Towards Environment-Aware 6G Communications via Channel Knowledge Map

CKM can be classified based on different criteria. According to the types of wireless links, we may have BS-to-any (B2X) and any-to-any (X2X) CKM. B2X CKM aims to portray the site-specific channel knowledge between a BS and its surrounding 2D or 3D location space. This is the right CKM type for the conventional BS-centric communications. Since the location of a ground BS is typically fixed once deployed, B2X CKM only requires the 2D or 3D coordinate of the potential UE locations as the input. Note that the map area/space of B2X CKMs associated with different BSs may overlap, which is necessary to facilitate CKM-based cell association, handover, and inter-cell interference coordination (ICIC). On the other hand, X2X CKM aims to offer site-specific channel knowledge between any pair of transmitter-receiver locations for the area/space of interest. Therefore, X2X CKM is at least 4D or 6D. X2X CKM is important for contemporary wireless networks with device-to-device (D2D) communications, distributed antenna systems, and cell-free networking architectures, where both the transmitters and receivers might be at arbitrary locations. Besides, even for the conventional BS-centric communication, X2X CKM is also quite useful for advanced transmission technologies, such as in-band uplink-downlink full-duplexing, where an UE in uplink communication may utilize the CKM to predict its interference caused to a co-channel UE in downlink. Besides, due to the heterogenous nature of contemporary wireless networks, it is likely that both B2X and X2X CKMs need to be maintained to complement each other.

On the other hand, based on the types of channel knowledge to be learned, CKM may be classified as quasi-stationary or volatile. Quasi-stationary CKM aims to predict the channel knowledge that remains relatively stable, or only vary at large time scales. Typical examples include the path loss, shadowing, presence/absence of LoS links, number of deterministic paths and their powers/phases/delays/AoAs/AoDs. Quasi-stationary CKM may also include the large-scale location-specific (instead of site-specific only) channel statistical parameters like the angular and delay spread, Rician factors, etc. Quasi-stationary CKM requires relatively coarse localization accuracy, say comparable to the channel shadowing coherent distance. One might argue that as quasi-stationary channel knowledge only varies slowly, it only requires infrequent training, and thus incurs little overhead. This seems to render it unnecessary to build quasi-stationary CKM. While such arguments might be true for the simple point-to-point communication links, for future wireless networks involving large channel dimensions and diversified links, we may no longer take the large-scale channel knowledge for granted, as evident by the four typical channel types discussed in Section III-A. On the other hand, volatile CKM involves channel knowledge that varies more frequently, usually due to the varying of the propagation environment that is difficult to predict beforehand. Typical examples include the instantaneous CSI due to multi-path fading and Doppler frequency shift, the channel angular/delay profile or the channel impulse response. While it is challenging to construct volatile CKM for high mobile environment, with the continuous improvement of localization accuracy and advanced data acquisition and processing capabilities, it is becoming more feasible to develop volatile CKM for relatively static or periodically changing environment, like automatic factory in Industrial Internet of Things (IIoT) applications.

The data used to build a CKM can be acquired via offline numerical simulation and offline or online measurements. With offline simulation, location-specific channel knowledge could be generated based on ray tracing with the available physical environment information. While offline simulation-based data acquisition is cost-effective for implementation, the resulting data quality heavily depends on the accuracy and detailedness of the knowledge on the physical environment, which limits their accuracy. This thus calls for on-site channel measurements for data collection. On one hand, dedicated measurement devices such as ground or aerial vehicles could be dispatched offline for data collection exclusively. On the other hand, data could also be collected online concurrently when actual communication takes place. With contemporary wireless networks that involve extensive channel training, but with typically one-shot usage of the trained result, the system only needs to introduce one additional step to store the location-tagged channel training result for online CKM data collection. Compared to offline numerical-based simulation, the offline/online measurement-based data collection is expected to provide more accurate data that better reflects the reality of radio propagation environment, though with higher implementation cost. Thus, in practice, a combination of the three aforementioned data collection methods could be used to complement each other, e.g., building a coarse CKM based on offline numerical simulation and refining it with offline/online measurements. For all the data acquisition methods, a fundamental question is how dense and how frequent the data should be sampled along the space and time dimensions. Intuitively, this depends on the channel knowledge type of interest and its spatial/temporal coherence distance/time. For example, for large-scale channel knowledge such as CSM, data sampling on meter-level is typically sufficient, while that for volatile CKM is usually much shorter.

