Research on High-Speed Railway Communication Strategy Based on LIFI Technology

In order to improve the eect of high-speed railway communication, this paper studies the high-speed railway communication strategy combined with LIFI technology and improves the eect of high-speed railway communication through intelligent communication methods. Moreover, this paper studies the access point allocation problem in indoor LiFi/WiFi hybrid network, fully considers the dynamic change of network load, and introduces a dynamic load balancing algorithm that determines AP allocation in hybrid network based on the data transmission rate as the switching threshold. In addition, this paper proposes two AP allocationmethods, such as dynamic switching threshold andminimum data rate constraint, to construct a reliable high-speed railway communication model.­e experimental results verify that the high-speed railway communication strategy based on LIFI technology proposed in this paper can eectively improve the high-speed railway communication eect.


Introduction
e simulation modeling of the wireless channel is the premise of wireless communication system research. For special scenarios such as high-speed railways, the construction of simulation scenarios and the modeling of wireless channels are indispensable steps. Although the theoretical wireless channel model cannot perfectly reproduce the extremely complex channel state in the real environment, the wireless channel modeling parameters measured according to the real environment construction and statistical results can more truly re ect the rationality and e ectiveness of various key technologies in the high-speed rail environment. However, at present, most of the research on high-speed rail 5G wireless communication is based on a single scenario and only considers modeling the Rice channel with invariant parameters [1] or directly modeling the Rayleigh channel [2]. Neither of these two types of simulations can reveal the actual situation of the high-speed rail, and thus cannot re ect the real performance of the key technologies of high-speed rail wireless communication.
ere are abundant multipaths in the high-speed rail environment. To eliminate the multipath e ect, the current high-speed rail wireless communication technology usually adopts the orthogonal frequency division multiplexing (OFDM) technology to convert the frequency-selective channel into a series of parallel channels through serialparallel conversion. And the orthogonal frequency atness channel improves the decoding success rate and thus the e ective transmission e ciency. Because OFDM technology has the advantages of high spectrum utilization and antimultipath fading and can be combined with various multiple access technologies, OFDM technology will continue to be used in high-speed rail wireless communication systems. However, considering the fast time variability of the channel, relatively many channel parameters need to be estimated. In order to simplify the estimated parameters, there are usually two models that transform the direct estimation problem of time-varying channels into the estimation problem of nite parameters: one is Gauss-Markov model (Gauss-Markov), also known as autoregressive (Auto.regressive), AR model [3]; the other is the Basis Expansion Model (BEM) [4], this Generally speaking, STBC requires that the element spacing of the eNodeB antenna array is large enough to ensure the fading independence of multipath components to obtain the maximum diversity gain. For example, for a uniform linear array (ULA), the element spacing should be 5 times the carrier wavelength to 10 times [11]. In contrast, beamforming techniques require the element spacing of the antenna array to be half a wavelength to implement the selection of the angular domain. Since these two technologies have different requirements on the spacing of antenna array elements, combining STBC with beamforming technology faces challenges. Reference [12] proposed a beamforming mechanism combined with STBC, and reference [13] proposed a beam switching mechanism combined with STBC.
e performance of the beamforming mechanism combined with STBC is analyzed under some specific conditions. For example, reference [14] assumes imperfect channel information at the transmitter, reference [15] assumes a correlated MIMO fading channel, and reference [16] assumes e scenario of cooperating base station. e methods for generating the beamforming weight vector in the combination algorithm given in the above literature are all based on the decomposition of the covariance matrix of the channel. is method is obtained by using the largest eigenvector of the channel covariance matrix as the beamforming weight vector. Higher signal-to-noise ratio. It is a relatively simple method to calculate the beamforming weight vector based on DOA [17]. Reference [18] proposed that in a wireless environment with a small angular spread, the beam should not be directed toward the user DOA in order to ensure the maximum array gain, but a pair of beams should be generated to symmetrically fall on both sides of the user DOA. In reference [19], this combining mechanism is applied to the indoor communication environment; instead of using only a single antenna array above [20], divides an antenna array into two sub-arrays, and calculates the achievable array gain, followed by this assumption is also extended to systems with more than three transmit antennas.
is paper combines LIFI technology to study the highspeed railway communication strategy, improves the highspeed railway communication effect through intelligent communication methods, and promotes the improvement of high-speed railway communication technology.

