DEEP CONVOLUTIONAL NEURAL NETWORK POWERED TERAHERTZ ULTRA-MASSIVE MULTI-INPUT-MULTI-OUTPUT CHANNEL ESTIMATION METHOD

Abstract

A THz UM-MIMO channel estimation method based on the DCNN comprises the steps: the hybrid spherical and planar-wave modeling (HSPM), by taking a sub-array in the antenna array as a unit, employing the PWM within the sub-array, and employing the SWM among the sub-arrays; estimating the channel parameters between the reference sub-arrays at Tx and Rx through a DCNN, including the angles of departure and arrival, the propagation distance and the path gain; deducing the channel parameters between the reference sub-array and other sub-arrays by utilizing the obtained channel parameters and the geometrical relationships among sub-arrays, and recovering the channel matrix; wherein accurate three-dimensional channel modeling is achieved by the HSPM, which possesses high modeling accuracy and low complexity.

Claims

1. A THz UM-MIMO CE method based on a DCNN, comprising the following steps: step i), the HSPM modeling, which takes the sub-array as a unit, using the PWM in the sub-array, and models the channel among sub-arrays by the SWM; step ii), using the first sub-array at the transceiver end as a reference sub-array, using a DCNN to estimate the departure angle, the angle of arrival, the propagation distance and the path gain between the reference sub-arrays according to real values, element imaginary values and element absolute values of the channel observation matrix; step iii), deriving channel parameters between the reference sub-array and the remaining sub-arrays by using the channel parameters obtained in step ii) and the geometric relationship between the sub-arrays, and reconstruct the channel matrix.

2. The THz UM-MIMO CE method according to claim 1, wherein step i) comprises the following steps: a) dividing the antennas at Tx and Rx into K.sub.t and K.sub.r sub-arrays, respectively, and different sub-arrays have the same multi-path number A the amplitude of the channel gain between different sub-arrays is the same, while the phase of the channel gain is changed due to different geometric distances and transceiver angles, to obtain the block structured channel model: $H_{HSPM} = \overset{N_{p}}{\underset{p = 1}{.Math.}} [\begin{matrix} .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{11}} {a_{rp}^{11} (a_{tp}^{11})}^{H} & .Math. & .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{1 K_{t}}} {a_{rp}^{1 K_{t}} (a_{tp}^{1 K_{t}})}^{H} \\ .Math. & .Math. \\ .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{K_{r} 1}} {a_{rp}^{K_{r} 1} (a_{tp}^{K_{r} 1})}^{H} & .Math. & .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{K_{r} 1}} {a_{rp}^{K_{r} K_{t}} (a_{tp}^{K_{r} K_{t}})}^{H} \end{matrix}],$ wherein: N.sub.p is the number of multi-path in the channel, α.sub.p.sup.11 is the channel gain of the p-th path, .sub.λ is the signal wavelength, D.sub.p.sup.k.sup.r.sup.k.sup.t is the transmission distance of the p-th path, a.sub.rp.sup.k.sup.r.sup.k.sup.t and a.sub.tp.sup.k.sup.r.sup.k.sup.t the antenna array guide vector for the transceiver end, respectively; and b) receiving a signal according to a channel model, wherein the received signal Y=W.sup.HH.sub.HSPMF+N, W is a received combining matrix comprising an analog combining matrix and a digital combining matrix, and F is a transmitting beamforming matrix comprising an analog beamforming matrix and a digital beamforming matrix, and N is a noise signal.

3. The THz UM-MIMO CE method according to claim 2, wherein the block structured channel model refers to: each block $.Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{k_{r} k_{t}}} {a_{rp}^{k_{r} k_{t}} (a_{tp}^{k_{r} k_{t}})}^{H}$ is a channel matrix between sub-arrays at the Tx and Rx, the channel matrix is the PWM matrix within the block, and among different sub-arrays, the channel is spherical-wave modeled, resulting in different transceiver angles, and thus the a.sub.rp.sup.k.sup.r.sup.k.sup.t and a.sub.tp.sup.k.sup.r.sup.k.sup.t are different.