In terms of the mathematical representation of CKM, the simplest approach is table-based, i.e., all collected data is stored in a table, whose query input is the 2D/3D UE location for area/space B2X CKM and 4D/6D transmitter-receiver location pairs for area/space X2X CKM. On the other hand, for volatile CKM or CKM where wideband frequency-dependent knowledge if of interest, the time and frequency should be also added as query input dimensions. Thus, we may have CKMs of up to 8 input dimensions. To reconstruct the entire CKM based on the finite number of collected data, the interpolation techniques such as Kridging-based method could be applied [13]. However, table-based CKM requires high storage capacity and query response time due to the “curse of dimensionality” issue. For example, for an 8D CKM with even very coarse 10-level quantization on each dimension, we would already have $10^{8}$ entries to store and query the table. Besides, table-based CKM does not really try to learn the intrinsic structure or feature of the data, but instead simply “memorize” it. Therefore, an alternative approach is to learn the CKM based on deep neural networks (DNNs), with the transmitter/receiver location, time, and frequency as the DNN input, and the channel knowledge to be predicted as the output. As a result, the powerful deep learning algorithms such as backpropagation can be applied to train the coefficients of the DNN with the collected data. Compared to table-based CKM, DNN-based CKM resolves the “curse of dimensionality” problem and is more responsive for channel knowledge prediction once properly trained. Note that there are different variations for the DNN-based CKM architecture. For example, for CKM with frequency-selective channel knowledge, if we are only interested in the channel knowledge of a few sub-bands, then we may move the frequency dimension from the DNN input to output by expanding the output dimension accordingly. One such example is discussed in Section IV-A.

Channels for yet-to-reach locations: Existing methods for CSI acquisition mainly rely on pilot-based channel training, which, however, is possible only when the communication devices actually reach the locations for the channels to be trained. In contemporary wireless networks, it is often useful to acquire the channel knowledge for those wireless links with yet-to-reach or never-to-reach locations to gain a global channel view, so as to enable predictive network planning or resource allocation. This is possible by invoking the channel prediction capability of CKM. For example, by utilizing CKM together with UE trajectory prediction, foresighted handover mechanisms could be developed to resolve issues like “ping-pong” handover in complex environment. As another example, with CKM over the 3D aerial space, a cellular-connected UAV may plan its flying trajectory to avoid coverage holes in the sky [14]. Such tasks are difficult, if not impossible, to accomplish if purely relying on conventional training-based channel estimation.

Channels for non-cooperative nodes: Typical examples for channels with non-cooperative nodes include: (i) the wireless links between secondary transmitters and primary receivers in cognitive radio systems; and (ii) the wireless links between legitimate transmitters and adversary eavesdroppers for physical layer security communications. The non-cooperative nature of such nodes makes it difficult to acquire their CSI with conventional channel training approach, though they are essential for interference mitigation and security protection. CKM offers a promising solution to resolve this issue, by predicting the channel knowledge mainly based on their location information, which is more readily available, say with passive sensing for eavesdroppers or directly retrieved from the network database for primary receivers.

Channels with large dimensions: The overhead of channel training and feedback, in terms of communication resources like training energy, pilot sequence length, and response delay, increase drastically with channel dimensions. 6G is expected to involve communications with extremely large antenna arrays, super high frequency (up to Tera-Hertz), wide bandwidth, and ultra-massive connectivity, but at the same time targeting for sub-milliseconds (ms) communication latency. This brings a big challenge for conventional pilot-based channel training. On the other hand, CKM offers a promising solution to facilitate or even obviate real-time CSI acquisition, which is quite appealing for simultaneously achieving high-capacity and low-latency communications.

