CN101580064B - Adaptive control method for controlling vibration of vehicle - Google Patents

Abstract

The invention discloses an adaptive control method for controlling the vibration of a vehicle, comprising the step of adjusting the weight coefficient of a wave filter based on an LMS adaptive algorithm to lead the quadratic performance index of target parameters to be minimal. The method can be used for controlling the target parameters such as the vertical vibration acceleration of a vehicle body, a dynamic load between wheels and a road surface, the dynamic deflection of a suspension system, and the like, and controlling the vibration of the vehicle in real time, and can simply and effectively improve the driving smoothness of the vehicle.

Description

The self-adaptation control method that is used for Vehicular vibration control

Technical field

The present invention relates to a kind of control method, relate in particular to a kind of self-adaptation control method that is used for Vehicular vibration control.

Background technology

Automobile vibration is to influence vehicle running smoothness and road-holding property key factor.Vehicle vibration damping mainly uses suspension system, generally is made up of elastic element and damping element.In order to the disturbance force that buffering and absorption produce because of Uneven road, the roll force that produces when bearing motor turning simultaneously.And the two is contradiction in automobile design, is difficult to satisfy simultaneously this requirement based on the conventional suspension systems of classical theory of vibration isolation.

In the prior art, the automotive suspension vibration control system picks up vehicle body absolute velocitye, vehicle body to the relative velocity of wheel, the signals such as acceleration/accel of vehicle body by sensor mostly, through Computer Processing and send the instruction control.Regulate the damping coefficient or the control effort of shock absorber by actuating units such as electric hydraulic control valve or stepping motors.

There is following shortcoming at least in above-mentioned prior art:

In real system implementation algorithm need square, computing such as average or differential, calculation of complex, efficient are low.

Summary of the invention

The purpose of this invention is to provide a kind of self-adaptation control method that is used for Vehicular vibration control simply, efficiently.

The objective of the invention is to realize through following technical scheme:

The self-adaptation control method that is used for Vehicular vibration control of the present invention comprises and adopts the LMS adaptive algorithm that target component is controlled, and makes the quadratic performance index of said target component reach minimum through the weight coefficient of adjusting filter.

Technical scheme by the invention described above provides can be found out; The self-adaptation control method that is used for Vehicular vibration control of the present invention; Owing to adopt the LMS adaptive algorithm that target component is controlled, make the quadratic performance index of said target component reach minimum through the weight coefficient of adjusting filter.When being used for Vehicular vibration control, can improve the ride comfort of running car simply, efficiently.

Description of drawings

Fig. 1 is the principle signal of LMS adaptive line combiner in the specific embodiment of the present invention;

Fig. 2 is the principle signal of LMS adaptive transversal filter in the specific embodiment of the present invention;

Figure 3a is a single-frequency excitation sprung weight acceleration time domain response curve in the specific embodiment of the present invention;

Figure 3b is single-frequency excitation sprung weight acceleration/accel frequency response curve figure in the specific embodiment of the present invention;

Figure 4a is single-frequency excitation dynamic wheel load time-domain response curve figure in the specific embodiment of the present invention;

Figure 4b is single-frequency excitation dynamic wheel load frequency response curve figure in the specific embodiment of the present invention;

Figure 5a is single-frequency excitation suspension dynamic deflection time-domain response curve figure in the specific embodiment of the present invention;

Figure 5a is single-frequency excitation suspension dynamic deflection frequency response curve figure in the specific embodiment of the present invention;

Fig. 6 is LMS adaptive control error e(n in the specific embodiment of the present invention) the change procedure diagram of curves;

Fig. 7 is LMS adaptive control weight convergence conditional curve figure in the specific embodiment of the present invention;

Fig. 8 is a sprung weight acceleration power spectral density diagram of curves under the road excitation in the specific embodiment of the present invention;

Fig. 9 is dynamic wheel load power spectral density plot figure under the road excitation in the specific embodiment of the present invention;

Figure 10 is road excitation lower suspension dynamic deflection power spectral density plot figure in the specific embodiment of the present invention.