4. The THz UM-MIMO CE method according to claim 1, wherein the DCNN comprises an input layer, seven convolutional layers (CV), four max-pooling layers (MP), a flattening layer, and a fully connected (FC) output layer, wherein the input layer receives the element real value Re{Y} of the channel observation matrix Y, the element imaginary value lm{Y} and the element absolute value |Y|; the CV respectively comprise 16, 32, 64, 128, 62, 32, 16 convolution kernels, wherein a convolution filter with a size of 3×3 is deployed to extract features; zero padding (ZP) and batch normalization (BN) are deployed in the convolution process; the pooling layer uses the maximum pooling method to extract the maximum value in the 2×2 size pool to reduce the number of network dimensions, and simplify the training process. The flattening layer rearranges the neurons into one dimension and is connected to the FC output layer for outputting the estimated emission angle, the angle of arrival, the propagation distance, and the path gain.

5. The THz UM-MIMO CE method according to claim 1, wherein the step iii) comprises: a) calculating parameters of the line-of-sight path: respectively projecting the line-of-sight path to the X-Y plane and the Y-Z plane, and deriving a transceiver angle between the remaining sub-arrays under the line of sight $θ_{t}^{k_{t} k_{r}} = \arccos (\frac{D_{xy}^{11} \cos θ_{t}^{11}}{D_{xy}^{k_{r} k_{t}}}), θ_{r}^{k_{t} k_{r}} = θ_{r}^{11} + \arcsin (\frac{Δ d_{x} \cos θ_{t}^{11}}{D_{xy}^{k_{t} k_{r}}}),$ $ϕ_{t}^{k_{t} k_{r}} = \arccos (\frac{D_{yz}^{11} \cos ϕ_{t}^{11}}{D_{yz}^{k_{t} k_{r}}}), ϕ_{r}^{k_{t} k_{r}} = - ϕ_{r}^{11} + \arcsin (\frac{Δ d_{z} \cos θ_{t}^{11}}{D_{yz}^{k_{t} k_{r}}})$ and a propagation distance $D^{k_{t} k_{r}} = \frac{D_{yz}^{k_{t} k_{r}}}{\cos θ_{t}^{11}},$ wherein: θ.sub.t.sup.k.sup.t.sup.k.sup.r and ϕ.sub.t.sup.k.sup.t.sup.k.sup.r are the receiving azimuth angle and the elevation angle of the k.sub.r sub-array at the k.sub.t transmitting end, while θ.sub.r.sup.k.sup.t.sup.k.sup.r and ϕ.sub.r.sup.k.sup.t.sup.k.sup.r the k.sub.r sub-array at the k.sub.t receiving end, respectively, $D_{xy}^{11} = D^{1 1} \cos ϕ_{t}^{11} D_{yz}^{11} = D^{11} \cos θ_{t}^{1 1}, D_{xy}^{k_{t} k_{r}} = \sqrt{{(Δ d_{x}^{k_{t} k_{r}})}^{2} + {(D_{xy}^{11})}^{2} - 2 Δ d_{x}^{k_{t} k_{r}} D_{xy}^{11} \sin θ_{t 1}^{11}}, D_{yz}^{k_{t} k_{r}} = \sqrt{{(Δ d_{z}^{k_{t} k_{r}})}^{2} + {(D_{yz}^{11})}^{2} - 2 Δ d_{z}^{k_{t} k_{r}} D_{yz}^{11} \sin ϕ_{r}^{11}},$ wherein: D.sup.11 is the distance between the transmitting end and the reference sub-array of the receiving end, and Δd.sub.x.sup.k.sup.t.sup.k.sup.r the relative displacement along the X axis of the k.sub.r sub-array at the k.sub.t transmitting end, here D.sup.11 represent the distance between the reference sub-arrays at Tx and Rx. Δd.sub.x.sup.k.sup.t.sup.k.sup.r refer to the relative