IMU噪声模型

    xiaoxiao2022-07-02  124

    目录

    陀螺噪声模型白噪声(White Noise)零偏(Bias)白噪声的累积,随机游走,随机游走系数各种噪声性能指标的单位换算Dirac函数的单位Allan variance

    陀螺噪声模型

    陀螺测量模型: (1) x m ( t ) = x ( t ) + b ( t ) + n ( t ) x_m(t) = x(t)+b(t)+n(t) \tag{1} xm(t)=x(t)+b(t)+n(t)(1)

    x m x_{\rm m} xm为测量输出, x x x为真值, b b b为零偏, n n n为速率噪声,假设为高斯白噪声。

    白噪声(White Noise)

    假设测量模型中的 n ( t ) n(t) n(t) 为零均值高斯白噪声,其均值(mean)为零: E [ n ( t ) ] = 0 E[n(t)]=0 E[n(t)]=0

    其自相关函数为Delta函数,即白噪声在不同时刻的值统计无关: E [ n ( t 1 ) n ( t 2 ) ] = σ 2 δ ( t 1 − t 2 ) E[n(t_1)n(t_2)]=\sigma^2\delta(t_1-t_2) E[n(t1)n(t2)]=σ2δ(t1t2)

    自相关函数的数学定义: R ( τ ) = ∫ − ∞ ∞ x ( t ) x ( t − τ ) d t R(\tau)= \int_{-\infty}^{\infty} x(t)x(t-\tau) \rm{d} t R(τ)=x(t)x(tτ)dt

    离散形式: n [ k ] = 1 Δ T ∫ t 0 t 0 + Δ T n ( t ) d t n [ k ] ∼ N ( 0 , σ d 2 ) σ d = σ 1 Δ T n[k]=\frac {1}{\Delta T} \int_{t_0}^{t_0+\Delta T} n(t) \mathrm{d}t \\ n[k]\sim N(0,\sigma_\mathrm{d}^2)\\ \sigma _\mathrm{d}=\sigma\frac{1}{\sqrt{\Delta T}} n[k]=ΔT1t0t0+ΔTn(t)dtn[k]N(0,σd2)σd=σΔT 1

    其中 Δ T \Delta T ΔT是采样时间(即平均时间)。这里假设理想低通滤波器的带宽为 f = 1 Δ T f=\frac {1}{\Delta T} f=ΔT1。根据采样定理,这个频率应该是2倍于感兴趣的信号最高频率,即有用信号最高频率不能超过 f = 1 2 Δ T f=\frac {1}{2\Delta T} f=2ΔT1,高于这个频率的信号和噪声全部被滤除。以角速率陀螺的测量输出为例, σ d \sigma_d σd的单位为 ° / h \degree/h °/h Δ T \Delta{T} ΔT的单位是 h h h,所以 σ \sigma σ的单位为 ° / h \degree/\sqrt h °/h

    离散过程是求采样周期 Δ T \Delta T ΔT内的平均输出: n [ k ] = 1 Δ T ∫ t 0 t 0 + Δ T n ( t ) d t n[k]=\frac {1}{\Delta T} \int_{t_0}^{t_0+\Delta T} n(t) \mathrm{d}t\\ n[k]=ΔT1t0t0+ΔTn(t)dt

    其均值为 E ( n [ k ] ) = 1 Δ T ∫ t 0 t 0 + Δ T E [ n ( t ) ] d t = 0 \begin{aligned} E(n[k])&=\frac {1}{\Delta T} \int_{t_0}^{t_0+\Delta T} E[n(t)] \mathrm{d}t\\ &=0 \end{aligned} E(n[k])=ΔT1t0t0+ΔTE[n(t)]dt=0