The specific embodiment

The self-adaptation control method that is used for Vehicular vibration control of the present invention; Its preferable specific embodiment is; Comprise and adopt the LMS adaptive algorithm that target component is controlled, make the quadratic performance index of said target component reach minimum through the weight coefficient of adjusting filter.

Described quadratic performance index can comprise with the next item down or multinomial: error signal mean square value, average power.

Said filter is LMS adaptive transversal filter or LMS adaptive line combiner.

Said LMS algorithm can be regulated said weight coefficient according to the negative gradient of single error signal variance.

You can take a single error signal squared gradient

as the mean square error function gradient

estimate.

Can be through this method: the dynamic load between bouncing of automobile body acceleration/accel, wheel and road surface, the dynamic deflection of suspension system to estimating with the next item down or multinomial target component; And improve the ride comfort of running car through control to above-mentioned target component.

When improving the ride comfort of running car through control to multinomial target component, the bouncing of automobile body acceleration/accel evaluating of attaching most importance to.

The present invention adopts adaptive control technology; State and parameter through on-line identification controlled object constantly in control process; The parameter of timely adjusting control device; Eliminate because unknown, time becomes and the adverse effect that the performance of control system is produced such as non-linear, thereby controlled system is operated under the well behaved level.

Auto-adaptive filtering technique is debated in the knowledge process in system, and the number of required input Serial No. was few when system parameter was done certain variation, and required total calculated amount (in multiplication and division, add, subtract, square etc. operation times) few, have tangible rapidity.

Adaptive filter algorithm has a variety of, LMS(Least Means Squares) the algorithm computing is simple.The present invention is based on of the application of the horizontal sef-adapting filter of LMS algorithm,, propose the suspension system of LMS adaptive control to vehicle suspension 1/4 model in signal conditioning.

Combine accompanying drawing that scheme of the present invention is carried out detail analysis through specific embodiment below:

As shown in Figure 1, the onrecurrent sef-adapting filter also can be described as the adaptive line combiner, for by element x < > 0 <> , x < > 1 <> , Λ, x < > L-1 <> The signal vector of forming, one group of adjustable power correspondingly is w < > 0 <> , w < > 1 <> , Λ, w < > L-1 <> For one group of fixing weights, its output is the linear combination of input component.For single input system, incoming signal can be thought the different time serieses constantly of same signal, and as shown in Figure 2, sef-adapting filter can be realized by adaptive line combiner and unit delay unit, thereby constitute adaptive transversal filter.

As shown in Figure 2, be y(n if establish the linear estimate of sef-adapting filter), the true response of simultaneity factor is d(n), then have

e ²(n)＝[d(n)-y(n)] ² (1)

Be the square error that becomes with sequential n.Definition

ε(n)＝E{e ²(n)} (2)

Wherein, E{} represents the expectation value of amount in the {}, i.e. ensemble average.Since average when ensemble average is not, so ε (n) is the function of moment t, be defined as error of mean square (MSE).

According to horizontal sef-adapting filter structure, have

$y\left(n\right)=\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{w}_i\left(n\right)\&CenterDot;x(n-i)=W^T\left(n\right)\&CenterDot;X\left(n\right)---\left(3\right)$

Wherein, W < > T <> (n)=[w < > 0 <> (n), Λ, w < > L-1 <> (n))

X ^T(n)＝[x(n)，Λ，x(n-L+1)] (5)

Because y(n) be incoming signal x(n-i) linear function (w < > i <> (n) be coefficient in the linear equation), so by x(n-i) filter output y(n) process be called linear filtering.

The adaptive signal processing algorithm generally is used for obtaining relevant adjustable parameter, and making with these parameters is objective function minimum under certain criterion of variable, and the mean square value of evaluated error is a kind of criterion commonly used.Therefore, distinguish from criterion, adaptive signal processing method can divide two types of :(1) the random statistical method; (2) accurate method.In accurate method, criterion has comprised the true definite input data that obtained, and in the random statistical method, criterion has comprised the statistical nature of input data.The derivation of algorithm is the basis with the ensemble average (being error of mean square) of evaluated error square.

