An indoor networking planning method and device based on visual dynamic simulation

Abstract

The application relates to the technical field of communication network planning and design, and specifically provides an indoor networking planning method and device based on visual dynamic simulation, which comprises the following steps: S1, a house type grid image of a target indoor space is acquired, pixel-level semantic segmentation and space topology reconstruction are performed through a hierarchical visual large model, and structured engineering parameters containing wall categories, space connectivity relationships and corresponding frequency band signal attenuation parameters are output; S2, based on the structured engineering parameters, similar historical cases are retrieved in combination with a pre-constructed trusted networking knowledge graph, and at least one wireless access point deployment candidate scheme is generated by using a multi-objective optimization algorithm; S3, the candidate scheme is input into a physical simulation engine for electromagnetic field distribution calculation, and the calculation result is input into a visual cognitive model to generate a visualized coverage evaluation report; a reinforcement learning intelligent agent automatically adjusts deployment parameters according to the evaluation report, and the physical simulation and visual cognitive process are repeated until a preset convergence condition is met; and S4, a final optimized deployment point and corresponding signal coverage visualized result are output. Compared with the prior art, the application can automatically understand complex indoor structures, efficiently utilize historical data and integrate simulation and optimization.

Description

Technical Field

[0001] This invention relates to the field of communication network planning and design technology, and specifically provides an indoor network planning method and device based on visual dynamic simulation. Background Technology

[0002] Currently, operators generally adopt a sequential process of "engineer on-site survey + experience-based judgment + commercial simulation tool verification" in planning indoor wireless networks for commercial customers. This model has three major technical drawbacks: (1) Spatial information acquisition is costly and has a very low degree of structure. Engineers must carry equipment such as laser rangefinders and signal scanners to conduct on-site surveys, manually marking wall materials and obstacle locations on paper or CAD floor plans, which takes an average of over 2 hours per project. Existing image processing tools (such as OpenCV edge detection) can only extract simple lines and cannot distinguish materials with different radio frequency penetration characteristics, such as reinforced concrete walls, lightweight brick walls, and glass curtain walls, nor can they generate structured engineering files (such as wall attenuation coefficients and dielectric constants) for electromagnetic simulation. This bottleneck leads to a heavy reliance on manual secondary processing for subsequent simulation modeling, resulting in significant error accumulation.

[0003] (2) Historical successful cases cannot be reused on a large scale; The operator's archives contain hundreds of thousands of delivered network deployment plans, but these are all unstructured PDF sketches or CAD drawings. Device coordinates are not aligned with the actual physical space, and corresponding measured performance data is lacking. Existing technologies (such as rule-based expert systems) can only generate coarse-grained plans based on preset conditions (such as "deploy one AP per 50 square meters"), failing to automatically extract scenario-based, high-precision deployment strategies from massive amounts of heterogeneous historical data. This results in new project planning still heavily relying on a few top-tier engineers, leading to significant fluctuations in plan quality and an inability to support rapid response during the pre-sales phase.

[0004] (3) Simulation verification and planning decision-making are completely separated. Existing commercial electromagnetic simulation software requires several minutes to tens of minutes for a single high-precision simulation, and necessitates repeated manual parameter configuration. It is only used as a "compliance verification" tool after design completion and cannot be deeply integrated into the planning and iteration process. In the pre-sales stage, in pursuit of speed, detailed simulations are often abandoned, relying solely on experience-based estimations, directly resulting in a first-time installation compliance rate consistently hovering below 85%. Furthermore, simulation results are output in the form of dense numerical matrices, which are difficult for non-professionals to interpret and cannot directly assist marketing decisions.

[0005] In summary, the industry urgently needs an intelligent networking planning method that can automatically understand complex indoor structures, efficiently utilize historical data, and integrate simulation and optimization. Summary of the Invention

[0006] This invention addresses the shortcomings of the prior art by providing a highly practical indoor network planning method based on visual dynamic simulation.

[0007] A further technical objective of this invention is to provide a reasonably designed, safe, and applicable indoor networking planning device based on visual dynamic simulation.

