Indoor access point optimized placement method and system based on multi-modal large language model

Abstract

The application relates to an indoor access point optimization placement method and system based on a multimodal large language model, which comprises the following steps: analyzing a building plan by using a multimodal large language model to generate a current parent population containing S solutions; calculating the coverage value of each solution in the current parent population and the corresponding path loss graph, and storing them in a historical data set; inputting the access point position, coverage value and path loss graph of the current parent population into the multimodal large language model to perform mutation and crossover operations to generate a child population; selecting an optimal solution from the child population by using the multimodal large language model based on the access point position, coverage value and path loss graph in the historical data set; and fine-tuning the access point position by using the multimodal large language model according to the path loss graph of the optimal solution, updating the current parent population and iteratively optimizing until the termination condition is met. The method integrates numerical information and visual information, and improves the coverage performance and optimization efficiency of the access point placement.

Description

Technical Field

[0001] This invention relates to the field of wireless communication network optimization technology, specifically to a method and system for optimizing the placement of indoor access points based on a multimodal large language model. Background Technology

[0002] In wireless communication networks, the optimized placement of indoor access points is a crucial step in ensuring network coverage quality. Traditional wireless network optimization methods (such as convex optimization, heuristic algorithms, and reinforcement learning) mainly rely on numerical information for iterative calculations, making it difficult to effectively utilize multimodal information in the environment. Traditional methods cannot directly integrate such information, leading to optimization results that are prone to getting trapped in local optima, and they also incur high computational costs, especially the high-frequency calls to ray tracing models, which consume a significant amount of computational resources.

[0003] Current technologies do not fully exploit the value of visual information. Furthermore, traditional evolutionary algorithms suffer from high initialization randomness, a lack of semantic guidance in their breeding strategies, and reliance on extensive ray tracing calculations during the selection process, leading to low optimization efficiency. Therefore, effectively integrating multimodal information to improve the environmental awareness and computational efficiency of optimization algorithms has become a pressing issue. Summary of the Invention

[0004] This invention provides a method and system for optimizing indoor access point placement based on a multimodal large language model. By integrating multimodal information to guide the evolutionary algorithm, it improves the coverage performance and optimization efficiency of access point placement, and solves the problems of insufficient utilization of multimodal information and high computational complexity in traditional methods.

[0005] To achieve the above objectives, the first aspect of the present invention provides a method for optimizing the placement of indoor access points based on a multimodal large language model, comprising the following steps:

[0006] The building floor plan is analyzed using a multimodal large language model to generate an initial population containing S solutions, denoted as the current parent population. Each solution contains N access point locations.

[0007] The ray tracing model is used to calculate the coverage value and the corresponding path loss map of each solution in the current parent population, and the access point location, coverage value and path loss map are stored in the historical dataset.

[0008] Repeat the following steps until the termination condition is met:

[0009] Input the access point location, coverage value, and path loss map of the current parent population into the multimodal large language model, perform mutation and crossover operations, and generate the offspring population;

[0010] Calculate the coverage value and path loss graph of each solution in the offspring population and update the historical dataset; based on the information in the updated historical dataset, use a multimodal large language model to select the optimal solution set from the current parent population and offspring population;

[0011] Based on the path loss graph of the selected optimal solution set, the access point position is fine-tuned through a multimodal large language model to generate a new solution set; the new solution set is then updated to the current parent population for the next iteration.

[0012] Furthermore, when generating the initial population, the multimodal large language model generates S diverse initial solutions based on the dimensions of the building floor plan, wall distribution, and signal occlusion constraints. Each diverse initial solution includes multiple access point locations, which satisfy the requirements of being far from the walls and having a uniform spatial distribution. The access point locations of the diverse initial solutions are output in list form.

[0013] Furthermore, methods for performing mutation operations include:

[0014] By using a multimodal large language model to analyze the path loss map of the solution in the parent population, high loss regions with signal loss values ​​exceeding a preset threshold are identified.

[0015] Locate the access point position associated with the high-loss region;

[0016] Based on the location of the access point, select k associated access points and move their coordinates toward a low-loss region where the signal loss value is lower than a preset threshold to generate a variant solution.

[0017] Furthermore, methods for performing crossover operations include:

[0018] Two solutions are randomly selected from the solution set generated by the mutation operation;

[0019] The multimodal large language model is based on the coverage contribution matrix, and selects the dominant access point position with high coverage contribution from the two solutions respectively;

[0020] Swap the locations of the advantageous access points and verify that the swapped locations do not overlap within the building plane, generating a sub-solution that satisfies the spatial constraints.

[0021] Furthermore, methods for selecting the optimal solution from the offspring population using multimodal large language models include:

[0022] Analyze the distribution of access points, coverage values, and signal strength characteristics of path loss maps for each solution in the historical dataset to establish a mapping relationship between solution attributes and coverage performance.

[0023] Based on the mapping relationship, the coverage performance of each solution in the offspring population is predicted, and the solution with the best coverage performance is selected as the output.

[0024] Furthermore, methods for fine-tuning the access point location using a multimodal large language model include:

[0025] Analyze the signal strength distribution characteristics in the path loss map of the selected optimal solution to identify high signal loss regions and low signal loss regions.

[0026] The access point located in the high signal loss region is moved to the adjacent low signal loss region to generate a new, fine-tuned solution.

[0027] Furthermore, the termination condition is met when the number of iterations reaches a preset threshold or the coverage value converges, where the coverage value is defined as the proportion of grids in the target plane where the path loss is less than the preset threshold.

[0028] To achieve the above objectives, a second aspect of the present invention provides an indoor access point optimization placement system based on a multimodal large language model, the system comprising the following modules:

[0029] The initialization module is used to analyze building floor plans using a multimodal large language model and generate an initial population containing S solutions, denoted as the current parent population. Each solution contains N access point locations.

[0030] The data storage module is used to calculate the coverage value and its corresponding path loss map for each solution in the current parent population using the ray tracing model, and to store the access point location, coverage value and path loss map to the historical dataset;

[0031] The loop module is used to repeatedly execute the following steps until a termination condition is met:

[0032] The breeding module is used to input the access point location, coverage value, and path loss map of the current parent population into the multimodal large language model, perform mutation and crossover operations, and generate offspring populations.

[0033] The selection module is used to calculate the coverage value and path loss graph of each solution in the offspring population and update the historical dataset; based on the information in the updated historical dataset, the optimal solution set is selected from the current parent population and offspring population using a multimodal large language model;

[0034] The local search module is used to fine-tune the access point position of the selected optimal solution set based on the path loss graph of the multimodal large language model, and generate a new solution set; the new solution set is then updated to the current parent population for the next iteration.