Dynamic Load Balancing of LiFi/WiFi Hybrid Network.
Considering the application scenario of indoor visible light communication, a hybrid network model of indoor LiFi and WiFi is constructed, as shown in Figure 1. 16 LiFi APs and 1 WiFi AP are deployed, and each LiFi AP consists of multiple LEDs. We assume that all photoelectric detectors (PD) are vertically upward facing the ceiling, and the incident angle of the optical transmission link is equal to the radiation angle. Each LiFi AP covers a limited area, which is an attocell network. All LiFiAPs reuse the same frequency band. Combined with the actual needs of high-speed railway communication, the network system is constructed, and multiple sensors are used for data processing. e next part of this paper will study the algorithm to improve the dynamic load balancing of system operation. Subsequent studies consider the e ect of co-channel interference.
High-speed trains run fast. e fast mobility of highspeed rail makes it di cult to estimate the channel state information, and high-speed movement will also bring about a large Doppler shift. However, due to the regularity of high-speed rail travel, the information about the location is well estimated. With accurate location information, the transmit beam of the roadside base station can be directly aimed at the train station, so that the system capacity from the base station to the train will be maximized. e optical channel model does not consider the in uence caused by the lter of the front-end equipment, only considers the channel gain of light in free space, and uses the optical propagation model under the line-of-sight link (LOS). e expression for the DC gain between AP and the target user is: Among them, m is the Lambertian radiation series, A p is the e ective receiving area of the receiver photodetector, r i is the horizontal distance between the user and AP i , and h is the vertical distance between the user and AP i .
LEDs work in the linear region, and the output optical power is proportional to the input voltage. High-speed rail communication transmission has its own unique characteristics: base stations are evenly distributed on the track roadside, and on-board radios are distributed on the top of the train. Two-hop communication mode is adopted. Passengers communicate. Intensity modulation and direct detection are used to ensure that only positive and realvalued signals are transmitted to the receiver. A DC bias voltage xoc is added to the modulated electrical signal. e conversion between the average electrical power and average optical power of the signal conforms to the relationship ζ P opt / P i , ζ is de ned as the DC bias factor, where P opt is the average optical power of the LiFiAP, which is proportional to xnc. P is the electrical power of the signal. For a given user u connected to the a-th LiFi AP, the signal-tointerference ratio of the link can be expressed as: Among them, η pd is the response rate of the receiving end PD, B L is the modulation bandwidth of the LiFi network, N 0 is the noise power spectral density, H u,a is the channel gain between user u and LiFiAP a, and H u,else is the channel gain between user u and the interfering LiiAP.
Shannon capacity is used to calculate the available data rate between user u and LiFiAP a, namely: In the LiFi network, due to the dense arrangement of access points, it can provide higher regional frequency spectral e ciency performance, but it also causes serious cochannel interference problems. e introduced partial frequency reuse scheme can e ectively suppress the CCI problem, but in addition, the LiFi network also has the problem that the light is easily blocked and the network is interrupted. In this chapter, the method of time division multiplexing (TDMA) is used to realize multiuser access. Using a fairer scheduling algorithm, users served by LiFiAP share equal time resources.
In general, an indoor WiFi channel can be modeled as a Rayleigh fading channel, and the typical path loss expression is [example%].
Among them, d is the distance between the transmitter and receiver in m; the carrier frequency f is 2.4 GHz. e topology of a communication network plays a major role in determining the communication mode between interconnected devices. Speci cally, the fault tolerance and fault diagnosis characteristics of the system are determined by the connectivity of nodes in the network; the regularity of the network greatly a ects the complexity and reorganization characteristics of algorithms such as communication routing.
A, B, and C are constants whose values depend on the communication model. For communication models such as indoor o ce scenes, when only line-of-sight propagation is considered, A 18.7, B 46.8, C 20, X is the shadow e ect loss, and a Gaussian random variable with zero mean and standard deviation σ 3dB. Since there is no interference between the WiFi and LiFi channels, and only one WiFiAP is deployed, the signal-to-noise ratio (SNR) of the WiFiAP signal received by the user is:

Advances in Multimedia
Among them, P R is the transmit power of the WiFi AP; the expression of the path gain G des is G des � 10 − PL(dB)/10 ; the expression of the noise power is P n � k B TB R , k B is the Boltzmann constant, k B � 1.38 × 1023 J/K, T is the temperature of the surrounding environment, B R is the communication bandwidth available to the WiFi network. At state n, the data rate available to user u served by WiFi AP a is: Considering the case of multiple users, we assume that electric power is equally allocated to each subcarrier, and carrier resources are dynamically allocated to each user for signal transmission. Using a proportional fair scheduling algorithm, users served by WiFiAP share equal bandwidth resources. en the bandwidth B a � B x /NR is allocated to each user by the WiFi AP, where NR is the number of users served by the WiFi AP.

Dynamic Load Balancing Switching Scheme.
e overhead caused by handover in the indoor network is at millisecond level, which is much smaller than the interval time T p between two states. Poisson distribution can be introduced to model and describe, and different types of switching costs are modeled as independent and identically distributed Poisson random variables. t ij is defined as the switching cost of switching from AP i to AP j , and λ ij � E[t ij ] is the mean value of switching costs. e probability mass function of the Poisson distribution is expressed as: Due to the loss of throughput between the AP and the user caused by the switching, the switching efficiency between the two states is introduced, and the expression is: Among them, C L is the set of AP's of LiFi, and C R is the set of APs of WiFi. e data rate considering the switching overhead is the product of the switching efficiency and the data rate of the communication link, and the expression is: We assume that the complete set of users is U, the set of users served by LiFi is U L , and the set of users served by WiFi is U R . N L is the number of users in set U L , and N R is the number of users in set U R . S is the number of randomly generated users. a u ′ represents the AP assigned to user u in the n − 1 state. e central unit can calculate R n u,a and Γ n u,a from the channel state information between user u and LiFi and WiFi APs. In each state, in order to make full use of the higher spectrum resources of the LiFi network, users are preferentially allocated to LiFiAPs. erefore, the initial value of the state is Considering the handover overhead, we assume that the LiFi AP that enables user u to obtain the highest data rate of the communication link is labeled w L,u , as follows: Users on the LiFiAP equally allocate time resources, and the data rate of the optical link available to each user is: Among them, N w L,u is the number of users served by LiFi AP w L,u .
In the method FT, the user satisfies the condition of Ω u < c and is reassigned to the WiFiAP, wherein the handover threshold c is a fixed value. e WFiAP label is represented by w R,u . e allocation of user APs in state n is: In method DT, the central unit dynamically determines the switching threshold c s according to the number of users. e user meets the Ω u < c s condition and is assigned to the WiFiAP. e allocation of user AP's in state n is: In method MDRC, the handover threshold y is equal to the data rate requirement value A. In order to determine the WiFi users, the central unit finds the user u with the smallest data rate among all the users. If the user satisfies Ω u < c and the number of users served in the WiFi network does not exceed 10 (NR ≤ 10), it is determined as a WiFi user. As the number of users in the LiFi network decreases, the potential data rate of users changes accordingly. e value of the potential data rate Q a of the remaining user u in the LiFi network is recalculated e user AP allocation in state n is: According to the AP allocation method FT, DT, or MDRC, the serving AP of the user can be determined, and the UL, w L,u , and UR can be determined. e central processing unit connects the corresponding users to the LiFi network according to U and w L,u and connects the corresponding users to the WiFi network according to UR.
According to the AP allocation method, the AP serving the user in state n can be determined, and the data rate available to user u is: Among them, N a u is the number of users served by LiFi AP a, and N& is the number of users served by WiFi AP. Given the data rate requirement A in the hybrid system, the outage probability of the system is de ned as the percentage of the total working status in which the user does not meet the data rate requirement in the system, which is expressed as:

System
Model. An indoor dynamic scenario is considered, multiple LiFi APs and one WiFi AP are deployed, which is denoted by AP i , where i 0 represents that the access point is WiFiAP, i 0 represents that the access point is WiFiAP, and N i is the number of users in the AP i coverage unit. e WiFi AP covers all indoor areas, each LiFi AP covers a limited area, and the user is in the mixed network coverage area as shown in Figure 2.
In the LiFi network, the time division multiplexing (TDMA) method is used to realize multiuser access. Using a fairer scheduling algorithm, users served by LiFiAP share equal time resources. e data rate of user u served by LFiAP is: e DC-biased optical orthogonal frequency division multiplexing is used as the downlink information transmission scheme to ensure that the real-valued signal is transmitted to the receiving end. erefore, at least half of the subcarriers must be used to achieve Hermitian symmetry, where N i is the number of users served by the LiFiAP. e data rate available to users u served by WiFiAPr is: e utility function is designed for the user-available data rate. is chapter only considers downlink data transmission. According to the existing data rate utility model, a uni ed data rate utility model can be constructed for the LiFi/WiFi hybrid network as follows: Among them, ψ a and ψ b are two data rate thresholds. When the user's available data rate is less than ψ a , the user's data rate communication requirements cannot be met. However, when the data rate exceeds ψ b , it can no longer improve user satisfaction. e data rate is between the two thresholds, and there is a positive correlation between user satisfaction and the data rate.

Access Point Selection Scheme Based on Reinforcement
Learning. Reinforcement learning is an online learning algorithm that learns an optimal action strategy to obtain the maximum reward by continuously improving the agent's behavior in the interaction between the agent and the environment.
When the train passes through the mountains, the highspeed rail tunnel can ensure the stable and high-speed operation of the train, and the tunnel scene has become one of the typical high-speed rail scenes. Common tunnels are structurally limited, with tunnels ranging in size from a few hundred meters to several kilometers, depending on the surrounding geography. Considering the unique structure of high-speed rail tunnels, the radio wave propagation inside the tunnel is very di erent from other high-speed rail scenarios. e wireless signal will interact with the abundant e ective scatterers inside the tunnel, and multiple re ections, di ractions, and transmissions will become the main propagation mechanisms inside the tunnel. erefore, it is necessary to realize the information transfer between the agent and the environment. e process of the agent interacting with the environment is divided into the following steps: (1) e agent perceives the environmental state s t at time t. (2) For state s t and expected immediate reward R t , the agent performs action a t according to a certain strategy. (3) e action performed by the agent acts on the environment. e environment changes to update the state to s t+1 and gives an immediate reward R t to feedback to the agent, and the time t is updated to t + 1. By repeating the above steps continuously, the agent nally learns the optimal action strategy under the given goal.
We assume that the environment of the agent at time r can be represented by the state s t , s t ∈ S, and S is a nite state set. e agent can take action a t at time t, a t ∈ A, and A is a

