Performance evaluation of vertical handover in Internet of Vehicles

Handover (HO) in the vehicular network is challenging to perform since the mobility of vehicles changes. Many research works have studied this issue and performed handover without any degradation in network metrics. In the study of Chang et al. (2019), a cluster-based handoff, and dynamic edge-backup node (DEBCK) is proposed where the vehicles on the road lane were clustered, and the backup node provides handoff. Here, the cluster head performs the handoff and the backup mobile edge vehicle. The three main parameters that were taken into account for handoff are storage, communication, and energy. The main drawback of this work is poor handoff performance of backup mobile edge and cluster head, and failure to perform handoff whenever there is a need. In the study of Jubara (2020) a procedure for HO was proposed with the aim of minimization of delay in HO. A cross-layer protocol in an adaptive L4 HO procedure begins to estimate signal strength and if the quality of the signal was poor, then the link between user and base station disconnects. Then the Stream Control Transmission Protocol (SCTP) is assigned to the new IP and it is updated to the layers. However, the signal strength was not the only significant metric to make HO decisions. Due to the mobility of the vehicle and moving pattern on the road lane, HO of moving vehicles was proposed (Choi et al., 2018). According to the idea of this work, a group of users consists of a mail leader, sub-leader, and follower. The sub-leader was selected based on the maximum number of connections. In case if more than one vehicle has similar characteristics then, a sub-leader was selected at random. Initially, the vehicle computes reference signal received power (RSRP), reference signal received quality (RSRQ), link quality, and is reported for HO decision. A decision tree was built for HO decision-making using RSRP measurement. But the vehicles HO in a group requires frequent computation in the group, as well as measurement, and hence the computation will be higher in this work. The network layer-based L2 extension HO scheme was proposed (Naeem et al., 2019) and the architecture consists of an access router (AR), roadside unit (RSU), and vehicles. This work defines two HO schemes as inter-AR HO and intra-AR HO. The key goal of this scheme was to minimize latency and improve the packet delivery ratio. A fuzzy logic model and Elman Neural Network (ANN) was designed to decide along with the assurance of QoS (Naeem et al., 2019). For HO decisions, the parameters that are taken into account as cost, transmission range, velocity, load, and capacity. Even though this work performs better, the time for HO decision consumes time which increases the delay in the HO that may cause packet drop and degrades packet delivery ratio. The paper (Singh et al., 2020) concentrates on handover as well as routing. A handoff protocol was proposed that computes link expiration time (LET) for detecting the connectivity between vehicles. The partner selection protocols enable a selection of optimal partner nodes (PN). Initially, the route was determined from GPS information and then the partner in the routes was selected from the vehicular LET using the traffic information. The vehicle with a high LET will be selected as the optimal PN in the route. In this work, only a single metric was taken into account for selecting a route between source and destination. However, if an opposite moving vehicle with high LET cannot be selected as PN and hence it requires considering other parameters too. In the study of Leu et al. (2019), and enhanced Access Network Discovery and Selection Function (ANDSF) was presented to perform a BS selection in the network. This algorithm combines with multilayer perceptron (MLP). The parameters were load, signal strength, throughput, and delay. The traditional workflow of the ANDSF is illustrated in Figure 1.

Figure 1:

The ANDSF was equipped within the EPC which was started to be used in 3G and also on advanced radio access networks. This server was employed to discover information, manage policies, select policies, manage rules, and others. The user equipment can be a sensor, vehicle, or any other device that can access radio technology. The server first discovers the device and then performs a change in the connectivity. The procedure works by the developed set of rules and policies.

The vertical HO was performed using multi-criteria methods by taking into account the significant parameters such as QoS, delay, cost, and others (Hamurcu and Eren, 2020). Due to the consideration of multiple metrics for HO decision using enhanced Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) fuzzy logic (Embus et al., 2020). The working of this combination of algorithm works as per the following steps:

Step 1: creates decision matrix using the parameters that were involved for HO decision. The computation was executed for each available network in the coverage area.