[0008] The technical solution adopted by this invention to solve its technical problem is: An indoor network planning method based on visual dynamic simulation has the following steps: S1, Visual Analysis; The system acquires a grid image of the target indoor space, performs pixel-level semantic segmentation and spatial topology reconstruction using a hierarchical visual large model, and outputs structured engineering parameters including wall type, spatial connectivity, and corresponding frequency band signal attenuation parameters. S2, Solution Generation; Based on the structured engineering parameters, and combined with the pre-built trusted networking knowledge graph to retrieve similar historical cases, a multi-objective optimization algorithm is used to generate at least one candidate solution for wireless access point deployment. S3, Dual-loop simulation optimization; The candidate schemes are input into the physical simulation engine to calculate the electromagnetic field distribution, and the calculation results are input into the visual cognition model to generate a visual coverage evaluation report. The reinforcement learning agent automatically adjusts the deployment parameters according to the evaluation report and repeats the physical simulation and visual cognition process until the preset convergence condition is met. S4. Output the final optimized deployment locations and the corresponding signal coverage visualization results.

[0009] Furthermore, in the hierarchical visual large model described in step S1, an encoder-decoder segmentation network based on a hybrid architecture of Transformer and CNN is used to identify indoor elements and output a pixel-level classification mask; a topology reconstruction network based on a graph neural network is used to extract the polygonal boundary coordinates of the room and the connectivity width of doors / windows from the classification mask to generate a spatial connectivity weighted graph.

[0010] Furthermore, the construction of the trusted networking knowledge graph in step S2 is as follows: The system collects floor plans, equipment deployment coordinates, and measured performance data of historical networking schemes, and maps the scheme sketches to a standard coordinate system using a spatial registration algorithm. A pre-trained deep neural network performance evaluator is used to score the quality of historical schemes, filtering out low-quality schemes. For the retained high-quality schemes, a fixed-dimensional networking strategy feature vector is generated using a cross-modal Transformer encoder. Using the strategy vector, scene labels, equipment models, and performance indicators as node attributes, a directed knowledge graph is constructed that includes similarity relationships between cases, relationships between cases and equipment adoption, and performance superiority / inferiority relationships between cases.

[0011] Furthermore, the multi-objective optimization algorithm employs a two-stage cascade strategy: Phase 1: Using the fast propagation model as the evaluation function, the NSGA-III algorithm is used to perform a global Pareto front search on the number of devices, coverage, and robustness. The population initialization step retrieves the knowledge graph based on the visual features of the current apartment type and calls similar case deployment schemes to generate the initial chromosome. The second stage involves selecting a preferred scheme from the Pareto front solution set, fixing the number and model of equipment, using the particle swarm optimization algorithm to perform local fine optimization only on the equipment coordinates and transmission power, and calculating the fitness value using high-fidelity simulation results from the finite-difference time-domain method.

[0012] Furthermore, in step S3, the physical simulation engine adopts a GPU-accelerated parallel time-domain finite difference solver, with the structured engineering parameters and deployment parameters as inputs and the electric field intensity amplitude of three-dimensional discrete grid points as outputs. The visual cognition model includes: a conditional generative adversarial network, used to render discrete field strength data into a smooth, gradient color heatmap; and an AI evaluation network based on a visual Transformer fine-tuning, used to automatically segment continuous blind zone regions with signal strength below a preset threshold from the heatmap, identify potential interference regions with similar signal strength at multiple access points, and independently score key regions marked by the user.

[0013] Furthermore, in the reinforcement learning agent, the state space of the near-end policy optimization algorithm is a concatenation of the quantized blind zone distribution vector output by the AI ​​evaluation network and the key region score. The action space includes coordinate displacement commands for the access point within the permitted deployment area, commands to increase or decrease transmit power, and commands to switch working channels; The reward function is designed as a weighted sum of the reduction in blind zone area and the improvement in critical area score, minus the penalty for the increase in total transmit power and the penalty for changes in deployment location.