displacement of the k.sub.r sub-array at Rx to the k.sub.t sub-array at Tx along the X-axis. Δd.sub.z.sup.k.sup.t.sup.k.sup.r is the relative displacement along the Z-axis of the k.sub.r sub-array at Rx to the k.sub.t sub-array at Tx; D.sub.yz.sup.11=D.sup.11 cos θ.sub.t.sup.11, $D_{xy}^{k_{t} k_{r}} = \sqrt{{(Δ d_{x}^{k_{t} k_{r}})}^{2} + {(D_{xy}^{11})}^{2} - 2 Δ d_{x}^{k_{t} k_{r}} D_{xy}^{11} \sin θ_{t 1}^{11}};$ b) calculating the plane equation of the reflecting surface: solving the equation A.sub.px+B.sub.py+C.sub.pz+D.sub.p=0, by means of the reflecting surfaces by means of the coordinates of the transceiver end and the law of reflection, wherein: each parameter is: $A_{p} = \frac{\sin ϕ_{tp}^{11} (C_{p}^{11} \cos θ_{sp}^{11} \cos ϕ_{sp}^{11} - B_{p}^{11} \sin ϕ_{tp}^{11})}{\cos ϕ_{sp}^{11} \sin ϕ_{tp}^{11} (A_{p}^{11} \cos θ_{sp}^{11} - B_{p}^{11} \sin θ_{p}^{11})}, B_{p} = \frac{\sin ϕ_{tp}^{11} (A_{p}^{11} \sin ϕ_{sp}^{11} - C_{p}^{11} \sin θ_{sp}^{11} \cos ϕ_{sp}^{11})}{\cos ϕ_{sp}^{11} \sin ϕ_{tp}^{11} (A_{p}^{11} \cos θ_{sp}^{11} - B_{p}^{11} \sin θ_{sp}^{11})}, C_{p} = 1,$ the lower corner mark .sub.p index the propagation path; and c) obtaining parameters of the non-line-of-sight path after obtaining the plane equation of the reflecting surface, $θ_{tp}^{k_{t} k_{r}} = \arcsin [\frac{x_{p}^{k_{r} k_{r}}}{\sqrt{{(x_{p}^{k_{r} k_{r}})}^{2} + {(y_{p}^{k_{r} k_{r}})}^{2}}}], ϕ_{tp}^{k_{t} k_{r}} = \arcsin [\frac{z_{p}^{k_{r} k_{r}}}{\sqrt{{(x_{p}^{k_{r} k_{r}})}^{2} + {(y_{p}^{k_{r} k_{r}})}^{2} + {(z_{p}^{k_{r} k_{r}})}^{2}}}], θ_{rp}^{k_{t} k_{r}} = \arcsin [\frac{x_{p}^{k_{r} k_{r}} - R_{k_{r} x}}{\sqrt{{(x_{p}^{k_{r} k_{r}} - R_{1 x})}^{2} + {(y_{p}^{k_{r} k_{r}} - R_{1 y})}^{2}}}], ϕ_{rp}^{k_{r} k_{r}} = \arcsin [\frac{z_{p}^{k_{r} k_{r}} - R_{k_{r} z}}{\sqrt{{(x_{p}^{k_{r} k_{r}} - R_{k_{r} x})}^{2} + {(y_{p}^{k_{r} k_{r}} - R_{k_{r} y})}^{2} + {(z_{p}^{k_{r} k_{r}} - R_{k_{r} z})}^{2}}}] .$ $wherein$ $D_{p}^{k_{t} k_{r}} = \sqrt{{(x_{p}^{k_{r} k_{r}})}^{2} + {(y_{p}^{k_{r} k_{r}})}^{2} + {(z_{p}^{k_{r} k_{r}})}^{2}} + \sqrt{{(R_{k_{r} x} - x_{p}^{k_{r} k_{r}})}^{2} + {(R_{k_{r} y} - y_{p}^{k_{r} k_{r}})}^{2} + {(R_{k_{r} z} - z_{p}^{k_{r} k_{r}})}^{2}}, x_{p}^{k_{r} k_{r}}, y_{p}^{k_{r} k_{r}},$ z.sub.p.sup.k.sup.r.sup.k.sup.r are the coordinates of the reflection point of the remaining sub-array on the reflecting surface (R.sub.k.sub.r.sub.x, R.sub.k.sub.r.sub.y, R.sub.k.sub.r.sub.z)=(D.sup.11 sin θ.sub.t1.sup.11 cos ϕ.sub.t1.sup.11+d.sub.rx.sup.k.sup.r, D.sup.11 cos θ.sub.t1.sup.11 cos ϕ.sub.t1.sup.11, D.sup.11 sin ϕ.sub.t1.sup.11−d.sub.rz.sup.k.sup.r), and d.sub.rx.sup.k.sup.r is the distance from k.sub.r th sub-array of the receiving end to the reference sub-array along the X-axis, d.sub.rz.sup.k.sup.r is the distance from the k.sub.r th sub-array of the receiving end to the reference sub-array along the Z-axis.