    其方差为 σ d 2 = E ( ( n [ k ] ) 2 ) = E ( ( 1 ( Δ T ) 2 ∫ t 0 t 0 + Δ T ∫ t 0 t 0 + Δ T n ( τ ) n ( t ) d τ d t ) = 1 ( Δ T ) 2 ∫ t 0 t 0 + Δ T ∫ t 0 t 0 + Δ T E [ n ( τ ) n ( t ) ] d τ d t = σ 2 ( Δ T ) 2 ∫ t 0 t 0 + Δ T ∫ t 0 t 0 + Δ T δ ( t − τ ) d t d τ = σ 2 ( Δ T ) 2 ∫ t 0 t 0 + Δ T 1 d t = σ 2 Δ T \begin{aligned} \sigma_d^2&=E\left((n[k])^2\right) \\ &=E \left( (\frac {1}{(\Delta T)^2} \int_{t_0}^{t_0+\Delta T} \int_{t_0}^{t_0+\Delta T} n(\tau) n(t) \mathrm{d}\tau \mathrm{d}t \right) \\ &=\frac {1}{(\Delta T)^2} \int_{t_0}^{t_0+\Delta T} \int_{t_0}^{t_0+\Delta T} E[n(\tau) n(t)] \mathrm{d}\tau \mathrm{d}t \\ &=\frac {\sigma^2}{(\Delta T)^2} \int_{t_0}^{t_0+\Delta T} \int_{t_0}^{t_0+\Delta T} \delta(t - \tau) \mathrm{d}t \mathrm{d}\tau \\ &=\frac{\sigma^2}{(\Delta T)^2} \int_{t_0}^{t_0+\Delta T} 1 \mathrm{d}t \\ &=\frac{\sigma^2}{\Delta T} \end{aligned} σd2=E((n[k])2)=E(((ΔT)21t0t0+ΔTt0t0+ΔTn(τ)n(t)dτdt)=(ΔT)21t0t0+ΔTt0t0+ΔTE[n(τ)n(t)]dτdt=(ΔT)2σ2t0t0+ΔTt0t0+ΔTδ(tτ)dtdτ=(ΔT)2σ2t0t0+ΔT1dt=ΔTσ2

    可见平均时间越长,采样信号的方差越小,这与直觉一致。

    其中 ∫ t 0 t 0 + Δ T δ ( t − τ ) d t = ∫ t 0 − τ t 0 − τ + Δ T δ ( m ) d m = 1 \begin{aligned} \int_{t_0}^{t_0+\Delta T} \delta (t - \tau) \mathrm{d}t &= \int_{t_0-\tau}^{t_0-\tau+\Delta T} \delta (m) \mathrm{d}m =1 \end{aligned} t0t0+ΔTδ(tτ)dt=t0τt0τ+ΔTδ(m)dm=1

    由于 τ ∈ [ t 0 , t 0 + Δ T ] \tau \in [t_0,t_0+\Delta T] τ[t0,t0+ΔT],所以上面的积分范围为 − Δ T ≤ t 0 − τ ≤ 0 -\Delta T\leq t_0-\tau \leq0 ΔTt0τ0,而 0 ≤ t 0 − τ + Δ T ≤ Δ T 0 \leq t_0-\tau+\Delta T \leq \Delta T 0t0τ+ΔTΔT,所以上式=1。

    零偏(Bias)

    传感器测量误差中的缓变部分建模为“布朗运动(Brownian motion)”,也称为“维纳过程(Wiener process)”,或“随机游走(random walk)”(离散域中)。 b ˙ ( t ) = w ( t ) w ( t ) ∼ N ( 0 , σ b ) E [ w ( t 1 ) w ( t 2 ) ] = σ b 2 δ ( t 1 − t 2 ) \dot{b}(t)=w (t)\\ w(t)\sim N(0,\sigma_b)\\ E[w(t_1)w(t_2)]=\sigma_b^2\delta(t_1-t_2) b˙(t)=w(t)w(t)N(0,σb)E[w(t1)w(t2)]=σb2δ(t1t2)

    其中 ω ( t ) \omega(t) ω(t)为零均值高斯白噪声。离散形式: b d [ k ] = b d [ k − 1 ] + w d [ k ] w d [ k ] ∼ N ( 0 , σ b Δ T ) b_d[k]=b_d[k-1]+w_d [k]\\ w_d[k]\sim N(0,\sigma_b \sqrt{\Delta T}) bd[k]=bd[k1]+wd[k]wd[k]N(0,σbΔT )

