雷达信号检测问题与奈曼-皮尔逊准则
雷达信号检测问题属于一个二元检测问题,即或者为目标加噪声情况,或者为仅含有噪声情况。对观察信号空间 D D D的划分问题,即将 D D D空间划分为 D 1 D_1 D1 (有信号)和 D 0 D_0 D0 (无信号)两个子空间,并满足 D = D 0 ∪ D 1 D=D_0\cup D_1 D=D0∪D1, D 0 ∩ D 1 = ∅ D_0\cap D_1=\emptyset D0∩D1=∅。子空间 D 0 D_0 D0和 D 1 D_1 D1称为判决域。若某个观测量 ( x ∣ H i ) ( i = 0 , 1 ) (x|{{H}_{i}})(i=0,1) (x∣Hi)(i=0,1)落入 D 1 D_1 D1域,就判决假设 H 1 H_1 H1成立,否则就判决假设 H 0 H_0 H0成立。
在不考虑杂波与干扰的问题下,认为接收回波中仅可能存在信号加噪声或纯噪声这两种情况。将仅含有噪声的情况设为
H
0
H_0
H0假设,含有信号加噪声的情况为
H
1
H_1
H1假设,有
{
H
0
:
x
(
t
)
=
n
(
t
)
H
1
:
x
(
t
)
=
s
(
t
)
+
n
(
t
)
\left\{ \begin{aligned} & {{H}_{0}}:x(t)=n(t) \\ & {{H}_{1}}:x(t)=s(t)+n(t) \\ \end{aligned} \right.
{H0:x(t)=n(t)H1:x(t)=s(t)+n(t)
其中
n
(
t
)
n(t)
n(t)为均值为0,方差为
σ
n
2
\sigma_n^2
σn2的高斯白噪声,
s
(
t
)
s(t)
s(t)为确知信号。对二元检测问题,存在两种正确判决和两种错误判决:
P
d
=
P
(
H
1
∣
H
1
)
P
n
=
P
(
H
0
∣
H
0
)
P
m
=
P
(
H
0
∣
H
1
)
P
f
a
=
P
(
H
0
∣
H
0
)
\begin{aligned} & {{P}_{d}}=P({{H}_{1}}|{{H}_{1}}) \\ & {{P}_{n}}=P({{H}_{0}}|{{H}_{0}}) \\ & {{P}_{m}}=P({{H}_{0}}|{{H}_{1}}) \\ & {{P}_{fa}}=P({{H}_{0}}|{{H}_{0}}) \\ \end{aligned}
Pd=P(H1∣H1)Pn=P(H0∣H0)Pm=P(H0∣H1)Pfa=P(H0∣H0)
其中,
P
d
{{P}_{d}}
Pd为检测概率,
P
n
{{P}_{n}}
Pn为正确不发现概率,
P
m
{{P}_{m}}
Pm为漏警概率,
P
f
a
{{P}_{fa}}
Pfa为虚警概率。
检测概率与漏警概率的关系为:
P
d
+
P
m
=
P
(
H
1
∣
H
1
)
+
P
(
H
0
∣
H
1
)
=
1
{{P}_{d}}+{{P}_{m}}=P({{H}_{1}}|{{H}_{1}})+P({{H}_{0}}|{{H}_{1}})=1
Pd+Pm=P(H1∣H1)+P(H0∣H1)=1
正确不发现概率与虚警概率的关系为:
P
n
+
P
f
a
=
P
(
H
0
∣
H
0
)
+
P
(
H
1
∣
H
0
)
=
1
{{P}_{n}}+{{P}_{fa}}=P({{H}_{0}}|{{H}_{0}})+P({{H}_{1}}|{{H}_{0}})=1
Pn+Pfa=P(H0∣H0)+P(H1∣H0)=1
假设
H
0
{{H}_{0}}
H0出现的先验概率为
P
(
H
0
)
P({{H}_{0}})
P(H0),
H
1
{{H}_{1}}
H1出现的先验概率为
P
(
H
1