To formula (1), (2), (3), find the solution W < > T <> (n), make ε (n) minimum.Because ε (n) is each coefficient w < > i <> Function, can be written as ε (W).Can know by above-mentioned formula

ε(W)＝E[e ²(n)]＝E{[d(n)-W ^TX(n)] ²} (6)

Wherein, W < > T <> Be W < > T <> Slightly writing (n), the situation when studying network characteristic at present and being in plateau.

Launch by formula (6),

ε(W)＝E{d ²(n)}-2W ^TE{d(n)X(n)}+W ^TE{X(n)X ^T(n)}W (7)

The right side first term is manipulated signal d(n in the formula (7)) mean square power, second expression d(n), x(n) the cross-correlation amount, be expressed as Rxd.Rxd is a time-varying vector (being a constant vector when stable state), and its each element is expressed from the next:

$R_{\mathrm{xd}}=\left[\begin{array}{c}E\left\{d\left(n\right)x\left(n\right)\right\}\\ E\left\{d\left(n\right)x(n-1)\right\}\\ \mathrm{\&Lambda;}\\ E\left\{d\left(n\right)x(n-L+1)\right\}\end{array}\right]=\left[\begin{array}{c}{R}_{\mathrm{xd}}\left(0\right)\\ R_{\mathrm{xd}}\left(1\right)\\ \mathrm{\&Lambda;}\\ R_{\mathrm{xd}}(L-1)\end{array}\right]---\left(8\right)$

R wherein < > Xd <> (m) be cross-correlation coefficient, be defined as

R _xd(m)＝E{d(n)x(n-m)} (9)

In actual process, real R < > Xd <> (m) value is a unknown number often, need estimate.The general method of estimation that adopts is: for steady ergodic stochastic signal x, y, and its corresponding R < > Xy <> (m) can estimate by following formula

$R_{\mathrm{xy}}\left(m\right)=\frac{1}{k-\left|m\right|}\underset{i=0}{\overset{k-\left|m\right|-1}{\mathrm{\&Sigma;}}}x(n-i)\&CenterDot;y(n-|m|-i)---\left(10\right)$

K data number when calculating in the formula, general run of thins get k＞＞| m|.

The E{X(n)X of formula (7) right side in last < > T <> (n)} represents x(n) the autocorrelation battle array, be designated as R < > Xx <>

R _xx＝E{X(n)X ^T(n)} (11)

For stationary signal, R < > Xx <> Irrelevant with n, R < > Xx <> Can be written as

$R_{\mathrm{xx}}=E\left\{\left[\begin{array}{c}x\left(n\right)\\ x(n-1)\\ M\\ x(n-L+1)\end{array}\right]\right[x\left(n\right),x(n-1),\mathrm{\&Lambda;},x(n-L+1)\left]\right\}---\left(12\right)$

Then according to formula (8), R < > Xx <> Can be written as

$R_{\mathrm{xx}}=\left[\begin{array}{cccc}{\mathrm{\&phi;}}_x\left(0\right)& {\mathrm{\&phi;}}_x\left(1\right)& \mathrm{\&Lambda;}& {\mathrm{\&phi;}}_x(L-1)\\ {\mathrm{\&phi;}}_x\left(1\right)& {\mathrm{\&phi;}}_x\left(0\right)& \mathrm{\&Lambda;}& {\mathrm{\&phi;}}_x(L-2)\\ \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}\\ {\mathrm{\&phi;}}_x(L-1)& {\mathrm{\&phi;}}_x(L-2)& \mathrm{\&Lambda;}& {\mathrm{\&phi;}}_x\left(0\right)\end{array}\right]---\left(13\right)$

φ in the formula < > x <> (l) be l rank cross-correlation coefficient.