    即: σ b d = σ b Δ T \sigma _{bd}=\sigma_b \sqrt{\Delta T} σbd=σbΔT

    离散过程: ∫ t 0 t 0 + Δ T b ˙ ( t ) d t = ∫ t 0 t 0 + Δ T w ( t ) d t b ( t 0 + Δ T ) − b ( t 0 ) = ∫ t 0 t 0 + Δ T   w ( t ) d t b d [ k ] − b d [ k − 1 ] = ∫ t 0 t 0 + Δ T w ( t ) d t \begin{aligned} \int_{t_0}^{t_0+\Delta T} \dot b(t) dt &= \int_{t_0}^{t_0+\Delta T} w(t) \mathrm{d}t \\ b(t_0+\Delta T) - b(t_0) &= \int_{t_0}^{t_0+\Delta T} \ w(t) \mathrm{d}t\\ b_d[k]-b_d[k-1] &=\int_{t_0}^{t_0+\Delta T} w(t) \mathrm{d}t \end{aligned} t0t0+ΔTb˙(t)dtb(t0+ΔT)b(t0)bd[k]bd[k1]=t0t0+ΔTw(t)dt=t0t0+ΔT w(t)dt=t0t0+ΔTw(t)dt w [ k ] = ∫ t 0 t 0 + Δ T w ( t ) d t w[k]=\int_{t_0}^{t_0+\Delta T} w(t) \mathrm{d}t w[k]=t0t0+ΔTw(t)dt 离散白噪声均值: E ( w [ k ] ) = E ( ∫ t 0 t 0 + Δ T w ( t ) d t ) = ∫ t 0 t 0 + Δ T E [ w ( t ) ] d t = 0 \begin{aligned} E\left (w[k]\right ) &= E \left(\int_{t_0}^{t_0+\Delta T} w(t) \mathrm{d}t \right) \\ &=\int_{t_0}^{t_0+\Delta T} E[w(t)] \mathrm{d}t\\ &=0 \end{aligned} E(w[k])=E(t0t0+ΔTw(t)dt)=t0t0+ΔTE[w(t)]dt=0 方差: σ b d 2 = E ( ( w [ k ] ) 2 ) = E ( ( ∫ t 0 t 0 + Δ T w ( t ) d t ) 2 ) = E ( ∫ t 0 t 0 + Δ T ∫ t 0 t 0 + Δ T w ( τ ) w ( t ) d τ d t ) = ∫ t 0 t 0 + Δ T ∫ t 0 t 0 + Δ T E [ w ( τ ) w ( t ) d τ d t = σ b 2 ∫ t 0 t 0 + Δ T ∫ t 0 t 0 + Δ T δ ( t − τ ) d τ d t = σ b 2 Δ T \begin{aligned} \sigma_{bd}^2 &=E\left( (w[k])^2 \right)\\ &=E\left( \left( \int_{t_0}^{t_0+\Delta T} w(t) \mathrm{d}t \right)^2 \right)\\ &=E \left( \int_{t_0}^{t_0+\Delta T} \int_{t_0}^{t_0+\Delta T} w(\tau) w(t) \mathrm{d}\tau \mathrm{d}t \right) \\ &=\int_{t_0}^{t_0+\Delta T} \int_{t_0}^{t_0+\Delta T} E[w(\tau)w(t) \mathrm{d}\tau \mathrm{d}t \\ &=\sigma _b^2\int_{t_0}^{t_0+\Delta T} \int_{t_0}^{t_0+\Delta T} \delta (t-\tau) \mathrm{d}\tau \mathrm{d}t \\ &=\sigma _b^2 \Delta T \end{aligned} σbd2=E((w[k])2)=E(t0t0+ΔTw(t)dt)2=E(t0t0+ΔTt0t0+ΔTw(τ)w(t)dτdt)=t0t0+ΔTt0t0+ΔTE[w(τ)w(t)dτdt=σb2t0t0+ΔTt0t0+ΔTδ(tτ)dτdt=σb2ΔT