[0014] Furthermore, in step S4, the equipment material library is automatically associated based on the final optimized solution to generate a visualized marketing deliverable that includes an equipment list, installation point map, expected signal coverage heat map, and return on investment.

[0015] An indoor networking planning device based on visual dynamic simulation includes: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to execute an indoor networking planning method based on visual dynamic simulation.

[0016] Compared with existing technologies, the indoor network planning method and device based on visual dynamic simulation of the present invention have the following outstanding advantages: (1) The "physical-visual loop" simulation evaluation optimization architecture is proposed for the first time; Breaking away from the traditional sequential model where "simulation is only for verification and optimization relies on manual intervention," this approach couples the precise FDTD physical model with the cGAN visual perception model through a closed-loop reinforcement learning mechanism. The visual model's semantic understanding of the heatmap directly drives optimization actions, while the physical model provides quantitative evidence for visual evaluation. The two are mutually causal and co-evolve. This mechanism enables the system to possess the cognitive ability of human experts to "simulate first, identify blind spots, and adjust points," while increasing the iteration speed by two orders of magnitude.

[0017] (2) To realize the symbolic storage and accurate reasoning of networking experience; By employing a data distillation framework, unstructured historical schemes are transformed into 128-dimensional computable policy vectors, and a knowledge graph with semantic relationships is constructed. This achieves, for the first time, full automation of the entire process of "similarity retrieval → scheme reuse → performance comparison" for network case studies. Compared to traditional rule-based or simple regression methods, the schemes in this invention have higher initial population quality, and the Pareto front convergence speed is improved by approximately 40%.

[0018] (3) The engineering application benefits are significant; Completely eliminating manual surveying and trial-and-error based on experience, the average planning cycle for a single project has been reduced from 83 minutes to 6.5 minutes, and the solution output achieves second-level response, meeting the needs of rapid pre-sales support.

[0019] (4) Strong technical versatility and scalability; This invention is not only applicable to Wi-Fi indoor networking, but can also be quickly migrated to smart home IoT deployment and even smart city light pole site planning by changing the propagation model and material library, and has cross-scenario universality. Attached Figure Description

[0020] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an architecture diagram of an indoor networking planning method based on visual dynamic simulation; Figure 2 It is a dynamic simulation closed-loop diagram of an indoor networking planning method based on visual dynamic simulation. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are merely 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 are within the scope of protection of the present invention.

[0023] The following is a preferred embodiment: Example

[0024] like Figure 1 , 2 As shown in this embodiment, an indoor network planning method based on visual dynamic simulation has the following steps: S1, Visual Analysis; The system acquires a raster image of the target indoor space, performs pixel-level semantic segmentation and spatial topology reconstruction using a hierarchical visual large model, and outputs structured engineering parameters including wall type, spatial connectivity, and corresponding frequency band signal attenuation parameters.

[0025] Among them, the hierarchical visual large model has an encoder-decoder segmentation network based on a hybrid architecture of Transformer and CNN, which is used to identify at least 15 types of interior elements, including reinforced concrete walls, lightweight brick walls, glass curtain walls, metal doors, and wooden doors, and outputs pixel-level classification masks; and a topology reconstruction network based on graph neural networks is used to extract the polygonal boundary coordinates of the room and the connectivity width of doors / windows from the classification mask, and generate a spatial connectivity weighted graph.

[0026] S2, Solution Generation; Based on the structured engineering parameters, and combined with the pre-built trusted networking knowledge graph to retrieve similar historical cases, a multi-objective optimization algorithm is used to generate at least one candidate solution for wireless access point deployment. The trusted networking knowledge graph is constructed in the following ways: Collect floor plans, equipment deployment coordinates, and measured performance data of historical networking schemes, and map the scheme sketches to the standard coordinate system using a spatial registration algorithm; use a pre-trained deep neural network performance evaluator to score the quality of historical schemes and filter out low-quality schemes; for the retained high-quality schemes, generate fixed-dimensional networking strategy feature vectors through a cross-modal Transformer encoder; and construct a directed knowledge graph containing similarity relationships between cases, relationships between cases and equipment adoption, and performance superiority / inferiority relationships between cases, using strategy vectors, scene labels, equipment models, and performance indicators as node attributes.