6. A THz UM-MIMO CE system realizing the said THz UM-MIMO CE method according to either claim 1 comprising: a RF-chain, an analog beamformer, an analog combiner, and a digital combiner, wherein the RF-chain of the transmitting end receives the baseband pilot signal to perform digital beamforming to obtain a radio frequency signal, and the analog beamformer performs analog beamforming according to a preset beam codebook signal and transmits the analog beamforming signal to the channel; and the analog combiner at Rx end combines the received signals and then outputs the received signals to the digital beam forming unit; the digital beamformer performs beamforming processing according to the combined analog signal to obtain a digital baseband signal, and the baseband end processes the digital baseband signal to complete the CE.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0013] FIG. 1 is a schematic diagram of a system according to the present invention;

[0014] FIG. 2 is a structural diagram of the DCNN according to the present invention;

[0015] FIG. 3A is a schematic diagram of the effect of an embodiment: the error of different channel models at different communication distances is exploited;

[0016] FIG. 3B is a schematic diagram of the effect of an embodiment: the error of different channel models at different sub-array spacing is determined

[0017] FIG. 3C is a schematic diagram of the effect of an embodiment: the error of different channel models at different carrier frequencies;

[0018] FIG. 4 is a schematic performance comparison diagram of the method of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

[0019] As shown in FIG. 1, a THz UM-MIMO CE system based on a DCNN according to an embodiment includes: a radio frequency (RF)-chain, an analog beamformer, an analog combiner, and a digital combiner. The RF-chain at the transmitter (Tx) transmits the baseband pilot signal and perform digital beamforming, to obtain the RF signal. The analog beamformer performs analog beamforming according to a predefined beam codebook, and transmits the analog signal to the channel. The analog combiner at the receiver (Rx) combines the received signals and outputs it to the digital combiner. The digital combiner performs signal processing on the combined analog signal to obtain a baseband signal. Finally, the baseband processes the digital signal to complete the CE.

[0020] The present embodiment relates to a THz UM-MIMO CE method based on a DCNN is composed of the following steps:

[0021] Step 1, as shown in FIG. 1, take a sub-array as a unit, the PWM is adopted in the sub-array, and the SWM is used between the sub-arrays to complete the HSPM, specifically it is composed of: [0022] 1) dividing the antennas at Tx and Rx into K.sub.t and K.sub.r sub-arrays, respectively, and different sub-arrays have the same number of multi-path N.sub.p. The channel gains between different sub-arrays have the same amplitude, while the phase of the channel gain is changed due to different geometric distances and transceiver angles. The HSPM can be expressed as

[00001] $H_{HSPM} = {.Math.}_{p = 1}^{N_{p}} [\begin{matrix} .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{11}} {a_{rp}^{11} (a_{tp}^{11})}^{H} & .Math. & .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{1 K_{t}}} {a_{rp}^{1 K_{t}} (a_{tp}^{1 K_{t}})}^{H} \\ .Math. & .Math. \\ .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{K_{r} 1}} {a_{rp}^{K_{r} 1} (a_{tp}^{K_{r} 1})}^{H} & .Math. & .Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{K_{r} 1}} {a_{rp}^{K_{r} K_{t}} (a_{tp}^{K_{r} K_{t}})}^{H} \end{matrix}],$

[0023] where N.sub.p is the number of multi-paths in the channel, α.sub.p.sup.11 is the channel gain of the p.sup.th path, λ is the signal wavelength, D.sub.p.sup.k.sup.r.sup.k.sup.t is the transmission distance of the p.sup.th path, α.sub.rp.sup.k.sup.r.sup.k.sup.t and α.sub.tp.sup.k.sup.r.sup.k.sup.t are the array steering vectors at Tx and Rx, respectively, whose value are determined by the propagation angles.