    标准差: σ b d = σ b Δ T \sigma_{bd}=\sigma _b \sqrt {\Delta T} σbd=σbΔT

    σ b d \sigma_{bd} σbd的单位是 d e g deg deg Δ T \Delta{T} ΔT的单位为 h h h,所以 σ b \sigma_b σb的单位是 ° / h \degree/\sqrt{h} °/h

    表1:IMU噪声模型参数总结

    参数符号单位陀螺噪声谱密度 σ g \sigma_g σg r a d s 1 H z \frac{\mathrm{rad}}{\mathrm{s}} \frac{1}{\sqrt{\mathrm{Hz}}} sradHz 1加计噪声谱密度 σ a \sigma_a σa m s 2 1 H z \frac{m}{s^2} \frac{1}{\sqrt{Hz}} s2mHz 1陀螺随机游走 σ b g \sigma_{bg} σbg r a d s 2 1 H z \frac{rad}{s^2} \frac{1}{\sqrt{Hz}} s2radHz 1加计随机游走 σ b a \sigma_{ba} σba m s 3 1 H z \frac{m}{s^3} \frac{1}{\sqrt{Hz}} s3mHz 1IMU采样速率 1 Δ T \frac{1}{\Delta T} ΔT1 H z Hz Hz

    白噪声的累积,随机游走,随机游走系数

    随机游走(Random Walk)定义为白噪声的时间积分: θ ( t ) = ∫ 0 t n ( τ ) d τ \begin{aligned} \theta(t) &= \int_{0}^{t} n(\tau) \mathrm{d}\tau \end{aligned} θ(t)=0tn(τ)dτ

    均值: E ( θ ) = E ( ∫ 0 t n ( τ ) d τ ) = ∫ 0 t E ( n ( τ ) ) d τ = 0 \begin{aligned} E\left (\theta\right ) &= E \left (\int_{0}^{t} n(\tau) \mathrm{d}\tau \right) \\ &=\int_{0}^{t} E(n(\tau)) \mathrm{d}\tau \\ &=0 \end{aligned} E(θ)=E(0tn(τ)dτ)=0tE(n(τ))dτ=0

    方差: σ θ 2 ( t ) = E ( ( ∫ 0 t n ( τ ) d τ ) 2 ) = E ( ∫ 0 t ∫ 0 t n ( τ 1 ) n ( τ 2 ) d τ 1 d τ 2 ) = ∫ 0 t ∫ 0 t E ( n ( τ 1 ) n ( τ 2 ) ) d τ 1 d τ 2 = σ 2 ∫ 0 t ∫ 0 t δ ( τ 1 − τ 2 ) d τ 1 d τ 2 = σ 2 t \begin{aligned} \sigma_\theta^2(t) &=E\left( \left( \int_{0}^{t} n(\tau) \mathrm{d}\tau \right)^2 \right)\\ &=E \left( \int_{0}^{t} \int_{0}^{t} n(\tau_1) n(\tau_2) d\tau_1 \mathrm{d}\tau_2 \right) \\ &=\int_{0}^{t} \int_{0}^{t} E(n(\tau_1)n(\tau_2)) \mathrm{d}\tau_1 \mathrm{d}\tau_2 \\ &=\sigma ^2\int_{0}^{t} \int_{0}^{t} \delta (\tau_1-\tau_2) \mathrm{d}\tau_1 \mathrm{d}\tau_2\\ &=\sigma ^2 t \end{aligned} σθ2(t)=E((0tn(τ)dτ)2)=E(0t0tn(τ1)n(τ2)dτ1dτ2)=0t0tE(n(τ1)n(τ2))dτ1dτ2=σ20t0tδ(τ1τ2)dτ1dτ2=σ2t σ 2 \sigma^2 σ2具有速率的量纲。

    标准差: σ θ ( t ) = σ t \sigma_\theta(t)=\sigma \sqrt {t} σθ(t)=σt

    以上结果说明,随机游走的均值为0,而其方差和标准差是积分时间的函数,方差与积分时间成正比,标准差与积分时间的方根成正比。由于 σ θ \sigma_\theta σθ的单位为 d e g deg deg t t t的单位为 h h h,所以上式中 σ \sigma σ的单位为 ° / h \degree/\sqrt h °/h σ \sigma σ也称为随机游走系数(RWC, Random Walk Coefficient),是一个新定义的量,它在数值上与白噪声的标准差相同,但有不同的含义。随机游走参数指的就是随机游走系数。

    注意要用基本单位推出导出单位。

    各种噪声性能指标的单位换算

    角度随机游走(ARW)的单位采用 ° / h \degree / \sqrt{h} °/h ,陀螺零偏稳定性的单位采用 ° / h \degree / h °/h。用来表达噪声大小的其它测量量是功率谱密度(PSD,power spectral density,单位为 ( ° / h ) 2 / H z (\degree / h)^2 / Hz (°/h)2/Hz)和FFT噪声密度(单位为 ° / h / H z \degree / h / \sqrt{Hz} °/h/Hz )。PSD为功率谱,FFT为幅度谱。不同噪声指标之间的转换公式: A R W ( ° / h ) = 1 60 P S D ( ( ° / h ) 2 / H z ) A R W ( ° / h ) = 1 60 F F T ( ° / h / H z ) \begin{aligned} ARW(\degree/\sqrt{h})&=\frac{1}{60} \sqrt{PSD((\degree / \mathrm{h })^ \mathrm{2} / \mathrm{Hz})} \\ ARW(\degree/\sqrt{h})&=\frac{1}{60} FFT(\degree / h / \sqrt{Hz}) \end{aligned} ARW(°/h )ARW(°/h )=601PSD((°/h)2/Hz) =601FFT(°/h/Hz )

    Dirac函数的单位

    根据定义, δ \delta δ函数描述了某种理想物理量的密度分布的概念,因此其单位为: [ δ ( τ ) ] = 1 [ x ] [\delta(\tau)]=\frac{1}{[x]} [δ(τ)]=[x]1

    其中 [ x ] [x] [x]为该物理量的分布域单位。当 x x x为时间时,其单位为: [ δ ( τ ) ] = 1 sec = Hz [\delta(\tau)]=\frac{1}{\textrm{sec}}=\textrm{Hz} [δ(τ)]=sec1=Hz

    所以角速率噪声方差的单位为: [ E [ n r ( t + τ ) n r ( t ) ] ] = [ σ r c 2 δ ( τ ) ] = rad 2 sec 2 [E[n_r(t+\tau)n_r(t)]]=[\sigma_{r_c}^2\delta(\tau)]=\frac{{\textrm{rad}}^2}{{\textrm{sec}}^2} [E[nr(t+τ)nr(t)]]=[σrc2δ(τ)]=sec2rad2

    [ σ r c 2 ] = ( r a d s e c ) 2 s e c = r a d 2 s e c = ( r a d s e c ) 2 1 H z \begin{aligned} [\sigma_{rc}^2]&=\left(\frac{rad}{sec}\right)^2sec \\ &=\frac{{rad}^2}{sec} \\ &=\left(\frac{rad}{sec}\right)^2\frac{1}{Hz} \end{aligned} [σrc2]=(secrad)2sec=secrad2=(secrad)2Hz1

    标准差的单位为: [ σ r c ] = r a d s e c s e c = r a d s e c = r a d s e c 1 H z \begin{aligned} [\sigma_{rc}]&=\frac{rad}{sec} \sqrt{sec} \\ &=\frac{rad}{\sqrt{sec}} \\ &=\frac{rad}{sec} \frac{1}{\sqrt{Hz}} \end{aligned} [σrc]=secradsec =sec rad=secradHz 1

    Allan variance

    计算Allan variance: σ 2 ( τ ) = 1 2 τ 2 ⟨ ( θ k + 2 m − 2 θ k + m + θ k ) 2 ⟩ \sigma^2(\tau)=\frac{1}{2\tau^2}\langle(\theta_{k+2m}-2\theta_{k+m}+\theta_k)^2 \rangle σ2(τ)=2τ21(θk+2m2θk+m+θk)2

    其中 τ = m τ 0 \tau=m\tau_0 τ=mτ0 ⟨ ⟩ \langle \rangle 是ensemble average(总体平均)。总体平均可以展开为:

    kalibr

    最新回复(0)