The processing steps illustrated above for enhanced TOPSIS using fuzzy were able to overwhelm the problems in conventional RSS-based HO. Each step includes multiple criteria, these steps were not parallel, i.e. on each HO request, all the process requires to be performed and the decision was made after ranking.

The IoV environment that uses different types of radio access network due to the coverage range of each radio access. However, the vehicles have in-built DSRC for short-range data transmission, while the destination vehicle moves far from the source, then a route has to be preferred for data transmission.

Vehicles perform routing by selecting relay vehicles between the source and destination since the DSRC range was small and hence it is not able to connect longer distance vehicles. In the study of Nguyen and Jung (2020), Ant Colony Optimization (ACO) algorithm is proposed with the idea of coloring vehicles. This algorithm presents two processes as solution construction and pheromone update. The idea of coloring was to give similar colors for the vehicles that have the same destination. As per the pheromone value, the route was selected in this work. However, this work failed to consider the significant parameters of the vehicles for the computation of the pheromone value that decides the transmission route. In the study of Al-Kharasani et al. (2020), a cluster-based adept cooperative algorithm (CACA) is proposed focusing on the QoS metrics. As per this work, clustering formation is done and a cluster head was selected. This work follows Optimized Link State Routing (OLSR) protocol with the Multi-Point Relay (MPR). This selection takes into account mobility factors, distance range, and quality of path (QoP). The vehicles that satisfy these parameters were selected as MPR and then the intersection vehicles were eliminated. The selection of MPR was not efficient, since the vehicles move at high speed. A protocol design was proposed, i.e. partner selection protocol that considers Vehicle Link Expiration Time (VLET) (Ndashimye et al., 2020). In this work the handoff means a vehicle disconnects from a partner node and joins a new partner node (PN), the partner node enables to perform data transmission. The only measure that was used in the selection of PN was not efficient since there are other significant metrics as signal strength which was also essential in node selection. A cross-layer design was proposed (Leu et al., 2019) that selects an optimal route based on the metrics forwarding probability, bandwidth, and link duration. The forwarding probability for the vehicle was formulated by considering velocity, distance, and communication range. The link duration was mathematically calculated as communication link lifetime that takes into account vehicle velocity, GPS location, and communication range. Then, the third parameter of bandwidth was calculated from link gain, noise power, and channel bandwidth.

The relay node selection was presented in the study of Cao et al. (2019) for relay selection using the estimation of curving rate. A double direction relay node selection was involved when the request to broadcast (RTB) was 1 and then it select relay from the estimation of curving rate, delivery ratio, one-hop delay, and message dissemination speed. The curving rate was formulated from the road length and the range of the vehicle. The computation of each parameter one after the other for route selection was time consuming and it leads to higher packet drop.

Routing is also performed using optimization algorithms. In the study of Leu et al. (2019), a hybrid optimization algorithm is proposed combining monarch butterfly and gray wolf optimization for route selection. The parameters that were taken into account for route selection are different costs computed for congestion, collision, travel, and QoS. For QoS prediction, fuzzy membership functions were applied. Initially, the butterfly algorithm was involved and then the gray wolf was performed for position updates and selecting optimal paths. The traditional issue in gray wolf optimization is its poor performance, and low accuracy. Fuzzy logic was also used to select routes by estimating link quality and achievable throughput. The link quality was based on the position, direction, and expected transmission count. As per the fuzzy weight, the output of the selection of next hop relay was performed. However, this work failed to tolerate the mobility issues concerning vehicular communication.

The proposed heterogeneous IoV network is designed with vehicles consisting of a 5G base station, LTE base station, RoadSide Unit (RSU), and vehicles. The entities that participates in this system are defined below.