)
P({{H}_{1}})
P(H1),观测信号
x
(
t
)
x(t)
x(t)在
H
1
{{H}_{1}}
H1假设下的条件概率密度函数为:
p
(
x
∣
H
0
)
=
1
(
2
π
)
σ
n
e
−
x
2
2
σ
n
2
p(x|{{H}_{0}})=\frac{1}{\sqrt{(2\pi )}{{\sigma }_{n}}}{{e}^{-\frac{{{x}^{2}}}{2\sigma _{n}^{2}}}}
p(x∣H0)=(2π)σn1e−2σn2x2
x
(
t
)
x(t)
x(t)在
H
0
{{H}_{0}}
H0假设下的条件概率密度函数为:
p
(
x
∣
H
1
)
=
1
(
2
π
)
σ
n
e
−
(
x
−
s
)
2
2
σ
n
2
p(x|{{H}_{1}})=\frac{1}{\sqrt{(2\pi )}{{\sigma }_{n}}}{{e}^{-\frac{{{(x-s)}^{2}}}{2\sigma _{n}^{2}}}}
p(x∣H1)=(2π)σn1e−2σn2(x−s)2
设定判决门限
V
T
{{V}_{T}}
VT,若输入信号的包络超过了该门限,则认为有目标,否则认为没有目标。故根据
H
1
{{H}_{1}}
H1假设下的条件概率密度函数可得检测概率为:
P
d
=
∫
V
T
∞
p
(
x
∣
H
1
)
d
x
{{P}_{d}}=\int_{{{V}_{T}}}^{\infty }{p(x|{{H}_{1}})dx}
Pd=∫VT∞p(x∣H1)dx
根据
H
0
{{H}_{0}}
H0假设下的条件概率密度函数可得虚警概率为:
P
f
a
=
∫
V
T
∞
p
(
x
∣
H
0
)
d
x
{{P}_{fa}}=\int_{{{V}_{T}}}^{\infty }{p(x|{{H}_{0}})dx}
Pfa=∫VT∞p(x∣H0)dx
判决门限
V
T
{{V}_{T}}
VT的选择与所采用的最佳检测检测准则有关。常用的最佳检测准则有最大后验概率准则、贝叶斯准则、极小极大化准则、最小错误概率准则、奈曼-皮尔逊准则等。在雷达信号检测领域,由于预先未知目标出现的概率,很难确定一次漏警所造成的损失。在一定的虚警概率下,使漏警概率最小或使检测概率最大,为奈曼-皮尔逊准则。
数学上,用如下表达式来描述奈曼-皮尔逊准则:
P
e
=
P
m
+
λ
0
P
f
a
=
P
(
H
0
∣
H
1
)
+
λ
0
P
(
H
1
∣
H
0
)
{{P}_{e}}={{P}_{m}}+{{\lambda }_{0}}{{P}_{fa}}=P({{H}_{0}}|{{H}_{1}})+{{\lambda }_{0}}P({{H}_{1}}|{{H}_{0}})
Pe=Pm+λ0Pfa=P(H0∣H1)+λ0P(H1∣H0)
由于检测概率与漏警概率之间满足
P
d
+
P
m
=
1
{{P}_{d}}+{{P}_{m}}=1
Pd+Pm=1,有
P
e
=
1
−
P
d
+
λ
0
P
f
a
=
1
−
P
(
H
1
∣
H
1
)
+
λ
0
P
(
H
1
∣
H
0
)
{{P}_{e}}=1-{{P}_{d}}+{{\lambda }_{0}}{{P}_{fa}}=1-P({{H}_{1}}|{{H}_{1}})+{{\lambda }_{0}}P({{H}_{1}}|{{H}_{0}})
Pe=1−Pd+λ0Pfa=1−P(H1∣H1)+λ0P(H1∣H0)
为提高判决质量,减小噪声干扰随机性的影响,一般需要对接收信号进行多次观测采样,对于
N
N
N次独立取样,输入信号为
N
N
N维空间,接收样本矢量表示为
x
=
[
x
1
,
x
2
,
.
.
.
,
x
N
]
T
\mathbf{x}={{[{{x}_{1}},{{x}_{2}},...,{{x}_{N}}]}^{\text{T}}}
x=[x1,x2,...,xN]T
当输入为
x
(
t
)
=
s
(
t
)
+
n
(
t
)
x(t)=s(t)+n(t)
x(t)=s(t)+n(t)时,
N
N
N个取样点的联合概率分布密度函数为
p
(
x
1
,
x
2
,
⋯
,
x
N
∣
H
1
)
p({{x}_{1}},{{x}_{2}},\cdots ,{{x}_{N}}|{{H}_{1}})
p(x1,x2,⋯,xN∣H1),当输入为
x
(
t
)
=
n
(
t
)
x(t)=n(t)
x(t)=n(t)时,其联合概率分布密度函数为
p
(
x
1
,
x
2
,
⋯
,
x
N
∣
H
0
)
p({{x}_{1}},{{x}_{2}},\cdots ,{{x}_{N}}|{{H}_{0}})
p(x1,x2,⋯,xN∣H0),检测概率和虚警概率分别为
P
d
=
∬
D
1
⋅
⋅
⋅
∫
p
(
x
1
,
x
2
,
⋅
⋅
⋅
⋅
x
N
∣
H
1
)
d
x
1
d
x
2
⋅
⋅
⋅
d
x
N
{{P}_{d}}=\iint_{{{D}_{1}}}{\cdot }\cdot \cdot \int{p}({{x}_{1}}\text{,}{{x}_{2}}\text{,}\cdot \cdot \cdot \cdot {{x}_{N}}\mid {{H}_{1}})\text{d}{{x}_{1}}\text{d}{{x}_{2}}\cdot \cdot \cdot \text{d}{{x}_{N}}
Pd=∬D1⋅⋅⋅∫p(x1,x2,⋅⋅⋅⋅xN∣H1)dx1dx2⋅⋅⋅dxN
P f a = ∬ D 1 ⋅ ⋅ ⋅ ∫ p ( x 1 , x 2 , ⋅ ⋅ ⋅ x N ∣ H 0 ) d x 1 d x 2 ⋅ ⋅ ⋅ d x N {{P}_{fa}}=\iint_{{{D}_{1}}}{\cdot }\cdot \cdot \int{p}({{x}_{1}}\text{,}{{x}_{2}}\text{,}\cdot \cdot \cdot {{x}_{N}}\mid H_0)\text{d}{{x}_{1}}\text{d}{{x}_{2}}\cdot \cdot \cdot \text{d}{{x}_{N}} Pfa=∬D1⋅⋅⋅∫p(x1,x2,⋅⋅⋅xN∣H0)dx1dx2⋅⋅⋅dxN
则总错误概率与联合概率分布密度函数的关系为
P
e
=
1
−
∬
D
1
⋅
⋅
⋅
∫
[
p
(
x
1
,
x
2
,
⋅
⋅
⋅
,
x
N
∣
H
1
)
−
λ
0
p
(
x
1
,
x
2
,
⋅
⋅
⋅
x
N
∣
H
0
)
]
d
x
1
d
x
2
⋅
⋅
⋅
d
x
N
{{P}_{\text{e}}}=1-\iint_{{{D}_{1}}}{\cdot }\cdot \cdot \int{\left[ p\left( {{x}_{1}},{{x}_{2}},\cdot \cdot \cdot ,{{x}_{N}}\mid {{H}_{1}} \right)-{{\lambda }_{0}}p\left( {{x}_{1}},{{x}_{2}},\cdot \cdot \cdot {{x}_{N}}\mid {{H}_{0}} \right) \right]}\text{d}{{x}_{1}}\text{d}{{x}_{2}}\cdot \cdot \cdot \text{d}{{x}_{N}}
Pe=1−∬D1⋅⋅⋅∫[p(x1,x2,⋅⋅⋅,xN∣H1)−λ0p(x1,x2,⋅⋅⋅xN∣H0)]dx1dx2⋅⋅⋅dxN
观察空间的划分应保证总错误概率
P
e
{{P}_{\text{e}}}
Pe最小,即后面的积分值最大。因此有
p
(
x
1
,
x
2
,
⋅
⋅
⋅
,
x
N
∣
H
1
)
−
λ
0
p
(
x
1
,
x
2
,
⋅
⋅
⋅
x
N
∣
H
0
)
≥
0
p\left( {{x}_{1}},{{x}_{2}},\cdot \cdot \cdot ,{{x}_{N}}\mid {{H}_{1}} \right)-{{\lambda }_{0}}p\left( {{x}_{1}},{{x}_{2}},\cdot \cdot \cdot {{x}_{N}}\mid {{H}_{0}} \right)\ge 0
p(x1,x2,⋅⋅⋅,xN∣H1)−λ0p(x1,x2,⋅⋅⋅xN∣H0)≥0
的点判定为有信号,否则判定为无信号。即
p
(
x
1
,
x
2
…
x
N
∣
H
1
)
p
(
x
1
,
x
2
…
x
N
∣
H
0
)
{
≥
λ
0
<
λ
0
\frac{p({{x}_{1}}\text{,}{{x}_{2}}\ldots {{x}_{N}}\mid {{H}_{1}})}{p({{x}_{1}}\text{,}{{x}_{2}}\ldots {{x}_{N}}\mid {{H}_{0}})}\left\{ \begin{aligned} & \ge {{\lambda }_{0}}\quad \\ & <{{\lambda }_{0}}\quad \\ \end{aligned} \right.
p(x1,x2…xN∣H0)p(x1,x2…xN∣H1){≥λ0<λ0
参考文献:
[1] 陈伯孝等. 现代雷达系统分析与设计[M]. 陕西:西安电子科技大学出版社. 2012.