Channels with severe hardware/processing limitations: The large channel dimension together with the imperative needs for energy- and cost-effective implementation of 6G requires using simple hardware and low-complexity signal processing techniques, such as analog or hybrid analog/digital beamforming with only one or very few radio frequency (RF) chains, low-resolution analog-to-digital converters (ADCs), or even fully passive transmission/reflection with reconfigurable intelligent surfaces [15]. Such restrictive transceiver architectures exacerbate the training and feedback overhead of conventional pilot-based training for channels with large dimensions. In this case, CKM-assisted wireless network is expected to provide a promising solution by exploiting the environmental-awareness for reducing or even avoiding training overhead.

Depending on the specific applications and performance requirements, CKM can be utilized in two different ways: training-free communications and CKM-assisted channel training.

With CKM-enabled training-free communications, the channel knowledge predicted by the CKM is directly used for the system design and performance optimization, without requiring the conventional pilot-based channel training. This is usually applicable when the CKM is sufficiently accurate and/or when only coarse performance guarantee is needed. Some possible examples include power allocation, interference mitigation, cell association, and user pairing based on large-scale channel gains (e.g., path loss and shadowing). In certain scenarios, the CKM-predicted channel knowledge is firstly used to reconstruct the channel before being used, an example of which is discussed in Section IV-B. CKM-enabled training-free communication is especially appealing for low-complexity and low-latency communications since it eliminates the overhead and processing delay associated with channel training and feedback.

On the other hand, with CKM-assisted channel training, CKM is used as additional side information to enhance the performance of conventional pilot-based channel training, as illustrated in Fig 2. This is usually the case when only coarse CKM is available and/or when strict performance guarantee is needed. Besides CKM, another side information that is useful for channel training is the (possibly coarse) location information of the transmitter and/or receiver, which is readily available with variety of localization techniques today, such as GPS, vision/light-based localization. Note that not only the communication channel, but also refined node locations can be estimated with training. Since side information at least does not hurt, it is expected that CKM-assisted channel training with environmental-awareness in Fig.2(b) is superior than the conventional pilot-based channel training in Fig. 2(a), in terms of the performance enhancement and overhead reduction. For instance, for the mmWave beamforming example shown in Fig. 1(b), the conventional environmental-ignorant beam training would require exhaustive beam sweeping over the entire 3D space. By contrast, the searching space can be significantly reduced with CKM-assisted beam training, thanks to its environmental-awareness.

D2D communication enables devices that are in close proximity to exchange information directly, without having to traversing the BS or core network. Consider an area consisting of $K$ D2D user pairs that need to share $N$ frequency sub-bands, where $N<K$. One fundamental problem is sub-band assignment, i.e., choosing one out of the $N$ sub-bands for each D2D pair, such that the achievable sum rate of all D2D users is maximized. Clearly, the solution depends on the available channel knowledge, not just of those direct links between a D2D transmitter and its desired receiver, but also those cross links, since D2D users assigned to the same sub-band would interfere with each other. Therefore, with pilot-based channel training, a total of $K^{2}N$ channels need to be estimated, which would incur high training overhead and complex handshaking processes when $K$ and/or $N$ is large. This issue could be circumvented by CKM-enabled environment-aware D2D sub-band assignment.

The 3D physical environment of the studied urban area for D2D communications is shown in Fig. 3(a). A X2X CGM is constructed to predict the channel gains for each of the $N$ sub-bands, with the transmitter-receiver location pair given as the input. The CGM is trained by a fully connected feedforward DNN with one input layer of dimension $4$, corresponding to the transmitter-receiver pair locations, one output layer of dimension $N$, corresponding to the predicted channel gains of the $N$ sub-bands, and 7 hidden layers that have 64, 128, 256, 512, 256, 128, 64 neurons, respectively. The standard ReLU activation is used for each hidden layer, and the DNN is trained to minimize the mean squared error (MSE) between the predicted and true channel gains. The commercial ray tracing software Remcom Wireless Insite²²2https://www.remcom.com/wireless-insite-em-propagation-software is used to simulate the actual radio propagations and generate the training and test data. Fig. 3(b)-(d) shows the resulting DNN-based CGM associated with one example transmitter location and sub-band. As a benchmark comparison, we also consider the resulting CGM with the classical path loss model based channel gains, where the modelling parameters including the path loss exponent, intercept, and frequency-dependent factor are curve-fitted based on the same training data as the DNN-based approach. It is observed from Fig. 3 that the DNN-based CGM matches quite well with the true map, and they both follow the layout of the physical environment in Fig. 3(a). By contrast, the fitted path loss model-based CGM results in concentric contours for channel gains, which is far from the reality, due to the restriction of the theoretical path loss model that only depends on node distances, rather than their specific locations.

Fig. 4 shows the achievable sum rate with D2D sub-band assignment based on perfect CSI, CGM, and fitted path loss model, respectively, where $K=30$ and $N=12$. Note that since the sub-band assignment is a combinatorial optimization problem, which is NP hard in general, we apply a greedy-based assignment algorithm to all the three schemes, i.e., the $K$ D2D users are considered successively, and for each D2D user under consideration, the sub-band that would lead to the highest sum rate is selected. It is observed that the CGM-based assignment significantly outperforms that based on fitted path loss model. This is expected since the former is able to predict the channel gains more accurately, as evident from Fig. 3. The figure also shows that the CGM-based scheme performs closely to that based on perfect CSI, yet without requiring any channel training. This demonstrates the great potential of CKM-based resource allocation for wireless networks.

For the second example, we consider a mmWave communication system in an urban environment shown in Fig. 5(a). To enable training-free mmWave beamforming, a B2X CPM is constructed to predict the essential information of the significant channel paths. In our numerical results, the output of the CPM include the path gain, phase, zenith AoD $\theta$ and azimuth AoD $\phi$ of the three strongest paths. Thus, the input and output dimensions of the CPM is 2 and 12, respectively. The Remcom commercial ray tracing software Wireless Insite is used to generate the groundtruth data, which is stored in a table since the query input dimension is small. Based on the finite ray tracing data samples, the inverse distance weighting (IDW) method of the $K$ nearest neighbors (KNN) is used to construct the entire CPM of all locations. With $K=3$ nearest neighbors used, Fig. 5 shows the resulting azimuth AoD maps of the strongest path. As a comparison, the location-based AoD map is also shown, i.e., the AoDs are simply calculated based on the relative positions of the BS and receiver, while irrespective of the propagation environment. It is observed that the KNN-based CPM matches quite well with the ground-truth map. By contrast, the location-based azimuth AoD map is given by the simple rays, which is far from reality due to its ignorance of the actual radio propagation environment.

Fig. 6 shows the average communication rate versus the number of BS transmit antennas with mmWave beam selection based on perfect CSI, CPM, and UE locations, respectively. The average is taken with respect to $1000$ randomly selected UE locations that have at least one detectable channel path. The BS is assumed to be equipped with $N_{z}\times N_{y}$ uniform planar array (UPA) with adjacent elements separated by half wavelength, where $N_{z}$ and $N_{y}$ represent the number of vertical and horizontal elements, respectively. We assume that $N_{z}$ is fixed to $10$ while $N_{y}$ varies. The low-complexity codebook-based analog beamforming is performed, where the codebook is generated based on the transmit array response vectors with uniformly sampled values over $\sin(\phi)$ and $\cos(\theta)$, with resolution $1/(2N_{y})$ and $1/(2N_{z})$, respectively. For CPM-based scheme, the channel is first reconstructed based on the predicted path information, according to which the best beam is selected. Besides the ideal case with errorless UE location information, we also consider the practical cases with mean UE localization errors of 1 meter (m) and 5 m. It is firstly observed that the CPM-based beam selection, which is environment-aware, significantly outperforms the location-based scheme. It is also observed that even with a practical localization error of 1 m, the CPM-based scheme, which is training-free, can already achieve a significant portion of the performance of that with perfect CSI, e.g., around $90\%$ when the number of transmit antennas is $400$. Fig. 6 also reveals that for both CPM-based and location-based beam selection, as the number of transmit antennas increases, the beamforming performance becomes more sensitive to the UE localization accuracy, which is expected due to the sharper beam formed with larger antenna arrays.