Definition 1: Vehicle – the vehicle moves on a restricted path, i.e. on-road lane in which the path is pre-defined in a map. The moving speed of the vehicle depends on the vehicle. Vehicles have in-build GPS, using which their latitude and longitude information is gathered. The location of the vehicle and the speed of the vehicle is dynamic. Vehicles use DSRC and other advance Ran for data transmission. It transmits safety and non-safety data.

Definition 2: RSU – RSU is employed in IoV for performing communication with the infrastructure. This entity is static in the environment and also it enables DSRC for vehicles.

Definition 3: 5G mmWave base station (BS) – the BS is static and this allows to perform high speed–short-range communication. It can solve the lack of spectrum issue.

Definition 4: LTE BS – this BS is also static and it allows long-distance communication with higher bandwidth and comparatively high spectrum efficiency.

The proposed system model is depicted in Figure 2, which composes all the above-defined entities into the system. The road lane has ‘n’ number of moving vehicles in their direction on the road. In this work, the handover is a decision that will be taken by the vehicle only when the current base station link is not good. But in case of sudden need in transmitting a safety application, it makes network selection process at that moments along with the consideration of data type as one of the parameters. Handover decision is the decision by which the need for handover is determined and it performs handover to the available network. For handover decision dynamic Q-learning in which the threshold is set as per the environment. If the handover has to be performed, it then selects a network from fuzzy-convolution neural network (F-CNN). For network selection, the fuzzy rules are defined and used in CNN. Then routing takes place by using an optimization algorithm called jellyfish algorithm that selects V2V pairs between source to destination and so, it is called V2V chain routing.

Figure 2:

Handover decision by dynamic Q-learning, the dynamic means to use threshold concerning the available network. Dynamic Q-learning algorithm determines the need for handover by evaluating vehicle speed and signal strength. We set the threshold for signal strength using Shannon entropy rule as shown in the following equation: ()(1)S(ss)=E[−log(P(ss))](1)where S(ss) denotes the Shannon entropy for signal strength that composes of ss values for DSRC, mmWave, and LTE that range between (‒30 to ‒70 dBm). P(ss) denotes the probability of the signal strength (Figure 3).

Figure 3:

Let Q(S,A) represent state S and action A based on the Q-values. Each state S will have two parameters and this Q(S,A) is determined and updated in the rule. The temporal difference update rule is as follows: ()(2)Q(S,A)+α(R+γQ(S′,A′)−Q(S,A))→Q(S,A)

The term Q(S^΄,A^΄) defines next state and action R is the reward given by the agent, γ is the discount factor that is [0–1], then α is the learning rate [0–1], i.e. it denotes the step length to estimate the (S,A). The action is taken using ϵ-greedy policy where ϵ represents epsilon. The pseudo-code for dynamic Q-learning is given below to decide the decision for handover:

Network selection is the process of selecting a network from the available RANs. F-CNN algorithm is applied for network selection. The CNN is designed with layers of convolution, max-pooling, and fully connected layers. The layers are employed with fuzzy rules that are defined from the metrics signal strength, the distance between BS and vehicle, vehicle density in serving BS, data type (safety or non-safety), and line of sight. The definition for each metric is depicted below.

Definition 1: Signal strength – signal strength defines the SNR which gives the number of signals. A channel will compose noise as well as signal, the high the noise, the channel is unfit for transmission. The SNR (S_r) is determined from signal power P_s, and noise P_N respectively. The formulation is: ()(3)Sr=PsPn

Definition 2: Distance between BS and vehicle – the distance between BS and a vehicle is estimated using Euclidean distance. This measure defines the stability of the link, as the distance increases the link will be unstable and when the distance decreases the link will be stronger. Euclidean distance is computed using the following equation: ()(4)D(LBS,LV)=(x−x1)2+(y−y1)2

For computing distance, the coordinate points of the BS and vehicle is used. Distance D(LBSLV) is determined from the BS location coordinates of (x,y), and vehicle location coordinates of (x₁,y₁), respectively. The location of BS is fixed and so it requires to know only the vehicle coordinate for distance estimation.

