Smart reflecting surface assisted visible light communication train adaptive handover method

Abstract

The application discloses a kind of intelligent reflecting surface assisted visible light communication train adaptive inter-zone switching method, including the IRS-VLC train-ground communication network model formed by vehicle-mounted PD, illumination source, IRS;The method of the present application completes four links of state prediction, boundary update, parameter optimization and execution in each decision cycle;First, the position and speed in the next short time domain are predicted using train position state sequence;Then, the geometric boundaries of switching trigger HOT, switching failure HOF and ping-pong PP are updated according to the predicted state;Next, the switching margin HOM and trigger delay TTT are adaptively adjusted by genetic algorithm, so that the switching parameters adapt to the real-time running state of the train;Finally, when the train enters the next decision time, the inter-zone switching decision is executed according to the new switching parameters, and the new observed state is fused through the sliding window, and the next cycle is entered.The present application can accurately represent the change of train running state.

Description

Technical Field

[0001] This invention relates to the field of intelligent urban rail transit, and in particular to an adaptive cross-zone handover method for visible light communication trains assisted by an intelligent reflective surface. Background Technology

[0002] With the rapid development of urban rail transit towards intelligence and digitalization, communication-based train control (CBTC) systems have become core equipment for ensuring the safe and efficient operation of trains. The vehicle-to-ground wireless communication network undertakes the bidirectional transmission of train control commands, operational status data, and multimedia information; its network performance directly determines operational safety and efficiency. Traditional vehicle-to-ground communication mainly employs radio frequency (RF) communication technologies such as Wireless Local Area Network (WLAN), Long Term Evolution (LTE), and 5G. In the relatively enclosed and structurally narrow environment of subway tunnels, existing RF communication technologies face severe challenges such as multipath fading, electromagnetic interference, spectrum resource depletion, and information security vulnerabilities, making it difficult to meet the high reliability requirements of future smart urban rail transit for vehicle-to-ground communication. Therefore, exploring a new vehicle-to-ground communication technology with high security and interference resistance is an inevitable trend for ensuring the sustainable development of urban rail transit.

[0003] In recent years, visible light communication (VLC) has become a key candidate technology for indoor positioning, green communication, and intelligent transportation due to its advantages such as abundant spectrum resources, absence of electromagnetic interference, and high security. The evenly spaced lighting sources installed on the walls of subway tunnels and their relatively enclosed information transmission environment provide natural infrastructure conditions for building VLC vehicle-to-ground communication networks. However, due to the linear transmission characteristics of visible light and the limitations of the receiver's field of view (FOV), the effective communication coverage of a single light source is limited. Trains frequently traverse different coverage areas while running in tunnels, causing severe fluctuations in the received signal, affecting the precise control of the train, and even endangering driving safety. Traditional handover mechanisms based on received signal strength have a delayed response and fixed handover parameters, making it difficult to effectively balance handover delay and the ping-pong (PP) effect, easily leading to high handover failure (HOF) and link interruptions. To address this challenge, domestic and international scholars have conducted research on VLC handover optimization through multiple paths, from multi-connection architectures and intelligent predictive decision-making to emerging physical layer control technologies. The literature on CoMP-based dynamic handover for vehicular VLC networks and Proactive link handover deploying coordinated transmission for indoor visiblelight communications (VLC) networks employs a coordinated multi-point transmission or multi-connection mechanism, utilizing link redundancy to achieve a "connect first, then disconnect" smooth handover process. The literature on Collaborative online learning-based distributed handover scheme in hybrid VLC / RF 5G systems introduces collaborative online learning to dynamically adjust handover parameters, improving handover efficiency in heterogeneous VLC / RF networks. However, these methods do not improve the problem of limited single-point coverage; frequent handovers significantly increase system signaling overhead. Furthermore, artificial intelligence technology has also shown great potential in VLC inter-cell handover.The literature *Data-driven link assignment with QoS guarantee in mobile RF-optical HetNet of things* utilizes reinforcement learning and Long Short-Term Memory (LSTM) networks to predict link outages and optimize throughput and latency. The literature *Deep reinforcement learning-based adaptive handover mechanism for VLC in a hybrid 6G network architecture* and *Prediction-model-assisted reinforcement learning algorithm for handover decision-making in hybrid LiFi and WiFi networks* further introduce deep reinforcement learning and trajectory prediction to enhance the foresight and adaptability of decision-making. However, these works are mostly geared towards instantaneous link quality optimization, which can easily induce unnecessary handovers during channel fluctuations, potentially exacerbating HOF and PP.

[0004] Intelligent reflecting surfaces (IRS), as an emerging technology that can reshape the wireless propagation environment by adjusting the state of reflecting units, have shown significant application potential in the field of visible light communication in recent years. In terms of communication enhancement, the paper "Intelligent reflecting surface-aided indoor visible light communication systems" introduces IRS into indoor VLC systems, demonstrating that proper configuration of the mirror array can effectively overcome line-of-sight obstruction and significantly improve the receiver signal-to-noise ratio and achievable data rate. In terms of position awareness, the paper "Simultaneous position and orientation estimation in single optical IRS-assisted visible light systems using single LED and single PD" achieves simultaneous estimation of the receiver's three-dimensional position and orientation in a minimalist hardware architecture using a single LED light source, photodetector, and single optical IRS. The paper "Simultaneous position and orientation estimation for reconfigurable intelligent surfaces-assisted visible light communications" further extends pose estimation to general RIS-assisted VLC systems, establishing a joint estimation framework for position and orientation and deriving the theoretical lower limit of accuracy. During the handover process, the IRS significantly expands the effective coverage of a single light source and greatly improves the signal quality in edge areas, thereby reducing the number of handovers during train operation from the root and effectively suppressing the occurrence of HOF and PP.The literature *Enabling efficient blockage-aware handover in RIS-assisted mmWave cellular networks* proposes a blockage-aware handover algorithm that directly avoids handover interruptions by utilizing indirect links constructed using IRS. The literatures *An equivalent model for handover probability analysis of IRS-aided networks* and *Discrete-time modeling and handover analysis of intelligent reflecting surface-assisted networks* establish handover probability analysis models and discrete-time modeling methods, respectively. The literature *Optical RIS-collaborative beamforming: a high-efficiency covertness enhancement scheme for VLC system* studies an optical IRS-collaborative beamforming scheme, while the literature *Handover management through reconfigurable intelligent surfaces for VLC under blockage conditions* explores the feasibility of using IRS to alleviate VLC link congestion. These studies have achieved preliminary results in improving VLC network handover performance and link reliability, and have explored the application potential of IRS in handover. However, existing studies have given little consideration to the drastic fluctuations in communication links under conditions of limited light source coverage, and have failed to fully assess the evolution mechanism of HOF and PP in IRS-assisted networks. This makes it difficult for handover parameters to adapt to the rapid operation of trains in tunnel scenarios, and the reliability of inter-zone handover in IRS-assisted VLC networks is still not guaranteed. Summary of the Invention

[0005] The purpose of this invention is to address the train handover problem in tunnel VLC networks by providing an intelligent reflector-assisted adaptive handover method for visible light communication trains. First, a liquid crystal IRS is introduced to dynamically regulate the spatial distribution of the optical signal, expanding the effective coverage of a single light source and reducing the handover frequency from a physical perspective. Second, a handover boundary geometric model is established, transforming the complex decision-making problem into intuitive spatial geometric constraints, providing clear decision boundaries for parameter adjustment. Finally, an LSTM network is used to predict the train's position and speed, and the handover margin (HOM) and trigger time delay (TTT) are adaptively adjusted using the geometric model as constraints, making the handover parameters adapt to the real-time operating state of the train, thereby reducing the probability of handover failure and ping-pong handover.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A method for adaptive handover of visible light communication trains with intelligent reflector-assisted handover is disclosed, which uses an IRS to construct an IRS-VLC vehicle-to-ground communication network model. The IRS-VLC vehicle-to-ground communication network model includes an onboard digital camera (PD), a lighting source installed along the tunnel wall as a service access point, and an IRS installed on the tunnel wall opposite the access point. The access point and the onboard PD form a Loss-of-Sight (LoS) direct link model, and the IRS and the onboard PD form an IRS reflective link model. The onboard PD, the lighting source, and the IRS form the IRS-VLC vehicle-to-ground communication network model. The instantaneous received optical power of the onboard PD is composed of the superposition of the LoS direct link and the reflective link assisted by the LC-IRS. The IRS-assisted visible light communication train adaptive handover method is based on the RS-VLC vehicle-to-ground communication network model. Within each decision cycle, it completes four stages: state prediction, boundary update, parameter optimization, and execution. First, it uses the train's position and state sequence to predict the position and speed in the next short time domain. Then, it updates the geometric boundaries of the handover trigger (HOT), handover failure (HOF), and ping-pong turnaround (PP) based on the predicted state. Next, it adaptively adjusts the handover margin (HOM) and trigger delay (TTT) using a genetic algorithm to adapt the handover parameters to the train's real-time operating state. Finally, when the train enters the next decision time, it executes the handover decision according to the new handover parameters and merges the new observation state through a sliding window to enter the next cycle.

