Machine Learning Based Hybrid Precoding for MmWave MIMO-OFDM with Dynamic Subarray

In the fifth generation mobile communications, the application of millimeter-wave (mmWave) is vital by the virtue of providing high data rate and large bandwidth [1, 2, 3]. However, mmWave channel suffers from a severe path loss, and the traditional fully-digital precoding with massive antennas to mitigate this issue is extremely power consuming [4]. Therefore, hybrid precoding has been proposed to achieve the large array gains with the reduced hardware cost and power consumption [5, 6, 7, 8]. By far, existing broadband hybrid precoding schemes usually focus on fully-connected array (FCA), while seldom consider partially-connected subarray (PCS) like fixed subarray (FS) and dynamic subarray (DS). Therefore, hybrid precoding with PCS in broadband channel is an interesting topic to explore.

Most prior work is based on narrowband mmWave channels [11, 10, 9]. Specifically, a compressive sensing-based hybrid precoding has been proposed in [11], where the channel sparsity is ingeniously exploited to design hybrid precoding with the aid of orthogonal matching pursuit (OMP) algorithm. Moreover, a constant envelope hybrid precoding scheme is proposed, where two cost-efficient sub-connected hybrid architectures are considered to optimize the hybrid precoding under per-antenna constant envelope constraints [10]. To improve bit-error-rate, an over-sampling codebook-based hybrid minimum sum-mean-square-error precoding is designed [9]. On the other hand, mmWave channels appear to have the frequency-selective fading, where OFDM is usually adopted to combat the time dispersion channels [12, 13, 14]. Specifically, an insightful broadband hybrid precoder based on limited-feedback codebook has been proposed for FCA [12]. By exploiting the channel correlation information among different subcarriers, a broadband hybrid precoding is proposed for FS and DS [13]. Finally, [14] has theoretically shown the optimality of frequency flat precoding by proving that dominant subspaces of the frequency domain channel matrices of different subcarriers are equivalent. However, this conclusion is based on the purely sparse channels with discrete angles of arrival (AoA) and angles of departure (AoD), while the explicit precoder solution is not provided.

In this paper, we propose a machine learning based hybrid precoding scheme for mmWave MIMO-OFDM systems with DS, where a shared agglomerative hierarchical clustering (shared-AHC) algorithm is proposed for DS grouping to improve SE performance. First, we propose a PCA-based analog precoder scheme for FS by abstracting the low-dimensional signal space of frequency-flat precoder for given subarray from the high-dimensional signal space of optimal frequency-selective fully-digital precoders using PCA. Besides, the optimality of the proposed PCA-based hybrid precoder design is theoretically proven and verified by simulations. Second, we propose the shared-AHC algorithm inspired by cluster analysis in the field of machine learning for antenna grouping. By implementing shared-AHC algorithm, the SE performance of PCS can be further enhanced for effective antenna grouping adapting to the spatial features of the frequency-selective channels. Finally, we consider the practical passive/active antennas for EE performance analysis. Simulation results confirm the better spectral efficiency (SE) and energy efficiency (EE) performance achieved by the proposed scheme than existing schemes. Meanwhile, DS has the overwhelming advantage for both active and passive antennas.

Notations: Following notations are used throughout this paper. $\mathbf{A}$ is a matrix, $\mathbf{a}$ is a vector, $a$ is a scalar, and $\mathcal{A}$ is a set. Conjugate transpose and transpose of $\mathbf{A}$ are $\mathbf{A}^{H}$ and $\mathbf{A}^{T}$, respectively. The $(i,j)$th entry of $\mathbf{A}$ is $[\mathbf{A}]_{i,j}$, and $[\mathbf{A}]_{i,:}$ ($[\mathbf{A}]_{:,j}$) denotes the $i$th row ($j$th column) of $\mathbf{A}$. Frobenius norm is denoted by $||\cdot||_{F}$. $|\mathbf{A}|$, $|\mathcal{A}|$, $|\mathbf{a}|$, and $|a|$ are the determinant of a square matrix $\mathbf{A}$, cardinality of a set $\mathcal{A}$, $\ell_{2}$-norm of a vector $\mathbf{a}$, and modulus of a number $a$, respectively. The $i$th largest singular value of a matrix $\mathbf{A}$ is defined as $\lambda_{i}(\mathbf{A})$. Additionally, $\text{blkdiag}(\mathbf{a}_{1},\cdots,\mathbf{a}_{K})$ is a block diagonal matrix with $\mathbf{a}_{i}$ ($1\leq i\leq K$) on its diagonal blocks.