Advances in Multimedia
limited set of actions. e agent perceives the current environment through the state s, selects and executes an action a t , and after the action acts on the environment, the state is updated to s t+1 with probability P s t ,a t (s t+1 ). In this action, the agent receives a reward R t from the environment. e system repeats this process of perception, action selection, and reward acquisition. e goal of the agent is to nd the optimal action policy π(a|s) that maximizes the reward obtained in the future. Combined with the optimization requirements of high-speed rail operation, the strategy is formulated and analyzed. Given a policy π: S ⟶ A, the cumulative expected return over an in nite time period can be de ned as: Among them, R t (s t , a t ) is the immediate reward obtained by performing action a t in state s t . E is the expectation operator, and c t ∈ [0, 1) is the discount factor. It re ects the importance of future returns to the current moment. is parameter has a certain real-time nature, and it is self-evident that it has a practical role in high-speed rail communication, especially in the process of high-speed rail operation, the parameter analysis has a high requirement for the timeliness of data analysis, so this is also an important reason for choosing this parameter in this paper. When y 0, it means that future rewards have no e ect on the current state, and when y is close to 1, it means that the importance of future rewards is close to that of immediate rewards. Formula (20) can be rewritten in the form of the Bellman formula. Solving the optimal action policy π * is transformed into the problem of solving the Bellman optimality criterion: Among them, P s,a (s ′ ) is the transition probability from the initial state s ′ to the new state s′ due to performing action a. Under the condition of known R(s, a) and P s,a in the traditional dynamic programming method, formula (21) can solve the optimal policy by an iterative method. However, since the statistical data in the wireless network environment in the dynamic switching scenario is di cult to obtain, the corresponding dynamic transfer function P s,a cannot be obtained. However, the Q-learning algorithm does not need to know P s,a , which provides a solution to the handover problem in the wireless network environment.
Under a given strategy π, the corresponding Q-value is expressed as: We assume that Q * (s, a) represents the maximum value in the Q-function, and it determines the optimal action a for each possible (s, a) combination pair, then we have: V * (s) can be replaced by max a∈A Q * (s, a), and the Q-learning algorithm selects the optimal Q * (s, a) in each iteration. e update of the Q-function in the learning process must satisfy the following formula: α represents the learning rate. e literature has con rmed that if each (s, a) combination pair is visited in nite times, then the learning rate α will approach 0. en, as time r approaches in nity, (s, a) will converge to Q * (s, a) with probability 1.
Based on the simulated channel models in di erent scenarios, the statistical characteristics of open, viaduct, and cutting scenarios are simulated, analyzed, and compared. Di erent HSR channel parameters in di erent scenarios, such as di erent values of Rice factor, will have an important impact on the evaluation of channel characteristics. Empty scenes, viaducts, and cuttings have di erent geographic structures and therefore di erent numbers of e ective scatterers in the simulation model.
Taking into full consideration the large number of handover actions that may be caused by the characteristics of the hybrid network of LiFi AP and WFi AP, the user's immediate return is de ned as the product of the handover loss factor β and the user's available data and utility function: e switching loss factor β is de ned as the degree of loss to the user's available data rate caused by switching between two APs, β ∈ [0, 1]. e smaller β is, the system evaluation standard (instant reward function) considers that the loss caused by handover to the system is larger, and try to avoid the occurrence of handover in the network. When the user's action a at time t does not trigger the switching action, β 1.

Reinforcement Learning Access Point Selection Algorithm
Based on Case Reasoning 2.5.1. Overview of Case Reasoning Methods. A typical problem-solving process based on case-based reasoning can be summarized into four steps: retrieve, reuse, revise, and retain, as shown in Figure 3.  Figure 3: Problem-solving process based on case reasoning. 6 Advances in Multimedia

Case Reasoning Scheme Design.
We assume that the cases in the case base are represented by Z x , Z k : Y k , Q k , U k , k 1, 2, . . . , K, where K is the total number of cases stored in the case base; Y k represents the feature vector of case k. e case matching degree can be calculated through the feature vector between cases, Q k represents the solution of case k, which stores the Q-value table obtained after the learning of each case, U k represents the utility value of case k. e utility value is used to evaluate the pros and cons of the solution of the case. Combined with the case study process in this paper, set case k as the research object for process analysis, then combined with the above algorithm requirements, the following case analysis is carried out. e eigenvectors of case k are: (26) Among them, L represents the number of case features. e similarity between the new case Y j and the case Y k is de ned as: Among them, ω l is the weight coe cient of the rst feature, satisfying L l 1 ω l 1, sim(y l j , y l k ) is the similarity of the l-th feature, and y l k,w represents the w-th feature component in the I th feature in case k. It uses Euclidean distance as a measure of similarity, which is expressed as: (28) e similarity value x is calculated by formula (27), and the matching case can be obtained as: We assume that the statistical distribution of network load in each period does not uctuate too much. e similarity of time periods is expressed as: sim y l j , y l k 0, p j p k , e utility function of the signal-to-interference ratio is set as: In the nonstationary high-speed rail simulation channel model, based on the assumption of uncorrelated scattering and antenna stationarity, the statistical characteristics of the channel in di erent scenarios can be deduced, and the rst tap of the channel model is emphatically analyzed, including the line-of-sight component and the single re ection component. e setting of this utility function is the same as the expression of Shannon's capacity, so the utility value of the signal-to-interference ratio between the user and the AP link is proportional to the maximum data rate available to the user. en the feature vector of the case can be expressed as: ③ Q-value table fusion is: the feature vector is obtained at the beginning of each time slot, and then the source case most similar to the target case is found from the case library. Changes in Q-value are made based on the similarity between cases. e Q-value fusion method of the target case and the matching case is: We add the (s, a) combination pair that exists in the Qvalue  Advances in Multimedia 7