[0027] The multi-objective optimization algorithm employs a two-stage cascade strategy: Phase 1: Using the fast propagation model as the evaluation function, the NSGA-III algorithm is used to perform a global Pareto frontier search on the number of devices, coverage, and robustness. The population initialization step retrieves the knowledge graph based on the visual features of the current apartment type and calls similar case deployment schemes to generate the initial chromosome.

[0028] The second stage involves selecting a preferred scheme from the Pareto front solution set, fixing the number and model of equipment, using the particle swarm optimization algorithm to perform local fine optimization only on the equipment coordinates and transmission power, and calculating the fitness value using high-fidelity simulation results from the finite-difference time-domain method.

[0029] S3, Dual-loop simulation optimization; The candidate schemes are input into the physics simulation engine to calculate the electromagnetic field distribution, and the calculation results are input into the visual cognition model to generate a visual coverage evaluation report. The reinforcement learning agent automatically adjusts the deployment parameters according to the evaluation report and repeats the physics simulation and visual cognition process until the preset convergence condition is met.

[0030] The physical simulation engine uses a GPU-accelerated parallel time-domain finite difference solver. The input is structured engineering parameters and deployment parameters, and the output is the electric field intensity amplitude of three-dimensional discrete grid points.

[0031] The visual cognition model includes: a conditional generative adversarial network, used to render discrete field strength data into a smooth, gradient color heatmap; and an AI evaluation network based on a visual Transformer fine-tuning, used to automatically segment continuous blind zone regions with signal strength below a preset threshold from the heatmap, identify potential interference regions with similar signal strength at multiple access points, and independently score key regions marked by the user.

[0032] In the reinforcement learning agent using the near-end policy optimization algorithm, the state space is the concatenation of the quantized blind zone distribution vector output by the AI ​​evaluation network and the key area score; the action space includes coordinate displacement instructions of the access point within the allowed deployment area, transmission power increase / decrease instructions, and working channel switching instructions; the reward function is designed as a weighted sum of the reduction in blind zone area and the improvement in key area score, minus the total transmission power increment penalty and the deployment position change penalty.

[0033] S4. Output the final optimized deployment locations and the corresponding signal coverage visualization results; Based on the final optimized solution, the system automatically associates the equipment and material library to generate a visual marketing deliverable that includes an equipment list, installation location map, expected signal coverage heat map, and return on investment. Example

[0034] The raw historical data contains a large number of low-quality, outdated, and even erroneous networking schemes. To extract high-value knowledge from them, a two-stage data distillation pipeline was constructed.

[0035] Phase 1: Solution validity filtering and automatic labeling; To define a "high-quality solution", the following conditions must be met simultaneously: 1) Customer acceptance passed; 2) The coverage rate of key areas measured in the later stage is > 98%; 3) The return on investment for equipment is higher than the average level.

[0036] Using the Swin Transformer pre-trained on ImageNet and indoor scene datasets, features are extracted from the aligned floor plan-scheme overlay to obtain a 1024-dimensional visual feature vector Vf.

[0037] The scheme structure (AP location, model) is encoded into a fixed-length sequence Sc.

[0038] Construct a deep neural network as the evaluator. The input is the fused features of Vf and Sc. The training label y comes from the measured data: calculate the SSIM and key area coverage difference between the simulated predicted heatmap and the measured heatmap. If SSIM > 0.75 and the difference < 10%, then y = 1 (good quality); otherwise, y = 0 (bad quality). Aim for the evaluator to achieve AUC = 0.94 on the validation set.

[0039] The trained evaluator scores all historical solutions, automatically filters out the bottom 30% of low-confidence solutions, and automatically labels the retained solutions with "scene tags" (such as "open office", "multi-story loft", "dense workstations").