We consider an mmWave massive 3-dimensional (3D) MIMO system, where both the BS and user employ the uniform planar array (UPA), and OFDM is adopted to combat the frequency-selective fading channels. The BS is equipped with $N_{t}=N_{t}^{v}\times N_{t}^{h}$ antennas and $N_{t}^{\rm RF}\ll N_{t}$ chains, where $N_{t}^{v}$ and $N_{t}^{h}$ are the numbers of vertical and horizontal transmit antennas, respectively. The user is equipped with $N_{r}=N_{r}^{v}\times N_{r}^{h}$ antennas, where $N_{r}^{v}$ and $N_{r}^{h}$ are the numbers of vertical and horizontal receive antennas, respectively. We consider the downlink transmission, and the received symbols at the user can be written as [11]

where $1\leq k\leq K$ with $K$ being the number of subcarriers, $\mathbf{F}_{\rm BB}[k]\in\mathbb{C}^{N_{t}^{\rm RF}\times N_{s}}$, $\mathbf{F}_{\rm RF}\in\mathbb{C}^{N_{t}\times N_{t}^{\rm RF}}$, $\mathbf{W}[k]\in\mathbb{C}^{N_{r}\times N_{s}}$, $\mathbf{H}[k]\in\mathbb{C}^{N_{r}\times N_{t}}$, $\mathbf{x}[k]\in\mathbb{C}^{N_{s}\times 1}$, and $\mathbf{n}[k]\in\mathbb{C}^{N_{r}\times 1}$ are the digital precoder, analog precoder, fully-digital combiner, channel matrix, transmitted signal, and noise associated with the $k$th subcarrier, respectively, and $N_{s}$ is the number of data streams for each subcarrier. Noise $\mathbf{n}[k]$ satisfies $\mathbf{n}[k]\sim\mathcal{CN}(0,\sigma_{n}^{2})$, and transmitted signal $\mathbf{x}[k]$ satisfies $\mathbb{E}[\mathbf{x}[k]\mathbf{x}^{H}\![k]]\!\!=\!\!\frac{P}{KN_{s}}$, where $P$ is average total transmit power.

The frequency-domain channel $\mathbf{H}[k]$ can be expressed as $\mathbf{H}[k]=\sum_{d=0}^{D-1}\mathbf{H}_{d}[d]e^{-j\frac{2\pi k}{K}d}$ [12], where $D$ is the maximum delay spread of the discretized channels, and $\mathbf{H}_{d}[d]\in\mathbb{C}^{N_{r}\times N_{t}}$ is the delay-$d$ channel matrix. We consider the clustered channel model [11], where the channel is composed by $N_{\rm cl}$ clusters of multipaths with $N_{\rm ray}$ rays in each cluster. Thus the delay-$d$ channel matrix can be written as