With formula (8), (12) substitution formula (7),

ε(W)＝E[d ²(n)]-2W ^TR _xd+W ^TR _xxW (14)

By formula (14) visible, error of mean square ε (W) obviously is the weights (functions of filter coefficient)W.If get L=2, error of mean square is the quadratic function of weights, and formula (7) becomes

$\mathrm{\&epsiv;}\left(W\right)=\mathrm{\&epsiv;}({\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1,w_2)={\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1^2{\mathrm{\&phi;}}_x\left(0\right)+2w_1{\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2{\mathrm{\&phi;}}_x\left(1\right)+{\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2^2{\mathrm{\&phi;}}_x\left(0\right)$

${-2w}_1{R}_{\mathrm{xd}}\left(0\right)-{2w}_2{R}_{\mathrm{xd}}\left(1\right)+E\left[d^2\left(n\right)\right]---\left(15\right)$

Utilize differential zero setting method to find the solution, order

$\left\{\begin{array}{c}\frac{\&PartialD;}{\&PartialD;{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1}\mathrm{\&epsiv;}({\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1,w_2){|}_{{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1,{\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2={\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1^{*},{\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2^{*}}=0\\ \frac{\&PartialD;}{{\&PartialD;\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2}\mathrm{\&epsiv;}({\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1,w_2){|}_{{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1,{\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2={\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1^{*},{\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2^{*}}=0\end{array}\right.---\left(16\right)$

Get

$\left\{\begin{array}{c}{2\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1^{*}{\mathrm{\&phi;}}_x\left(0\right)+{2\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2^{*}{\mathrm{\&phi;}}_x\left(1\right)-2R_{\mathrm{xd}}\left(0\right)=0\\ {2\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2^{*}{\mathrm{\&phi;}}_x\left(0\right)+2{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1^{*}{\mathrm{\&phi;}}_x\left(1\right)-2R_{\mathrm{xd}}\left(1\right)=0\end{array}\right.---\left(17\right)$

Obviously when if weights are got the L rank, can be derived as matrix form and be:

R _xxW ^*＝R _xd (18)

Promptly $W^{*}=R_{\mathrm{xx}}^{-1}{R}_{\mathrm{xd}}---\left(19\right)$

By formula (19) W that solves < > * <> Assurance is the extreme value of ε (W), but whether minimal value also should be verified.Ask for this reason

$\frac{{\&PartialD;}^2}{{\&PartialD;\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100865744E00012.png,H2009100865744E00015.png"\; state="\{\{state\}\}">w</figure-callout>}}_1^2}\mathrm{\&epsiv;}\left(W\right){|}_{W=W^{*}}={2\mathrm{\&phi;}}_x\left(0\right)$

$\frac{{\&PartialD;}^2}{{\&PartialD;\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="H2009100865744E00013.png,H2009100865744E00014.png"\; state="\{\{state\}\}">w</figure-callout>}}_2^2}\mathrm{\&epsiv;}\left(W\right){|}_{W=W^{*}}={2\mathrm{\&phi;}}_x\left(0\right)---\left(20\right)$

For stationary signal x(n)

φ _x(0)＝E{x(n-l)x(n-l)}＝E{x ²(n)} (21)

Be φ < > x <> (0) be x(n) mean square value, so the value perseverance just.Then by formula (20) visible, φ < >x <>(0) at W < >* <>The place Perseverance just.Be that ε (W) function is that a centre is to recessed parabolic curved surface, formula (19) separate the permanent minimal solution that is.Accurately ask this equation, need know R < > Xx <> And R < > Xd <> The priori statistical information, when these can't be known in advance, then seek numerical solution.

The LMS adaptive algorithm is to make quadratic performance index (error signal mean square value or average power) reach minimum through the weight coefficient of adjusting filter, is an a kind of special gradient decline type algorithm.Real gradient descent algorithm will be regulated the filter weight coefficient according to the negative gradient of error of mean square; Because error of mean square is unknown usually in real work, so LMS algorithm solution to this problem is to regulate weight coefficient according to the negative gradient of single sample variance.

According to the steepest descent method, the next time the weight coefficient vector W (n +1) should be equal to the right of the current coefficient vector plus a percentage of the mean square error in the negative gradient of the function is

$W(n+1)=W\left(n\right)-\mathrm{\&mu;}{\&dtri;e}^2\left(n\right)---\left(22\right)$

μ is called convergence coefficient for the gain constant of control adaptive speed and stability in the formula.