[0024] The block structure of the HSPM H.sub.HSPM refers to: each block

[00002] $.Math. α_{p}^{11} .Math. e^{- j \frac{2 π}{λ} D_{p}^{k_{r} k_{t}}} {a_{rp}^{k_{r} k_{t}} (a_{tp}^{k_{r} k_{t}})}^{H}$

is a PWM matrix between the sub-arrays at Tx and Rx. Among between different sub-arrays, the SWM is considered, resulting in different transceiver angles, and thus α.sub.rp.sup.k.sup.r.sup.k.sup.t and α.sub.tp.sup.k.sup.r.sup.k.sup.t are different. [0025] 2) According to the channel model, the received signal can be expressed as Y=W.sup.HH.sub.HSPMF+N, W is the received combining matrix comprising the analog combining and digital combining matrices, and F is a transmitted beamforming matrix comprising an analog beamforming matrix and a digital beamforming matrix, and N is a noise signal.

[0026] Step 2, using the first sub-array at Tx and Rx as the reference sub-arrays, and perform parameter estimation by using the DCNN, including the departure angles, the angles of arrival the propagation distance, and the path gain.

[0027] 2.1 The DCNN estimates the departure angles (θ.sub.t, p, ϕ.sub.t, p), the angle of arrival (θ.sub.r, p, ϕ.sub.r, p), the propagation D.sub.p.sup.k.sup.r.sup.k.sup.t, and the path gain |α.sub.p.sup.11| according to Re{Y}, Re{Y} and |Y|.

[0028] As shown in FIG. 2, there are totally 15 layers in the DCNN, which includes an input layer, seven convolutional layers (CV), four max-pooling layers (MP), a flattening layer, and a fully connected (FC) output layer, wherein the input layer receives the element real value Re{Y} of the channel observation matrix Y, the element imaginary value lm{Y} and the element absolute value |Y|; the CV respectively comprise 16, 32, 64, 128, 62, 32, 16 convolution kernels, wherein a convolution filter with a size of 3×3 is deployed to extract features; zero padding (ZP) and batch normalization (BN) are deployed in the convolution process; the pooling layer uses the maximum pooling method to extract the maximum value in the 2×2 size pool to reduce the number of network dimensions, and simplify the training process. The flattening layer rearranges the neurons into one dimension and is connected to the FC output layer for outputting the estimated emission angle, the angle of arrival, the propagation distance, and the path gain.

[0029] Step 3, deriving channel parameters between the reference sub-array and the remaining sub-arrays by using the geometric relationship between the sub-arrays obtained in Step 2, and restoring the channel matrix, specifically comprising:

[0030] 1) Calculating parameters of the line-of-sight path: projecting the line-of-sight path as shown in FIG. 1 to the X-Y plane and the Y-Z plane, respectively, and deriving the

[00003] $angles θ_{t}^{k_{t} k_{r}} = \arccos (\frac{D_{xy}^{11} \cos θ_{t}^{11}}{D_{xy}^{k_{t} k_{r}}}), θ_{r}^{k_{t} k_{r}} = θ_{r}^{11} + \arcsin (\frac{Δ d_{x} \cos θ_{t}^{11}}{D_{xy}^{k_{t} k_{r}}}),$ $ϕ_{t}^{k_{t} k_{r}} = \arccos (\frac{D_{yz}^{11} \cos ϕ_{t}^{11}}{D_{yz}^{k_{t} k_{r}}}), ϕ_{r}^{k_{t} k_{r}} = - ϕ_{r}^{11} + \arcsin (\frac{Δ d_{z} \cos θ_{t}^{11}}{D_{yz}^{k_{t} k_{r}}})$

and the propagation distance

[00004] $D^{k_{t} k_{r}} = \frac{D_{yz}^{k_{t} k_{r}}}{\cos θ_{t}^{11}},$

wherein: θ.sub.t.sup.k.sup.t.sup.k.sup.r and ϕ.sub.t.sup.k.sup.t.sup.k.sup.r denote the received azimuth and elevation angles of the k.sub.t sub-array at Tx end and the k.sub.r sub-array at Rx, respectively.