[0007] As an improvement to the above technical solution, in the LoS direct link: Let the transmitted optical power of the service access point be P. t,j The train's current position is denoted as X, the effective receiving area of ​​the onboard PD is A, and the corresponding LosS DC channel gain is denoted as... ; The LC-IRS-assisted reflection link is as follows: Suppose that the nth liquid crystal cell in the IRS array serving the service access point is under control voltage u n The state coefficients are as follows This characterizes the unit's overall modulation capability of the reflection link under its current operating state; the optical signal transmitted by the serving access point reaches the vehicle-mounted PD after being reflected by this unit, and its auxiliary reflection link gain is expressed as... ; The reference reflection link gain is determined by spatial geometry without considering the liquid crystal electronic control state; Considering polarization loss factor Reflection efficiency function related to incident angle The corrected expression is: ; in, The effective area of ​​the nth LC-IRS cell. The distance between the access point and the nth unit. The distance between the nth unit and PD. For the transmission angle from the service access point to the IRS unit, The incident angle at the IRS unit. The reflection angle from the IRS unit to the vehicle-mounted PD. The angle of incidence at the vehicle-mounted PD; If the IRS arrays serving this area have a total of N j If there are 1 effective unit, then at train position X, the total reflected and received optical power after IRS assistance is: ; During train operation, the total effective optical power actually received by the onboard PD is composed of the LOS direct component and the LC-IRS auxiliary reflection component. .

[0008] It should be noted that: Access point: An access point is a standardized access node used by various network devices and business systems to establish network connections and realize data interaction. It is the boundary interaction carrier for information sending and receiving and communication transmission, and carries the basic communication functions of device networking, command transmission and data interaction. Service access point; a communication node that is open to the outside world by an entity with network connectivity and information output capabilities, providing network services such as data transmission, command interaction, and business invocation through this node; Target access point: is a secondary interactive node in the communication link that is downstream of the service access point and is to be used to establish a communication connection. It is used to receive data and instructions sent by the service access point and complete the next stage of network transmission and information interaction.

[0009] As an improvement to the above technical solution, when the total received optical power of the target access point reaches a preset handover margin relative to the serving access point... When this occurs, the system triggers a handover decision; the triggering condition formula is as follows: ; Let t1 be the time when the trigger condition formula first becomes true, and x1 be the corresponding train position; let T be the system startup duration. j,k The TTT timer; if the condition is within the time window [t1, t1+T] j,k If this holds true continuously within a given time period, then at time t2 = t1 + T... j,k Issue a switching command; If execution delay is taken into account The train will eventually be The switching is complete; the instantaneous velocity of the train along the track direction is v(t), then the time and position relationship from triggering to the completion of the switching can be uniformly expressed as: In the formula, x3 represents the position of the train when the switch is completed; If we assume that the train speed remains approximately constant during the short switching interval, and denote v1 = v(t1), then we have ; To ensure that the handover can be completed within the effective overlapping coverage area, the cumulative displacement of the train within [t1,t3] should not exceed the remaining length of the overlapping coverage area corresponding to the current position. ,Right now ; Depend on Therefore, the feasible upper bound of TTT is satisfied. .

[0010] As an improvement to the above technical solution, during the TTT timing phase or execution delay phase after handover triggering, if the total received optical power corresponding to the target access point drops to the receive threshold P before the handover is completed... th The following are the criteria for determining HOF. Conversely, if the train successfully completes the transfer to AP at time t3... k The switching occurs during the minimum dwell time T. p If, due to transient channel fluctuations, the reverse handover condition is reactivated, a PP (Pre-Transfer) event is determined to have occurred. The criterion for this is expressed as follows: .

[0011] As an improvement to the above technical solution, when switching from the service access point to the target access point, the parameter vector to be optimized is: The corresponding switching parameter optimization problem can be expressed as: ; in, and AP j To AP k The upper bounds of HOF and PP probabilities during the handover process, with w1 and w2 as weighting coefficients, are used to characterize the system's response to two types of abnormal handover risks. and These are the lower and upper bounds of the switching margin, respectively; T max This represents the maximum allowed trigger delay of the system.

[0012] As an improvement to the above technical solution, when updating the geometric boundaries of the handover trigger (HOT), handover failure (HOF), and ping-pong (PP) based on the predicted state, the steps for IRS-assisted handover boundary updates are as follows: For the g-th schedulable IRS array at the service access point, let its equivalent reflection center be X. g =(x g , yg For events i∈{HOT, PP}, the initial boundary under a given working mode m∈{L, N} is written as the squared distance ratio to the trajectory. In the formula, Events under pattern m The corresponding squared distance ratio factor; Mode L is for combined direct and reflected link assistance, and Mode N is for reflected link assistance only; when When the initial boundary is a circle, its expression is:

[0013] when At that time, the initial boundary degenerates into the perpendicular bisector of the line connecting the target access point and the IRS center, and its expression is:

[0014] For the HOF boundary under IRS assistance, the modeling process and formula Consistent; During operation, the train may also lose its auxiliary link due to changes in the geometric relationship with the serving IRS array; at this time, the total received optical power will change abruptly, which may directly trigger HOT or cause HOF. When the train moves away from the IRS array, the auxiliary reflection gain falls below the minimum effective threshold. At that time, the gain interruption boundary can be obtained as follows:

[0015] in The geometric constants related to the IRS reflection characteristics are determined by the minimum effective reflection gain threshold. The relevant geometric constants, unit reflectivity, and path loss index together determine and satisfy the following conditions. ,in The reference reflection link gain per unit distance; When the train moves outside the effective angle range of the IRS, the auxiliary reflection link is interrupted; Let the normal direction of the g-th IRS array be... The effective half-width is The corresponding angular break boundary is represented by two rays.

[0016] From the above equation, it can be seen that the two rays begin with the first... IRS array locations As the endpoint, along the array normal direction With effective half angle The determined boundary direction extends to enclose the effective angular service area of ​​the IRS.

[0017] As an improvement to the above technical solution, when updating the geometric boundaries of the handover trigger (HOT), handover failure (HOF), and ping-pong (PP) based on the predicted state, it is necessary to correct the IRS-assisted handover boundary. The steps of this correction method are as follows: For circular elements, the effective portion is denoted as an arc tuple. X c R and R represent the center and radius of the circle, respectively. and Indicates the starting and ending polar angles of the effective circular arc; For line elements, the effective part is denoted as a line segment tuple. X s and X e Let d be an endpoint of the line segment. s and d e It is a distance scalar in a linear projection coordinate system.

[0018] Let the equation of the constrained line be , represents the signed distance function of a point on a given curve. The sign of the function can be used to determine whether the switching boundary point is located within the tunnel operating area; If the initial boundary point is located outside the rectangular region, an empty set is output; for a circular initial boundary, the effective arc interval on the same side of the constraint line can be represented as: ; in, This represents the distance from the center of the circle to the constraining line. The polar angle at the intersection of the line and the circle. The midpoint of the corresponding short arc. As a reference point indicating the effective side, this formula is used to exclude arc segments outside the tunnel operating area or on the side where the constraint is ineffective; Considering the minimum effective reflection gain constraint, let the constraint circle be... , representing the effective service range of the IRS in a given operating state, if the initial handover boundary is {X c If R}, then the distance between the centers of the two circles is The effective arc range of the initial circular boundary within the constraint circle can be represented as: ; Among them, the angle of intersection Derived from the Law of Cosines: ; Here It represents the azimuth angle of the line connecting the centers of the circles, accurately eliminating trajectory segments that do not meet energy or distance constraints; and finally obtaining the effective geometric boundaries of HOT, HOF, and PP events under actual physical constraints.

[0019] As an improvement to the above technical solution, after the handover is triggered, the change in signal quality within the TTT duration determines whether HOF occurs; after the handover is completed, the dwell time T p The change in internal signal quality determines the probability of PP occurrence. Therefore, the methods for estimating the probabilities of HOF and PP are as follows: Discrete sampling is performed on the switching boundary to obtain a finite set of location points; Along each line segment or arc on the boundary, at equal intervals Take points in sequence until the remaining length of the segment is less than The set of switching boundary points after discretization is as follows: For the train's location point X0, the distance to the risk arc The distance is ; in Point X0 is relative to the center X of the circle. c The polar angle; Similarly, from point X0 to line segment The distance is: ; Where X f It is the distance from point X0 to the straight line containing the line segment. The dangling foot.

