Sensor-based predictive communication for highly dynamic multi-hop vehicular networks

We introduce a sensor-aided predictive algorithm for multi-hop link quality estimation. The proposed concept uses vehicle sensor data to improve a link adaptation and end-to-end path selection for a vehicular multi-hop data transmission. The obtained results show that the proposed concept allows better multi-hop link quality estimation and significantly improves end-to-end transmission characteristics in dynamic vehicular environments.


I. INTRODUCTION
The process of driving tasks automation and vehicle-tovehicle (V2V) communications have become key trends in shaping automotive industry of the future.These technological advances combined with various types of mission critical and functional safety applications impose high requirements on the quality of the end-to-end communication between vehicles [1].However, high dynamics of vehicles, acting as communicating partners, combined with unique properties of vehicular environment such as an impact of Doppler effects, presence of scatterers at both sides of the communication link pose a range of distinct link quality estimation challenges compared to existing cellular-based solutions.
In wireless communications, the optimal transmission scheme is adaptively selected based on the estimated channel state information (CSI).The time required for CSI estimation and dynamics of the channel eventually result in a selection of a suboptimal data transmission scheme and in an overall network performance degradation, also known as CSI aging [2].
The problem of channel estimation and optimal end-toend path selection in dynamic multi-hop networks is even more challenging.A large number of possible links and hops to be monitored by a decision making node results in an increased mismatch between the estimated and actual CSI at the transmission time.
The importance of minimizing CSI aging is a recognized problem in literature [3].It has been also shown [4] that technological achievements of modern vehicles lead by developments in automated driving solutions can be beneficially used to improve communication between two vehicles due to a better understanding of the nature of communication link properties.The benefits of sensor-aided prediction for direct link vehicular communications were highlighted in [5].To the best of our knowledge, existing works do not address the the possibility of applying sensor-based vehicular prediction algorithms for a multi-hop link quality forecasting and an endto-end path selection in highly dynamic vehicular networks.
In current work we extend the concept of sensor-aided predictive communications, originally developed for a directlink [5], to a multi-hop V2V scenarios.The key contributions of this paper are: • We evaluate the benefit of sensor-based predictive communications in vehicular multi-hop link-qualityestimation tasks, • We present different ways how the concept of predictive communication can be applied to multi-hop V2V, • We verify the applicability of the proposed approach via simulation in a select vehicular scenario with realistic sensors and environment properties.

AGING
In this section we highlight the problem of an CSI aging in direct-link and multi-hop vehicular communications.
Let us consider a multi-hop decode-and-forward V2V communication environment.Due to high dynamics of vehicles, surrounded by other objects, the communication link between two partners experiences time-varying large and small scale fading effects.We also assume each relay to employ the single carrier frequency division multiplexing (SC-FDM) with the LTE-A uplink reference symbol structure and a Zero-forcing (ZF) equalization strategy [6].

A. Single-hop CSI aging
Given the aforementioned assumptions, a received symbol vector y at subcarrier k can be expressed as [7] where H k,k ∈ C Nr×Nt is a channel matrix for N t transmit and the N r receive antennas between k-th and k -th subcarrier, W k ∈ C Nt×N l is an N l -layer precoding matrix at k-th subcarrier, and s k ∈ C N l is a vector of data symbols with σ 2 s being the average power transmitted on each layer.The component k =k H k,k W k s k represents the impact of the so-called Inter-Carrier Interference (ICI) with the power σ 2 ICI , and n ∈ C Nr denotes zero mean additive white Gaussian noise with variance σ 2 n on antenna n r .Since the pilot symbols are separated in time, the timevarying nature of the channel results in an inevitable channel estimation error between the true channel H and the estimated channel Ĥ at data positions: with σ 2 e being the mean squared error (MSE) at each element.Inserting (1) into (2) and denoting e s k = s k − ŝk , where ŝk is the ZF estimate of s k , we compute the layer-dependent symbol estimation MSE covariance matrix [7]: where σ d is data symbol power and Ψ ∈ C N l ×N l is the inverse of squared estimated effective channel.From (3) the following estimated signal-to-interferenceplus-noise ratio (SINR) at m-th layer is defined [6]: with ψm being m, m-th element of Ψ.
Additionally we assume that precoded symbols are transmitted in blocks using a set of M modulation and coding schemes (MCS), C = {C 1 , C 2 , . . ., C M }, characterized by SINR thresholds S and spectral efficiencies R. The block is assumed to be transmitted efficiently if the estimated SINR γm exceeds the threshold S j with the selected MCS index j.
Due to channel aging, the actual link quality may deviate from underestimated ĵ < j or overestimated ĵ > j applied MCS indexes.In the latter case the link will not be able to successfully transmit data and will need an additional time for the failure detection, MCS reconfiguration and data retransmission.As a result, in both cases the the spectral efficiency of the link becomes underutilized.

B. Multi-hop CSI aging
The problem of delayed channel feedback is even more significant in dynamic multi-hop vehicular environments.
To highlight the problem, let us consider a SISO-based multi-hop communication, where the end-to-end link selection and resource allocation is done at the node initiating the transmission, denoted as A. Also each node is only capable to signal pilot symbols for channel quality estimation only within its scheduled resource blocks.In this case the end-toend multi-hop feedback delay will have twofold contributions.On one hand, estimation of a reliable CSI-feedback at each link takes time: where d ps is a time interval between two consecutive pilot symbols and k avg is a number of consecutive channel estimations needed to obtain reliable channel feedback.
On the other hand the maximum network-relevant delay d nr,est (N ) between a specific node N and the decision making node A is caused by a periodicity of feedback reports d pFB to the decision making node, time per each hop i N to deliver d d,iN and process d pr,iN the channel relevant information over I N hops between A and N : Finally, the vector of maximum feedback delays for the shortest path between the decision making node and each of L nodes of interest will depend on vectors of scheduled feedback periodicities d pFB and multi-hop feedback delivery times: After the required CSI-information has been received by the decision making node, the path selection and resource allocation process will be affected by feedback delays d L .
In contrast to the path selection and resource allocation, the efficiency of the actual end-to-end data transmission at any given link B between node N and its direct neighbor will be also affected by the maximum delay d nr,est (N ) to deliver and process CSI feedback from N to A plus the additional delay to deliver and process data from A to N over I N links: Finally, the dynamics of the SINR aging compared to actual SINR γ(B) for the fixed path at a given link B will be Based on (4), ( 7) and ( 8), the resulting impact of CSI-aging on an end-to-end link efficiency in a multi-hop network will be characterized by the following parameters: • applied channel estimation techniques • relation between the dynamics of the link quality change and the feedback delay • statistical distribution of the feedback delays • applied method of the feedback delivery • ability to predict the expected CSI state

COMMUNICATIONS
The need for channel prediction to compensate a feedback delay is a known problem [3].Existing approaches collect multiple past channel estimates to predict future channel conditions [8].Nevertheless, most of them are applicationor scenario-limited.For example, the performance of spline interpolation and averaging highly depends on the dynamics of environment, while historical averaging fails completely in dynamic environments with direct link communications [8].

A. Sensor-based prediction in V2V communications
Recently it has been shown, that the ability of vehicles to collect surrounding information via on-board sensors brings unique context-aware advantages for direct V2V communications [5].In this paper we extend the sensor-based prediction scheme to the multi-hop vehicular communication according to Fig. 1 and evaluate corresponding applicability regions.
Let us assume a sensor-equipped vehicle capable to fuse available sensor data, process it and conduct feature extraction.This information can be transferred to the communication system, which will learn about the vehicle's position, dynamics, presence of objects with distinct properties and will enable estimation of expected environmental states.Since the exchange of local information about environment is the inherent property of V2V communications, we assume that each vehicle is capable to obtain dynamic and static properties of the surrounding environment including other vehicles with the precision of the available off-the-shelf sensor equipment.
Let us further assume a V2V network with M vehicles involved in multi-hop communication tasks, as shown in Fig. 2. We further assume that the information about environment is periodically exchanged among the vehicles.Finally at the time instant t the following processed information will be available at the decision making node: • a vector of M estimated positions and directions pvh combined with absolute velocities of each vehicle vdv , • a matrix of estimated positions P Ô and estimated dimensions D Ô of Ô detected scattering objects, • a vector of delayed CSI-feedbacks f CSI (γ) from L scheduled links between M nodes, • vectors of corresponding feedback delays d L and data transmission delays d nr,data (N ) defined in ( 7) and ( 8), • a vector of confidence intervals for received feedback parameters, which depend on properties of detected objects.Based on this information, a range of predictive algorithms can be realized at the decision making node by solving an optimization problem and applying a proper constraint.To mention a few: the minimization of the multi-hop CSI-aging impact, minimization of required feedback over the network, selection of the most efficient or stable end-to-end path, etc.
In the current work we limit our scope on applying the predictive multi-hop communication to minimize the impact of CSI-aging on the end-to-end data transmission efficiency between the decision making node A and the node D.
Let us define the optimal efficiency at time t of the endto-end transmission as the spectral efficiency R j,Dw of the weakest link D w between a node D w−1 and D w which has the highest SINR out of all weakest links in all possible paths between A and D. As it was discussed in Chapter II, the actually selected spectral efficiency for each link deviates from the optimal due to multi-hop CSI-aging.Then, if the end-toend path was optimally selected and following (8), the nonpredictive approach according to Section II-B results in the following SINR mismatch: In contrary, for the predictive communication the following algorithm is applied at the decision making node, see Fig. 1: 1) based on available information about environment, such as pvh , vdv , P Ô , DO , and received delayed CSIfeedbacks from each link of interest, predict the expected CSI-states and confidence intervals for every link d out of L at the time of potential link use, 2) find the optimal end-to-end path from A to D given the predicted CSI-states and confidence intervals, Plugging ( 8) in (11) the expanded form will be: (d d,iD w +d pr,iD w )).
(12) As can be seen from ( 12), the sensor-based predictive approach differs by the prediction deviation parameter α(D w ) and by the periodicity coefficient d α(Dw) .

IV. SIMULATION AND RESULTS
In this chapter we define the scenario of interest, describe the simulation setup and present and discuss obtained results.

A. Scenario definition
In order to conduct simulations we first consider the oversimplified configuration with only four vehicles involved in the communication, denoted as A, B, C, D, as shown in Fig. 2. We assume vehicle A to be a source, which collects CSI feedback, and vehicle D a sink of the multi-hop transmission, respectively, while B and C are possible relays.Further we assume only vehicle D to have non-zero velocity of v D = 55 m/s and limit our setup to maximum allowed number of hops being two.We also assume that over the simulation time period the moving vehicle passes next to the obstacle which results in a sharp NLOS-to-LOS transition for the B-D link.

B. Simulation setup
The simulation was conducted in two phases.First, the link level channel mismatch power levels σ 2 e discussed in Section II-A were estimated for a typical range of absolute and relative velocities and typical link level channel conditions for the select scenario, namely Highway: free flow [9].The initial simulations were performed in a Matlab-based V2V linklevel simulation environment, which is developed from the Vienna LTE-A uplink simulator [10].For this purpose a set of changes was introduced to reflect key properties of direct link V2V environment, for the link-level implementation details and simulation setup, see [5].For all simulation scenarios we consider a 5.9 GHz carrier frequency, which results in the most challenging Doppler effects among widely used frequencies for V2V communications.Besides these modifications, we estimated the impact of Doppler effects on the SINR based on available sensor information and sensor accuracy [5].
After the link level properties for a select scenario were obtained, the results in terms of estimated mean squared error (MSE) powers σ 2 e per Doppler shift, Rician K-factor1 and sensor-accuracy were plugged into the specifically developed large scale multi-hop simulation environment.At this point the evolution of large scale parameters for each object was found.The distance dependent pathloss function for V2V Highway scenario defined in [11] was coupled with the impact of three largest Fresnel zones.Then the correlation properties of slow fading process were coupled with the select scenario properties and with an evolution of each link over space according to [12].The prediction-relevant properties, such as velocity, positioning and object recognition error parameters are derived from the existing sensor-fusion solutions, available in [13].Modifications of the simulator and key parameters of both simulation setups are summarized in Table I.

C. Simulation results
To highlight the benefit of predictive algorithms and to understand the regions of their applicability, in the first simulation setup we evaluate the network efficiency loss in terms of SINR mismatch between the actual value and the values available at node A via delayed CSI feedback of the weakest link over different levels of channel variations a γ .For this purpose we fix the multi-hop path to be A-B-D and vary the steepness of the signal attenuation change by changing maximum NLOS attenuation level of the scatterer.
Two different configurations of CSI-feedback updates were considered in this setup: (a) non-predictive periodic CSIfeedback with maximum feedback delays set to d pfb = [7,27] ms; (b) predictive periodic CSI-feedback with positioning inaccuracy modeled by Gaussian process with deviation set to σ = [0.25,0.5] m.In the latter case, the prediction and expected steepness of the SINR change for the NLOS-LOS transition of the weakest link is found based on difference between the actual SINR level at NLOS region and expected SINR at the time of LOS.The predicted SINR for the given link as well as the start and end of the NLOS-LOS transition is estimated based on sensor-aided information about environment, available at vehicles, as discussed in Section III-A.The results of this simulation setup are shown in Fig. 3.The predictive algorithm shows superior performance compared to the non-predictive algorithms.Additionally, it can be seen that position uncertainty of detected objects (vehicles and obstructing objects) influences the efficiency of predictive approach, while an increase in channel variations deteriorates performance of both algorithms.
In the second simulation setup we demonstrate how the predicted information about future CSI at each link reduces the number of feedback messages required to limit the SINR mismatch.For this purpose we consider a non-predictive case with periodic CSI-feedback as in the first simulation setup and predictive approach with the adaptive periodicity of the CSIfeedback generation d pfb,pred , as described in Section III-A.First, the SINR mismatch at different levels of SINR variation rates are calculated for various possible feedback intervals, as described in the previous simulation setup.Then certain tolerable threshold is arbitrary selected based on allowed mean SINR mismatch.Finally, the predicted feedback periodicity is obtained from the intersecting point of the arbitrary threshold and one of feedback periodicity curves at position of the expected SINR variation rate.Fig. 3 illustrates the predictive selection of d pfb,pred = {7, 27} for the arbitrary threshold of 2 dB (which approximately corresponds to one MCS step).The simulation results presented in Fig. 4 show that predictive algorithms are capable to significantly reduce the number of CSI-feedback over the wide range of transition regions compared to non-predictive periodic CSI-feedback.

V. CONCLUSIONS AND FUTURE WORK
We presented the concept of vehicle sensor-aided predictive link quality estimation for multi-hop V2V communications in dynamic environments.The obtained results show the applicability of the proposed concept in multi-hop V2V networks, where nodes are capable to obtain and exchange information about the surrounding environment.In future work, other road scenarios will be investigated.Besides this, further analysis of The transition region is defined by the range of attenuation change, which is the difference between maximum and minimum attenuation over the transition process the impact of scattering reflections, and number of nodes on sensor-aided prediction in multi-hop transmission will be done.

3 )
start transmission and, if applicable, adjust feedback periodicity requests d pFB,pred (D w ) = d pFB • d α(Dw) according to coefficient d α(Dw) dependent on prediction deviation parameter α(D w ), 4) adjust confidence intervals based on the new feedback.Now the predictive SINR mismatch at the weakest link D w depends on prediction deviation per unit of time α(D w ) and time d pred (D w ) to reach the link D w , which is similar to the non-predictive d nr,data (D) case but has variable feedback periodicity parameter d pFB,pred (D w ) instead of fixed d pFB (D): Δγ prd (D w ) = α(D w )d pred (D w ).

Fig. 3 .
Fig.3.Impact of the link variation rate on the SINR mismatch, when a multi-hop feedback is delayed.The link variation rate is defined as the SINR variation a rate of the weakest link in the multi-hop setup

Fig. 4 .
Fig. 4. Average number of feedback messages during NLOS-LOS transition.The transition region is defined by the range of attenuation change, which is the difference between maximum and minimum attenuation over the transition process