[0035] Furthermore, the reproduction module integrates mutation and crossover operators for a multimodal large language model, wherein,

[0036] The mutation operator in the breeding module performs the following: analyzes the high-loss region in the path loss graph and locates the associated access points; moves at least k associated access points toward the low-loss region to generate a mutation solution.

[0037] The crossover operator in the breeding module performs the following steps: randomly selects two mutated solutions, selects the dominant access point in each solution based on the coverage contribution, swaps the positions of the dominant access points, and generates offspring solutions after verifying that there is no overlap within the building plane.

[0038] Furthermore, the selection module performs the following steps: establishing a mapping relationship between solution attributes and coverage performance in the historical dataset; predicting the coverage performance of solutions in the offspring population; and selecting the solution with the highest coverage value as the output.

[0039] The beneficial effects of this invention are:

[0040] Compared with existing technologies, the present invention provides an indoor access point optimization placement method and system based on multimodal large language model, which solves the core problems of insufficient utilization of multimodal information and low computational efficiency in traditional wireless network optimization by deeply integrating multimodal large language model (MLLM) and evolutionary algorithm (EA). Specifically, to address the shortcomings of traditional numerical optimization methods (such as convex optimization, heuristic algorithms, and reinforcement learning) in effectively integrating visual information (such as building floor plans and path loss maps), which can easily lead to local optima, this invention designs four types of MLLM enhancement operators: 1) In the initialization stage, MLLM directly parses the visual semantics of the building floor plan (such as wall distribution and spatial structure) to generate an initial solution with both high coverage potential and spatial diversity, avoiding the blindness of random initialization; 2) In the breeding stage, MLLM jointly analyzes numerical information (AP positions, historical coverage values) and the visual features of the path loss map (high-loss red areas, low-loss blue areas) to guide mutation operations to optimize weak coverage areas and crossover operations to fuse advantageous AP positions and avoid overlap, significantly improving the quality of solution generation; 3) In the selection stage, MLLM constructs an implicit mapping model of coverage performance based on historical datasets, only calling ray tracing calculations for the predicted optimal solution, reducing a significant amount of computational overhead compared to the traditional EA strategy that requires full population evaluation; 4) In the local search stage, MLLM fine-tunes AP positions based on the continuous gradient information of the original path loss map, accurately improving coverage performance. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0042] Figure 1 This is the overall flowchart of the MEvo access point algorithm.

[0043] Figure 2This is a diagram illustrating two contrasting scenarios for indoor access point placement, in which... Figure 2 (a) To ensure proper placement of the plan, Figure 2 (b) is a diagram of an unreasonable placement scheme.

[0044] Figure 3 This is a schematic diagram of the traditional evolutionary algorithm process.

[0045] Figure 4 This is a schematic diagram of the complete process of the MEvo access point algorithm.

[0046] Figure 5 These are two-dimensional schematic diagrams of eight test scenarios, among which... Figure 5 (a) is a planar schematic diagram of the TS1 test scenario. Figure 5 (b) is a planar schematic diagram of the TS2 test scenario. Figure 5 (c) is a planar schematic diagram of the TS3 test scenario. Figure 5 (d) is a planar schematic diagram of the TS4 test scenario. Figure 5 (e) is a planar schematic diagram of the TS5 test scenario. Figure 5 (f) is a planar schematic diagram of the TS6 test scenario. Figure 5 (g) is a planar schematic diagram of the TS7 test scenario. Figure 5 (h) is a planar schematic diagram of the TS8 test scenario.

[0047] Figure 6 This is a diagram illustrating the coverage values ​​of five algorithms across eight test scenarios.

[0048] Figure 7 This is a graph showing the convergence behavior of five algorithms in TS5 and TS8 scenarios, based on the number of coverage evaluation iterations. Figure 7 (a) is a graph showing the convergence behavior of the coverage evaluation based on the ray tracing model in the TS5 scenario. Figure 7 (b) is a graph showing the convergence behavior of the coverage evaluation times based on the ray tracing model in the TS8 scenario.

[0049] Figure 8 This is a schematic diagram of the path loss of five algorithms in the TS5 case, where Figure 8 (a)-(e) are path loss diagrams of five algorithms in TS5: MEvoAP, LMCO, Local search, DE and GA.

[0050] Figure 9 This is a schematic diagram of the path loss of five algorithms in the TS8 scenario, among which... Figure 9 (a)-(e) are path loss diagrams of five algorithms in TS8: MEvoAP, LMCO, Local search, DE and GA.

[0051] Figure 10 This is a diagram showing the coverage values ​​of the MEvo access point and its four variants in eight test scenarios.

[0052] Figure 11 This is a diagram illustrating the coverage values ​​of different multimodal large language models integrated into the MEvo access point across eight test scenarios. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0054] According to embodiments of the present invention, it should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the following manufacturing method, in some cases the steps shown or described may be performed in a different order than that shown here.

[0055] This embodiment relates to the field of wireless communication network optimization technology, specifically an indoor access point placement optimization method using a multimodal large language model-enhanced evolutionary algorithm, which is suitable for improving the coverage performance of indoor wireless communication networks.

[0056] Traditional wireless communication network optimization methods struggle to effectively integrate numerical and visual information when dealing with complex indoor environments, resulting in limitations in coverage uniformity and interference resistance. As the complexity of indoor wireless environments increases, efficiently utilizing multimodal information to guide access point placement has become a key challenge for improving network performance.

[0057] This implementation proposes an indoor access point optimization method based on a multimodal large language model, which guides access point placement optimization by fusing multimodal information. The method uses an evolutionary algorithm as the optimization framework, leveraging a multimodal large language model to process visual information such as floor plans and path loss maps, as well as numerical information such as access point location and coverage values. Four core operators provide intelligent guidance for the optimization process. The flowchart for MEvo access points is shown below. Figure 1 As shown, the MEvo access point refers to an indoor wireless access point (AP) location deployment scheme optimized by an evolutionary algorithm enhanced by a multimodal large language model, including the following steps S100 to S300:

[0058] Step S100: Analyze the building floor plan using a multimodal large language model to generate an initial population containing S solutions, denoted as the current parent population. Each solution contains N access point locations.

[0059] The initial population is generated using a multimodal large language model (MLLM). Specifically, a building floor plan (including wall location information) is input into the MLLM, which automatically generates an initial population containing S solutions based on its pre-trained wireless communication domain expertise and spatial reasoning capabilities. Let be the current parent population. Each solution consists of the two-dimensional coordinates of N access points (APs), formatted as a list. ,in, This indicates the horizontal coordinates of the access point (AP) on the building floor plan. This represents the vertical coordinates of the access point (AP) on the building floor plan. Finally, the current parent population will be obtained. .