Definition 3: Vehicle density – the density of vehicle V_D denotes the number of vehicles that are connected with that particular BS. ()(5)VD=∑(NCL,NNL)(5)where N_CL and N_NL represents the number of connected links and number of new links.

Definition 4: Data type – the data type in vehicles are two, they are safety and non-safety. In this work, safety is denoted as 0 and non-safety as 1. The safety messages will be of traffic information, high-speed vehicle information. This type of data has a higher priority in transmission than the non-safety data.

Definition 5: LoS – line of sight defines the direct contact between the vehicle and BS without any obstacles that block the signals. For transmission, LoS is only preferred and the signals in Non-LoS are not preferred.

The above five metrics involve the development of fuzzy rules. The fuzzy logic deals with the decision-making by the defined rules as shown in Table 1. The mmWave signals will be chosen for any type of traffic, but only when the LoS is present since blockage of mmWave leads to poor performance, in case of blockage the vehicle selection will be 4G LTE.

The fuzzy logic method operates with the IF-THEN rules in the interference engine. The input is in crisp values that are converted into a fuzzy set. As per the fuzzy rule, the interference engine constructs membership function between [0,1]. The fuzzy logic operations are built into CNN. Figure 4 depicts the constructed fuzzy logic with CNN. The output high (H), medium (M), and low (L) denotes as follows: ()(H,M,L)→(mmWave,LTE,DSRC)

Figure 4:

A pseudo-code below is illustrated based on the workflow of this fuzzy-CNN algorithm:

The use of CNN will give results for multiple vehicles at the same time by parallel processing. The proposed fuzzy-CNN is composed of 32 rules, which are defined from five parameters. Since the CNN can process in parallel, the 32 rules will be processed in the convolution layer. According to the selected network, the requested vehicle will handover from the current network to the target network.

The process of routing is carried out using jellyfish optimization algorithm where the vehicles are formed like V2V pairs, hence the name V2V chain routing. The routes are selected by computing the objective function using channel metrics (SNR (s_r), link quality (l_q)), vehicle metrics (Speed (s_p), Relative direction (R_d)), and vehicle performance metrics (Delay (D_l), throughput (T_p)).

A time control mechanism is used to switch between active or passive movements in this algorithm. The time control c(t) is formulated and computed using the following equations: ()(6)c(t)=|(1‒tMaxite)×(2×rand(0,1)‒1)|

when rand(0,1) > (1‒c(t)), then passive motion: ()(7)rand(0,1)<(1‒c(t))thenactivemotion

Here the jellyfish are assumed as vehicles and the ocean is assumed as road lane where the vehicle moves in different speed.

The ocean current direction represented as OC→ and it is mathematically given as below:

Let: ()(8)OC→=1Vp=X*‒ecμ

Then: ()(9)OC→=X*‒dff(9)where ‘V_p’ is the vehicle density. X^* denotes the best location, μ is the mean location, and e_c is the attraction factor, here the attraction of on destination. Then, the objective function is defined to select a best route. This function OF is formulated as follows: ()(10)OF(Ms)=∑(sr,lq)(sp,Rd)(Dl,Tp)

Ms represents a set of parameters in which the delay and speed must be minimum and all the other parameters can be a maximum value for the selection of the routes. Here the OF is applied for the complete route, since this work selects an optimal route from the available routes. The metrics are estimated from the channel: ()(11)lq=1Pf×Pr(11) ()(12)Rd=2rsinsin2(Δla2)+cos(lav)×cos(lanpi)×sin2(Δln2)(12) ()(13)Dl=PLb(13)where P_fP_r represents the number of transmitted and the received packets in the same link between two vehicles, (la,ln) represents the latitude and longitude, (la_v,ln_v) represents the vehicle location, and (la_np, ln_np) represents the next hop location and r is the radius. P_L, b represents packet length and bit rate, i.e. transmission speed in bits per second that are used to compute the delay estimation.

