer gain may be determined as follows:
4.4.2. Observer Convergence Simulation Verification
In order to guarantee that the concomitant observer can achieve convergence to the real NSV model, so that the concomitant observer states can be used to perform the FTC law, different command signals, including step one as well as sine one, are used to simulate the observer convergence.
Figure 11 and
Figure 12 show the error curves between the observer and the actual NSV model when given step and sine signals respectively, so that the observer convergence can be verified.
In
Figure 11 and
Figure 12, the given roll angle reference signal has a maximum magnitude value of 12°. The maximum observation error of the roll angle is about 2°, whose amplitude is small and can be converged to 0 in about 5s. The maximum amplitude of the error curve of the roll angle rate is about 7°/s, which can be converged to the actual system as well. Though this rate is not the final output of the NSV model, the error amplitude and convergence results also meet the observer design requirements. The estimated error of the sideslip angle is about 0.5° or less and can eventually converge. The estimated error amplitude of the yaw angular velocity is within 0.5°/s, which also meets the convergence requirement. These simulation results can be used to verify the algorithm in Theorem 1.
4.5. Response Comparisons of Sensor and NT-DRAS Fault Conditions
4.5.1. FTC Strategy Illustrations
To compare the effectiveness of the FTC, the following different strategies are given:
Strategy-1(S-1). A baseline controller-based control strategy (the NSV state feedback-based control law without observer design).
Strategy-2(S-2). A fixed structure observer-based LQR-FTC strategy.
Strategy-3(S-3). A concomitant observer-based LQR-FTC strategy.
4.5.2. Response Comparison Using S-1 and S-3 under the Changing Fault Conditions in Case 2 and Case 3
Figure 13 shows a comparison of the tracking performance using S-1 and S-3. The concomitant observer dynamic response and the real system responses using the baseline control law and the FTC law are presented in this figure.
As shown in
Figure 13, even though some current NT-DRAS working channels are suffering performance degradations or failures, and in addition to the sensor faults, the NSV roll response command can still be tracked well using S-3. Compared with the concomitant observer response, there is only a little response lag. However, by using S-1, the roll state yields no response. Meanwhile, the sideslip response can retain a small oscillation value by using S-3 while the sideslip response diverges by using S-1. The comparison shows that S-3 is still effective under simultaneous NT-DRAS and sensor fault conditions. The quantization statistics in
Table 2 also indicate the same conclusion.
4.5.3. Response Comparison Using S-2 and S-3 under Changing Fault Conditions from Case 1 to Case 4
Hybrid State Generation Mechanism in S-3
Figure 14 shows the hybrid state generation principle when using the concomitant observer-based FTC method (S-3).
As shown in
Figure 14, under sensor fault cases 2–4, the hybrid states can be integrated by using the concomitant observer states and the NSV left normal states. By changing the corresponding dimension of the concomitant observer, only if the unknown states due to sensor failures are obtained can the observation errors of the left normal states be avoided.
To verify the advantages of S-3 and compare it with S-2, the fault conditions are assumed to change. In every specific case, the concomitant observer can change its corresponding dimension to obtain the missing unknown states. Composing the concomitant estimated states with the NSV left normal states can result in the concomitant observer-based FTC law, as shown in Equation (26). The missing roll rate is estimated as
, while the missing roll and yaw rates are estimated as
and
, respectively, in Case 2. The missing roll and yaw rates and roll angle are estimated as
,
, and
, respectively, in Case 4. The concomitant observer changes its corresponding dimensions with the failed sensor while the observer gain matrix also changes to fit with the specific cases. However, S-2 always uses the originally determined gain matrix, even under changing conditions.
The control gain under the normal condition is solved as
in Equation (23), while the FTC gain under Case 2 is solved as
in Equation (24). The FTC gain under Case 3 is solved as
in Equation (27). Since in Case 4, some current NT-DRAS channel performance indices are too low, the solved FTC gain based on these indices cannot achieve the desired control performance. Therefore, the LQR-based flight control gain adjustment is not used for Case 4; rather, the NT-DRAS channel switching strategy is used so that NT-DRASs performance health status matrix
can be updated to a new one as
. Therefore, the control gain under Case 4 can be chosen as
.