where $p_{i,l}^{d}[d]=\sqrt{N_{t}N_{r}/(N_{\rm cl}N_{\rm ray})}\alpha_{i,l}p(dT_{s}-\tau_{i,l})$ is the delay-domain channel coefficient, $\tau_{i,l}$, $\alpha_{i,l}$, and $p(\tau)$ are the delay, the complex path gain, and the pulse shaping filter for $T_{s}$-spaced signaling, respectively. Thus the relationship between the frequency-domain channel coefficiency and the delay-domain channel coefficiency is $p_{i,l}[k]=\sum_{d=0}^{D-1}p_{i,l}[d]\exp(-j2\pi kd/K)$. In (2), $\mathbf{a}_{t}(\phi^{t}_{i,l},\theta^{t}_{i,l})$ and $\mathbf{a}_{r}(\phi^{r}_{i,l},\theta^{r}_{i,l})$ are the steering vectors of the $l$th path in the $i$th cluster at the transmitter and receiver, respectively. In the steering vectors, $\phi^{t}_{i,l}$ and $\theta^{t}_{i,l}$ are the azimuth and elevation angles of the $l$th ray in the $i$th cluster for AoDs, and $\phi^{r}_{i,l}$ and $\theta^{r}_{i,l}$ are the azimuth and elevation angles of the $l$th ray in the $i$th cluster for AoAs. Therefore, the transmit steering vectors for the UPA at the BS can be expressed as $\mathbf{a}_{t}(\phi^{t}_{i,l},\theta^{t}_{i,l})=[1\ \ \ \cdots\ \ \ e^{-j2\pi(m\frac{d_{h}}{\lambda}\sin(\theta^{t}_{i,l})\cos(\phi^{t}_{i,l})+n\frac{d_{v}}{\lambda}\sin(\phi^{t}_{i,l}))}\ \ \ \cdots\\ e^{-j2\pi((N_{t}^{h}-1)\frac{d_{h}}{\lambda}\sin(\theta^{t}_{i,l})\cos(\phi^{t}_{i,l})+(N_{t}^{v}-1)\frac{d_{v}}{\lambda}\sin(\phi^{t}_{i,l}))}]^{T}/\sqrt{N_{t}}$ [11], where $\lambda$ is the carrier wavelength, and $d_{v}$ and $d_{h}$ are the distances between adjacent antenna elements in vertical and horizontal direction, respectively. Similarly, we can also obtain $\mathbf{a}_{r}(\phi^{r}_{i,l},\theta^{r}_{i,l})$ with the same form.

The achievable rate of the mmWave MIMO heavily depends on the transmit hybrid precoder, which can be obtained by solving the following optimization problem [12]

where $\mathcal{F}_{\rm RF}$ is a set of feasible RF precoder satisfying constant-modulus constraint. The coupling between $\mathbf{F}_{\rm RF}$ and $\{\mathbf{F}_{\rm BB}[k]\}_{k=1}^{K}$ and the constant-modulus constraint of $\mathcal{F}_{\rm RF}$ lead to the challenging hybrid precoder design.

In this section, we derive hybrid precoders for the PCS, in which only a subset of antennas are connected to each RF chain. Our goal is to design the optimal frequency-flat RF precoder from the fully-digital frequency-selective precoder.

In Section III, we have found the SE performance of FS heavily depends on $\{\mathbf{F}_{\mathcal{S}_{r}}\}_{r=1}^{N_{t}^{\rm RF}}$. This observation motivates us to optimize the antenna grouping $\{\mathcal{S}_{r}\}_{r=1}^{N_{t}^{\rm RF}}$ to further improve the SE performance when DS is considered.

The DS problem can be formulated as follows

This optimization problem is a combinational optimization problem, which requires an exhaustive search to reach the optimal solution. To obtain the optimal solution, the number of all possible combinations for exhaustive search can be $\frac{1}{(N_{t}^{\rm RF})!}\sum_{n=0}^{N_{r}^{\rm RF}}(-1)^{N_{t}^{\rm RF}-n}\binom{N_{t}^{\rm RF}}{n}n^{N_{t}}$ according to [18], which is a very large number. To illustrate, when $N_{t}=64$ and $N_{t}^{\rm RF}=4$, the number of all possible combinations can be up to $1.4178\times 10^{37}$.

Therefore, a low-complexity algorithm need to be develop to solve problem (14). Specifically, we use the Minkowski $\ell_{2}$-norm [19] to estimate the square of the singular value of the matrix $\mathbf{F}_{\mathcal{S}_{r}}$ by $\lambda_{1}^{2}(\mathbf{F}_{\mathcal{S}_{r}})=\lambda_{1}(\mathbf{R}_{\mathcal{S}_{r}})\approx\frac{1}{|\mathcal{S}_{r}|}\sum_{i=1}^{|\mathcal{S}_{r}|}\sum_{j=1}^{|\mathcal{S}_{r}|}|[\mathbf{R}_{\mathcal{S}_{r}}]_{i,j}|=\frac{1}{|\mathcal{S}_{r}|}\sum_{i\in\mathcal{S}_{r}}\sum_{j\in\mathcal{S}_{r}}|[\mathbf{R}_{F}]_{i,j}|$, where $\mathbf{R}_{\mathcal{S}_{r}}=\mathbf{F}_{\mathcal{S}_{r}}\mathbf{F}_{\mathcal{S}_{r}}^{H}$ and $\mathbf{R}_{F}=\mathbf{F}\mathbf{F}^{H}$.

To reduce the complexity while achieve the good SE performance, we consider the antenna grouping from the viewpoint of clustering analysis in machine learning. To be specific, we propose a shared-AHC algorithm as listed in Algorithm 1, which is developed from the AHC algorithm in machine learning to group the antennas into different subarrays associated with different RF chains. Traditional AHC algorithm is a clustering algorithm that builds a cluster hierarchy from the bottom up. It starts by adding all data to multiple clusters, followed by iteratively pair-wise merging these clusters until only one cluster is left at the top of the hierarchy. The shared-AHC algorithm is different from the traditional AHC algorithm [15] in two distinguished aspects. First, the aim of clustering in our antenna grouping problem is to build $N_{t}^{\rm RF}$ clusters instead of only one cluster in conventional AHC algorithm. Second, the pair-wise merging criterion in the proposed algorithm is “shared”, while the conventional AHC algorithm only considers the target cluster. To further illustrate the “shared” mechanism, we introduce the metric of mutual correlation $g(\mathcal{S}_{n},\mathcal{S}_{m})$ between the cluster $\mathcal{S}_{n}$ and $\mathcal{S}_{m}$

In each clustering iteration, we first focus on a cluster $\mathcal{S}_{n}$, and find a cluster $\mathcal{S}_{m}$ maximizes $g(\mathcal{S}_{n},\mathcal{S}_{l})$ among all possible $\mathcal{S}_{l}$. If the cluster $\mathcal{S}_{n}$ also maximizes $g(\mathcal{S}_{m},\mathcal{S}_{l})$ among all possible $\mathcal{S}_{l}$, we merge $\mathcal{S}_{n}$ and $\mathcal{S}_{m}$. Otherwise, the cluster $\mathcal{S}_{n}$ and cluster $\mathcal{S}_{m}$ are not merged and algorithm goes into the next iteration. Therefore, our proposed algorithm is featured as “shared”, since two clusters mutually share the maximum correlation in the sense of (15). This process is realized in Algorithm 1.

The implementation of PCS not only reduces the hardware complexity, but also improves the EE. In this section, we analyze the EE of the designs. Define the EE as $\eta=RB/P$, where $B$ is the bandwidth of the channel, and $P$ is the total power consumption of the system.

Different connection patterns between the phase shifters and antennas can influence the power consumption. Because the number of phase shifters is different in different connection patterns. In this system, FCA use up to $N_{t}N_{t}^{\rm RF}$ phase shifter for each RF chain connecting to every antennas. While the PCS use $N_{t}$ phase shifters.

Furthermore, the different antenna architectures should also be taken into account regarding the total power. Specifically, we consider the hybrid MIMO system using passive antennas and active antennas as shown in Fig. 10 of [20]. Both of them consist electronic components such as digital-analog convertors (DAC), power amplifiers (PA), local oscillators (LO), and mixers etc. The main difference between active and passive antenna architecture lies in the number of the PAs. In passive antennas, the number of PAs is the same as that of the RF chains. While for active antennas, the number of PAs is the same as that of antennas. This difference can lead to different power consumption because the PAs are heavily power-consuming. Thus we will analyze the power consumption of the two different antenna architectures, respectively.