Equation 10 defines the objective function through which the optimal route is selected using jellyfish optimization algorithm.

The performance of the proposed HO, network selection, and routing algorithms are evaluated in the next section.

The proposed work is simulated using OMNeT++. Table 2 shows the simulation parameters assumed in our proposed work.

The comparative analysis gives the obtained results in comparative graphs. The proposed work is compared with previous works that use conventional RSS-based selection, TOPSIS, ANDSF, and V2I-MoloHA methods relating to handover, network selection, and routing issues. It is a multi-criteria decision-making algorithm that processes with more than one criterion. The parameters that are considered for the evaluation are mean handover, handover failure, throughput, and delay.

Mean handover and handover failure

The mean handover is the number of successful handovers of a vehicle from one network to another. Handover failure is defined as the number of unsuccessful handovers that happen due to poor decision-making.

The lesser mean handover denotes the better performance of the proposed algorithm as it has minimized the number of unnecessary handovers in the network. In previous work of TOPSIS, it was used for the selection of network that fails to perform proper ranking. Similarly, the use of parameters for the selection of network was either based on vehicle characteristic or environmental characteristic which leads to select the best target network that eventually increases mean handover along with the increase in the handover failure.

The proposed dynamic Q-learning algorithm can learn the vehicle environment in a particular surrounding. The prediction of handover requirement from the vehicle speed and signal strength is efficient. Further to the prediction, we perform a selection of networks using the F-CNN algorithm for selecting a network by analyzing the metrics of the particular vehicle. The process of prediction and network selection in this work tends to improve the performance of the handover-based metrics.

Figures 5 and 6 illustrate the mean handover and handover failure concerning the increase in vehicle speed. The improvement in the performances of HO failure rate and mean handover is due to the handover decisions made by dynamic Q-learning algorithm and appropriate selection of networks due to fuzzy-CNN. The mean handover in the proposed work decreases with the increase in vehicle speed and hence, suitable for large-scale environments. Besides, the decrease in mean handover reduces the HO failure counts. In general with the increase in vehicle speed, the handover failure occurs but as the proposed work uses Q-learning for predicting the requirement of handover before that of the network selection it can take an absolute decision at the increase of vehicle speed. The main reasons behind the degradation of handover are illustrated below.

Figure 5:

Figure 6:

Selection of parameters to select the suitable network which requires considering vehicle metrics as well as the BS metrics:

The number of handovers increases due to the absence of prediction of the vehicle regarding the need for handover. This leads to an increase the number of unnecessary handovers which also requires large resource blocks for performing the computations.

The handover failure rate HO_FR is computed mathematically based on the below equation: ()(14)HOFR=HOFHOS+HOF

The termsHO_F and HO_S represents the number of handover failure and handover success, respectively. According to the count of these measurements, the handover failure rate is determined. The handover failure is caused because of the poor handover decision; hereby the proposed work first predicts the handover requirement from the vehicle request by learning the environment and then if the decision is to perform handover, it selects a target network. From the comparative graph, the average value of failure rate in proposed is 0.015, while the previous work achieves 0.13, 0.04, 0.07, and 0.03 in conventional, TOPSIS, ANDSF-HO, and V2I-MoLoHA, respectively. The minimization of handover failure reflects on absolute handover decision. Similarly, the reduction in the number of handover shows that the unnecessary handover is reduced by efficient prediction and network selection in proposed.

Table 3 gives a comparison on the average values estimated from the performance of conventional method, TOPSIS, ANDSF-HO, and V2I-MoLoHA in terms of number of handover and handover failure. Then the improvement percentage of handover efficiency is depicted in the above table. The handover efficiency impact on other network parameters that enhances overall network efficiency.

Table 3.

Comparison of HO efficiency.