In the actual control process, in order to reduce the solution W (n) the amount of computation required for each iteration, to meet the system's real-time, to take a single error sample squared gradient as the mean square error function gradient

is estimated that there are

$\widehat{\&dtri;}{e}^2\left(n\right)=-2e\left(n\right)X\left(n\right)---\left(23\right)$

By formula (1), (3), (22), (23) can sum up rule and receive out the LMS adaptive filter algorithm and do

$\left\{\begin{array}{c}y\left(n\right)=W^{T}X\left(n\right)\\ e\left(n\right)=d\left(n\right)-y\left(n\right)\\ W(n+1)=W\left(n\right)+2\mathrm{\&mu;e}\left(n\right)X\left(n\right)\end{array}\right.---\left(24\right)$

In the convergence analytic process of LMS adaptive filter algorithm, consider how whether W reach the W(0 by initial setting) iteration becomes W < > * <> Can know that by the derivation of LMS adaptive algorithm the algorithm can be considered the steepest descent method that expectation is approximately instantaneous value; Therefore some average characteristics of LMS adaptive algorithm identical with steepest descent method still, but fluctuation can appear in the characteristic of its process, and this has introduced difficulty to analysis.

Because the variation of power W and produce X(n) model system in the variation of parameter or environment generally all than X(n) variation slow, therefore can establish the incoming signal X(n of LMS filter) with the weights W of LMS < > i <> Uncorrelated, can prove and work as μ in certain span, identical with change procedure of weighing in the steepest descent method and convergence situation, but the mathematical expectation absolute convergence of the weight vector W of LMS algorithm is to optimum right vector W < > * <> But in realistic simulation computation process, show; This incoherent supposition is not the sufficient condition of LMS adaptive algorithm convergence; At incoming signal X(n) and the power W of LMS algorithm or incoming signal between when bigger correlativity is arranged; Under the bigger prerequisite of error of mean square, weight vector W also can converge to optimum right vector W < > * <>

The weight coefficient number L of correct handling LMS sef-adapting filter, the overshoot coefficient M of LMS algorithm and the triangular relation of convergence coefficient μ are to guarantee the stability of control algorithm.The overshoot coefficient M of LMS algorithm can be expressed as

$M=\frac{\mathrm{<figure-callout\; id="4"\; label="N"\; filenames="H2009100865744E00013.png,H2009100865744E00015.png"\; state="\{\{state\}\}">N</figure-callout>}}{4\mathrm{\&tau;}}---\left(25\right)$

τ in the formula---time constant, μ is inversely proportional to convergence factor.

Can know that by the LMS algorithm speed of μ value direct control self adaptation convergence process is by formula (25) know that again too fast convergence rate can make the overshoot coefficient increase, it is out of control to form overshoot.In addition, the number that reduces weight coefficient helps system stability, causes available information few but L crosses the young pathbreaker, influences the simulation precision of weight vector to controlled system.Therefore guaranteeing suitably to increase the number L and the convergence coefficient μ value of weight coefficient, progressively to seek quick and stable control process under the stable prerequisite of LMS adaptive control system.

According to the LMS adaptive filter algorithm; To two-freedom vehicle suspension simplified model; Design LMS adaptive controller; The sprung weight acceleration/accel is as controlled leading indicator; According to preferred filter order and convergence coefficient; Under the input stimulus of two kinds of signals of single-frequency and road surface, carried out the simulation calculation analysis.

In emulation, it is 10mm that incoming signal adopts amplitude, and frequency is the sinusoidal excitation signal of 2Hz, and controller is outputed to the link between the power actr, like instruments such as D/A converter, low-pass filter, power amplifiers, is reduced to a ratio amplifying element.

Scheme 3a, figure 3b, scheme 4a, scheme 4b, scheme 5a, scheme 5b for controlling front and back sprung weight acceleration/accel, dynamic wheel load, suspension dynamic deflection performance ratio.

Visible by figure 3a, figure 3b, sprung weight acceleration/accel index is controlled in 5s under the lower amplitude effectively, and prolongation in time, and effect also can be more remarkable.At low-frequency resonance band place, amplitude is significantly reduced, and proves that the ride comfort of vehicle is improved preferably through its frequency domain figure explanation.

Visible by figure 4a, figure 4b, the dynamic wheel load index that characterizes vehicle handling quality has also obtained better controlled.

