1. Introduction
The suspension system is one of the key structural and functional components for vehicles [
1,
2]. Compared with passive and active suspension systems, the semi-active suspension system is regarded as a more promising candidate due to its cost-effectiveness, especially for its ability to improve both ride comfort and handling stability close to the active one [
3]. Recently, semi-active suspension systems featuring magnetorheological (MR) dampers have received much attention [
4]. The MR damper has many clear advantages, such as a simple structure, low energy consumption, and wide range. However, it should be noted that two key questions still require further attention. The first is the compensation issue dealing with the response delay of the MR damper, and the other is the accurate control coping with modelling error and real-time variety of the output force [
5]. Thus, to obtain better suspension performances, the modelling error, real-time variety, and response delay of the MR damper must all be considered in designing the controller. To regulate the MR damper to give the objective output force, the control current must be solved according to the expected control force, the movement state, and the characteristics model of the MR damper.
Many researchers have tried to remedy the time delay, modelling error, and time-varying mechanical characteristics of the MR semi-active suspension system. There are two main methods for dealing with the time delay problem. In the 1950s, a so-called Smith predictor was designed and widely used to cope with the system time delay. The Smith predictor worked by introducing a compensation block in the feedback loop to offset the influence of time delay on the system [
6]. However, the effectiveness of the Smith predictor depended on the modelling accuracy [
7]. Recently, another time delay compensation control method has become increasingly popular [
8]. The Lyapunov–Krasovskii generalized function containing the time delay of the system was constructed. Then, its integral terms were eliminated using different integral inequalities to transform the integral of a vector product into the product of vector integrals to obtain an enlarged linear quadratic matrix. Moreover, the linear matrix inequality (LMI) solver was used to calculate the solution, ensuring that the linear quadratic matrix was negative definite. The different Lyapunov–Krasovskii generic derivatives required linearization using one or more integral inequalities [
9,
10]. Therefore, the control solution conservativeness from the specific Lyapunov–Krasovskii general function was different, which in turn led to different time delay compensation control effectiveness [
11]. Since bounding methods introduce some conservativeness, many studies have been conducted to address this problem, such as the Park inequality method, Moon inequality method, and Jensen inequality method [
12].
Investigation into the time-varying mechanical characteristics of the MR damper showed that it was caused by temperature change, component ageing, and ferromagnetic particle deposition [
13]. Several studies have been conducted to address the influence of temperature change. Dong et al. [
14] developed a sliding mode fault-tolerant controller dealing with the influence of temperature change on the MR damper as one of the model uncertainties. The method used a simple model to improve robustness at the cost of a declined suspension performance. It was noted that the temperature was taken as a variable in modelling the MR damper [
15]. Although the accuracy of the MR damper model was improved, the forwards and inverse models became complex. Among these above factors, the levels of component ageing and ferromagnetic particle deposition were difficult to measure. Obviously, in addition to temperature, taking component ageing and ferromagnetic particle deposition into account has to bring about more complexity.
Hence, there is a demand for effective countermeasures of both response time delay and time-varying mechanical characteristics to polish the MR semi-active suspension. This demand motivates us to develop an augmented Kalman estimator and equivalent replacement-based Taylor series-linear quadratic Gaussian (ERBTS-LQG) controller to provide a high observation accuracy for the real-time variety of the MR damper output force and improve the time delay compensation performance.
The main contributions of this research include two aspects: (1) an augmented Kalman estimator is established to observe the compound real-time variety of the MR damper output force; and (2) a novel equivalent replacement technology is presented to ensure that the Taylor series-LQG-based regulator is feasible and precisely tracks the control objective.
The rest of this paper is organized as follows:
Section 2 addresses the nonlinear model of the MR semi-active suspension system. In
Section 3, the ERBTS-LQG control scheme is presented.
Section 4 provides a superiority verification of the proposed control scheme. Finally, the research is concluded in
Section 5.
3. Scheme for Time Delay Compensation Control
The scheme of the augmented Kalman estimator and the ERBTS-LQG controller is illustrated in
Figure 7.
The proposed scheme involves three key steps.
- (1)
The system state equation is augmented by the differential equation of the compound time variation of the MR damper output force.
- (2)
The corresponding augmented Kalman estimator is designed to observe the system state vector and the compound real-time variety.
- (3)
The ERBTS-LQG controller is developed by replacing a part of the squared time delay control force in the suspension comprehensive performance index with the corresponding predictive control force.
3.1. Design of the Augmented Kalman Estimator
For the sake of convenience, write
as follows:
Here, we name the compound real-time variety of the MR damper output force.
To ensure the augmented system state equation with the characteristic of the minimum system,
is written in the differential equation as follows [
19]:
where
is the related parameters of
, and
indicates the positive infinitely small quantity.
Substituting Equations (12) and (13) into Equation (11) yields the augmented system state equation as follows:
Considering that the acceleration is easy to detect, and are chosen as the outputs of the MR semi-active suspension system. In this work, to improve observation accuracy, integrals of the two accelerations are chosen as the inputs of the augmented Kalman estimator.
The corresponding output equation is expressed as
The augmented Kalman estimator is designed as follows:
where
stands for the estimate of
;
represents the unique solution of the following Riccati equation.
where
is the measurement noise vector.
Therefore, using the newly developed Kalman estimator, in addition to the system state vector , the compound real-time variety of the MR damper output force is estimated.
3.2. Design of the ERBTS-LQG Controller
Usually,
is expressed as
where
refers to the active control force.
To obtain better suspension performance, the constant time delay of the MR damper must be considered in the controller sign. When the Taylor series-LQG strategy is applied, the active predictive control force
and the corresponding
are expressed as follows:
Substituting Equation (21) into Equation (14) yields the state equation as follows:
On the basis of
,
, and Equation (10), the cost function of the ideal time delay compensation system is written as
In Equation (24), does not satisfy the second LQG controller design condition that .
To solve the abovementioned problem, the approximate transformation can be expressed as follows:
Although is avoided, that method will give an overly large value of . This factor could be due to the following two reasons. The first is that, by using Equation (25), the control objective becomes inaccurate. The other is that Equation (25) brings a considerably small R that cannot effectively constrain .
To obtain a better time delay compensation control effect, Corollary 1 is proposed to effectively constrain the predictive control force to precisely track the control objective .
Corollary 1. when sufficiently large, positive satisfies and the following active forceprecisely tracks the control objective .
where is the unique solution of the Riccati equation as follows: Proof. Implementing the Fourier transform on
yields
where
and
is the circular frequency. □
Then, the 2-norm value of
can be written as follows:
Writing Equation (32) into the time domain form yields
According to Equations (21) and (33), the following equation can be obtained.
Hence, all the design conditions of the LQG controller are satisfied, and precisely tracks . The proof is completed.
By using Equations (26), (29) and (34), a part of the squared time delay control force in the suspension comprehensive performance index is equivalently replaced by the corresponding part of the predictive control force. Because this replacement is completely equivalent, this novel time delay compensation control method is called the equivalent replacement-based Taylor series-LQG control approach. It should be noted that the above operation is proposed first in this work and is difficult to realize by using the control method.
Therefore, the corresponding semi-active predictive control force
is as follows:
Because
, the predictive control current
is calculated as follows:
where
is the maximum control current to the MR damper.
In Equation (37), by replacing with , we can avoid the issue of not being calculated when .
4. Superiority Verifications
From the point of view of vehicle weight and operating conditions, the abovementioned MR semi-active suspension system can be fitted on sport or utility vehicles.
The driving condition of v = 50 m·s
−1 over the D-class road (D-cr) is investigated, which corresponds to a poorly constructed road with a relatively low velocity [
20]. Under the v = 50 km/h over the D-class road, the value of
presented in Equation (2) is set as 1024 × 10
−6 m
−1. From Equation (2), we can obtain the displacement input curve in
Figure 8.
The other vehicle parameters are listed in
Table 2.
4.1. Observation Accuracy of the Augmented Kalman Estimator
To clearly display the work effectiveness of the augmented Kalman estimator, four cases are checked:
Case 1: A small modelling error with ;
Case 2: A large modelling error with ;
Case 3: A considerably large real-time variety with ;
Case 4: A considerably large and fast real-time variety with .
The observation error and the corresponding root mean square (RMS) error are defined as follows:
where
is the observation of
and
denotes the function calculating the root mean square.
The observation effectiveness of the augmented Kalman estimator is described in
Figure 9 and
Figure 10. In these figures, “OBF” and “EOB” stand for the observation of
and the observation error, respectively.
The observation errors, the RMSEs, and the corresponding suspension comprehensive performance indexes of those five cases are listed in
Table 3.
Figure 9 and
Figure 10 show that all
curves track the
curves well under the above investigation cases. Moreover,
Table 3 shows that the root mean square of
is at almost the same level of less than 2. When a modelling error of 5% exists, the worst observation accuracy occurs, and the RMSE equals 0.0549. Compared with the integral of
, the RMSE is just 0.0027. This is a very high observation accuracy. Even though there are 10% modelling errors and 30% real-time variations, compared with the integral of
, the RMSE is just 0.0031. In addition, subjected to those investigation cases, the maximum difference in the suspension comprehensive performance index is equal to 0.0048.