[0020] Based on the above geometric mapping, from Switch to The upper bound of the handover failure probability is: ; Where N is the total number of discrete points. It is an indicator function that takes 1 if the condition is met and 0 otherwise; it indicates that if the distance the train may travel before the switch is completed is greater than the shortest distance from the trigger point to the HOF boundary, then the trigger point is considered to have HOF risk. Under the same boundary model, let the position of trigger point X0 after the switch completion after experiencing TTT and execution delay be denoted as . Then the upper bound of the PP probability is: .

[0021] As an improvement to the above technical solution, a genetic algorithm is used to adaptively adjust the switching margin (HOM) and trigger delay (TTT) to adapt the switching parameters to the real-time operating status of the train. The genetic algorithm is used by the ISSA-LSTM prediction module, and the steps involved in this task are as follows: Define the historical position x and velocity v of the train over the past m consecutive time steps as the input feature matrix S. t =[(x,v) t-m+1 ,…, (x,v) t] T After deep feature extraction by the LSTM network, the fully connected layer maps and outputs the state prediction vector p steps ahead. ; ISSA first uses the Chebyshev chaotic map to generate an initial solution, the expression of which is: In the formula, Z p Let p be the p-th chaotic variable, and c be the Chebyshev mapping order; Then, a dynamic nonlinear scaling factor is introduced to achieve adaptive adjustment of the search strategy, the expression of which is: ; Where n is the current iteration number, N max The maximum number of iterations, A random number between [0, 1]; Finally, Cauchy mutation and tent mapping perturbation are introduced into the global optimal position update strategy. After each iteration, the current global optimal solution is perturbed, expressed as: ; Where Cauchy(0,1) represents a standard Cauchy distribution random number, whose heavy-tailed characteristic is beneficial for increasing the search step size and enhancing the ability to escape local optima; Tent is the tent map, used to enhance local neighborhood search, and its expression is: .

[0022] As an improvement to the above technical solution, the initial population size of the genetic algorithm is: The maximum number of iterations is The crossover probability is The mutation probability is The threshold for early termination is ; Prediction module output and Used for online updates of boundaries and risk assessments, driving adaptive adjustments of HOM and TTT; For each set of switching parameters Calculate the upper bound of the HOF / PP probability and construct the fitness function. The larger the fitness, the lower the switching risk.

[0023] Compared with the prior art, the advantages and positive effects of this invention are: The introduction of IRS in this invention effectively expands the system coverage and provides a reliable guarantee for cross-area handover of visible light communication trains in tunnel scenarios; the proposed handover algorithm can significantly reduce HOF and PP risks and effectively guarantee the communication continuity of trains under rapid movement. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A schematic diagram of a VLC vehicle-to-ground communication network structure assisted by an IRS; Figure 2 This is a schematic diagram of a vehicle-to-ground communication model based on LC-IRS. Figure 3 This is a schematic diagram of the geometric model for inter-region handover; Figure 4 A schematic diagram of a prediction-based switching parameter optimization model; Figure 5 a is a schematic diagram of the randomly initialized population distribution. Figure 5 b is a schematic diagram of the optimized initial population distribution; Figure 6 This is a schematic diagram illustrating the dynamic changes of the scaling factor. Figure 7 This is a schematic diagram of the ISSA-LSTM algorithm flow. Figure 8 This is a schematic diagram of the adaptive handover algorithm. Figure 9 A schematic diagram comparing the convergence characteristics of different optimization algorithms; Figure 10 a is the convergence curve of location prediction trained by the ISSA-LSTM model. Figure 10 b is the velocity prediction convergence curve trained by the ISSA-LSTM model; Figure 11 Figure 'a' is a diagram comparing the predicted and actual positions of the train's operating status. Figure 11 b is a schematic diagram comparing the predicted and actual speed values ​​of the train's operating status. Figure 12 a is a 2D heatmap without IRS assistance, and 12b is a 2D heatmap with IRS assistance; Figure 12 c is the 3D heatmap without IRS assistance. Figure 12 d represents the 3D heat map assisted by IRS; Figure 13 Comparison diagram of received optical power distribution under VLC coverage; Figure 14 Figure a shows a comparison of the PP probability performance of different handover algorithms. Figure 14b is a comparison chart of HOF performance for different switching algorithms; Figure 15 This is a comparison chart showing the switching anomaly probability of various optimization algorithms at different speeds. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art to all other embodiments obtained without creative effort should be included within the protection scope of the present invention.

[0027] like Figures 1 to 15 As shown, this invention proposes an intelligent reflector-assisted adaptive train handover method for visible light communication to address the train handover problem in tunnel VLC networks. First, a liquid crystal IRS is introduced to dynamically regulate the spatial distribution of the optical signal, expanding the effective coverage of a single light source and reducing the handover frequency from a physical perspective. Second, a geometric model of the handover boundary is established, transforming the complex decision problem into intuitive spatial geometric constraints, providing clear decision boundaries for parameter adjustment. Finally, an LSTM network is used to predict the train's position and speed, and the handover margin (HOM) and trigger delay (TTT) are adaptively adjusted based on the geometric model constraints, making the handover parameters adapt to the real-time operating state of the train, thereby reducing the probability of handover failure and ping-pong handover.

[0028] 1. VLC vehicle-to-ground communication network model To address the limited coverage of single light sources in subway tunnels, Liquid Crystal IRS (LC-IRS) technology was introduced to construct an IRS-VLC vehicle-to-ground communication network model. A composite channel model was established, including a line-of-sight (LoS) direct link and an IRS reflection link. The switching anomaly judgment criteria and boundary conditions were quantified based on train operation characteristics, and the problem was transformed into a multi-objective optimization problem of switching parameters with minimizing the anomaly probability as the core, laying the foundation for the design of the switching algorithm.

[0029] 1.1 IRS-assisted VLC vehicle-to-ground communication network structure To address the relatively enclosed and narrow structural characteristics of subway tunnels, LED lighting sources are installed at equal intervals along the left side of the tunnel wall to construct a VLC (Vehicle-to-Ground) communication system. Due to the inherent line-of-sight propagation characteristics of VLC and the limited field of view (FOV) of onboard digital photodiodes (PDs), the communication coverage of traditional single-light-source systems is severely limited. Therefore, a programmable LC-IRS (Liquid-Local Infrared Regulator) is integrated into the tunnel wall opposite the Access Point (AP) to achieve beam shaping. The IRS-assisted VLC vehicle-to-ground communication network structure is as follows: Figure 1 As shown.

[0030] The IRS unit employs a multilayer liquid crystal composite structure, comprising, from top to bottom, an optical substrate, a transparent electrode, a liquid crystal layer, and a reflective electrode, with a silicon-based driving control circuit integrated at the bottom. The phase and amplitude of the reflected beam are modulated by the electro-optic effect of the liquid crystal layer. The spatial light field distribution is reconstructed through dynamic programmable configuration, improving the quality of non-line-of-sight links and providing both coverage extension and interference suppression.

[0031] 1.2 Vehicle-to-ground communication channel modeling based on LC-IRS The instantaneous received optical power of a vehicle-mounted PD is composed of the sum of the direct Loss (LoS) link and the reflected link assisted by the LC-IRS, such as... Figure 2 As shown. To accurately describe the characteristics of this composite channel, a mathematical model needs to be established relating the train's spatial location, IRS control parameters, and received optical power.

[0032] (1) LoS direct link Set up access point (AP) j The emitted optical power is P t,j The train's current position is denoted as X, and the effective receiving area of ​​the onboard PD is A. Under unobstructed LoS link conditions, the direct optical power received by the onboard unit satisfies the standard Lambertian radiation model, and the corresponding LoS DC channel gain is denoted as... Its value is determined by the spatial distance between the transmitter and receiver, the emission angle, the incident angle, the optical gain, and the detector's field of view; when the incident angle exceeds the upper limit of the field of view, the link gain is zero.

[0033] (2) LC-IRS-assisted reflection link In train operation scenarios, the rapidly changing channel conditions place extremely high demands on IRS configuration updates. LC-IRS employs a dynamic configuration scheme, adjusting the underlying control voltage to change the equivalent refractive index of the liquid crystal cells, thereby achieving dynamic reconstruction of the incident beam propagation direction and energy distribution. This is implemented by servicing the AP... j The nth liquid crystal cell in the IRS array is controlled by voltage u n The state coefficients are as follows This is used to characterize the unit's overall modulation capability of the reflection link under the current operating state.

[0034] APj The emitted optical signal reaches the vehicle-mounted PD after being reflected by this unit. The auxiliary reflection link gain can be expressed as: (1) In the formula, This is the reference reflection link gain determined by spatial geometry, without considering the liquid crystal control state. Polarization loss factor is also considered. Reflection efficiency function related to incident angle The corrected expression is: (2) in, The effective area of ​​the nth LC-IRS cell. for Distance to the nth unit The distance between the nth unit and PD. For the transmission angle from the service access point to the IRS unit, The incident angle at the IRS unit. The reflection angle from the IRS unit to the vehicle-mounted PD. The angle of incidence is the angle at which the vehicle-mounted PD is located.