[0040] Phase Two: Construction of Cross-Modal Knowledge Graph The goal is to enable machines to abstract universal networking "patterns" or "strategies" from high-quality solutions. Taking the topology of the apartment layout (nodes representing rooms, edges representing connectivity) and room features as input, the output is a probability distribution map of AP deployment. Based on a visual Transformer architecture, supervised training is performed on a massive number of high-quality solutions to make the model output as close as possible to the deployment location distribution of real high-quality solutions. Ultimately, the model can encode any apartment layout into a 128-dimensional "networking strategy feature vector".

[0041] Each high-quality solution is instantiated as a "case" node in the knowledge graph, constructing a trusted networking knowledge graph. Node attributes include: a 128-dimensional policy vector, scene labels, a device list, and final performance metrics (coverage area, average rate, etc.). Graph relationships include: Based on cosine similarity calculation using 128-dimensional vectors, cases with a similarity greater than 0.85 are connected. A connection is made between the "Case" node and the "Device Model" node, which is also suitable for connecting the "Case" node and the "Scene Tag" node. Based on performance metrics, similar cases with significant performance differences are connected to form a decision path.

[0042] Through graph construction, the system achieves symbolic and relational storage and efficient reasoning for complex network topologies. The current graph contains tens of thousands of nodes and relationships, supports millisecond-level similar case retrieval, and achieves a Top-5 retrieval accuracy of 93.2%.

[0043] For the mapping of hierarchical visual perception and interior structural engineering parameters: To support detailed simulation, a depth of understanding of apartment layouts far exceeding that of traditional methods is required, so a hierarchical visual model was designed.

[0044] Level 1: Pixel-level semantic segmentation ViT-ArchNet-Seg; Input a floor plan and output the category of each pixel (15 categories in total: various walls, doors and windows, columns, stairs, office furniture, etc.). A hybrid encoder-decoder architecture combining Transformer and CNN is used, achieving an mIoU of 0.91 on a self-built labeled dataset.

[0045] Level 2: Using Layout2Graph networks for instance parsing and topology reconstruction; Input the segmentation result and output the room instance polygon (boundary coordinates accurate to centimeters), including: room attributes (area, function type, identified by name and inferred by area rules) and spatial topology graph - a directed graph, where nodes are rooms and edge weights represent the width of connected doors or the width of open spaces, which is the critical path information for signal propagation.

[0046] Level 3: Attenuation factor mapping and engineering parameter assignment; The visually identified wall structure is converted into engineering parameters required for wireless channel simulation. Based on the built-in "material-attenuation correlation library", signal attenuation attributes are automatically assigned to each wall.

[0047] Based on the wall type (such as "reinforced concrete load-bearing wall", "lightweight brick partition wall", "glass curtain wall"), thickness, and database experience values, a typical penetration loss value (dB) is assigned to each wall.

[0048] After the system identifies a "standard reinforced concrete load-bearing wall (24cm)", it automatically associates the penetration loss value (2.4GHz: 20~25 dB; 5GHz: 25~30 dB); if it identifies a "single-pane glass door", it associates a lower loss value (2.4GHz: 3~6 dB; 5GHz: 5~8 dB). These parameters directly serve as key inputs to the subsequent signal propagation model.

[0049] For solving the optimal deployment point in a hybrid multi-objective optimization approach: Network planning is essentially a combinatorial optimization problem with multiple objectives and constraints. This invention designs a two-stage hybrid optimization solver.

[0050] Phase 1: Global solution search based on knowledge graph and NSGA-III; The goal is to quickly find a set of high-quality Pareto non-dominated solutions from a massive number of possibilities; By retrieving the K most similar historical cases to the current apartment layout from a vector library, and using the equipment configurations of these cases as a blueprint, crossover and mutation are performed to form an initial population. This strategy fully utilizes historical knowledge and significantly accelerates convergence.

[0051] In addition to cost (number of devices) and coverage, a third objective is introduced—solution robustness. By slightly perturbing the AP positions (simulating installation errors) and quickly re-simulating, the standard deviation of coverage fluctuation is calculated; the smaller the fluctuation, the higher the robustness.

[0052] The NSGA-III algorithm is used to handle three or more targets. Its reference point mechanism can better maintain the diversity of the population in the three-dimensional target space, ensuring that a set of candidate solutions that are balanced in terms of cost, performance, and robustness is finally obtained.

[0053] Phase Two: Local Precision Optimization Based on PSO and High-Fidelity Simulation After selecting a preferred solution (such as the lowest cost or best performance) from the Pareto frontier, the second stage is used for "fine-tuning" optimization.

[0054] With a fixed number and model of access points (APs), only their precise (x, y, z) coordinates and emission power are optimized. These four parameters of each AP constitute a particle, and the parameters of all APs are concatenated to form a high-dimensional particle.

[0055] This phase employs the FDTD high-fidelity simulation engine for accurate evaluation. A custom-developed FDTD solver is used on the GPU with a mesh accuracy of 2.5cm to accurately simulate the propagation of 5GHz band signals in complex indoor environments, including multipath effects. Constraints such as AP deployment on walls / ceilings and minimum spacing between devices are considered, and a penalty function is incorporated into the fitness calculation.

[0056] A single simulation using the Finite-Difference Time-Domain (FDTD) method takes approximately 1 minute (for a 300㎡ apartment, accelerated by an Ascend 910B). This time consumption is reduced to a few dozen iterations for a select few solutions in the second stage, keeping the total time within an acceptable range. The global search in the first stage uses a millisecond-level fast propagation model (such as the COST-231 multi-wall model) to ensure overall efficiency.

[0057] For physical-visual cyclic simulation evaluation and autonomous optimization; To achieve the high reliability goal of "design meets standards", this invention constructs a cyclic evaluation and optimization mechanism that combines physical simulation and visual cognition.

[0058] Based on vector data and AP parameters derived from apartment layout understanding, simulation scripts are automatically generated. Employing the parallel FDTD algorithm, a single full-band simulation of a typical 300-square-meter apartment can be completed within one minute on the Ascend 910B, outputting the precise three-dimensional field strength distribution E(x,y,z,f) for the entire space.

[0059] The raw field strength data output from FDTD is input into a conditional generative adversarial network (cGAN) to transform it into a smooth, gradually changing color heatmap that conforms to human visual perception, with blind spots clearly marked. Subsequently, an AI visual assessment model (fine-tuned based on ViT) performs semantic parsing on the heatmap, enabling it to identify: Global blind zone: A continuous area where the signal strength is below a threshold.

[0060] Potential interference zone: an area where multiple APs have similar signal strength, which may lead to frequent terminal switching.

[0061] Key area coverage quality: Individually score the "key rooms" marked by users.

[0062] The model's assessment results (blind spot location, area, severity level) are quantified into a structured assessment report.

[0063] Based on the current scheme's evaluation report summary (blind zone distribution vector, key area score), the Near-End Policy Optimization (PPO) agent outputs AP parameter adjustment instructions (such as "move AP_3 0.5 meters in the X direction", "reduce AP_5's transmit power by 3dB", "stagger the channels of AP_1 and AP_4"). After the actions are executed, a new scheme is obtained, and simulation evaluation is performed again to obtain a new state and reward.

[0064] The reward function is: R = w1⋅ΔA (blind zone) + w2⋅ΔS (critical zone) − w3⋅ΔP (total power) − w4⋅ΔC (scheme change); The process iterates until the reward shows no significant increase for several consecutive times, ultimately achieving a completely self-optimizing solution without any intervention. Example

[0065] An indoor networking planning device based on visual dynamic simulation includes: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to execute an indoor networking planning method based on visual dynamic simulation.

[0066] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0067] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An indoor network planning method based on visual dynamic simulation, characterized in that, It has the following steps: S1, Visual Analysis; The system acquires a grid image of the target indoor space, performs pixel-level semantic segmentation and spatial topology reconstruction using a hierarchical visual large model, and outputs structured engineering parameters including wall type, spatial connectivity, and corresponding frequency band signal attenuation parameters. S2, Solution Generation; Based on the structured engineering parameters, and combined with the pre-built trusted networking knowledge graph to retrieve similar historical cases, a multi-objective optimization algorithm is used to generate at least one candidate solution for wireless access point deployment. S3, Dual-loop simulation optimization; The candidate schemes are input into the physical simulation engine to calculate the electromagnetic field distribution, and the calculation results are input into the visual cognition model to generate a visual coverage evaluation report. The reinforcement learning agent automatically adjusts the deployment parameters based on the evaluation report and repeats the physical simulation and visual cognition process until the preset convergence condition is met. S4. Output the final optimized deployment locations and the corresponding signal coverage visualization results.