The suspension dynamic deflection is visible by figure 5a, figure 5b, and amplitude reduces not obvious, because what in the algorithm control of dynamic deflection is taked is indirect control, because the existence of unsprung weight, and especially for low-frequency excitation, the significantly reduction of the dynamic deflection very difficulty that will become.

When the frequency that changes sinusoidal signal, carried out exciting control process from 2Hz～25Hz, all receive similar effect.In control computation process; The output of controller all restrains gradually; Just control slightly difference of effect; This mainly is that ADAPTIVE LMS ALGORITHM can be according to different input signals automatic compensation weight coefficient; On statistical significance, have the ability that weight vector is approached to theoretical optimal value, this proves that further the stability of this algorithm is higher.

Convergence process such as Fig. 6, shown in Figure 7 about adaptive filtering error change procedure and weights in the analogue computing process; In the adjustment process of LMS adaptive controller; Beginning e(n) and W(n) variation is bigger; But the trend of convergence is gradually arranged; The output error and the weights of controller tend to be steady basically before and after 4s, and the LMS filter controller is accomplished the adjustment of weight coefficient basically, and system outlet is near stable.

Obtain on the basis of better effects in the control of single-frequency exciting, the simplification suspension system Properties Control under the excitation of road pavement model has been done following analysis.Think that in implementing the simulation control process actr can provide good power width of cloth characteristic in the research frequency range; It is made as linear element; Promptly import and be output into direct ratio and do not have time-delay; And instruments such as D/A converter, low-pass filter and power amplifier are reduced to the linear scaling amplifier, with the basic feature of outstanding control algorithm.

Under the The controller effect; The performance evaluation index of vehicle suspension such as Fig. 8 before and after the control～shown in Figure 10; Visible through the acceleration power spectral density curve; In 0～25Hz frequency band; Vibrating effect is controlled greatly, and especially about 2Hz, amplitude differs more than 10 times before and after the control; The control effect is remarkable, proves that further this controller has the ability that can improve vehicle ride comfort.

For the dynamic wheel load index; Visible by its power spectral density plot; The place reduces about 1/3 approximately at the 12Hz resonance peak; Under with the prerequisite of sprung weight acceleration signal as error input signal; By the vehicle suspension live load control indexes of random road surface signal excitation under this situation; Reason has two: one, and the LMS algorithm shows that indirectly the application force to unsprung weight tends towards stability after adjustment output is stable, so index is stable after the live load control; The 2nd, suitably choose LMS convergence of algorithm coefficient, relax its rate of convergence, though the controlled effect of sprung weight acceleration/accel index does not reach the best, dynamic wheel load control effect is effective.

Effect significantly improves at low-frequency range suspension dynamic deflection as shown in figure 10 before and after the suspension dynamic deflection control indexes, compares with the single-frequency exciting, and it is excellent that effect becomes.It is mainly due under the random road surface excitation, and the mis-behave of passive suspension is serious, and has the LMS algorithm of adaptive ability still to keep exporting preferably control ability, so effect is obvious.

Among the present invention; The automobile suspension system self-adaptation control method is to two-freedom vehicle suspension system model; Simulation Control through ADAPTIVE LMS ALGORITHM; Under the excitation of single-frequency and test pavement simulating signal; Sprung weight acceleration/accel, dynamic wheel load and suspension dynamic deflection have all obtained improvement to a certain degree; Particularly obviously reduced the vertical direction acceleration/accel of sprung weight; Other two indexes is also obtained certain control effect, verified the feasibility of vehicle suspension system LMS self adaptation active control strategies.

In real system implementation algorithm do not need square, average or differentiate, have simple and easy and high efficiency.Need not be average, the gradient component has comprised a big noise contribution certainly, but in the carrying out of adaptive process; The actual effect of playing a LPF; Along with the carrying out of process, this noise must obtain decay, thereby it is more suitable in the automobile suspension system random vibration control.

The above; Only for the preferable specific embodiment of the present invention, but protection scope of the present invention is not limited thereto, and any technical personnel of being familiar with the present technique field is in the technical scope that the present invention discloses; The variation that can expect easily or replacement all should be encompassed within protection scope of the present invention.