[0035] As shown in equation (2), the gain of the LC-IRS auxiliary link is determined by the spatial geometric mapping at the transceiver end and the electronic control response of the liquid crystal cell. Since the LED emits incoherent light, there is no coherent interference effect at the receiver for the multi-cell reflected signals, and the total received power of the system is equivalent to the linear superposition of the power of each link. If the IRS array serving this area has N j If there are 1 effective unit, then at train position X, the total reflected and received optical power after IRS assistance is: (3); During train operation, the total effective optical power actually received by the onboard PD is composed of the LOS direct component and the LC-IRS auxiliary reflection component. (4).

[0036] 1.3 IRS-assisted handover process In VLC networks, handover is used to ensure service continuity during train operation. In VLC systems, handover decisions typically reference the A3 event logic in mobile communications. For example... Figure 3 As shown, when the target access point AP k Total received optical power relative to the serving access point (AP) j Reaching the preset switching margin When this occurs, the system triggers a handover decision. The triggering condition formula is: (5) Let t1 be the time when equation (5) first becomes true, and x1 be the corresponding train position. Let T be the system startup duration. j,k The TTT timer; if the condition is within the time window [t1, t1+T] j,k If this holds true continuously within a given time period, then at time t2 = t1 + T... j,k Issue a switch command. Further consider execution latency. The train will eventually be The switchover is complete. If the instantaneous velocity of the train along the track is v(t), then the time-position relationship from triggering to the completion of the switchover can be uniformly expressed as: (6) In the formula, x3 represents the position of the train when the switch is completed. If the train speed is assumed to remain approximately constant during the short switching interval, and v1 = v(t1), then we have .

[0037] To ensure that the handover can be completed within the effective overlapping coverage area, the cumulative displacement of the train within [t1,t3] should not exceed the remaining length of the overlapping coverage area corresponding to the current position. ,Right now (7) From equation (7), we can obtain that the feasible upper bound of TTT satisfies (8)

[0038] The TTT value is constrained by both the train speed and the remaining length of the current overlapping coverage area: the higher the train speed or the shorter the overlapping coverage area, the smaller the allowed TTT; otherwise, the train may leave the effective coverage area before the handover is completed, thus increasing the risk of handover failure. During the TTT timing phase or execution delay phase after handover triggering, if the total received optical power corresponding to the target access point drops to the receive threshold P before the handover is completed... th The following are the criteria for determining HOF: (9) Conversely, if the train successfully completes the transfer to AP at time t3... k The switching occurs during the minimum dwell time T. p If the reverse handover condition is reactivated due to transient channel fluctuations, a PP (Pre-Transfer) event is determined to have occurred. The criterion for this can be expressed as: (10) The quantitative judgment criteria of HOF and PP reveal the coupling relationship between handover parameters and abnormal risks under train movement and link fluctuations. Fixed parameter strategies are difficult to adapt to dynamic operating scenarios and channel changes, so it is necessary to construct an adaptive handover parameter optimization algorithm. Cross-area handover includes three stages: link measurement, handover decision, and handover execution. The decision stage is the core of determining handover reliability; the essence of parameter optimization is to balance trigger timeliness and process stability in this stage, ensuring that the handover is completed within the effective coverage area while suppressing ping-pong handover induced by instantaneous signal fluctuations.

[0039] 1.4 Problem Description Switching margin With trigger duration T j,k Together, these factors determine the switching trigger position x1, the command issuance position x2, and the switching completion position x3. A smaller HOM or a shorter TTT helps to trigger the switching earlier, but if the parameter settings are too sensitive, they are easily affected by instantaneous power fluctuations, thus triggering unnecessary reverse switching. Conversely, a larger HOM or a longer TTT, while beneficial for suppressing frequent switching, can lead to delayed switching decisions and increase the risk of HOF (Head-Off Failure). Therefore, there is a significant trade-off between HOM and TTT, and their values ​​should be dynamically adjusted according to the train's position and operating status.

[0040] Considering that the handover processes between adjacent access point pairs are independent, the overall network handover parameter optimization problem can be decomposed into multiple local sub-problems. Taking the serving access point (AP) as an example... j To the target access point AP k Taking the switching process as an example, the parameter vector to be optimized is defined as: (11)

[0041] The corresponding switching parameter optimization problem can be expressed as: (12) in, and AP j To AP k The upper bounds of HOF and PP probabilities during the handover process, with w1 and w2 as weighting coefficients, are used to characterize the system's response to two types of abnormal handover risks. and These are the lower and upper bounds of the switching margin, respectively; Tmax is the maximum allowed trigger delay of the system.

[0042] 2. Switch boundary geometry model In IRS-VLC mobile networks, optimizing handover performance relies on a precise description of the Handover Trigger (HOT), HOF, and PP locations. The optical power received by the onboard PD can be represented as a continuous function of the train's real-time spatial coordinates. Therefore, traditional power domain handover decision conditions can be mapped to spatial geometric constraints, thus constructing a geometric model of the handover boundaries for HOT, HOF, and PP. To uniformly describe the boundary morphology, circles and straight lines are used as basic elements. The circular boundary is represented as {X... c ,R c} represent the coordinates of the center and radius of the circle, respectively; the straight line Record .

[0043] 2.1 Switching Boundaries Without IRS Assistance When the vehicle-mounted PD only receives the LosS signal from the LED, the trajectory of the HOT and PP reaching a given threshold can be determined by the ratio of received optical power at adjacent access points, while the HOF boundary corresponds to the target AP's received optical power dropping to the system's lowest receive threshold P. th The critical position. Under conditions of no LC-IRS assistance and adjacent access points using the same transmit power and optical configuration, the decision boundary between HOT and PP can be further simplified to the train position to AP. j With AP k The square of the distance d is the ratio of the trajectory. For switching events i∈{HOT,PP}, their initial boundaries can all be unified as (13) In the formula, This is the squared distance ratio factor for the corresponding events; when i corresponds to the HOT event, Set up the service access point (AP). j and target access point AP k The physical location coordinates are denoted as X. j =(x j ,y j ) and X k =(x k ,y k According to the Apollonius circle property, when At that time, the initial boundary For a circular trajectory, its expression is: (14) when When the boundary is switched, it can be equivalent to a straight line, and its expression is: (15) In two-dimensional space, when When, the region satisfying the event criterion is located inside the circle; when At this time, the region satisfying the event criterion is located outside the circle. For HOF events, the decision condition is that the received optical power of the target AP is lower than the minimum receive threshold. The following corresponds to the boundary as the target AP. k The equal-power receiving trajectory. Since the incident angle and optical gain change gradually near the boundary, the boundary can be approximated as a circle centered on the target AP, denoted as . (16) in, To meet The equivalent radius corresponding to the time. When the train crosses this boundary, the LoS received optical power provided by the target access point will be lower than the threshold, thus posing a risk of handover failure.

[0044] 2.2 IRS-assisted switching boundaries When the vehicle-mounted PD receives cooperative communication from the IRS, mode L uses a combination of direct and reflected link assistance, while mode N uses only reflected link assistance. For the serving access point (AP)... j The schedulable g-th IRS array is denoted by its equivalent reflection center as X. g =(x g ,y g For events i∈{HOT,PP}, the initial boundary under a given working mode m∈{L,N} can be written as a squared distance ratio trajectory. (17) In the formula, Events under pattern m The corresponding squared distance ratio factor. When When the initial boundary is a circle, its expression is: (18)

[0045] when At that time, the initial boundary degenerates into the perpendicular bisector of the line connecting the target access point and the IRS center, and its expression is: (19) The modeling process for the HOF boundary under IRS assistance is consistent with Equation (16). During operation, the train may lose the auxiliary link due to changes in the geometric relationship with the serving IRS array. At this time, the total received optical power will change abruptly, potentially triggering HOT or causing HOF. When the train moves away from the IRS array, causing the auxiliary reflection gain to fall below the minimum effective threshold... At that time, the gain interrupt boundary can be obtained as (20) in The geometric constants related to the IRS reflection characteristics are determined by the minimum effective reflection gain threshold. The relevant geometric constants, unit reflectivity, and path loss index together determine and satisfy the following conditions. ,in Let be the reference reflection link gain per unit distance. The auxiliary reflection link is interrupted when the train moves outside the effective angle range of the IRS. Let the normal direction of the g-th IRS array be . The effective half-width is The corresponding angular break boundary is represented by two rays. (twenty one) From equation (21), it can be seen that the two rays begin with the first... IRS array locations As the endpoint, along the array normal direction With effective half angle The determined boundary direction extends to enclose the effective angular service area of ​​the IRS.

[0046] 2.3 Switch Boundary Correction The boundary trajectories generated by the initial geometric model cover all theoretical locations where switching occurs. These need to be corrected by considering the tunnel's physical space and light propagation constraints. The final effective switching boundary is a set of feasible trajectories under multiple constraints. For circular elements, the effective portion is denoted as an arc tuple. X c R and R represent the center and radius of the circle, respectively. and This represents the starting and ending polar angles of the effective arc. For straight line elements, the effective portion is denoted as a line segment tuple. X s and X e Let d be an endpoint of the line segment. s and d e It is a distance scalar in a linear projection coordinate system.

[0047] First, although the tunnel cross-section is a circular arch, the actual operating area of ​​the train trajectory and switching control can be simplified to a rectangular region. Let the equation of the constraint line be... , represents the signed distance function of a point on a given curve. The sign of the function indicates whether the switching boundary point is located within the tunnel operating area; if the initial boundary point is outside the rectangular area, an empty set is output. For a circular initial boundary, its effective circular arc interval on the same side of the constraint line can be represented as... (twenty two) in, This represents the distance from the center of the circle to the constraining line. The polar angle at the intersection of the line and the circle. The midpoint of the corresponding short arc. This formula is used to exclude arc segments outside the tunnel operating area or on the side where the constraint is ineffective, as a reference point indicating the effective side.

[0048] Secondly, consider the minimum effective reflection gain constraint. Let the constraint circle be... This represents the effective service range of the IRS in a given operating state. If the initial handover boundary is {X} c If R ∈ R, then the distance between the centers of the two circles is... The effective arc range of the initial circular boundary within the constraint circle can be represented as: (twenty three) Among them, the angle of intersection Derived from the Law of Cosines (twenty four) Here This represents the azimuth angle of the line connecting the centers of the circles, accurately eliminating trajectory segments that do not meet energy or distance constraints. By comprehensively considering constraints such as tunnel operating space, effective angle range, and minimum effective reflection gain, physically inaccessible trajectory segments can be accurately eliminated, ultimately obtaining the effective geometric boundaries of HOT, HOF, and PP events under actual physical constraints.

[0049] 2.4 Probability Estimation of HOF and PP After handover is triggered, the change in signal quality during the TTT duration determines whether HOF (House of Failure) occurs; after handover is completed, the dwell time T... p Changes in internal signal quality determine the probability of a handover event (PP). By analyzing the shortest geometric distance from the handover trigger point to each risk boundary, the risk of different handover events can be assessed.

[0050] To address the challenge of directly handling continuous boundaries in numerical computation, the boundary is discretely sampled to obtain a finite set of location points. Along each line segment or arc on the boundary, points are sampled at equal intervals. Take points in sequence until the remaining length of the segment is less than The set of switching boundary points after discretization is as follows: For the train's location point X0, the distance to the risk arc The distance is (25) in Point X0 is relative to the center X of the circle. c The polar angle. Similarly, the distance from point X0 to line segment X0. The distance is (26) Where X fIt is the distance from point X0 to the straight line containing the line segment. The foot of the perpendicular. Based on the above geometric mapping, from... Switch to The upper bound of the handover failure probability is (27) Where N is the total number of discrete points. It is an indicator function that takes 1 if the condition is met and 0 otherwise. Equation (27) shows that if the distance the train may travel before the switch is completed is greater than the shortest distance from the trigger point to the HOF boundary, then the trigger point is considered to have HOF risk. Under the same boundary model, the position of the trigger point X0 after the switch is completed after experiencing TTT and execution delay is denoted as Then the upper bound of the PP probability is (28) The method for assessing the upper bound of the switching anomaly probability is based on worst-case considerations, that is, assuming that the train points to the nearest risk boundary along the shortest path, thus deriving a probability index with high robustness. As can be seen from equations (27) and (28), the location of the switching boundary and the upper bound of the HOF / PP risk are both related to the train's position and speed.

[0051] 3. Optimization of handover parameters based on location prediction Traditional handover parameters are mostly statically configured based on the current network state, making it difficult to adapt to the rapid dynamic changes in optical link quality in mobile scenarios, which can easily lead to HOF and PP problems. To address this, a train operation state prediction-driven adaptive optimization model for handover parameters is constructed: the train position and speed at the next decision moment are obtained through a time-series prediction network, and the HOM and TTT are dynamically optimized by combining the handover boundary geometry model. Compared to methods that directly predict handover parameters, this hierarchical architecture reduces the training complexity of the prediction model while ensuring the controllability of the parameter optimization results. Since the handover boundary morphology and the upper bounds of HOF and PP risks are strongly correlated with train position and speed, high-precision prediction is a core prerequisite for subsequent boundary updates and parameter optimization.

[0052] 3.1 Prediction Module Based on ISSA-LSTM In the IRS-VLC train handover scenario, train position, speed, and handover geometric boundaries exhibit significant spatiotemporal correlation. To obtain train state information at the next decision moment, an ISSA-LSTM prediction module combining the Improved Sparrow Search Algorithm (ISSA) and LSTM was designed. The overall parameter optimization model is as follows: Figure 4 As shown.

[0053] LSTM, through its gating mechanism, can characterize long-term dependencies in trajectory sequences, making it suitable for nonlinear time-series prediction of train operating states. To clarify the mapping relationship between the prediction model and the input feature space, the historical position x and velocity v of the train over the past m consecutive time steps are defined as the input feature matrix S. t =[(x,v) t-m+1 ,…,(x,v) t ] T After deep feature extraction by the LSTM network, the fully connected layer maps and outputs the state prediction vector p steps ahead. The predictive performance of LSTM is highly dependent on hyperparameter configuration. Traditional manual tuning and grid search methods are computationally expensive and prone to getting trapped in local optima. Current mainstream automatic hyperparameter optimization methods mainly include two categories: Bayesian optimization and swarm optimization. Bayesian optimization has poor adaptability when dealing with discrete hyperparameters and has high computational cost; swarm optimization algorithms do not rely on the gradient information of the objective function, have strong compatibility with discrete-continuous mixed variables, and are the mainstream technical approach for hyperparameter optimization of trajectory prediction models.

[0054] SSA (Search Algorithm for Detectors and Detectors) achieves an efficient balance between global exploration and local development through a hierarchical collaborative mechanism involving discoverers, participants, and scouts. This algorithm boasts low computational complexity and an easily modifiable structure, making it suitable for low-to-medium dimensional hyperparameter optimization scenarios in train trajectory prediction, while also meeting the efficiency and reliability requirements of onboard computing environments. However, the original SSA is susceptible to factors such as uneven initial population distribution, fixed role ratios, and premature convergence of the optimal individual in complex optimization problems, leading to local optima. To address these issues, improvements are made in three areas: population initialization, dynamic search adjustment, and optimal individual updating.

[0055] (1) Population initialization based on Chebyshev chaotic mapping The original SSA (Solution-Specific Algorithm) initializes the population randomly, which can easily lead to uneven distribution of individuals and reduce solution space coverage. To improve the uniformity of the initial population, the Chebyshev chaotic mapping is used to generate the initial solution, the expression of which is: (29) In the formula, Z p Let p be the p-th chaotic variable, and c be the order of the Chebyshev mapping. This mapping has good ergodicity and randomness, which helps to expand the initial search range and improve population diversity. Figure 5As can be seen from point a, random initialization exhibits significant local clustering and uneven coverage, making it difficult to guarantee the diversity of the initial population. In contrast, the Chebyshev chaotic mapping, with its excellent ergodicity, allows the population to be evenly distributed throughout the solution space. This characteristic enhances the algorithm's global exploration capability, preventing premature convergence to local optima, and also shortens the optimization iteration process, significantly improving the accuracy and stability of the prediction module.

[0056] (2) Dynamic nonlinear scaling factor In the SSA algorithm, the number of discoverers and scouts is usually fixed in advance, making it difficult to simultaneously address the global search in the early stages of iteration and the local convergence in the later stages. Therefore, a dynamic nonlinear scaling factor is introduced to adaptively adjust the search strategy, and its expression is: (30); Where n is the current iteration number, N max The maximum number of iterations, A random number between [0, 1]. Figure 6 As shown, this strategy enables the algorithm to maintain strong global search capabilities in the early stages of iteration and enhance local exploration capabilities in the later stages.

[0057] (3) Optimal individual perturbation based on Cauchy variation and Tent mapping To enhance the global exploration capability and local exploitation accuracy of the SSA algorithm, Cauchy mutation and tent mapping perturbation are introduced into the global optimal position update strategy. After each iteration, the current global optimal solution is perturbed, as expressed by: (31); Where Cauchy(0,1) represents a standard Cauchy distribution random number, whose heavy-tailed characteristic is beneficial for increasing the search step size and enhancing the ability to escape local optima; Tent is the tent map, used to enhance local neighborhood search, and its expression is: (32)

[0058] The key hyperparameters of the LSTM are encoded as position vectors of individual sparrows. The root mean square error (RMSE) of the validation set is used as the fitness function to iteratively optimize within the hyperparameter space. Once the termination condition is met, the optimal hyperparameter combination is output, thus constructing the final prediction model. The complete process is as follows: Figure 7 As shown, the model outputs predicted train position and speed at the next decision moment, providing input support for subsequent switching boundary updates and parameter adaptive optimization.

[0059] 3.2 Adaptive Switching Algorithm Flow The adaptive switching algorithm completes four stages in each decision cycle: state prediction, boundary update, parameter optimization, and execution. For example... Figure 8As shown, the system first uses the train position state sequence to predict the position and speed in the next short time domain; then, it updates the geometric boundaries of HOT, HOF and PP according to the predicted state; next, it adaptively adjusts HOM and TTT through a genetic algorithm; finally, when the train enters the next decision time, it executes the cross-region handover decision according to the new handover parameters, and merges the new observation state through a sliding window to enter the next cycle.

[0060] 3.3 Optimization of switching parameters based on the switching boundary geometry model The effective geometric boundaries of HOT, HOF, and PP are updated based on the predicted train position and speed values, and the risk upper bounds of HOF and PP under given switching parameters are further calculated. This optimization objective is influenced by the predicted train state, the shape of the IRS-assisted switching boundary, and the parameter vector. It has no closed-form analytical solution and is constrained by the feasible regions of HOM and TTT, classifying it as a constrained non-convex optimization problem. Therefore, a genetic algorithm (GA) is used to heuristically search the parameter space to find the parameter combination with the minimum risk. The algorithm pseudocode is shown in Table 1, with an initial population size of [missing information]. The maximum number of iterations is The crossover probability is The mutation probability is The threshold for early termination is Prediction module output and Used for online updates of boundaries and risk assessments, driving adaptive adjustments of HOM and TTT. For each set of switching parameters... Calculate the upper bound of the HOF / PP probability and construct the fitness function. The larger the fitness, the lower the switching risk.

[0061] Table 1. Optimization of switching parameters based on the switching boundary geometry model. .

[0062] 4. Simulation Experiment and Result Analysis Based on the "Metro Train Operation Simulation Test Platform" of Gansu Provincial Rail Transit Research Center and combined with real line data from Chengdu Metro Line 1, a VLC (Vehicle-to-Ground Communication) simulation scenario was constructed. The experiment first evaluated the prediction performance of ISSA-LSTM for train position and speed; then analyzed the impact of introducing IRS on the geometric distribution of VLC handover boundaries; finally, it compared and analyzed the performance advantages of the proposed handover algorithm in reducing the probability of handover anomalies and improving handover stability at different train operating speeds.

[0063] 4.1 Simulation Scenarios and Parameter Settings In the simulation experiment, the transmitting end reused the tunnel lighting source. The installation height of the visible light communication AP was set to 4.5 m, and 11 APs were evenly arranged at 10 m intervals along the train's running direction. The LC-IRS array was deployed on the opposite side of the tunnel wall corresponding to the midpoint of each pair of adjacent APs, with the array center installed at a height of 4.0 m. The array consisted of 8×8 square reflective elements, with each reflective element measuring 5 mm × 5 mm. The receiving end PD was installed on the top of the train's front, and its running trajectory was generated based on the actual route topology data of Chengdu Metro Line 1. The main parameters are shown in Table 2.

[0064] Table 2 Experimental parameters .

[0065] 4.2 Performance Validation of Prediction Model (1) ISSA optimization performance verification To verify the optimization performance of the proposed ISSA, representative swarm intelligence algorithms from different development stages were selected for horizontal comparison. Classic benchmark algorithms include Particle Swarm Optimization (PSO) and Grey Wolf Optimizer (GWO), while emerging algorithms include Dung Beetle Optimizer (DBO) and Saddle-Billed Optimization Algorithm (SBOA). To ensure fair comparison of different algorithms under the same computational cost, the population size was uniformly set to 30 and the maximum number of iterations to 50.

[0066] The fitness change curves of different optimization algorithms in the hyperparameter optimization process are as follows: Figure 9 As shown, ISSA-LSTM achieves faster fitness reduction in the early stages of iteration and stabilizes around the 20th iteration; in contrast, SSA-LSTM stalls after approximately the 23rd iteration. These results demonstrate that ISSA outperforms the original SSA in both convergence speed and optimization stability, making it more suitable for adaptive search of key hyperparameters in LSTM.

[0067] (2) Validation of ISSA-LSTM prediction performance To verify the joint prediction performance of ISSA-LSTM for train position and speed, a single-step state prediction experiment was designed. Experimental data was taken from a train driving simulation platform, and the input features included timestamps, cumulative mileage, instantaneous speed, and acceleration. The prediction model consisted of an input layer, an LSTM layer, a Dropout layer, and a fully connected output layer. The ISSA algorithm population size was set to 30, the maximum number of iterations was 50, and the optimal hyperparameters obtained were: learning rate 0.0012, number of hidden layer neurons 128, number of fully connected layer nodes 64, and dropout rate 0.15.

[0068] Figure 10 a, Figure 10 b represents the loss and accuracy curves for position and velocity prediction on the validation set. With increasing training epochs, the loss for position and velocity prediction gradually decreases and converges, while the accuracy continuously improves; the position prediction accuracy reaches 0.9974, and the velocity prediction accuracy reaches 0.9961 on the validation set. The results show that the hyperparameters optimized by ISSA effectively improve the model's learning ability and generalization performance, accurately capturing the nonlinear characteristics of train operation and ensuring high-precision and stable convergence.

[0069] (3) Analysis of Train Operation Status Prediction Results The train operation status prediction performance results obtained based on the ISSA-LSTM model training are as follows: Figure 11 a, Figure 11 As shown in b, within the 40s to 120s interval, speed, as a differential component, is extremely sensitive to high-frequency fine-tuning during the train's cruise phase, while position, as an integral component, exhibits strong smoothness. Comparing the actual trajectory with the predicted curve reveals that the train's actual operating state closely matches the predicted state, with a maximum position error of 2.26m and a maximum speed error of 1.35km / h, indicating that the model possesses good predictive performance.

[0070] (4) Performance comparison of different prediction models To verify the model's performance advantage, it was compared with Back Propagation (BP) neural networks, LSTM, Gated Recurrent Units (GRU), Bidirectional Long Short-Term Memory (Bi-LSTM) networks, and the unmodified SSA-LSTM model. All models underwent 10 independent experiments on the same test set, and the average of each metric was used as the final performance evaluation. Evaluation metrics included the coefficient of determination (COP). The mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are all calculated based on the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).

[0071] Table 3. Comparison of performance indicators for vehicle position prediction under different models .

[0072] Tables 3 and 4 show that ISSA-LSTM achieves optimal results in both position and velocity prediction tasks. Compared to SSA-LSTM, its position prediction MSE is reduced by approximately 56.8%, and its velocity prediction MSE is reduced by approximately 78.8%. Experimental results confirm that by refining hyperparameters through ISSA, the accuracy of LSTM in capturing the complex motion characteristics of rail transit can be significantly improved, providing a highly reliable position and velocity benchmark for subsequent adaptive optimization of switching parameters.

[0073] Table 4. Comparison of performance indicators for vehicle speed prediction under different models .

[0074] 4.3 IRS-assisted switching performance analysis (1) The influence of IRS on switching geometry boundary To analyze the regulatory effect of the IRS on the handover boundary of a VLC system, the impact on handover is evaluated by comparing the two-dimensional handover boundary with and without IRS assistance, and comparing it with the three-dimensional receive power topology. Figure 12 As shown in a, 12b, 12c, and 12d, without an IRS, the received power decreases rapidly with distance, exhibiting a "high in the center, low at the edge" distribution, with the handover boundary concentrated at the track center. Using -30dBm as the received signal quality assessment threshold, the boundary radius is approximately 6.8m, indicating that the edge is prone to prematurely entering the handover sensitive area, leading to frequent and abnormal handovers. After introducing an LC-IRS, the spatial power distribution is reconstructed, the power in the weak signal area is significantly improved, and the handover boundary expands to 12.2m, an increase of 82.34%. This effectively expands the handover timing adjustment margin, improves the quality of edge links, and reduces the risk of abnormal handovers.

[0075] Within the coverage area of ​​two adjacent APs, the curves showing the change in received optical power of the vehicle-mounted PD with the train's running position are as follows: Figure 13As shown, the signal evolution characteristics at the coverage edge under IRS assistance can be intuitively compared. Based on the receiving sensitivity index of the vehicle-mounted PD, -40dBm is set as the link receiving failure threshold. In the coverage overlap area, if the received power oscillates frequently near the threshold, the handover decision condition will be repeatedly activated, corresponding to the PP handover trigger point marked in the figure. Without IRS assistance, the optical power attenuation in the AP overlap area is severe and the signal fluctuation is obvious, which can easily lead to ping-pong handover or even link interruption. After introducing IRS, the reflected signal energy compensation effectively smooths the edge power fluctuation, and the number of PP handover triggers is significantly reduced. Tests were conducted in a tunnel section between adjacent subway stations with a length of about 1.5km. By improving the physical layer signal quality only through IRS, the number of ping-pong handovers in the section was reduced by about 44.7%, and the stability and reliability of cross-area handover were effectively improved.

[0076] (2) Comparison of handover anomaly probabilities of different handover algorithms To verify the effectiveness of the proposed adaptive handover algorithm based on position state prediction in suppressing handover anomalies, it was compared with the fixed parameter (FP) handover algorithm, the best cell (BC) instantaneous handover algorithm based on the strongest received signal strength, and the coordinated multi-point (CoMP) handover algorithm. In the experiment, the train speed was set to 30 km / h, and simulation analyses were conducted under both IRS-assisted and IRS-assisted conditions.

[0077] like Figure 14 a Figure 14 As shown in Figures b and c, regarding PP probability, the BC algorithm, due to its "strong signal priority" approach and lack of handover hysteresis protection, is prone to triggering invalid and repeated handovers in signal overlap areas, resulting in the worst performance. The CoMP algorithm, by leveraging multi-point cooperation to smooth signal fluctuations, achieves performance close to the proposed algorithm. Regarding HOF probability, the FP algorithm takes the highest value, but its fixed trigger parameters lead to significant decision lag issues in VLC microcell scenarios. The BC algorithm, with its "instantaneous entry" characteristic, can quickly complete handovers, making its performance closest to the proposed algorithm. The proposed algorithm performs best in both suppressing PP and reducing HOF, thanks to the combination of ISSA-LSTM high-precision position prediction and a geometric boundary model, enabling spatiotemporal dynamic adjustment of handover parameters. With IRS assistance, compared to the traditional FP algorithm, the proposed algorithm reduces PP and HOF probabilities by 77.78% and 93.33%, respectively, significantly improving the continuity and robustness of the vehicle-to-ground communication system.

[0078] (3) Comparison of switching performance under different speed conditions To verify the algorithm's switching performance in dynamic scenarios, train speeds were increased from 10 km / h to 80 km / h in an IRS-assisted tunnel environment, with the overall switching anomaly probability used as the evaluation metric. For example... Figure 15As shown, with increasing vehicle speed, link power fluctuations intensify, the handover trigger position deviates from the ideal geometric boundary, and the handover anomaly probability of all algorithms shows an upward trend. Traditional FP algorithms use fixed parameters, which cannot adapt to dynamic changes in the link during high-speed scenarios, resulting in significant decision-making lag after speed increases. The proposed algorithm combines state prediction with a handover geometric boundary model, allowing for dynamic parameter adjustment based on train operating status, ensuring reasonable handover timing at all speeds. At 80 km / h, its overall handover anomaly probability is reduced by approximately 88.89% compared to the FP algorithm, demonstrating a clear performance advantage in high-speed scenarios.

[0079] (4) Switching decision latency and real-time performance analysis The computational overhead of the proposed switching algorithm consists of two stages: state prediction and parameter optimization. Let the prediction network time window length be m, the number of hidden layer neurons be h, the optimization algorithm population size be N, the maximum number of iterations be ϖ, and the search dimension be d. Then the overall asymptotic time complexity of the algorithm, O(m·h² + N(d + logN)), can converge to the polynomial time level and will not experience an exponential computational explosion as the feature dimension or the number of network layers increases.

[0080] To verify the applicability of the proposed adaptive handover algorithm in high-speed train operation scenarios, the average handover decision latency of different handover algorithms was compared, as shown in Table 5. The FP and CoMP algorithms, which lack a prediction mechanism, only involve scalar threshold comparison, resulting in low computational complexity and minimal decision latency. The latency of prediction-based algorithms increases with model complexity, with BP, LSTM, and SSA+LSTM showing progressively larger decision latencies. The proposed algorithm achieves an average single handover decision latency of 2.96 ms to ensure high-precision state prediction and adaptive parameter optimization. When the train passes through the IRS-assisted visible light communication network at a maximum speed of 80 km / h to switch overlapping areas, the decision latency accounts for less than 1% of the actual physical handover time window, meeting the CBTC standard requirement of less than 100 ms for vehicle-to-ground wireless communication handover latency. Therefore, the proposed algorithm achieves a significant leap in handover performance with minimal computational overhead, meeting the real-time requirements of high-speed train operation.

[0081] Table 5 Comparison of average decision delay for different handover algorithms .

[0082] 5. Conclusion To address the frequent handover failures and ping-pong handovers during VLC train handover in tunnel scenarios, an IRS-assisted adaptive handover algorithm is proposed. This method reconstructs the spatial distribution of optical signals by introducing an LC-IRS, constructing a handover boundary geometric model and mapping the risk assessment of HOF and PP from the power domain to the spatial domain. Simultaneously, it combines ISSA-LSTM to predict train position and speed, and dynamically optimizes the handover margin and trigger duration based on this prediction, thus achieving adaptive handover control for moving scenarios. Simulation results show that the constructed ISSA-LSTM prediction model can accurately characterize changes in train operating status, with maximum position and speed errors of 2.26 m and 1.35 km / h, respectively, providing reliable input for handover parameter optimization. After introducing the IRS, the system handover geometric boundary diameter increases by 79.41%, indicating that the IRS can improve coverage continuity and expand the effective handover area. Compared with fixed-parameter handover methods, the proposed algorithm reduces the ping-pong handover probability and handover failure probability by 77.78% and 93.33%, respectively; and maintains a low overall abnormal handover probability even under high-speed operating conditions. The research results confirm that the introduction of IRS effectively expands the system coverage and provides a reliable guarantee for cross-area handover of visible light communication trains in tunnel scenarios; the proposed handover algorithm can significantly reduce HOF and PP risks and effectively guarantee the communication continuity of trains under rapid movement.

[0083] This embodiment does not limit the implementation process of the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for adaptive handover of visible light communication trains assisted by an intelligent reflective surface, characterized in that: An IRS-VLC vehicle-to-ground communication network model is constructed using an IRS. The IRS-VLC vehicle-to-ground communication network model includes an on-board digital camera (PD), a lighting source installed along the tunnel wall as an access point, and an IRS installed on the tunnel wall opposite the access point. The access point and the on-board PD form a Loss-of-Sight (LoS) direct link model, and the IRS and the on-board PD form an IRS reflection link model. The on-board PD, the lighting source, and the IRS form the IRS-VLC vehicle-to-ground communication network model. The instantaneous received optical power of the on-board PD is superimposed from the LoS direct link and the reflection link assisted by the LC-IRS. The IRS-assisted visible light communication train adaptive handover method is based on the IRS-VLC vehicle-to-ground communication network model. Within each decision cycle, it completes four stages: state prediction, boundary update, parameter optimization, and execution. First, it uses the train's position and state sequence to predict the position and speed in the next short time domain. Then, it updates the geometric boundaries of the handover trigger (HOT), handover failure (HOF), and ping-pong turnaround (PP) based on the predicted state. Next, it adaptively adjusts the handover margin (HOM) and trigger delay (TTT) using a genetic algorithm to adapt the handover parameters to the train's real-time operating state. Finally, when the train enters the next decision time, it executes the handover decision according to the new handover parameters and merges the new observation state through a sliding window to enter the next cycle.

2. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 1, characterized in that... : In the LoS direct link: Let the transmitted optical power of the service access point be P. t,j The train's current position is denoted as X, the effective receiving area of ​​the onboard PD is A, and the corresponding LosS DC channel gain is denoted as... ; The LC-IRS-assisted reflection link is as follows: Suppose that the nth liquid crystal cell in the IRS array serving the service access point is under control voltage u n The state coefficients are as follows This characterizes the unit's overall modulation capability of the reflection link under its current operating state; the optical signal transmitted by the serving access point reaches the vehicle-mounted PD after being reflected by this unit, and its auxiliary reflection link gain is expressed as... ; The reference reflection link gain is determined by spatial geometry without considering the liquid crystal electronic control state; Considering polarization loss factor Reflection efficiency function related to incident angle The corrected expression is: ； in, The effective area of ​​the nth LC-IRS cell. The distance between the access point and the nth unit. The distance between the nth unit and PD. For the transmission angle from the service access point to the IRS unit, The incident angle at the IRS unit. The reflection angle from the IRS unit to the vehicle-mounted PD. The angle of incidence at the vehicle-mounted PD; If the IRS arrays serving this area have a total of N j If there are 1 effective unit, then at train position X, the total reflected and received optical power after IRS assistance is: ; During train operation, the total effective optical power actually received by the onboard PD is composed of the LOS direct component and the LC-IRS auxiliary reflection component. .

3. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 2, characterized in that... : When the total received optical power of the target access point reaches a preset handover margin relative to the serving access point When this occurs, the system triggers a handover decision; the triggering condition formula is as follows: ; Let t1 be the time when the trigger condition formula first becomes true, and x1 be the corresponding train position; let T be the system startup duration. j,k The TTT timer; if the condition is within the time window [t1, t1+T] j,k If this holds true continuously within a given time period, then at time t2 = t1 + T... j,k Issue a switching command; If execution delay is taken into account The train will eventually be The switchover is complete; the instantaneous speed of the train along the track direction is v(t), then the time and position relationship from triggering to the completion of the switchover can be uniformly expressed as: In the formula, x3 represents the position of the train when the switch is completed; If we assume that the train speed remains approximately constant during the short switching interval, and denote v1 = v(t1), then we have ; To ensure that the handover can be completed within the effective overlapping coverage area, the cumulative displacement of the train within [t1,t3] should not exceed the remaining length of the overlapping coverage area corresponding to the current position. ,Right now ; Depend on Therefore, the feasible upper bound of TTT is satisfied. .

4. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 3, characterized in that... : During the TTT timing phase or execution delay phase following the handover trigger, if the total received optical power corresponding to the target access point drops to the receive threshold P before the handover is completed... th The following are the criteria for determining HOF. Conversely, if the train successfully completes the transfer to AP at time t3... k The switching occurs during the minimum dwell time T. p If, due to transient channel fluctuations, the reverse handover condition is reactivated, a PP (Pre-Transfer) event is determined to have occurred. The criterion for this is expressed as follows: .

5. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 4, characterized in that... : When switching from the service access point to the target access point, the parameter vector to be optimized is: The corresponding switching parameter optimization problem is expressed as follows: ; in, and AP j To AP k The upper bounds of HOF and PP probabilities during the handover process, with w1 and w2 as weighting coefficients, are used to characterize the system's response to two types of abnormal handover risks. and These are the lower and upper bounds of the switching margin, respectively; T max This represents the maximum allowed trigger delay of the system.

6. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 1, characterized in that... : When updating the geometric boundaries of the handover trigger (HOT), handover failure (HOF), and ping-pong (PP) based on the predicted state, the steps for IRS-assisted handover boundary adjustments are as follows: For the g-th schedulable IRS array at the service access point, let its equivalent reflection center be X. g =(x g , y g For events i∈{HOT, PP}, the initial boundary under a given working mode m∈{L, N} is written as the squared distance ratio to the trajectory. In the formula, For events under pattern m The corresponding squared distance ratio factor; Mode L is for combined direct and reflected link assistance, and Mode N is for reflected link assistance only; when When the initial boundary is a circle, its expression is: ； when At that time, the initial boundary degenerates into the perpendicular bisector of the line connecting the target access point and the IRS center, and its expression is: ； For the HOF boundary under IRS assistance, the modeling process and formula Consistent; During operation, the train may also lose its auxiliary link due to changes in its geometric relationship with the serving IRS array; At this point, the total received optical power will change abruptly, which may directly trigger HOT or cause HOF. When the train moves away from the IRS array, the auxiliary reflection gain falls below the minimum effective threshold. At that time, the gain interruption boundary can be obtained as follows: ; in The geometric constants related to the IRS reflection characteristics are derived from... The relevant geometric constants, unit reflectivity, and path loss index together determine and satisfy the following conditions. ,in The reference reflection link gain per unit distance; When the train moves outside the effective angle range of the IRS, the auxiliary reflection link is interrupted; Let the normal direction of the g-th IRS array be... The effective half-width is The corresponding angular break boundary is represented by two rays. ; From the above equation, it can be seen that the two rays begin with the first... IRS array locations As the endpoint, along the array normal direction With effective half angle The determined boundary direction extends to enclose the effective angular service area of ​​the IRS.

7. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 6, characterized in that... : When updating the geometric boundaries of the handover trigger (HOT), handover failure (HOF), and ping-pong (PP) based on the predicted state, it is necessary to correct the IRS-assisted handover boundaries. The steps of this correction method are as follows: For circular elements, the effective portion is denoted as an arc tuple. X c R and R represent the center and radius of the circle, respectively. and Indicates the starting and ending polar angles of the effective circular arc; For line elements, the effective part is denoted as a line segment tuple. X s and X e Let d be an endpoint of the line segment. s and d e This is a distance scalar in a linear projection coordinate system; Let the equation of the constrained line be , represents the signed distance function of a point on a given curve. The sign of the function determines whether the switching boundary point is located within the tunnel operating area; If the initial boundary point is located outside the rectangular region, an empty set is output; for a circular initial boundary, the effective circular arc interval on the same side of the constraint line is represented as: ； in, This represents the distance from the center of the circle to the constraining line. The polar angle at the intersection of the line and the circle. The midpoint of the corresponding short arc. As a reference point indicating the effective side, this formula is used to exclude arc segments outside the tunnel operating area or on the side where the constraint is ineffective; Considering the minimum effective reflection gain constraint, let the constraint circle be... This represents the effective service range of the IRS in a given operating state. If the initial handover boundary is {X} c If R}, then the distance between the centers of the two circles is The effective arc interval of the initial circle boundary within the constraint circle is represented as: ； Among them, the angle of intersection Derived from the Law of Cosines: ； It represents the azimuth angle of the line connecting the centers of the circles, accurately eliminating trajectory segments that do not meet energy or distance constraints; and finally obtaining the effective geometric boundaries of HOT, HOF, and PP events under actual physical constraints.

8. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 7, characterized in that... : After handover is triggered, the change in signal quality during the TTT duration determines whether HOF (House of Failure) occurs; after handover is completed, the dwell time T... p The change in internal signal quality determines the probability of PP occurrence. Therefore, the methods for estimating the probabilities of HOF and PP are as follows: Discrete sampling is performed on the switching boundary to obtain a finite set of location points; Along each line segment or arc on the boundary, at equal intervals Take points in sequence until the remaining length of the segment is less than ; The set of switching boundary points after discretization is For the train's location point X0, the distance to the risk arc The distance is: ； in Point X0 is relative to the center X of the circle. c The polar angle; Similarly, from point X0 to line segment The distance is: ； Where X f It is the distance from point X0 to the straight line containing the line segment. The foot hanging down; Based on the above geometric mapping, from Switch to The upper bound of the handover failure probability is: ； Where N is the total number of discrete points. It is an indicator function that takes 1 if the condition is true and 0 otherwise; it indicates that if the distance the train travels before the switch is completed is greater than the shortest distance from the trigger point to the HOF boundary, then the trigger point is considered to have HOF risk. Under the same boundary model, let the position of trigger point X0 after the switch completion following the TTT and execution delay be denoted as . Then the upper bound of the PP probability is: 。 9. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 8, characterized in that... : The switching margin (HOM) and trigger delay (TTT) are adaptively adjusted using a genetic algorithm to adapt the switching parameters to the real-time operating status of the train. The genetic algorithm used is an ISSA-LSTM prediction module. The steps involved in this task are as follows: Define the historical position x and velocity v of the train over the past m consecutive time steps as the input feature matrix S. t =[(x,v) t-m+1 ,…, (x,v) t ] T After deep feature extraction by the LSTM network, the fully connected layer maps and outputs the state prediction vector p steps ahead. ; ISSA first uses the Chebyshev chaotic map to generate an initial solution, the expression of which is: In the formula, Z p Let p be the p-th chaotic variable, and c be the Chebyshev mapping order; Then, a dynamic nonlinear scaling factor is introduced to achieve adaptive adjustment of the search strategy, the expression of which is: ； Where n is the current iteration number, N max The maximum number of iterations, A random number between [0, 1]; Finally, Cauchy mutation and Tent mapping perturbation are introduced into the global optimal position update strategy; after each iteration, the current global optimal solution is perturbed, and the expression is: ； Where Cauchy(0,1) represents a standard Cauchy distribution random number, whose heavy-tailed characteristic is beneficial for increasing the search step size and enhancing the ability to escape local optima; Tent is the tent map, used to enhance local neighborhood search, and its expression is: 。 10. The intelligent reflective surface-assisted adaptive handover method for visible light communication trains according to claim 9, characterized in that... : The initial population size of the genetic algorithm is The maximum number of iterations is The crossover probability is The mutation probability is The early termination threshold is ; Prediction module output and Used for online updates of boundaries and risk assessments, driving adaptive adjustments of HOM and TTT; For each set of switching parameters Calculate the upper bound of the HOF / PP probability and construct the fitness function. The larger the fitness, the lower the switching risk.

Images