The above investigation results demonstrate that the augmented Kalman estimator is a perfect solution for both the modelling error and real-time variety of the MR damper output force.
4.2. Work Effectiveness of the ERBTS-LQG Controller
It is assumed that the MR semi-active suspension suffers from both modelling error and real-time variety. To verify the superiority of the ERBTS-LQG controller, the sprung mass acceleration, the tire deflection, the suspension deflection, and the suspension comprehensive performance index are compared among the passive suspension, the time delay free semi-active suspension, and the MR semi-active suspension regulated by the proposed time delay compensation controller. These three suspensions are denoted by “PS”, “ISAS”, and “MRSAS”. The ideal active force (IAF), the ideal semi-active force (ISAF), and
curves in the time domain are illustrated in
Figure 11. The comparison curves of the suspension performances in the frequency domain are drawn in
Figure 12 and
Figure 13.
At the turning point when the ideal semi-active force is in need or disappears,
Figure 11 illustrates that
and the ideal semi-active force are almost synchronous. This means that after a time delay, the predictive semi-active force can accurately track the ideal semi-active force. Additionally,
is less than the ideal semi-active force. The reason can be explained as follows: the actual MR damper output force is constrained by the maximum control current of 3 A.
Figure 12 and
Figure 13 clearly illustrate that both the tire deflection and the suspension deflection of the MR semi-active suspension are almost the same as those of the ideal suspension. Moreover, compared with the passive suspension, the proposed ERBTS-LQG controller regulates the MR semi-active suspension to obviously reduce the suspension comprehensive performance index and the peak value of the PSD of the unsprung mass acceleration. Moreover, the PSD of tire deflection for ISAS and MRSAS are worse compared to the PS in the frequency range from 10 to 20 Hz. This is because the driving condition of v = 50 m·s
−1 over the ISO D-class road (D-cr) is regarded as a low-velocity driving condition. In this condition, the riding comfort is usually given priority. To obtain a better riding comfort, a larger weight coefficient of sprung mass acceleration is selected, which results in unsatisfactory improvement of tire deflection. This result is reasonable. The RMSs of those suspension performance indicators are listed in
Table 4.
In this table, compared with the passive suspension, the MR semi-active suspension and the ideal semi-active suspension decrease the unsprung mass acceleration by 21.16% and 24.66%, respectively. For the suspension comprehensive performance index, the corresponding decreases are 23.99% and 26.52%, respectively.
Figure 12 and
Figure 13 and
Table 4 show that the unsprung mass acceleration and the suspension comprehensive performance index of the ideal semi-active suspension are just slightly better than those of the MR semi-active suspension using the proposed ERBTS-LQG controller. The reason can be explained as follows. First, by using Equation (21), there is a system error in calculating the active predictive control force. Moreover, when an active force with negative damping characteristics is needed, unlike the ideal semi-active suspension, which can be given a control force equal to zero, the MR semi-active suspension inevitably suffers from an uncontrolled positive damping output force.
To check the adaptability of the proposed ERBTS-LQG controller, some other large time delay cases are investigated. These cases refer to time delays equal to 16 ms, 24 ms, and 32 ms. Comparison curves of the unsprung mass acceleration and the suspension comprehensive performance index are depicted in
Figure 14 and
Figure 15, respectively. Statistical data of the four suspension performance indicators are provided in
Table 5.
Figure 14 shows that when the time delay changes from 16 ms to 32 ms, the PSD of the unsprung mass acceleration only slightly increases around the natural frequency of wheel vibration. Similarly,
Figure 15 clearly illustrates that the suspension comprehensive performance index only slightly degrades. It can be seen in
Table 4 and
Table 5 that when the time delay increases from 8.5 ms to 16 ms, 24 ms, and 32 ms, the RMS of the unsprung mass acceleration is enlarged by 0.93%, 1.83%, and 3.01%, respectively. The corresponding increases in the suspension comprehensive performance index are 1.01%, 1.46%, and 2.49%, respectively.
The above results indicate that when subjected to a time delay from 8.5 ms to 32 ms, the ERBTS-LQG controller always regulates the MR semi-active suspension to achieve very near control effectiveness of the ideal time delay-free semi-active suspension. In particular, the proposed controller has excellent time delay compensation ability and adaptability.