[0060] During the generation process, MLLM must satisfy the following: rely solely on internal inference capabilities (not external optimization algorithms) to generate initial solutions with excellent coverage performance, i.e., the solutions must satisfy the principle of maximizing signal coverage; combine the wall distribution in the building plan to ensure that the AP location is far away from the walls to avoid signal obstruction, while also considering the physical feasible area constraints; the generated S solutions must be distributed and dispersed in the solution space to avoid clustering.

[0061] This process is guided by structured prompts in MLLM: the task description clearly defines the generation objective, the input information provides building structure data, the task requirements specify the number and distribution characteristics of solutions, and the output format mandates a uniform coordinate list. The final output is S sets of initial AP layout schemes that meet building constraints, have uniform spatial distribution, and possess high coverage potential.

[0062] Specifically, the multimodal large language model generates S diverse initial solutions from the building plan, each solution containing N access point coordinates. Ensure the initial population covers different spatial layout strategies. The mathematical expression is:

[0063]

[0064] Where F represents the architectural floor plan. Indicates the first An initial solution. , It is an initialization operator based on a multimodal large language model (MLLM). Its function is to take a building floor plan F as input and output a parent population containing S diverse initial solutions. .

[0065] Step S200: Use the ray tracing model to calculate the coverage value and the corresponding path loss map of each solution in the current parent population, and store the access point location, coverage value and path loss map to the historical dataset;

[0066] In this step, the coverage performance of the current parent population is evaluated and data is stored. The specific process is as follows:

[0067] Step S201: For each solution in the population (Representing N AP coordinates), the 3D ray tracing model (considering direct signal, primary reflection, and secondary reflection) is called to calculate coverage performance. For example, this model will target the plane (height) Discretized into R=1600 grids, using path loss values Determine the coverage status: If (threshold) The grid is then covered, and the final coverage value is... Defined as the proportion of the covered grid, its formula is:

[0068]

[0069] Where max is the objective function to maximize. This is the coverage performance value, which represents the proportion of meshes successfully covered in the target plane. For the first The coordinates of each AP, The horizontal coordinate is... Vertical coordinates The total number of access points. The total number of discretized grid cells. For grid indexing, For grid The path loss value is calculated using a 3D ray tracing model. To cover the judgment threshold, The feasible deployment area for the access point has been defined. This indicates an indicator function. If the condition is true, i.e. ,but ,otherwise , Indicates constraints.

[0070] Step S202, Synchronous Generation and Solution Corresponding path loss diagram The signal strength distribution is visualized using color gradients (e.g., red represents high loss and blue represents low loss).

[0071] Step S203: Combine the triplet The data is stored in the historical dataset to form a complete record containing AP locations, coverage performance, and spatial signal characteristics.

[0072] Step S300: Repeat the following steps until the termination condition is met:

[0073] Step S301: Input the access point location, coverage value and path loss map of the current parent population into the multimodal large language model, perform mutation and crossover operations, and generate the offspring population;

[0074] In this step, the current parent population The system takes a historical dataset (containing access point locations, coverage values, and path loss maps) as input. Mutation and crossover operations are performed. Mutation involves analyzing the path loss map and coverage value of each solution using a multimodal large language model to adjust the access point locations and generate new solutions. Crossover involves randomly selecting two mutated solutions, fusing the dominant access point locations, and generating a progeny population. This step will yield the offspring population. .

[0075] Step S302: Calculate the coverage value and path loss graph of each solution in the offspring population and update the historical dataset; based on the information in the updated historical dataset, use a multimodal large language model to select the optimal solution set from the current parent population and offspring population;

[0076] In the selection operator of this step, the input is the offspring population. Using historical datasets, a multimodal large language model was used to analyze the mapping relationship between the features and coverage values ​​of solutions in the historical datasets, starting from the offspring population. Select the most likely optimal solution from the options. Finally, the optimal candidate solution is output. .

[0077] Step S303: Based on the path loss map of the selected optimal solution set, fine-tune its access point position through a multimodal large language model to generate a new solution set; update the new solution set to the current parent population for the next iteration.

[0078] In the local search operator of this step, the input is the optimal candidate solution. and Its corresponding path loss diagram (as defined in step S202) (belonging to the same type of object), then the multimodal large language model is based on the path loss graph. Fine-tuning the optimal candidate solution The location of the access point is used to generate an optimized solution. Next, the relationship needs to be verified. If the coverage value of the optimized solution is greater than the coverage value of the solution before optimization, that is... So keep ,otherwise .

[0079] The output is the optimized solution. The data is then stored in the historical dataset. The current parent population is also updated, and if the optimized solution... The coverage value is better than that of the parent population. If the worst solution is found, then replace the worst solution; otherwise, replace the current parent population. It remains unchanged.

[0080] When the stopping condition is met (such as reaching the maximum number of coverage assessments), output the current parent population. The solution with the highest coverage value is selected as the optimal access point placement scheme.

[0081] In this embodiment, as described in step S301 above, a multimodal large language model (MLLM) is used to perform mutation and crossover operations on the parent population to generate the offspring population. This is specifically divided into two stages:

[0082] Mutation operation: Mutate each solution in the current parent population. and its coverage value Path loss diagram Input MLLM. MLLM analyzes signal distribution based on path loss maps (e.g., red high-loss areas, blue low-loss areas), identifies areas with weak coverage, and locates associated access points (APs). It then selects k key APs (those requiring weak signal improvement or to avoid strong signal overlap), adjusts their coordinates towards the low-loss areas, and generates... There are several variant solutions.

[0083] Crossover operation: Two solutions are randomly selected from the mutated solution set and input into the MLLM. The MLLM performs three steps:

[0084] Selection: Based on the coverage contribution and spatial distribution of APs, select an advantageous AP location from the two solutions;

[0085] Exchange: Swap the advantageous AP positions of both sides;

[0086] Verification: Check whether the new solution satisfies the no-overlap constraint within the building plane. If it violates the constraint, regenerate the solution.

[0087] Repeat this process until output is available. One valid child solution (format same as the mutated solution).

[0088] This process guides MLLM to integrate numerical information (AP coordinates, coverage values) and visual information (path loss map) through structured prompts, breaking through the limitation of traditional evolutionary algorithms that only use numerical data, and realizing multimodal-driven solution space exploration.