Given the above antenna architecture, the power consumption for FCA and PCS are respectively $P_{\rm FCA}^{p}=N_{t}N_{t}^{\rm RF}P_{\rm PS}\!+\!N_{t}^{\rm RF}(P_{\rm DAC}\!+\!P_{\rm mix}\!+\!P_{\rm PA}\!+\!P_{\rm LO})$ and $P_{\rm PCS}^{p}=N_{t}P_{\rm PS}\!+\!N_{t}^{\rm RF}(P_{\rm DAC}+P_{\rm mix}\!+\!P_{\rm PA}\!+\!P_{\rm LO})$. By contrast, the power consumption for FCA and PCS with active antenna architecture are $P_{\rm FCA}^{a}=N_{t}N_{t}^{\rm RF}P_{\rm PS}\!+\!N_{t}P_{\rm PA}\!+\!N_{t}^{\rm RF}(P_{\rm DAC}\!+\!P_{\rm mix}\!+\!P_{\rm LO}\!)$ and $P_{\rm PCS}^{a}=N_{t}P_{\rm PS}\!+\!N_{t}P_{\rm PA}\!+\!N_{t}^{\rm RF}(P_{\rm DAC}\!+\!P_{\rm mix}\!+\!P_{\rm LO})$. Moreover, according to the antenna architecture for fully-digital (FD), the power consumption is $P_{\rm FD}=N_{t}(P_{\rm PA}+P_{\rm DAC}+P_{\rm mix}+P_{\rm LO})$. Additionally, the power consumption of electronic components in the three architectures are phase shifter $P_{\rm PS}=30$ mW [21], DAC $P_{\rm DAC}=200$ mW [21], mixer $P_{\rm mix}=39$ mW [22], LO $P_{\rm LO}=5$ mW [21], and PA $P_{\rm PA}=138$ mW [23].

In this section, we investigate the SE and EE performance for the hybrid precoder design. For the channel model, we adopt Dirac delta function as the pulse shaping filter and a cyclic prefix with the length of $D=64$. The number of subcarriers is $K=512$. The transmission bandwidth is $B=500$ MHz. We consider that the path delay is uniformly distributed in $[0,DT_{s}]$ ($T_{s}=1/B$ is the symbol period). The number of the clusters is $N_{\rm cl}=8$, and azimuth/elevation AoAs and AoDs follow the uniform distribution $\mathcal{U}[-\pi/2,\pi/2]$ with angle spread of $7.5^{\circ}$. Within each cluster, there are $N_{\rm ray}=10$ rays. As for the antennas, we consider transmitter adopt $8\times 8$ UPA with hybrid precoder, the receiver adopt $2\times 2$ UPA with fully-digital combiner, and the distance between each adjacent antennas is half wavelength. Moreover, we consider the number of RF chains at transmitter is $N_{t}^{\rm RF}=4$ and the data stream is $N_{s}=3$. Additionally, we consider 4 types of classical FS patterns shown in Fig. 1, where antenna elements with the same color share the same RF chain.

Throughout this part, following baselines will be considered for performance benchmarks: Optimal fully-digital scheme considers the fully-digital MIMO system, where the SVD-based precoder/combiner is adopted as the performance upper bound. Simultaneous OMP (SOMP) scheme is an extension version of the narrow-band OMP-based spatially sparse precoding in [11]. In broadband, SOMP-based hybrid precoding scheme can simultaneously design the RF precoder/combiner for all subcarriers. Discrete Fourier transform (DFT) codebook scheme designs the RF precoder/combiner from the DFT codebook instead of steering vectors codebook in SOMP scheme [9]. Covariance eigenvalue decomposition (EVD) scheme estimates the covariance matrix of the channels using the mean of auto-correlation matrices at each subcarrier [13]. The RF precoder is designed based on the EVD of the covariance matrix of the channels.