Claims (6)

1. a self-adaptation control method that is used for Vehicular vibration control is characterized in that, comprises adopting the LMS adaptive algorithm that target component is controlled, and makes the quadratic performance index of said target component reach minimum through the weight coefficient of adjusting filter, specifically comprises:

The function of the weight coefficient of said filter is:

ε (W)=E{d < > 2 <> (n) }-2W < > T <> E{d(n)X(n)}+W < > T <> E{X(n)X < > T <> (n)}W formula 7

The right side first term is manipulated signal d(n in the formula 7) mean square power, second the expression d(n), x(n) the cross-correlation amount, be expressed as Rxd;

Rxd is a time-varying vector, and it is a constant vector when stable state, and its each element is expressed from the next:

$R_{\mathrm{xd}}=\left[\begin{array}{c}E\left\{d\left(n\right)x\left(n\right)\right\}\\ E\left\{d\left(n\right)x(n-1)\right\}\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ E\left\{d\left(n\right)x(n-L+1)\right\}\end{array}\right]=\left[\begin{array}{c}{R}_{\mathrm{xd}}\left(0\right)\\ R_{\mathrm{xd}}\left(1\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ R_{\mathrm{xd}}(L-1)\end{array}\right]$ Formula 8

R wherein < > Xd <> (m) be cross-correlation coefficient, be defined as:

R < > Xd <> (m)=E{d(n)x(n-m) } formula 9

In actual process, real R < > Xd <> (m) value is a unknown number often, need estimate that the method for estimation of employing is: for steady ergodic stochastic signal x, y, and its corresponding R < > Xy <> (m) estimate by following formula:

$R_{\mathrm{xy}}\left(m\right)=\frac{1}{k-\left|m\right|}\underset{i=0}{\overset{k-\left|m\right|-1}{\mathrm{\&Sigma;}}}x(n-i)\&CenterDot;y(n-|m|-i)$ Formula 10

In the formula 10, k is the data number when calculating, general run of thins get k＞＞|m|;

The E{X(n)X of formula 7 right sides in last < > T <> (n)} represents x(n) the autocorrelation battle array, be designated as R < > Xx <> ,

R < > Xx <>=E{X(n)X < > T <> (n) the } formula 11

For stationary signal, R < > Xx <> Irrelevant with n, R < > Xx <> Be written as:

$R_{\mathrm{xx}}=E\left\{\left[\begin{array}{c}x\left(n\right)\\ x(n-1)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x(n-L+1)\end{array}\right]\right[x\left(n\right),x(n-1),\&CenterDot;\&CenterDot;\&CenterDot;,x(n-L+1)\left]\right\}$ Formula 12

Then according to formula 8, R < > Xx <> Be written as:

$R_{\mathrm{xx}}=\left[\begin{array}{cccc}{\mathrm{\&phi;}}_x\left(0\right)& {\mathrm{\&phi;}}_x\left(1\right)& \&CenterDot;\&CenterDot;\&CenterDot;& {\mathrm{\&phi;}}_x(L-1)\\ {\mathrm{\&phi;}}_x\left(1\right)& {\mathrm{\&phi;}}_x\left(0\right)& \&CenterDot;\&CenterDot;\&CenterDot;& {\mathrm{\&phi;}}_x(L-2)\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ {\mathrm{\&phi;}}_x(L-1)& {\mathrm{\&phi;}}_x(L-2)& \&CenterDot;\&CenterDot;\&CenterDot;& {\mathrm{\&phi;}}_x\left(0\right)\end{array}\right]$ Formula 13

φ in the formula < > x <> (l) be l rank cross-correlation coefficient;

With formula 8,12 substitution formulas 7:

ε (W)=E[d < > 2 <> (n) < > T <> R < > Xd <> +WTR < > Xx <> W formula 14

Visible by formula 14, error of mean square ε (W) obviously be weights (function of filter coefficient)W is got L=2, and error of mean square is the quadratic function of weights, and formula 7 becomes:

$\mathrm{\&epsiv;}\left(W\right)=\mathrm{\&epsiv;}(w_1,w_2)=w_1^2{\mathrm{\&phi;}}_x\left(0\right)+2w_1{w}_2{\mathrm{\&phi;}}_x\left(1\right)+w_2^2{\mathrm{\&phi;}}_x\left(0\right)$

${-2w}_1{R}_{\mathrm{xd}}\left(0\right)-2w_2{R}_{\mathrm{xd}}\left(1\right)+E\left[d^2\left(n\right)\right]$ Formula 15

Utilize differential zero setting method to find the solution, order:

$\left\{\begin{array}{c}\frac{\&PartialD;}{{\&PartialD;w}_1}\mathrm{\&epsiv;}(w_1,w_2){|}_{w_1,w_2=w_1^{*},w_2^{*}}=0\\ \frac{\&PartialD;}{{\&PartialD;w}_2}\mathrm{\&epsiv;}(w_1,w_2){|}_{w_1,w_2=w_1^{*},w_2^{*}}=0\end{array}\right.$ Formula 16

:

$\left\{\begin{array}{c}{2w}_1^{*}{\mathrm{\&phi;}}_x\left(0\right)+{2w}_2^{*}{\mathrm{\&phi;}}_x\left(1\right)-2R_{\mathrm{xd}}\left(0\right)=0\\ {2w}_2^{*}{\mathrm{\&phi;}}_x\left(0\right)+2w_1^{*}{\mathrm{\&phi;}}_x\left(1\right)-2R_{\mathrm{xd}}\left(1\right)=0\end{array}\right.$ Formula 17

When weights are got the L rank, be derived as matrix form and be:

R < > Xx <> W < > * <>=R < > Xd <> Formula 18

Promptly $W^{*}=R_{\mathrm{xx}}^{-1}{R}_{\mathrm{xd}}$ Formula 19

The W that solves by formula 19 < > * <> Assurance is the extreme value of ε (W), but whether minimal value also should be verified, asks for this reason:

$\frac{{\&PartialD;}^2}{{\&PartialD;w}_1^2}\mathrm{\&epsiv;}\left(W\right){|}_{W=W^{*}}=2{\mathrm{\&phi;}}_x\left(0\right)$

$\frac{{\&PartialD;}^2}{{\&PartialD;w}_2^2}\mathrm{\&epsiv;}\left(W\right){|}_{W=W^{*}}=2{\mathrm{\&phi;}}_x\left(0\right)$ Formula 20

For stationary signal x(n):

φ < > x <> (0)=E{x(n-l)x(n-l)}=E{x < > 2 <> (n) the } formula 21

Be φ < >x <>(0) be x(n) mean square value, then visible so the value perseverance is just by formula 20, φ < >x <>(0) at W < >* <>The place

Perseverance just, promptly ε (W) function be a centre to recessed parabolic curved surface, formula 19 separate the permanent minimal solution that is, accurately ask this equation, need know R < >Xx <>And R < >Xd <>The priori statistical information, when these statistical informations can't be known in advance, then seek numerical solution;

Through this method to controlling with the next item down or multinomial target component: the dynamic load between bouncing of automobile body acceleration/accel, wheel and road surface, the dynamic deflection of suspension system;

And improve the ride comfort of running car through control to above-mentioned target component.

2. the self-adaptation control method that is used for Vehicular vibration control according to claim 1 is characterized in that described quadratic performance index comprises with the next item down or multinomial: error signal mean square value, average power.

3. the self-adaptation control method that is used for Vehicular vibration control according to claim 1 is characterized in that said filter is LMS adaptive transversal filter or LMS adaptive line combiner.

4. according to claim 1, the 2 or 3 described self-adaptation control methods that are used for Vehicular vibration control, it is characterized in that said LMS algorithm is regulated said weight coefficient according to the negative gradient of single error signal variance.

5 according to claim 4, wherein the vibration control for a vehicle adaptive control method, characterized in that said single error signal, taking the square of the gradient

as the mean square error function gradient

The estimates.

6. the self-adaptation control method that is used for Vehicular vibration control according to claim 5 is characterized in that, when improving the ride comfort of running car through control to multinomial target component, and the bouncing of automobile body acceleration/accel evaluating of attaching most importance to.

Images