[00005] $D_{xy}^{11} = D^{1 1} \cos ϕ_{t}^{11} D_{yz}^{11} = D^{11} \cos θ_{t}^{1 1}, D_{xy}^{k_{t} k_{r}} = \sqrt{{(Δ d_{x}^{k_{t} k_{r}})}^{2} + {(D_{xy}^{11})}^{2} - 2 Δ d_{x}^{k_{t} k_{r}} D_{xy}^{11} \sin θ_{t 1}^{11}},$ $D_{yz}^{k_{t} k_{r}} = \sqrt{{(Δ d_{z}^{k_{t} k_{r}})}^{2} + {(D_{yz}^{11})}^{2} - 2 Δ d_{z}^{k_{t} k_{r}} D_{yz}^{11} \sin ϕ_{r}^{11}},$

where D.sup.11 represent the distance between the reference sub-arrays at Tx and Rx. Δd.sub.x.sup.k.sup.t.sup.k.sup.r refer to the relative displacement of the k.sub.r sub-array at Rx to the k.sub.t sub-array at Tx along the X-axis. Δd.sub.z.sup.k.sup.t.sup.k.sup.r is the relative displacement along the Z-axis of the k.sub.r sub-array at Rx to the k.sub.t sub-array at Tx.

[00006] $D_{y z}^{1 1} = D^{1 1} \cos θ_{t}^{11}, D_{xy}^{k_{t} k_{r}} = \sqrt{{(Δ d_{x}^{k_{t} k_{r}})}^{2} + {(D_{xy}^{11})}^{2} - 2 Δ d_{x}^{k_{t} k_{r}} D_{xy}^{11} \sin θ_{t 1}^{11}} .$

[0031] 2) calculating the plane equation of the reflecting surface: solving the equation A.sub.px+B.sub.py+C.sub.pz+D.sub.p=0 of the reflecting surface by means of the coordinates of the transceiver end and the equation of the equation,

[00007] $A_{p} = \frac{\sin ϕ_{tp}^{11} (C_{p}^{11} \cos θ_{sp}^{11} \cos ϕ_{sp}^{11} - B_{p}^{11} \sin ϕ_{tp}^{11})}{\cos ϕ_{sp}^{11} \sin ϕ_{tp}^{11} (A_{p}^{11} \cos θ_{sp}^{11} - B_{p}^{11} \sin θ_{p}^{11})}, B_{p} = \frac{\sin ϕ_{tp}^{11} (A_{p}^{11} \sin ϕ_{sp}^{11} - C_{p}^{11} \sin θ_{sp}^{11} \cos ϕ_{sp}^{11})}{\cos ϕ_{sp}^{11} \sin ϕ_{tp}^{11} (A_{p}^{11} \cos θ_{sp}^{11} - B_{p}^{11} \sin θ_{sp}^{11})},$

C.sub.p=1, wherein the lower corner mark .sub.p indexes the propagation path, and the remaining parameter definitions are the same as the before.

[0032] 3) obtaining parameters of the non-line-of-sight path after obtaining the plane equation of the reflecting surface, specifically comprising:

[00008] $θ_{tp}^{k_{t} k_{r}} = \arcsin [\frac{x_{p}^{k_{r} k_{r}}}{\sqrt{{(x_{p}^{k_{r} k_{r}})}^{2} + {(y_{p}^{k_{r} k_{r}})}^{2}}}], ϕ_{tp}^{k_{t} k_{r}} = \arcsin [\frac{z_{p}^{k_{r} k_{r}}}{\sqrt{{(x_{p}^{k_{r} k_{r}})}^{2} + {(y_{p}^{k_{r} k_{r}})}^{2} + {(z_{p}^{k_{r} k_{r}})}^{2}}}],$ $θ_{rp}^{k_{t} k_{r}} = \arcsin [\frac{x_{p}^{k_{r} k_{r}} - R_{k_{r} x}}{\sqrt{{(x_{p}^{k_{r} k_{r}} - R_{1 x})}^{2} + {(y_{p}^{k_{r} k_{r}} - R_{1 y})}^{2}}}],$ $ϕ_{rp}^{k_{r} k_{r}} = \arcsin [\frac{z_{p}^{k_{r} k_{r}} - R_{k_{r} z}}{\sqrt{{(x_{p}^{k_{r} k_{r}} - R_{k_{r} x})}^{2} + {(y_{p}^{k_{r} k_{r}} - R_{k_{r} y})}^{2} + {(z_{p}^{k_{r} k_{r}} - R_{k_{r} z})}^{2}}}] .$ $where$ $D_{p}^{k_{t} k_{r}} = \sqrt{{(x_{p}^{k_{r} k_{r}})}^{2} + {(y_{p}^{k_{r} k_{r}})}^{2} + {(z_{p}^{k_{r} k_{r}})}^{2}} + \sqrt{{(R_{k_{r} x} - x_{p}^{k_{r} k_{r}})}^{2} + {(R_{k_{r} y} - y_{p}^{k_{r} k_{r}})}^{2} + {(R_{k_{r} z} - z_{p}^{k_{r} k_{r}})}^{2}}, x_{p}^{k_{r} k_{r}}, y_{p}^{k_{r} k_{r}}, z_{p}^{k_{r} k_{r}}$

denotes the coordinates of the remaining subarrays od the reflection plane, (R.sub.k.sub.r.sub.x, R.sub.k.sub.r.sub.y, R.sub.k.sub.r.sub.z)=(D.sup.11 sin θ.sub.t1.sup.11 cos ϕ.sub.t1.sup.11+d.sub.rx.sup.k.sup.r, D.sup.11 cos θ.sub.t1.sup.11 cos ϕ.sub.t1.sup.11, D.sup.11 sin ϕ.sub.t1.sup.11−d.sub.rz.sup.k.sup.r).

[0033] In a specific practical experiment, we set 1024 antennas and 4 subarrays at Rx, the communication distance is 20 m and the carrier frequency is 0.3 THz, the precision of the HSPM combined channel model is improved by 14 dB. As shown in FIG. 3A-C and FIG. 4 and Table 1, compared to the existing CE method, the CE method based on the DCNN can complete CE in 0.172 ms, and achieves an estimation precision of 5.2 dB.

TABLE-US-00001 TABLE 1 Method Computational Complexity Running Time (ms) OMP [16] custom-character ((N.sub.pN.sub.t).sup.2) 221 AMP [17] ((N.sub.pN.sub.t).sup.2) 372 CNN 07 (b(N.sub.tN.sub.r).sup.2) 3.64 RNN [40] (cC.sup.3KtK.sub.r) 0.085 DCNN (b(C.sup.2KtK.sub.r).sup.2 0.172

[0034] Compared with the prior art, the HSPM provided by the invention achieves high precision and low complexity of channel modeling, and compared with the PWM, the precision is improved by 14 dB, and compared with the SWM, the complexity is reduced by 99%. The proposed CE method based on the DCNN achieves the improvement of the CE precision of 5.2 dB, and at the same time, since only the channel parameters between the reference sub-arrays need to be estimated, the CE overhead of the method is reduced by 93%.

[0035] The above specific implementation may be locally adjusted by a person skilled in the art without departing from the principle and spirit of the present disclosure, and the protection scope of the present disclosure is subject to the claims and is not limited by the specific embodiments described above, and various implementations within the scope of the present disclosure are not limited by the present disclosure.

DEEP CONVOLUTIONAL NEURAL NETWORK POWERED TERAHERTZ ULTRA-MASSIVE MULTI-INPUT-MULTI-OUTPUT CHANNEL ESTIMATION METHOD

Inventors

Cpc classification

Classification Explorer

H04B7/0456

ELECTRICITY

Classification Explorer

H04B7/0452

ELECTRICITY

Classification Explorer

H04B7/043

ELECTRICITY

Classification Explorer

H04L25/0254

ELECTRICITY

Classification Explorer

H04B7/0617

ELECTRICITY

International classification

Classification Explorer

H04B7/0452

ELECTRICITY

Classification Explorer

H04B7/0426

ELECTRICITY

Classification Explorer

H04L25/02

ELECTRICITY

Abstract

Claims

Description