High-Speed Railway Communication
Strategy Model Based on LIFI Technology e entire simulation system adopts a modular design scheme. e simulation framework is shown in Figure 4, which mainly includes ve functional modules, including cell model, user model, MIMO transmission, link adaptation, and system performance statistics. e cell of this simulation system adopts the networking mode of BBU + RRU. In order to facilitate the analysis, the cell model in the simulation in this chapter uses one RRU and one cell, as shown in Figure 5.
e train travels to the junction of the RRU, as shown in Figure 6(a). At this time, due to the in uence of shadow fading, Doppler frequency o set, and multipath e ect, the quality of the received signal is seriously degraded. At this time, if Type-0 CoMP is adopted, the serving RRU and the cooperative RRU of the user transmit the same data to the user at the same time, which can improve the quality of the signal received by the user. When two RRUs send data, the same precoding codebook is used, and the joint precoding codebook P is determined according to the channel conditions H0 and H1 of the user to di erent RRUs.
Under the Type-1 model, as shown in Figure 6, di erent precoding codebooks P0 and P1 are determined according to the channels H0 and H1 to adjacent RRUs. e cooperative RRUs use their respective precoding codebooks to send their respective data S0 and S1, and the data do not need to be synchronized at this time. e signal received by the user is the set of data sent by the two RRUs. e two cooperating RRUs share control signals through the backhaul channel, but do not need to maintain synchronization. is solution can be better applied to the inter-cell CoMP where synchronization is di cult to achieve.
In the current high-speed railway communication network, more and more communication devices are controlled by computers, and because in the computer system, the complexity of hardware failures is high, and the software also has a certain degree of potential failure, above fault. e de nition of safety requires that in any fault condition, the   general computer system can guarantee that the output is in a safe state, which will be di cult to achieve. erefore, it is necessary to break through the traditional fault. Security concept, in order to nd a way to solve the failure of the computer system. Safety issues, that is, to recognize faults from the perspective of relative safety and safe question. e public network networking mode mainly means that communication can be performed on the base station site and base station equipment along the original high-speed line. Users on the high-speed railway line share this wireless network with other ordinary users along the line, and lowspeed services along the high-speed railway line will not be a ected by high-speed mobile sites. Moreover, high-speed routes do not use dedicated frequencies, and the entire network can use the same frequency con guration, reducing the number of frequency allocations. e schematic diagram of the public network networking mode is shown in Figure 7.
In the group call area design, it must be ensured that the base station area of the station covers more than 2000 meters on both sides of the station. Only in this way, the GSM-R group call area in the base station area of the station can be guaranteed. Figure 8 shows the networking mode of the dispatching communication system.  The location of the railway department

Advances in Multimedia
proposed in this paper according to the speed. When the speed is less than 80 km/h, the traditional antenna selection algorithm is used, otherwise, the LIFI-based antenna selection algorithm is used. On this basis, the effect of the high-speed railway communication strategy based on LIFI technology proposed in this paper is evaluated, and the evaluation results are obtained through the evaluation of multiple groups of experts, as shown in Table 1.
It can be seen from the above research that the highspeed rail communication strategy based on LIFI technology proposed in this paper can effectively improve the effect of high-speed rail communication.

Conclusion
High-speed rail wireless communication systems have their inherent advantages and disadvantages. For example, high moving speeds and channel feedback delays will make smart antenna-based beamforming techniques ineffective for transmission. In contrast, opportunistic beamforming technology based on non-smart antennas is more suitable for this scenario because it does not require real-time channel vector feedback and only requires channel SNR reporting. is paper combines LIFI technology to study the high-speed railway communication strategy and improves the high-speed railway communication effect through intelligent communication methods. Moreover, this paper studies the access point allocation problem in indoor LiFi/ WiFi hybrid networks, fully considers the dynamic changes of network load, and introduces a dynamic load balancing algorithm that determines AP allocation in hybrid networks based on the data transmission rate as the switching threshold.
e experimental study verifies that the highspeed rail communication strategy based on LIFI technology proposed in this paper can effectively improve the highspeed rail communication effect.

Data Availability
e labeled dataset used to support the findings of this study is available from the corresponding author upon request.

Conflicts of Interest
e authors declare no conflicts of interest.