拓展卡尔曼滤波计算推导

In this paper, we aim to exploit the radar’s sensing ability and strong sensing signals to interpret the reception of the adversary target and enhance the covertness. To achieve effective jamming, the radar is required to estimate the location of the aerial adversary target and thereby construct the channels related to the adversary target. Note that the aerial target typically moves at a high speed. To improve the covert transmission performance, we employ the extended Kalman filter (EKF) based tracking model in \cite{26} to track and predict the location of the adversary target based on the estimated velocity information. For self-consistency, we briefly introduce the principle here, with more details of EKF-based target tracking available from \cite{26}.

Since the BS and the radar work in a cooperative manner, we assume that the interference from the BS to the radar can be eliminated before radar signal processing. Therefore, the signal-to-noise ratio (SNR) of the matched-filtered echo signals can be expressed as:

$${\gamma }_r=\frac{{\rho }_0^2{d}_{rw}^{-4}{G}_{\text{MF}}{\mathbf{a}}_r^{\mathrm{H}}({\theta }_{rw},{\varphi }_{rw})\mathbf{Q}{\mathbf{a}}_r({\theta }_{rw},{\varphi }_{rw})}{{\sigma }_r^2},$$

where $G_{\text{MF}}$ is the matched-filtering gain, and ${\sigma }_r^2$ is the noise power at the radar.

For ease of expression, the tracking period $T$ is discretized into $N$ time slots, with the duration $\delta =\frac{T}{N}$. In the $n$-th time slot, the location of the adversary target is denoted as:

$${\mathbf{q}}_w[n]=[x_w[n],y_w[n],z_w[n]{]}^{\mathrm{T}},$$

and the corresponding speed is:

$${\mathbf{v}}_w[n]=[{\dot{x}}_w[n],{\dot{y}}_w[n],{\dot{z}}_w[n]{]}^{\mathrm{T}}.$$

The state of the adversary target:

$$\mathbf{\alpha }[n]=[{\mathbf{q}}_w^{\mathrm{T}}[n],{\mathbf{v}}_w^{\mathrm{T}}[n]{]}^{\mathrm{T}},$$

is unknown to the radar and requires to be estimated. In each time slot, the radar first predicts the adversary target’s state based on the state estimated in the previous time slot. Based on the predicted state, both the communication base and radar waveforms are designed. Subsequently, new measurements are taken from the echoes, allowing the target state to be further estimated or predicted.

The state prediction model can be expressed as:

$$\mathbf{\alpha }[n]=\mathbf{F}\mathbf{\alpha }[n-1]+{\mathbf{z}}_{\mathbf{\alpha }[n]},$$

where

$$\mathbf{F}=\left[\begin{array}{cc}{\mathbf{I}}_3& \delta {\mathbf{I}}_3\\ 0_3& {\mathbf{I}}_3\end{array}\right],$$

${\mathbf{I}}_3$ and $0_3$ are $3\times 3$ identity matrix and zero matrix, respectively, and ${\mathbf{z}}_{\mathbf{\alpha }[n]}\sim \mathcal{N}(0,{\mathbf{Q}}_{\mathbf{\alpha }})$ is the state evolution noise vector, with ${\mathbf{Q}}_{\mathbf{\alpha }}$ being the state evolution covariance matrix.

By predicting the state of the adversary target in slot $n$ as:

$$\hat{\mathbf{\alpha }}[n]=\mathbf{F}\hat{\mathbf{\alpha }}[n-1],$$

the prediction covariance matrix is given by:

$$\mathbf{C}[n|n-1]=\mathbf{FC}[n-1]{\mathbf{F}}^{\mathrm{T}}+{\mathbf{Q}}_{\mathbf{\alpha }},$$

where $\mathbf{C}[n-1]$ is the covariance matrix of $\hat{\mathbf{\alpha }}[n-1]$.

Denoting the measurement parameters as:

β [ n ] = { τ ^ [ n ] , v ^ [ n ] , sin ⁡ θ ^ , cos ⁡ θ ^ , sin ⁡ ϕ ^ } , \boldsymbol{\beta}[n] = \{\hat{\tau}[n], \hat{v}[n], \sin\hat{\theta}, \cos\hat{\theta}, \sin\hat{\phi}\}, β[n]={τ^[n],v^[n],sinθ^,cosθ^,sinϕ^​},

the measurement model can be written as:

$$\mathbf{\beta }[n]={\mathbf{g}}_n(\mathbf{\alpha }[n])+{\mathbf{z}}_{\mathbf{\beta }[n]},$$

where ${\mathbf{z}}_{\mathbf{\beta }[n]}$ is the Gaussian measurement noise.

Through linearizing the measurement model around the predicted state, we have:

$$\mathbf{\beta }[n]\approx {\mathbf{g}}_n(\hat{\mathbf{\alpha }}[n|n-1])+{\mathbf{G}}_n(\mathbf{\alpha }[n]-\hat{\mathbf{\alpha }}[n|n-1])+{\mathbf{z}}_{\mathbf{\beta }[n]},$$

where ${\mathbf{G}}_n=\frac{\partial {\mathbf{g}}_n}{\partial \mathbf{\alpha }[n]}{|}_{\hat{\mathbf{\alpha }}[n|n-1]}$ denotes the Jacobian matrix of ${\mathbf{g}}_n$ with respect to $\mathbf{\alpha }[n]$.

Based on the above, the state of the adversary target in slot $n$ is estimated as:

$$\hat{\mathbf{\alpha }}[n]=\hat{\mathbf{\alpha }}[n|n-1]+{\mathbf{K}}_n\left(\mathbf{\beta }[n]-{\mathbf{g}}_n(\hat{\mathbf{\alpha }}[n|n-1])\right),$$

where the Kalman gain matrix ${\mathbf{K}}_n\in {\mathbb{R}}^{6\times 5}$ is given by:

$${\mathbf{K}}_n=\mathbf{C}[n|n-1]{\mathbf{G}}_n^{\mathrm{T}}{\left({\mathbf{G}}_n\mathbf{C}[n|n-1]{\mathbf{G}}_n^{\mathrm{T}}+{\mathbf{Q}}_{\mathbf{\beta }[n]}\right)}^{-1}.$$

Hence, the posterior covariance matrix is:

$$\mathbf{C}[n]=(\mathbf{I}-{\mathbf{K}}_n{\mathbf{G}}_n)\mathbf{C}[n|n-1].$$

Note that the trace of $\mathbf{C}[n]$ characterizes the posterior mean square error (MSE) for tracking the state of the adversary target. Therefore, we adopt the trace of $\mathbf{C}[n]$ in each time slot as the sensing performance metric.