2. The indoor network planning method based on visual dynamic simulation according to claim 1, characterized in that, In the hierarchical visual large model described in step S1, an encoder-decoder segmentation network based on a hybrid architecture of Transformer and CNN is used to identify indoor elements and output a pixel-level classification mask. A graph neural network-based topology reconstruction network is used to extract the polygonal boundary coordinates of the room and the connectivity width of doors / windows from the classification mask, and generate a spatial connectivity weighted graph.

3. The indoor network planning method based on visual dynamic simulation according to claim 2, characterized in that, The construction of the trusted networking knowledge graph in step S2 is as follows: The system collects floor plans, equipment deployment coordinates, and measured performance data of historical networking schemes, and maps the scheme sketches to a standard coordinate system using a spatial registration algorithm. A pre-trained deep neural network performance evaluator is used to score the quality of historical schemes, filtering out low-quality schemes. For the retained high-quality schemes, a fixed-dimensional networking strategy feature vector is generated using a cross-modal Transformer encoder. Using the strategy vector, scene labels, equipment models, and performance indicators as node attributes, a directed knowledge graph is constructed that includes similarity relationships between cases, relationships between cases and equipment adoption, and performance superiority / inferiority relationships between cases.

4. The indoor network planning method based on visual dynamic simulation according to claim 3, characterized in that, The multi-objective optimization algorithm employs a two-stage cascaded strategy: Phase 1: Using the fast propagation model as the evaluation function, the NSGA-III algorithm is used to perform a global Pareto front search on the number of devices, coverage, and robustness. The population initialization step retrieves the knowledge graph based on the visual features of the current apartment type and calls similar case deployment schemes to generate the initial chromosome. The second stage involves selecting a preferred scheme from the Pareto front solution set, fixing the number and model of equipment, using the particle swarm optimization algorithm to perform local fine optimization only on the equipment coordinates and transmission power, and calculating the fitness value using high-fidelity simulation results from the finite-difference time-domain method.

5. The indoor network planning method based on visual dynamic simulation according to claim 1, characterized in that, In step S3, the physical simulation engine uses a GPU-accelerated parallel time-domain finite difference solver. The input is the structured engineering parameters and deployment parameters, and the output is the electric field intensity amplitude of three-dimensional discrete grid points. The visual cognition model includes: a conditional generative adversarial network, used to render discrete field strength data into a smooth, gradient color heatmap; and an AI evaluation network based on a visual Transformer fine-tuning, used to automatically segment continuous blind zone regions with signal strength below a preset threshold from the heatmap, identify potential interference regions with similar signal strength at multiple access points, and independently score key regions marked by the user.

6. The indoor network planning method based on visual dynamic simulation according to claim 5, characterized in that, The reinforcement learning agent uses a proximal policy optimization algorithm where the state space is a concatenation of the quantized blind zone distribution vector output by the AI ​​evaluation network and the key region score. The action space includes coordinate displacement commands for the access point within the permitted deployment area, commands to increase or decrease transmit power, and commands to switch working channels; The reward function is designed as a weighted sum of the reduction in blind zone area and the improvement in critical area score, minus the penalty for the increase in total transmit power and the penalty for changes in deployment location.

7. The indoor network planning method based on visual dynamic simulation according to claim 1, characterized in that, In step S4, the equipment material library is automatically associated based on the final optimized solution to generate a visual marketing deliverable that includes an equipment list, installation point map, expected signal coverage heat map, and return on investment.

8. An indoor network planning device based on visual dynamic simulation, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to perform the method according to any one of claims 1 to 7.

Images