[0089] In this embodiment, as described in step S302 above, due to the high complexity of the ray tracing model, this model is used to evaluate the offspring population. Calculating the coverage value for each solution requires significant computation time. Therefore, this step uses a multimodal large language model (MLLM) based on limited historical data to derive the mapping relationship between coverage values ​​and access point locations, and then uses a ray tracing model to select the most promising coverage evaluation solutions.

[0090]

[0091] in, For selection operators based on Multimodal Large Language Model (MLLM), For offspring populations, This is a historical dataset.

[0092] Understandably, the specific process is as follows:

[0093] Step S302-1: Transfer historical datasets (Including AP locations, coverage values, and path loss maps of evaluated solutions) and offspring population ( Input the list of AP locations for each candidate solution into MLLM;

[0094] Step S302-2: MLLM analysis of complex patterns in historical data (such as the correlation between AP spatial distribution and red / blue areas in the path loss map) to establish the mapping relationship between solution attributes (such as the location of high loss areas and AP clustering degree) and coverage performance;

[0095] Step S302-3: Compare the attributes of candidate solutions through logical reasoning (non-statistical calculation), identify synergistic effects in historical patterns (e.g., AP layouts that are far from red high-loss areas and are evenly distributed indicate high coverage), and predict the potential coverage performance of each solution.

[0096] Step S302-4: Select the candidate solution with the best prediction performance. The output is a list of coordinates. Only the solution is verified for true coverage using a ray tracing model.

[0097] This step uses MLLM as the intelligent agent, replacing traditional traversal evaluation with reasoning based on multimodal historical data (numerical + visual), thus reducing the computational load of coverage calculations for each generation. Compression is reduced from once to once, significantly improving optimization efficiency.

[0098] In this embodiment, as described in step S303 above, to improve the utilization of MEvoAP, the present invention is based on the most promising solution in the offspring population (i.e. ) path loss map (i.e. To further improve its coverage performance, a local search operator based on a multilevel linear mixture model (MLLM) is designed, with the following specific steps:

[0099] Step S303-1: Select the optimal candidate solution from step S302. (Contains N AP coordinates) ) and its path loss diagram (In the diagram, red represents the high-loss area and blue represents the low-loss area.) Input MLLM and execute the formula:

[0100]

[0101] in: Indicates the MLLM local search operator; The solution is to be fine-tuned; It is its path loss diagram; To output the new solution after fine-tuning.

[0102] Step S303-2: MLLM identifies signal problem areas (such as the red high-loss area) based on the path loss map, and determines the AP location that needs adjustment through spatial reasoning (programming code disabled), for example:

[0103] Move the AP located in the high-loss area (red dense area) to the adjacent low-loss area (blue area);

[0104] Optimize AP spacing to avoid coverage overlap or blind spots;

[0105] Step S303-3: Output the new solution (Format same as input solution). Then evaluate its coverage value. Only when If necessary, retain the fine-tuning; otherwise, maintain the original solution (ensuring non-decreasing coverage performance).

[0106] This step utilizes the visual analysis capabilities of MLLM to transform the gradient information of the path loss map into a position adjustment strategy, breaking through the numerical limitations of traditional local search and achieving fine-grained optimization driven by signal features.

[0107] In the indoor access point optimization placement method based on a multimodal large language model, the termination condition for iterative optimization is set to one of the following two: the number of iterations reaches a preset threshold (such as 100 coverage value evaluations based on the ray tracing model as set in the experiment), or the coverage value converges (i.e., the improvement in coverage value of the optimal solution of the parent population in continuous iterations is less than the set tolerance); where the coverage value As the core optimization objective, it is defined as the path loss value in the target plane (discretized into R = 1600) grids. Not exceeding the preset threshold The grid ratio, which is evaluated by a computationally intensive 3D ray tracing model, quantifies the signal coverage performance of the access point layout (the closer the value is to 1, the better the coverage). The dual design of the termination condition ensures both computational feasibility (hard limit on the number of evaluations) and optimization efficiency (early termination when the coverage value is stable).

[0108] Combination Figure 2 As shown, Figure 2 This paper presents a comparison of two indoor access point placement schemes, aiming to intuitively reveal the limitations of traditional numerical optimization methods and highlight the key role of multimodal visual information in access point placement optimization. Figure 2 (a) A reasonable placement plan, Figure 2 (b) is an unreasonable placement scheme. The core difference between the two lies in the spatial distribution characteristics of the access points.

[0109] From a visual perspective, Figure 2 In (a), the access points are evenly distributed and far from the wall, while Figure 2 In (b), the access points are densely distributed and close to the walls. This difference is directly related to the coverage performance of the wireless signal: the walls, as obstacles to signal propagation, will significantly increase path loss, compress the signal coverage of access points close to the walls, and easily form coverage blind spots; while evenly distributed access points can more evenly cover the target area, reducing signal overlap and blind spots. These visually identifiable features (such as the distance between access points and walls, and the uniformity of distribution) are difficult to fully capture using traditional numerical methods, because traditional methods only rely on numerical information such as access point coordinates and coverage values, ignoring the constraints of visual information such as the wall distribution in the floor plan.

[0110] Should Figure 2 The core value lies in providing rational support for "introducing multimodal information." The MEvo access point algorithm proposed in this invention is based on this observation. By integrating floor plans and access point locations through a multimodal large language model, it guides access points away from walls and maintains a uniform distribution during the initialization phase, thereby avoiding... Figure 2 The shortcomings of the scheme in Figure (b). Furthermore, Figure 2 The implicit "correlation between spatial structure and signal coverage" is transformed into the reasoning basis of the multimodal large language model. For example, in the initialization prompt based on the multimodal large language model, it is explicitly required that the multimodal large language model constrain the access point location according to the wall distribution (black lines) in the floor plan. This is precisely the basis for... Figure 2 The algorithmic implementation of the rationality of the scheme in Figure (a) is shown.

[0111] In this embodiment, Figure 2It serves as a bridge connecting "problem identification" and "solution." It not only reveals the shortcomings of traditional methods—namely, the inability to use visual information to determine the rationality of access point distribution—but also provides specific optimization directions for the design of MEvo access points: by leveraging the visual understanding capabilities of a multimodal large language model, observable optimization principles in the graph can be transformed into executable algorithmic rules, ultimately improving coverage performance.

[0112] like Figure 3 As shown, in this embodiment, Figure 3 The flowchart for traditional evolutionary algorithms (EAs) illustrates the iterative framework of initialization, evaluation, reproduction, selection, and updating.

[0113] flow chart Figure 3 Starting from the beginning, the core steps are as follows: initializing the parent population, evaluating the fitness of each solution in the parent population, generating the offspring population, evaluating the fitness of the offspring population, updating the parent population, until the stopping condition is met and the optimal solution is output. This process reflects the essential characteristics of EAs, namely, achieving global optimization by simulating the iterative process of natural selection and genetic variation. Its core advantage lies in its independence from gradient information, making it suitable for complex nonlinear and nonconvex optimization problems. This is also an important reason why EA was chosen as the optimization framework in this embodiment.

[0114] In detail, each step of traditional Expert Analyzers (EAs) relies on numerical information: the initialization phase constructs an initial solution by randomly generating numerical coordinates; the propagation phase generates new solutions based on numerical perturbations; and the selection phase filters high-quality solutions by comparing numerical fitness values. This singular reliance on numerical information prevents it from utilizing visual information such as floor plans and path loss maps, limiting its performance in complex scenarios such as indoor access point placement. Traditional EAs may generate access point locations close to walls during initialization, and the numerical perturbations in their propagation phase are also difficult to specifically avoid signal obstruction by walls.

[0115] Figure 3 and subsequent Figure 4 (The MEvo access point flowchart) presents a stark contrast, highlighting the improved logic of the MEvo access point: The MEvo access point retains the iterative framework of EAs, but introduces a multimodal large language model to process multimodal information in key steps. For example, it replaces the random initialization of traditional EAs with initialization based on floor plans using a multimodal large language model, and replaces numerical mutations with mutations based on path loss maps using a multimodal large language model. This improvement allows the algorithm to leverage structural constraints and signal features in visual information, thereby overcoming the performance bottleneck of traditional EAs.

[0116] This embodiment clarifies the intervention point and function of the multimodal large language model operator by comparing it with the traditional EA process, proving that the MEvo access point is not a complete negation of EA, but rather achieves a performance leap by integrating multimodal information while inheriting its global optimization advantages.

[0117] like Figure 4 As shown, in this embodiment, Figure 4 The core flowchart of the MEvo access point systematically demonstrates the fusion mechanism of the multimodal large language model and EA, serving as the core visual carrier for understanding the working principle of the MEvo access point.

[0118] flow chart Figure 4 Starting from the beginning, the process iterates through initialization, reproduction, selection, local search, and population update, ultimately terminating with a stopping condition and outputting the optimal solution. This differs from traditional Expert Advisor (EA) flowcharts. Figure 3 In comparison, its core innovation lies in deeply embedding the multimodal large language model into the four key links of EA, forming a closed-loop collaboration between the "multimodal large language model-based operator" and the EA framework, while clarifying the integration path of multimodal information.

[0119] From the perspective of module composition, Figure 4 The above labels indicate the operational stages of the four major multimodal large language model operators:

[0120] Multimodal large language model-based initialization: Instead of the traditional random initialization of EA, an initial population is generated based on the floor plan to ensure that the access points are far away from the walls and are evenly distributed;

[0121] Mutation and crossover operators based on multimodal large language model: During the breeding phase, the multimodal large language model analyzes the access point location and path loss map, and adjusts the access point location accordingly to reduce signal blind spots;

[0122] Selection operator based on multimodal large language model: Utilize historical data (numerical + visual) to predict the coverage performance of successor solutions, and only perform ray tracing evaluation on the best candidate solution to reduce computational cost;

[0123] Local search operator based on multimodal large language model: fine-tuning access point location based on path loss graph to further optimize coverage details.

[0124] Figure 4 The lower section clarifies the input types of multimodal information: numerical information includes access point locations and coverage values, while visual information includes floor plans and path loss maps. This information is uniformly processed by the multimodal large language model and transformed into the basis for optimization decisions, achieving cross-modal semantic alignment between visual and numerical information. By analyzing the red high-loss areas in the path loss map and combining this with the access point coordinates, the multimodal large language model infers the access point locations that need adjustment, thereby generating new optimization solutions.

[0125] From an iterative logic perspective, MEvo's access point process embodies a balance between global and local exploration: the multimodal large language model initialization and reproduction operators ensure population diversity, while the selection and local search operators focus on optimizing high-quality solutions. This balance allows it to escape local optima within a limited number of iterations while finely optimizing coverage performance, which is the core reason why its convergence speed and coverage value are superior to traditional algorithms in experiments.

[0126] Figure 4 The technological value lies in its intuitive presentation of the innovative approach of multimodal information-driven optimization: through the joint understanding of visual and numerical information by a multimodal large language model, MEvo access points break through the limitations of traditional EAs that rely solely on numerical information, making access point placement optimization more in line with the physical characteristics of the indoor environment, and providing a new paradigm for the optimization of complex wireless scenarios.

[0127] like Figure 5 As shown, where Figure 5 (a) is a planar schematic diagram of the TS1 test scenario. Figure 5 (b) is a planar schematic diagram of the TS2 test scenario. Figure 5 (c) is a planar schematic diagram of the TS3 test scenario. Figure 5 (d) is a planar schematic diagram of the TS4 test scenario. Figure 5 (e) is a planar schematic diagram of the TS5 test scenario. Figure 5 (f) is a planar schematic diagram of the TS6 test scenario. Figure 5 (g) is a planar schematic diagram of the TS7 test scenario. Figure 5 (h) is a planar schematic diagram of the TS8 test scenario.

[0128] In this embodiment, Figure 5 The floor plans of eight test scenarios (TS1-TS8) are presented, serving as the core experimental data carrier for verifying the generalization capability of MEvo access points. The scenario design covers indoor environments of varying sizes, structural complexities, and functional types. See Table 1:

[0129] Table 1. Setup of the eight test scenarios

[0130]

[0131] In terms of scenario characteristics, the eight test scenarios exhibit the following significant diversity:

[0132] Size differences: The area ranges from 25.0m × 23.2m for TS4 to 50.0m × 30.0m for TS8, covering small offices to large canteens;

[0133] Structural complexity: The student hall in TS1 is a regular rectangle with few walls; the museum in TS3 contains multiple irregular partitions; the automobile factory in TS6 has dense equipment areas and a complex signal propagation environment.

[0134] Access point quantity requirement: Depending on the size of the scenario, the number of access points N ranges from 4 in TS1-TS4 to 6 in TS5-TS6 and TS8, and then to 5 in TS7 (store), reflecting the different coverage requirements of different scenarios.

[0135] These scenarios were not randomly selected, but rather designed to comprehensively test the adaptability of MEvo access points in simple and complex, small and large, and regular and irregular environments. Traditional algorithms may perform reasonably well in regular scenarios like TS1, but in complex scenarios such as TS3 with many partitions, the inability to use visual information to identify partition locations can easily lead to access point placement falling into blind spots, resulting in a sharp drop in coverage performance.

[0136] From the perspective of experimental logic, this Figure 5 This provides a foundation for the generalization verification of MEvo access points. MEvo access points use a multimodal large language model to visually understand the floor plan of each scenario and dynamically adjust optimization strategies: in regular scenarios, they focus on the even distribution of access points; in complex scenarios, they prioritize avoiding signal obstruction by barriers; and in large scenarios, they emphasize the rationality of access point spacing.

[0137] Furthermore, the diversity of scenarios also provided support for ablation experiments. By comparing the performance differences of MEvo access points and variants of the multimodal large language model removal operator in different scenarios, the role of each operator in complex environments can be clarified.

[0138] like Figure 6 As shown, in this embodiment, Figure 6 The average coverage of the MEvo access point was compared with that of four benchmark algorithms (GA, DE, LocalSearch, and LMCO) in eight test scenarios, which is the core experimental result demonstrating the performance advantage of the MEvo access point.

[0139] Data trends show that MEvo access points perform best across all scenarios, and their core advantages are reflected in the following aspects:

[0140] Absolute coverage value: MEvo access points have the highest coverage value in all scenarios, leading GA by an average of about 15% and LMCO by about 5%-8%;

[0141] Scene adaptability: MEvo access points show the smallest performance drop in simple scenarios (such as TS1) and complex scenarios (such as TS3), while GA and DE show significant performance degradation in complex scenarios.

[0142] Advantages of LMCO: As the first access point placement method to use LLM, LMCO only utilizes LLM in the initialization phase, while MEvo access point achieves superior performance in all scenarios through full-process optimization of four multimodal large language model operators, proving the necessity of multi-stage multimodal optimization.

[0143] The technical interpretation of these results can be combined with the mechanisms of the MEvo access point: multimodal large language model-based initialization ensures the quality of the initial solution; the reproduction operator generates high-quality offspring through multimodal information; the selection operator reduces invalid computation; and the local search operator optimizes details. The synergy of these four mechanisms enables the MEvo access point to efficiently utilize environmental information in different scenarios, while traditional algorithms, due to their reliance on fixed numerical models, cannot adapt to different scenarios.

[0144] The experimental results directly support the core argument of this embodiment: the integration of multimodal information can significantly improve wireless optimization performance. The advantage of the MEvo access point does not stem from a single operator, but from the synergy of the four multimodal large language model operators and the EA framework, which provides quantitative evidence for the effectiveness of "multimodal optimization".

[0145] like Figure 7 As shown, Figure 7 The convergence behavior of five algorithms based on ray tracing models in TS5 and TS8 scenarios is presented for coverage evaluation iterations. Figure 7 (a) Convergence behavior of the coverage evaluation number based on the ray tracing model in the TS5 scenario. Figure 7 (b) For the convergence behavior of coverage evaluation iterations based on the ray tracing model in the TS8 scenario, MEvoAP starts at a high performance point in both TS5 and TS8, exhibiting the fastest convergence speed and maintaining its advantage throughout the optimization process. LMCO starts at a slightly lower point than MEvoAP but shows a strong convergence trend. Local search has the lowest starting point among the algorithms, exhibiting only moderate convergence ability. The other two algorithms (DE and GA) show significantly slower convergence speeds.

[0146] like Figure 8 and Figure 9 As shown, in this embodiment, Figure 8 (a)-(e) are path loss diagrams for five algorithms—MEvoAP, LMCO, Local search, DE, and GA—in TS5 (Office). Figure 9 (a)-(e) are path loss diagrams of five algorithms, MEvoAP, LMCO, Localsearch, DE and GA, in TS8 (canteen). The visual differences in signal distribution further verify the coverage performance advantage of MEvo access points.

[0147] The core comparative features of path loss maps are as follows:

[0148] MEvo access points: The blue area (low loss) accounts for the largest proportion, while the red area (high loss) is sparsely distributed and mainly located at the edge of the scene, with good signal coverage uniformity;

[0149] LMCO: The blue area accounts for the second largest proportion, but there are local red areas, indicating that its advantages in the initial stage have not been fully amplified through subsequent optimization;

[0150] Local search: The blue areas are concentrated in local regions and are unevenly distributed overall, reflecting its tendency to get trapped in local optima;

[0151] GA / DE: The high proportion and scattered distribution of red areas indicate that the access point locations generated by traditional EA do not effectively avoid the influence of walls and partitions.

[0152] The technical reason for these differences lies in the degree of utilization of multimodal information: MEvo access points continuously analyze the correlation between path loss maps and access point locations through multimodal large language models, and dynamically adjust strategies to expand the blue area; while GA / DE only relies on numerical information, cannot identify the causes of the red area, and the optimization direction is blind.

[0153] From a signal propagation perspective, the path loss map of the MEvo access point shows that its location effectively utilizes the reflection path. In the TS5 office scenario, the access point avoids dense partitions, allowing the signal to cover most areas with minimal reflections, reducing losses caused by multipath effects. In contrast, the GA access point is located close to partitions, requiring the signal propagation path to undergo multiple reflections or diffractions, resulting in a significant increase in path loss (e.g., ...). Figure 8 (b), as shown in the densely red areas in 9(b). Especially in the densely partitioned areas of the TS5 office ( Figure 5 (e) Because GA-deployed APs cannot perceive wall structures, their locations are concentrated at the edges of spaces or in blind spots of partitions, causing the signal energy to attenuate sharply when penetrating multiple walls (path loss often exceeds the threshold θ=70dB), forming large areas of uncovered blind spots. In contrast, MEvoAP uses a multimodal large language model to analyze the path loss map (visual information) in real time, actively avoiding partition occlusion areas during the optimization process: its MLLM-based breeding operator can identify high-loss red areas in the path loss map, guiding the AP position to migrate to low-loss blue areas; the MLLM-based local search operator makes micron-level adjustments to the AP position based on the path loss gradient to optimize the signal reflection path. For example, in the TS8 cafeteria ( Figure 9(a) MEvoAP deploys APs in the central area and away from load-bearing walls, utilizing open space to reduce the number of reflections, increasing the proportion of blue low-loss areas by 32.7% (compared to GA), directly demonstrating the coverage advantage.

[0154] The root cause of this performance difference lies in the collaborative utilization mechanism of multimodal information: MEvoAP's MLLM establishes an implicit mapping relationship between AP location coordinates (numerical information) and path loss distribution map (visual information) through end-to-end learning, enabling the optimization process to have spatial structure cognition capabilities; while traditional GA only relies on random mutation / crossing of coordinate values ​​and cannot understand physical constraints such as wall location and material properties, causing AP clusters to get stuck in structural local optima in complex scenarios.

[0155] The analysis of the above experimental results is as follows: Differential Evolution (DE) and Genetic Algorithm (GA) are both population-based evolutionary algorithms, requiring coverage evaluation of a batch of new solutions in each iteration. Since the number of coverage evaluations is limited, the number of iterations may constrain their performance. In contrast, local search methods are single-point based algorithms, generating only one initial solution, which is prone to getting trapped in local optima. Their overall performance is limited by the quality of the initial solution. LLMO, by utilizing a Language Model (LLM) to generate initial solutions, can generate high-quality initial solutions more efficiently than traditional algorithms, thereby improving overall search performance.

[0156] Furthermore, MEvoAP combines population-based EA with MLLM. By introducing selection operators based on multilevel linear mixture models (MLLM), MEvoAP requires fewer coverage evaluations per iteration than genetic algorithms (DE) and genetic algorithms (GA), thus significantly improving search efficiency. In addition, MEvoAP employs other MLLM-based operators to further enhance global search performance.

[0157] To verify the effectiveness of each MLLM-based operator in MEvoAP, ablation experiments were conducted. As shown in Table 2, four MEvoAP variants (labeled MEvoAP1 to MEvoAP4) were constructed, each of which removed one MLLM-based operator.

[0158] Table 2. Setup of the ablation experiment

[0159]

[0160] Specifically:

[0161] 1) MEvoAP1 uses random initialization instead of MLLM-based initialization;

[0162] 2) MEvoAP2 replaces the MLLM-based reproduction operator with the traditional evolutionary algorithm EA reproduction operator;

[0163] 3) MEvoAP3 uses a ray tracing model to evaluate the coverage values ​​of all child solutions;

[0164] 4) MEvoAP4 is a local search operator for random solutions.

[0165] Figure 10 The average coverage of MEvoAP and its four variants is shown in eight test scenarios. Figure 10 As shown, MEvoAP outperforms other algorithms in all test scenarios, demonstrating the synergistic benefits of the four MLLM-based operators. In contrast, MEvoAP1 and MEvoAP3 significantly lag behind MEvoAP, highlighting the crucial role of the MLLM-based initialization and selection operators. Furthermore, although MEvoAP2 and MEvoAP4 outperform MEvoAP1 and MEvoAP3, their performance is still inferior to MEvoAP, indicating that the MLLM-based breeding and local search operators also contribute to the overall performance of MEvoAP.

[0166] Compare the performance of various MLLMs integrated into MEvoAP, including GPT-4o, Claude 3.5, and qwen-VL-Max. Figure 11 Coverage values ​​of three MLLMs under the MEvoAP framework in eight test scenarios are presented. Results show that GPT-4o outperforms the other MLLMs in six out of eight scenarios, while Claude 3.5 achieves the highest coverage in two scenarios. In contrast, qwen-VL-Max does not outperform the other MLLMs in any scenario.

[0167] In summary, this invention addresses the problem that traditional wireless network optimization methods struggle to effectively utilize multimodal information (especially visual information), leading to local optima and low computational efficiency. It proposes an indoor access point (AP) optimization placement method and system based on the Multimodal Large Language Model (MLLM) Enhanced Evolutionary Algorithm (EA). The core invention lies in the innovative design of four types of MLLM enhancement operators through deep integration of the MLLM and EA frameworks: 1) MLLM-based initialization: Utilizing MLLM to directly parse the visual semantics of building floor plans (such as wall distribution and spatial structure) to generate initial solutions with both high coverage potential and spatial diversity, avoiding the blindness of random initialization; 2) MLLM-based breeding operator: Jointly analyzing numerical information (AP location, historical coverage values) and visual features of path loss maps (such as high-loss red areas and low-loss blue areas) to guide mutation operations for targeted optimization of weak coverage areas (locating high-loss areas, associating APs, and moving them to low-loss areas). 1) Cross-operation merges advantageous AP locations and avoids overlap (selecting swap points based on coverage contribution and verifying no overlap), significantly improving solution quality; 2) MLLM-based selection operator: constructs an implicit mapping model of coverage performance based on historical datasets (AP locations, coverage values, and path loss maps), only calling computationally expensive ray tracing evaluations for the predicted optimal solution, significantly reducing computational load compared to the traditional EA strategy that requires full population evaluation; 3) MLLM-based local search operator: fine-tunes AP locations based on continuous gradient information (non-binarized simplified map) of the original path loss map (e.g., moving APs in high-loss areas to adjacent low-loss areas), accurately improving coverage performance. Through the synergy of the above four operators, end-to-end integration of numerical information (AP locations, coverage values) and visual information (building floor plans, path loss maps) is achieved, breaking through the limitation of traditional optimization methods that rely solely on numerical data, significantly optimizing computational efficiency while improving coverage performance (the selection operator reduces the number of ray tracing evaluations). Experimental results demonstrate that MEvoAP outperforms the four baseline algorithms in eight types of indoor scenarios, thus verifying that multimodal information can effectively improve wireless network optimization performance.

[0168] According to another aspect of the embodiments of this application, an indoor access point optimization placement system based on a multimodal large language model is also provided. The system part of the technical solution of the present invention is as follows:

[0169] Initialization module: Input architectural floor plan, call multimodal large language model to generate There are N initial solutions (each solution represents the location of N access points), and the coordinates of the access points in each solution satisfy... , The spatial constraints are defined, and the initial layout is displayed through a visual interface. The coverage value and path loss map for each solution are obtained using ray tracing software.

[0170] The breeding module receives the access point locations, coverage values, and path loss maps of the current parent population and performs mutation and crossover operations using a multimodal large language model. During mutation, the module inputs the path loss map into the multimodal large language model and parses the access point indices corresponding to high-loss regions. During crossover, it generates an exchange strategy based on the coverage contribution matrix to ensure the diversity of offspring solutions.

[0171] The selection module: The multimodal large language model predicts coverage using historical data. The model adopts a Transformer architecture, and the access point coordinate sequence and path loss map features are encoded and input into the fully connected layer to predict the coverage value. The module selects the solution with the highest predicted coverage value from the offspring population and calls the ray tracing module for verification.

[0172] Local search module: Based on the semantic analysis of the path loss map using a multimodal large language model, it generates access point adjustment vectors with a step size that is an integer multiple of 0.5 meters. If the coverage value is improved after adjustment, the new solution is retained.

[0173] Data storage module: It uses an SQLite database to store historical solutions, coverage values ​​and path loss maps, and supports data retrieval by timestamp and scene type, providing data support for incremental learning of multimodal large language models.

[0174] This invention improves coverage performance, optimizes computational efficiency, and verifies robustness. In multiple test scenarios, the average coverage value of the method described in this invention is 12.3%-27.5% higher than that of the Genetic Algorithm (GA). For example, in the TS8 scenario (50m×30m cafeteria, 6 access points), the coverage value of the method described in this invention reaches 91.5%, which is 5.8% higher than the LMCO algorithm and 19.3% higher than GA.

[0175] The selection operator based on the multimodal large language model reduces the number of ray tracing calls from S times per iteration to 1 time, shortening the computation time by more than 60%. In the TS5 scenario (42m×32.3m office), the method of this invention converged to 85.7% coverage after 50 evaluations, while GA still only reached 72.1% after 100 evaluations.

[0176] Ablation experiments demonstrate that the synergistic effect of the four multimodal large language model operators is crucial: removing the initialization operator reduces the coverage value by 15.2%, and removing the selection operator increases the computation time by 2.3 times.

[0177] According to another aspect of the embodiments of this application, an electronic device is also provided, including a processor and a memory, wherein the processor is configured to implement the steps of the method when executing a computer program stored in the memory.

[0178] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0179] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0180] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0182] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing the placement of indoor access points based on a multimodal large language model, characterized in that, Includes the following steps: The building floor plan is analyzed using a multimodal large language model to generate an initial population containing S solutions, denoted as the current parent population. Each solution contains N access point locations. The ray tracing model is used to calculate the coverage value and the corresponding path loss map of each solution in the current parent population, and the access point location, coverage value and path loss map are stored in the historical dataset. Repeat the following steps until the termination condition is met: Input the access point location, coverage value, and path loss map of the current parent population into the multimodal large language model, perform mutation and crossover operations, and generate the offspring population; Calculate the coverage value and path loss graph of each solution in the offspring population and update the historical dataset; based on the information in the updated historical dataset, use a multimodal large language model to select the optimal solution set from the current parent population and offspring population; Based on the path loss graph of the selected optimal solution set, the access point position is fine-tuned through a multimodal large language model to generate a new solution set; the new solution set is then updated to the current parent population for the next iteration. When generating the initial population, the multimodal large language model generates S diverse initial solutions based on the dimensions of the building floor plan, the distribution of walls, and signal obstruction constraints. Each diverse initial solution includes multiple access point locations, which satisfy the requirements of being far from the walls and having a uniform spatial distribution. Output the access point locations of the diverse initial solutions in list form; Methods for performing mutation operations include: By using a multimodal large language model to analyze the path loss map of the solution in the parent population, high loss regions with signal loss values ​​exceeding a preset threshold are identified. Locate the access point position associated with the high-loss region; Based on the location of the access point, select k associated access points and move their coordinates toward a low-loss region where the signal loss value is lower than a preset threshold to generate a mutation solution. Methods for performing crossover operations include: Two solutions are randomly selected from the solution set generated by the mutation operation; The multimodal large language model is based on the coverage contribution matrix, and selects the dominant access point position with high coverage contribution from the two solutions respectively; Swap the locations of the advantageous access points and verify that the swapped locations do not overlap within the building plane, generating a sub-solution that satisfies the spatial constraints.

2. The indoor access point optimization placement method based on a multimodal large language model as described in claim 1, characterized in that, Methods for selecting the optimal solution from the offspring population using multimodal large language models include: Analyze the distribution of access points, coverage values, and signal strength characteristics of path loss maps for each solution in the historical dataset to establish a mapping relationship between solution attributes and coverage performance. Based on the mapping relationship, the coverage performance of each solution in the offspring population is predicted, and the solution with the best coverage performance is selected as the output.

3. The indoor access point optimization placement method based on a multimodal large language model as described in claim 1, characterized in that, Methods for fine-tuning access point locations using multimodal large language models include: Analyze the signal strength distribution characteristics in the path loss map of the selected optimal solution to identify high signal loss regions and low signal loss regions. The access point located in the high signal loss region is moved to the adjacent low signal loss region to generate a new, fine-tuned solution.

4. The indoor access point optimization placement method based on a multimodal large language model as described in claim 1, characterized in that, The termination condition is met when the number of iterations reaches a preset threshold or the coverage value converges, where the coverage value is defined as the proportion of grids in the target plane where the path loss is less than the preset threshold.

5. An indoor access point optimization placement system based on a multimodal large language model, characterized in that, The system includes the following modules: The initialization module is used to analyze building floor plans using a multimodal large language model and generate an initial population containing S solutions, denoted as the current parent population. Each solution contains N access point locations. The data storage module is used to calculate the coverage value and its corresponding path loss map for each solution in the current parent population using the ray tracing model, and to store the access point location, coverage value and path loss map to the historical dataset; The loop module is used to repeatedly execute the following steps until a termination condition is met: The breeding module is used to input the access point location, coverage value, and path loss map of the current parent population into the multimodal large language model, perform mutation and crossover operations, and generate offspring populations. The selection module is used to calculate the coverage value and path loss graph of each solution in the offspring population and update the historical dataset; based on the information in the updated historical dataset, the optimal solution set is selected from the current parent population and offspring population using a multimodal large language model; The local search module is used to fine-tune the access point position of the selected optimal solution set based on the path loss graph using a multimodal large language model, and generate a new solution set; the new solution set is then updated to the current parent population for the next iteration. When generating the initial population, the multimodal large language model generates S diverse initial solutions based on the dimensions of the building floor plan, the distribution of walls, and signal obstruction constraints. Each diverse initial solution includes multiple access point locations, which satisfy the requirements of being far from the walls and having a uniform spatial distribution. Output the access point locations of the diverse initial solutions in list form; Methods for performing mutation operations include: By using a multimodal large language model to analyze the path loss map of the solution in the parent population, high loss regions with signal loss values ​​exceeding a preset threshold are identified. Locate the access point position associated with the high-loss region; Based on the location of the access point, select k associated access points and move their coordinates toward a low-loss region where the signal loss value is lower than a preset threshold to generate a mutation solution. Methods for performing crossover operations include: Two solutions are randomly selected from the solution set generated by the mutation operation; The multimodal large language model is based on the coverage contribution matrix, and selects the dominant access point position with high coverage contribution from the two solutions respectively; Swap the locations of the advantageous access points and verify that the swapped locations do not overlap within the building plane, generating a sub-solution that satisfies the spatial constraints.

6. The indoor access point optimization placement system based on a multimodal large language model as described in claim 5, characterized in that, The selection module performs the following steps: establishing a mapping relationship between solution attributes and coverage performance in the historical dataset; predicting the coverage performance of solutions in the offspring population; and selecting the solution with the highest coverage value as the output.

Images