In Fig. 2, we compare the SE performance of the proposed hybrid precoding scheme with the baselines, where both FCA and PCS are investigated. In Fig. 2 (a), for FCA, our proposed PCA-based hybrid precoding scheme outperforms conventional DFT codebook-based hybrid precoding scheme and SOMP-based hybrid precoding scheme. Both the proposed PCA-based hybrid precoding scheme and covariance EVD-based hybrid precoding scheme have the very similar performance, and they suffer from negligible performance loss when compared to the optimal fully-digital scheme. This is because mmWave MIMO channels associated with different subcarriers share the same row space due to the common scatterers. Meanwhile, our proposed algorithm can exploit the principal components of the common row space to establish the hybrid precoder. The SOMP-based and DFT codebook-based hybrid precoding schemes work poorly, since their analog codebooks are limited to the steering vector forms. Finally, it can also be observed that the influence of quantization in phase shifters is negligible for our scheme. As for the PCS, Fig. 2 (b) shows that our scheme outperforms conventional covariance EVD-based hybrid precoding scheme with different FS patterns and DS. The antenna grouping scheme in [13] considers a greedy approach, which may lead to the imbalance antenna grouping by acquiring local optimal solution. By contrast, the shared-AHC algorithm for DS grouping introduces the mutually correlation metric (15), which can efficiently avoid this issue. Therefore, the proposed shared-AHC algorithm for DS grouping outperforms it counterpart in [13].

In Fig. 3, we compare the EE performance of the proposed PCA-based hybrid precoding scheme and the baselines with FCA and PCS, where both passive and active antenna architectures are investigated. Note that the power values of key electronic components can refer to Section V. In Fig. 3 (a), for passive antenna architecture, the EE performance of PCS by using the proposed PCA-based hybrid precoding scheme outperforms that of FCA by using the SOMP-based and DFT codebook-based hybrid precoding schemes. The reason is that PCS adopts a much smaller number of phase shifters than FCA. Moreover, DS outperforms the other FS patterns in SE, and it consumes the same power with the other FS patterns. Therefore, DS outperforms other four types of FS patterns. It is worth mentioning that the optimal fully-digital scheme has the worst EE performance, since the numbers of power-consuming PAs, DACs, and mixers are proportional to that of antennas. In Fig. 3 (b), for active antenna architecture, the advantage of EE performance for different FS patterns by using the proposed hybrid precoding scheme over the FCA with several typical hybrid precoding schemes and optimal fully-digital precoding scheme is not considerable. This is because active antenna architecture requires the power-hungry PAs for each antenna. Meanwhile, the advantage of the reduced power consumption of FS structure is greatly weakened by its disadvantage in SE performance when compared to FCA. Finally, the EE performance of DS with the proposed hybrid precoding scheme still has the obvious advantage over the baselines and four typical types of FS with the proposed scheme. This reveals the appealing advantage of DS in practical situation when both the power consumption and SE should be well balanced.

This paper has proposed a hybrid precoding scheme based on machine learning for broadband mmWave MIMO systems with DS. We first acquire the low-dimensional frequency flat precoder from the optimal frequency-selective precoders based on PCA for FS. Then, we extend the proposed PCA-based hybrid precoder design to the DS. We propose the shared-AHC algorithm inspired by cluster analysis in machine learning for antenna grouping to further improve the SE performance. Additionally, we analyze the EE performance for FCA, FS, and DS with passive and active antennas. Simulations further verify the proposed PCA-based hybrid precoding scheme has the better SE and EE performance than conventional schemes.

This work was supported by the National Natural Science Foundation of China (Grant Nos. 61471037, 61701027, and 61201181), the Beijing Natural Science Foundation (Grant No. 4182055), Huawei Innovation Research Program (HIRP), and Youth Project of China Academy of Information and Communications Technology.