Response comparison using S-2 and S-3 when given specific step signal.
Figure 15 shows a tracking response performance comparison result using S-2 and S-3 when given a step signal. The dynamic performance improvement when using concomitant observer-based FTC is illustrated in this figure.
It can be seen from
Figure 15 that when a step signal is given to the NSV model, both S-2 and S-3 can make the NSV respond to the given command effectively, although the response performances are different. The rising time of the NSV roll response is about 5 s using S-3, while it is about 10 s using S-2. The maximum oscillation value of the sideslip angle is 0.05° using S-3, while it is about 0.5° using S-2. The dynamic response results show that S-3 can achieve better FTC performance not only in terms of a quick response, but also in terms of stability and flight control quality. The quantization statistic results in
Table 3 also yield the same conclusion.
Response comparison using S-2 and S-3 under changing fault conditions from Case 1 to Case 4
For the changing and worsening cases of the NSV model, the fault changing times are set as 13 s (Case 1 to Case 2), 43 s (Case 2 to Case 3), and 73 s (Case 3 to Case 4). It is assumed that the FDD mechanism can detect the fault information immediately and the FTC measures can also be performed quickly.
Figure 16 shows a tracking response performance comparison when the NSV maintains a constant roll angle using S-2 and S-3. The tracking performance improvement when using concomitant observer-based FTC is illustrated in this figure.
As shown in
Figure 16, the fault change occurs when the NSV maintains a constant roll command angle of 12°. The roll angle fluctuations are small by using both S-2 and S-3 (see
Table 4: roll response performance). However, the NSV can achieve better stability by using S-3. It can be seen from the fault changing process from Case1 to Case 2 that S-3-based FTC can assure a smaller sideslip angle and better roll action performance. It can be seen from the fault changing process from Case 2 to Case 3 that the NSV can have a smaller sideslip angle by using S-3 compared with S-2 (a longer time is necessary for the sideslip angle converges to 0°). When the fault condition changes to Case 4, the fault NT-DRAS channels have to be switched to update the health indices. Though the roll angle fluctuation is bigger when using S-3, it can result in very small sideslip angle, i.e., within 0.05°, while it is more than 0.12° with S-2. It can be concluded that the concomitant observer-based S-3 FTC strategy can result in better tracking performance because of the hybrid state-based FTC control law. It can be seen from
Table 4 that the sideslip response fluctuation can be reduced about 50% by using S-3.
Figure 17 shows a tracking response performance comparison when the NSV has a dynamic changing roll action using S-2 and S-3. During this dynamic process, when the faults occur and the fault cases become worse, the FTC advantage of S-3 is illustrated in this figure.
As shown in
Figure 17, the fault change occurs when the NSV has a dynamic changing roll action (0°→12°→−12°→0°). From the dynamic process from Case 1 to Case 2, both S-2- and S-3-based FTC strategies can maintain the desired NSV tracking performance, while S-3-based FTC can achieve a smaller sideslip angle. The dynamic process from Case 2 to Case 3 shows that the NSV can achieve a quicker roll step signal response with no fluctuations by using S-3 compared with S-2 (an overshoot angle of more than 5° occurs with fluctuations). Meanwhile, the sideslip angle is within a limited amplitude (0.1°) when using S-3, while it fluctuates during within an interval of −0.6° to 0.4° when using S-2. When the fault condition changes to Case 4, all the faulty NT-DRAS channels are updated. The roll action is still quicker with no overshoot when using S-3 compared with S-2. Additionally, the maximum sideslip angle is within 0.1° when using S-3, while it is bigger than 0.4° when using S-2.
Table 5 shows that the roll response tracking error can be reduced by about 6.85% by using S-3. The sideslip response fluctuation can be reduced by about 81.98% by using S-3. It can concluded that the concomitant observer-based S-3 FTC strategy is more suitable for the changing fault conditions compared with the fixed structure-based S-2 FTC strategy.