Method	Average number of HO	Better efficiency	Average HOFR	Better efficiency
Conventional	5.51	55%	0.133	90%
TOPSIS	2.15	20%	0.041	40%
ANDSF-HO	3.57	30%	0.069	60%
V2I-MoLoHA	3.03	25%	0.029	20%
Proposed	1.30	–	0.01	–

Throughput, delay and packet loss

Throughput is one of the significant performances metric in a network and it is mathematically computed using the formula as follows: ()(15)T=PszRtt×1.2PL0.5

The throughput T is estimated from the packet size P_SZ, round trip time R_tt and packet loss PL.

Figures 7 and 8 show the graphs for throughput, and delay. From Figures 7 and 8, there is an increase in the throughput and decrease in delay when compared to the existing techniques. This is due to the optimal selection of routes using jellyfish optimization algorithm. The graph shows little increase, and drops in the delay. The end-to-end delay is determined in terms of transmission delay between the relay vehicles from the source vehicle to the destination vehicle. The end-to-end delay (EE_D) is determined as follows: ()(16)EED=Nb(N0,N1)R(N0,N1)+Nb(N1,N2)R(N1,N2)+⋯…+Nb(Nn,Nm)R(Nn,Nm)

Figure 7:

Figure 8:

Let the vehicles in a route be represented as N₀, N₁, N_2,…, N_n, N_m for which the number of bits in each node is denoted as Nb and the rate of transmission is R. The N₀ is the source vehicle node, N_m is the destination vehicle node, while the nodes are the relay. In accordance to the estimation of the delay a better efficiency in the route selection is analyzed. In a route, the delay occur between every pair of vehicle due to the use of signal strength and hence the delay is predicted for each pair and end-to-end delay from the source to destination is determined.

The comparative results depict that proposed work is better than the previous conventional method, TOPSIS, ANDSF-HO, and V2I-MoLoHA. Among all the previous work, the conventional method of using only signal strength results in poor performance due to the growth of multiple challenges in data transmission of vehicles. Table 4 illustrates the mean value determined for each work in the performance of throughput and delay. Based on the mean throughput and delay, the percentage of improvement is proposed than the existing works. As per the comparison, a minimum of 23% and a maximum of 46% is better performance than the previous methods in this network (Table 5).

Table 4.

Comparison of throughput and delay.

Method	Mean T (kbps)	Better efficiency	Average delay (ms)	Better efficiency
Conventional	13.7	46%	39	21%
TOPSIS	35.96	23%	30	12%
ANDSF-HO	25	34%	37	19%
V2I-MoLoHA	31.89	27%	34	16%
Proposed	58.89	–	18	–

Table 5.

Comparison of packet loss.

Method	Packet loss (%)	Better efficiency
Conventional	48	21%
TOPSIS	32.4	12%
ANDSF-HO	24	19%
V2I-MoLoHA	18.8	16%
Proposed	12	–

One of the major reasons for the increase in packet losses is due to the link degradation problems which occur mainly due to high vehicle density, poor signal quality. In our work, we have proposed a jellyfish optimization algorithm for the selection of routes taking into account vehicle metrics, channel metrics, and transmission metrics. Figure 9 shows the graphical plots where there is a decrease in the packet loss concerning the vehicle density due to consideration of multiple metrics for selecting the shortest path. From the figure, when there is an increase in the vehicle density, there are possibilities of an increase in data transmission due to which the packet loss can increase. However, in our work, the packet losses are minimized due to the selection of optimized routes. The previous works of TOPSIS, ANDSF-HO, and V2I-MoLoHA fails to select the best route among the available route between source and destination. Therefore, the deployment of an algorithm for selecting the best route minimizes packet loss. Even the vehicle density increases there is a reduction in packet loss since not all vehicles will use the route for transmission. That is to say, the vehicles nearby will not require data transmission. Due to this reason, the packet loss in the proposed work does not increase suddenly with the increase in the number of vehicles.

Figure 9: