A wireless channel scenario classification method based on an adaptive algorithm

Abstract

This invention discloses a wireless channel scene classification method based on an adaptive algorithm, which can effectively realize scene classification in practical communication systems. The method includes the following steps: 1) Collecting geographical environment information and engineering configuration parameters as a scene feature dataset; 2) Preprocessing the scene feature dataset through standardization and principal component analysis (PCA); 3) Adaptive classification using a self-organizing map (SOM) neural network algorithm; 4) Traversing SOM models of different grid sizes and calculating corresponding evaluation indicators to determine the model with the highest weighted score; 5) Loading the optimal model, ultimately dividing the main urban area of ​​Hangzhou into eight sub-scenes: high-density business district, high-density residential district, low-density business district, transportation hub, medium-density residential district, park, low-density residential district, and indoor. The final weighted score is 0.6, indicating that this method can effectively achieve adaptive scene classification, providing an important prerequisite for channel modeling.

Description

Technical Field

[0001] This invention belongs to the field of wireless channel research, and in particular relates to a wireless channel scene classification method based on an adaptive algorithm. Background Technology

[0002] With the continuous development of communication technology, research on 6G is actively being promoted. The technical vision of 6G includes full spectrum, full coverage, and all application scenarios, bringing with it a variety of complex propagation scenarios. Different propagation scenarios and physical environments often lead to significant differences in wireless channel characteristics, thus requiring the design of different channel models. In 6G communication, when users move rapidly and continuously across multiple sub-scenarios, accurately selecting the appropriate channel model and optimizing its parameters is crucial for achieving efficient communication. This process first requires accurate classification of different sub-scenarios and then optimizing the parameters of the channel model based on the classification results. However, traditional scenario classification methods often struggle to effectively capture these variations in different environments. For example, the scenario classification in the 3GPP TR 38.901 standard only includes highly generalized categories such as urban micro base stations in streets and canyons, urban macro base stations, indoor offices, rural macro base stations, and indoor factories. This classification method is overly simplistic and cannot accurately reflect the propagation characteristics of specific areas, potentially leading to significant errors in practical applications.

[0003] To address this issue, existing research has classified propagation scenarios based on physical environment parameters and manually defined ranges. These methods typically rely on partial channel measurement data and improvements to existing standards. However, in practical applications, the ranges of physical environment parameters often overlap between different sub-scenarios, making effective differentiation difficult. Furthermore, comparing and analyzing multiple physical environment features usually depends on manual judgment, which is very difficult and inaccurate in complex environments. Currently, some scholars are attempting to use artificial intelligence (AI) technology to solve this problem. However, existing AI methods generally suffer from two problems: First, most methods are supervised learning, requiring pre-defined categories and numbers of sub-scenarios, which is highly subjective, and fixed classification standards may not adapt to dynamically changing environments. Second, many methods rely on feature data based on channel measurements, which are costly and time-consuming to collect, limiting their application in real-world scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a wireless channel scenario classification method based on an adaptive algorithm to achieve refined propagation scenario classification in 6G communication, thereby providing a more accurate basis for channel modeling and network optimization.

[0005] Technical Solution: To address the aforementioned technical problems, this invention proposes a wireless channel scene classification method based on an adaptive algorithm, comprising the following steps:

[0006] Step 1: Determine the format, accuracy, and available physical environment parameters of the electronic map to be used;

[0007] Step 2: Based on the existing network test data, determine the main coverage area of ​​each cell and realize scene segmentation;

[0008] Step 3: Extract the required physical environment parameters based on the segmentation results and electronic map; construct a scene feature dataset by combining the physical environment parameters with the engineering configuration parameters;

[0009] Step 4: Preprocess the scene feature dataset, including data standardization and dimensionality reduction.

[0010] Step 5: Train the preprocessed data using the SOM neural network algorithm, traverse SOM networks of different topological sizes, and determine the optimal number of scene classifications.

[0011] Step 6: Based on the determined optimal number of classifications, use the SOM network to adaptively classify the scene.

[0012] Furthermore, step 1 includes the following steps:

[0013] Step 1.1: Using SHP format electronic maps ensures efficient reading and processing of map data and is compatible with QGIS, facilitating the extraction and analysis of geographic information;

[0014] Step 1.2: Determine the physical environment parameters that need to be extracted from the electronic map:

[0015] Average building height: Building height directly affects signal reflection and refraction, making multipath effects more pronounced. Tall buildings easily block and refract signals, affecting the signal propagation path;

[0016] Building area ratio: Signal blocking effects are more pronounced in densely built-up areas, easily increasing signal path loss. Densely built-up areas typically lead to stronger signal obstruction and scattering, thus affecting signal coverage and quality;

[0017] Road width: Open road areas generally have less signal obstruction, which is conducive to signal propagation and reduces multipath interference. Relatively narrow roads may cause strong refraction and reflection of signals, increasing signal attenuation;

[0018] Road area ratio: Generally, a larger road area ratio reduces signal attenuation and improves signal propagation quality. Areas with a high road area ratio may offer better signal transmission conditions;

[0019] Vegetation type: Different types of vegetation absorb and scatter wireless signals to varying degrees;

[0020] Vegetation area ratio: Large green areas usually cause stronger signal attenuation, especially in densely vegetated areas within the coverage area, where the impact on signal is particularly significant.

[0021] Water area percentage: Water surfaces can reflect and refract signals, causing changes in the signal path. Furthermore, water bodies absorb a certain amount of wireless signal energy, thus exacerbating signal attenuation.

[0022] Furthermore, step 2 specifically includes the following steps:

[0023] Step 2.1: Collect minimum road test (MDT) data from the existing network, including cell signal strength, reception quality, and coverage boundary information;

[0024] Step 2.2: Based on MDT data, analyze the signal strength and quality of each cell to determine the main coverage area;

[0025] Step 2.3: Divide the area using the main coverage area of ​​the cell.

[0026] Furthermore, step 3 specifically includes the following steps:

[0027] Step 3.1: Based on the geometric calculation function of QGIS, perform corresponding mathematical operations through the attribute calculator to calculate and extract the physical environment parameters of each scene segmentation area;

[0028] Step 3.2: Associate the extracted physical environment parameters with the corresponding scene segmentation regions to form a dataset containing complete physical environment features.

[0029] Obtain the following project parameters from the project parameter configuration table:

[0030] Azimuth angle: The pointing angle of the base station antenna, which affects the direction of signal propagation;

[0031] Downtilt angle: The tilt angle of the antenna, which determines the vertical range of signal coverage;

[0032] Coverage type: Specifies whether the base station covers indoors or outdoors, affecting the signal attenuation mode;

[0033] The extracted physical environment parameters are fused with the corresponding engineering configuration parameters using multi-dimensional features to construct a scene feature dataset;

[0034] The constructed scene feature dataset is exported as a CSV file to provide input data for subsequent analysis.

[0035] Furthermore, step 4 specifically includes the following steps:

[0036] Step 4.1: Because the value ranges and units of each feature are different (e.g., building height is in meters, area ratio is a proportion, antenna angle is in degrees), directly using these features for calculation may lead to uneven weighting of different features in distance calculation. Standardization maps the data to the range [0, 1], eliminating the impact of scale differences among features;

[0037]

[0038] Where x is the original data value, x ′ For the standardized data value, x min and x max These are the minimum and maximum values ​​of feature x, respectively;

[0039] Step 4.2: PCA removes redundant information and noise through dimensionality reduction, retaining the most representative features, thereby reducing computational complexity and improving the efficiency of the classification model. To ensure that the dimensionality-reduced data can fully retain the original information, principal components with a cumulative variance of 80% are selected for dimensionality reduction.

[0040] Z = X·W

[0041] Where Z is the data after dimensionality reduction, X is the original data matrix, and W is the principal component matrix;

[0042] Step 4.3: During preprocessing, data quality checks are also required to identify and correct outliers or missing values. Missing values ​​can be supplemented using interpolation; outliers can be removed using appropriate anomaly detection methods.

[0043] Furthermore, step 5 specifically includes the following steps:

[0044] Step 5.1: Select the silhouette coefficient and the Calinski-Harabasz Index (CHI) as cluster evaluation indicators. The silhouette coefficient measures intra-cluster compactness and inter-cluster separation, while the CHI measures inter-cluster dispersion and intra-cluster dispersion.

[0045] To evaluate the performance of the SOM classification model, it is first necessary to select appropriate evaluation metrics. In this step, the silhouette coefficient and the Kalski-Hallabus index (CHI) were chosen as clustering evaluation metrics. These two metrics measure the clustering effect from different perspectives, helping to ensure the accuracy and reasonableness of the classification results. The silhouette coefficient can be expressed as:

[0046]

[0047] Where a(i) represents the average distance from sample i to other samples of the same class, which measures intra-class compactness; b(i) represents the average distance from sample i to the nearest sample of a different class, which measures inter-class separation.

[0048] CHI can be represented as

[0049]

[0050] Where B(k) represents the inter-cluster dispersion, which is the sum of the dispersion of the centroids of all clusters to the centroids of the entire dataset; W(k) represents the intra-cluster dispersion, which is the sum of the dispersion of a data point to the centroid of its cluster; n is the number of samples in the dataset; and k is the number of clusters.

[0051] Step 5.2: Classify the data using the SOM algorithm. The training process includes initializing neuron weights, a competition phase, weight updates, and repeated iterations until the network weights converge, resulting in an effective classification result. The training of the SOM network involves the following steps:

[0052] (1) Initialize neuron weights: Randomly initialize the weights of neurons so that the weight vector of each neuron can represent the feature space of the input data.

[0053] (2) In the competition phase, a sample vector is randomly sampled from the preprocessed training dataset and its distance to each neuron is calculated. The neuron with the closest distance is selected as the winner neuron. Its weight vector will be used as the nearest neighbor matching of the input data in the feature space and drive the self-organization process of the network topology.

[0054] (3) Weight update: Update the weights of the winner's neurons to make them closer to the feature vectors of the input data points. At the same time, update the weights of other neurons in the winner's neighborhood so that the entire network gradually forms a topological structure.

[0055] (4) Repeat the iteration, repeat the competition phase and weight update until the predetermined number of iterations or neuron weight convergence is reached. After training, the SOM network will be able to effectively map and classify the input data.

[0056] Step 5.3: Traverse SOM network models of different sizes, select the best mesh size, and determine the most suitable number of scene classifications. The test range is 3 to 15 clusters.

[0057] To determine the optimal number of scene classifications, it is necessary to iterate through SOM network models of different sizes, select the best grid size, conduct iterative testing within the range of 3 to 15 clusters, and dynamically generate different grid sizes. By adjusting the grid size, different clustering results can be explored, and the most suitable classification structure can be determined.

[0058] Step 5.4: Based on the profile coefficient and CHI score, select the number of categories and grid size with the highest evaluation index score. Finally, determine the 1×8 grid size as the optimal classification structure and the number of categories as 8.

[0059] Taking into account both the contour coefficient and CSI scores, the highest number of scene classifications and mesh size are selected based on the evaluation metric scores. In this invention, the specific implementation of combining the contour coefficient and CSI is as follows:

[0060]

[0061] The scaling factor was set to 0.0002. The results showed that the weighted score was highest with a grid size of 1×8, meaning that the number of classifications was optimal with 8 clusters, indicating that this number of classifications was the most suitable for the current dataset.

[0062] Furthermore, step 6 specifically includes the following steps:

[0063] Step 6.1: Configure the SOM network and train it according to the optimal number of classes and grid size. The network selects the winning neuron through a competition phase and updates the weights until the network weights converge.

[0064] Step 6.2: After training, the SOM network assigns scene data to different categories based on its weights and topology. Each neuron represents a category, and scene data is classified into the corresponding category based on the similarity between its features and the neurons.

[0065] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0066] This invention proposes an adaptive classification method based on electronic maps. First, it utilizes electronic maps to collect geographic environmental information and base station deployment-related engineering configuration parameters as the foundation for scene feature construction. This avoids the complexity of relying on large amounts of channel measurement data in traditional methods, significantly reducing the manpower and resources required for channel measurement. Second, through preprocessing methods such as principal component analysis and normalization, this invention can extract key features, reduce data redundancy, and improve the efficiency and accuracy of subsequent classification tasks. Finally, this invention uses the SOM algorithm for adaptive scene classification. Without the need for predefined sub-scene types, it can more accurately adapt to environmental changes and dynamic scenes, providing an important prerequisite for scene classification and channel modeling in practical communication. Attached Figure Description

[0067] Figure 1 This is a flowchart of the wireless channel scene classification method based on an adaptive algorithm provided in Embodiment 1 of the present invention;

[0068] Figure 2 This is a schematic diagram of a portion of the electronic map in Embodiment 1 of the present invention;

[0069] Figure 3 This is a schematic diagram of the adaptive classification results scores of the SOM algorithm and two other common algorithms in Embodiment 1 of the present invention;

[0070] Figure 4 This is a schematic diagram of the power fitting results for a high-density commercial area in Embodiment 1 of the present invention;

[0071] Figure 5 This is a schematic diagram verifying the power fitting results of a high-density commercial area in Embodiment 1 of the present invention;

[0072] Figure 6 This is a schematic diagram of the simulation results of the spatial cross-correlation function of the sub-scene receiver in Embodiment 1 of the present invention;

[0073] Figure 7 This is a schematic diagram of the simulation results of the time autocorrelation function of the sub-scene in Embodiment 1 of the present invention;

[0074] Figure 8 This is a schematic diagram of the simulation results of the sub-scene frequency correlation function in Embodiment 1 of the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0076] See Figure 1 This embodiment provides a wireless channel scene classification method based on an adaptive algorithm, and the specific steps are as follows:

[0077] Step 1: Obtain an electronic map of the target recognition environment, determine the format and accuracy of the selected electronic map, and the physical environment parameters that can be extracted from the electronic map. The specific steps are as follows:

[0078] Specifically, in this embodiment, the selected SHP format electronic map can support the accurate extraction of buildings, roads, and other environmental features. Based on this, the physical environmental parameters to be extracted are further defined, including building height, road width, vegetation type, etc., which directly affect the propagation characteristics of wireless signals.

[0079] Step 2: Based on the MDT test data, analyze the coverage area of ​​each cell. By collecting data on wireless signal strength, quality, and other indicators, determine the main coverage area of ​​each cell and segment the scene. The specific steps are as follows:

[0080] Step 2.1: Collect minimum road test (MDT) data from the existing network, including cell signal strength, reception quality, and coverage boundary information;

[0081] Step 2.2: Based on MDT data, analyze the signal strength and quality of each cell to determine the main coverage area;

[0082] Step 2.3: Divide the area using the main coverage area of ​​the cell.

[0083] Step 3: Based on the physical environment parameters extracted from the electronic map and the relevant engineering configuration parameters, construct a complete scene feature dataset. The specific steps are as follows:

[0084] Step 301: Select the analysis area based on the scene segmentation results. Using the "Select" tool or the "Attribute Query" function, select the corresponding area from the electronic map for analysis based on the scene segmentation results obtained in the previous steps. Extract physical environment parameters. For each analysis area, use the QGIS tool to extract the required physical environment parameters, including:

[0085] Building height: The building height information is already in the data and can be extracted directly from the attribute table;

[0086] Building area percentage: The total area of ​​buildings is calculated using spatial analysis tools and compared with the total area of ​​the region;

[0087] Road width and area ratio: The road width is calculated using the "buffer zone" analysis tool, and the road ratio is obtained through area calculation;

[0088] Vegetation type and area percentage: Extract data on different types of vegetation through attribute tables or "select query" and calculate the area percentage;

[0089] Water area percentage: Select a water area and calculate its area percentage.

[0090] Step 302: Associate the extracted physical environment parameters with the corresponding scene segmentation regions to form a dataset containing complete physical environment features;

[0091] Step 303: Obtain the azimuth, downtilt, and coverage type parameters from the engineering parameter configuration table;

[0092] Step 304: Combine the physical environment parameters of each scene with the engineering configuration parameter table to form a higher-dimensional dataset, constructing a complete scene feature dataset. This dataset will serve as the basis for subsequent classification and analysis, providing input for subsequent classification algorithms.

[0093] Step 4 mainly involves preprocessing the scene feature data, including data standardization, PCA, and other operations. The specific steps are as follows:

[0094] Step 401: Data standardization is to normalize different feature dimensions to ensure that the value range of each feature is the same, thereby avoiding the impact of some features on the training effect of the model due to excessively large value ranges.

[0095] Step 402: PCA is used to reduce the dimensionality of the dataset. By calculating the covariance matrix of the data, it extracts 80% of the principal components and retains the data features in the direction of maximum variance. The steps of PCA are as follows:

[0096] (1) Calculate the covariance matrix of the dataset;

[0097] (2) Calculate the eigenvalues ​​and eigenvectors of the covariance matrix;

[0098] (3) Select the first k principal components based on their eigenvalues;

[0099] (4) Project the data into a new feature space.

[0100] Step 403: Check data integrity and outliers to ensure the dataset is complete and free of missing values. If missing values ​​exist, handle them using interpolation or by removing samples with missing values.

[0101] Step 5: Use the SOM neural network algorithm to perform preliminary classification on the preprocessed dataset, and determine the most suitable number of scene categories based on the evaluation scores. The specific steps are as follows:

[0102] Step 501: Select the silhouette coefficient and CHI as clustering evaluation indicators. These two indicators can measure the clustering effect from different perspectives and ensure the accuracy and rationality of the classification results.

[0103] Step 502: After data preprocessing, the SOM algorithm is used to classify the dataset. The SOM network is trained on the input data, automatically mapping the data to a low-dimensional space and classifying it into different scene categories. Through competition and cooperation between neurons, the network learns the topological structure of the data, thereby achieving effective classification.

[0104] Step 503: To determine the most suitable number of scene classifications, traverse SOM network models with different grid sizes, testing from 3 clusters to 15 clusters, and explore different classification results by adjusting the grid size;

[0105] Step 504: After completing the SOM network training, a weighted score is calculated by comprehensively considering the contour coefficient and CHI scores to determine the optimal number of classifications and grid size. During testing, a 1×8 grid size resulted in the highest weighted score, indicating the best classification performance with 8 clusters. Therefore, 8 clusters were ultimately determined as the most suitable number of scene classifications.

[0106] Step 6: Configure the SOM network based on the determined optimal number of classifications and mesh size, and perform adaptive classification of the scene. The specific steps are as follows:

[0107] Step 601: Configure the SOM network structure based on the determined number of categories and grid size. Set the learning rate to 0.05, the neighborhood function width (Sigma) to 0.1, and the number of training iterations to 5000. The network selects a "winning neuron" for each input data point through a competitive phase. The weights of this winning neuron are updated to align with the feature vector of the input data. Simultaneously, the weights of the neighboring neurons of the winning neuron are adjusted accordingly based on the learning rate. Through multiple iterations, the network gradually converges, and the weights more accurately reflect the feature distribution of the input data.

[0108] Step 602: After training, the SOM network assigns scene data to corresponding neuron nodes based on the weights and topology formed during training. Each neuron represents a category. The network calculates the similarity between each scene data point and the neuron weights, and assigns the scene data to the neuron node with the highest similarity, thereby completing the scene classification.

[0109] The classification results were verified using measurement data collected in actual communication networks as a reference to evaluate their feasibility. The verification method is as follows:

[0110] The path loss model to be used is determined as an important basis for measuring the effectiveness of the classification results.

[0111] PL(d,f)=A·log 10 (d)+B+C·log 10 (f c )

[0112] Where d is the distance between the transmitter and receiver antennas, in meters; f c A represents the carrier frequency, in GHz; B represents the percentage increase in path loss with increasing distance; C represents the constant term used to calibrate path loss; and D represents the percentage increase in path loss with increasing frequency.

[0113] The path loss of each sub-scenario is fitted using the measurement data obtained through MDT.

[0114] By comparing and analyzing the data with actual measurement data, the rationality and accuracy of the SOM network classification results can be verified.

[0115] Furthermore, physical environment parameters are statistically analyzed for the categorized sub-scenes, and their parameters are optimized using channel model simulation to verify the channel differences between different sub-scenes. This includes the following steps:

[0116] Based on the SOM classification results, key physical environment parameters for each sub-scene are statistically analyzed, such as building height, terrain type, and vegetation coverage. The physical characteristics of each sub-scene are analyzed. The channel model parameters for each sub-scene are optimized. The channel characteristics of each sub-scene are simulated, and the differences in channel characteristics caused by differences in physical environment between different sub-scenes are analyzed.

[0117] like Figure 3 As shown, to intuitively demonstrate the performance of K-means, GMM, and SOM algorithms in propagation scenario classification, the classification results of the three clustering algorithms were analyzed when the number of clusters was 8. It can be seen that SOM performed best, with a silhouette coefficient of 0.587, a Kaliski-Halaba index of 3028.68, and a weighted score of 0.597, indicating that the clustering effect was the most balanced and optimized at this point. Figure 4 As shown in the figure, the simulated value fitting curve and the measured value fitting curve are in high agreement, indicating that the selected logarithmic distance model can accurately characterize the signal propagation characteristics under different scenarios. Figure 5 As shown, the fitted curves exhibit good fitting results in all scenarios, further demonstrating the applicability of the model.

[0118] and, Figure 6 As shown in the diagram, transportation hubs typically have open structures with minimal obstruction, resulting in signals primarily traveling via direct and stable reflection paths. Consequently, channel response variations between adjacent locations are minimal, leading to high spatial cross-correlation. Low-density commercial areas and medium-density residential areas exhibit relatively regular building layouts, with consistent obstruction and scattering environments, which also contributes to spatial signal continuity. High-density commercial areas, due to their numerous high-rise buildings, experience significant non-line-of-sight propagation and strong multipath effects, making signal propagation paths highly sensitive to location changes and reducing spatial cross-correlation. While high-density residential areas have lower-rise buildings, their dense and complex layouts also lead to rapid signal changes. Although parks lack obstructing buildings, vegetation, topography, and pedestrians increase the uncertainty of signal scattering, similarly resulting in lower spatial correlation.

[0119] Figure 7 As can be seen from the data, transportation hubs, due to their open structure and minimal obstruction, exhibit relatively low temporal autocorrelation due to the smaller changes in the channel environment over time. Medium-density residential areas and low-density commercial areas have relatively stable environments with predictable building distribution, resulting in slow channel changes during movement and also demonstrating high temporal autocorrelation. High-density commercial areas and low-density residential areas exhibit moderate temporal autocorrelation; the former, despite its dense buildings, has a relatively fixed structure, while the latter, although experiencing obstruction changes, has relatively weak overall dynamism. Parks and high-density residential areas show low temporal autocorrelation. Parks, with their numerous randomly moving people and swaying vegetation, cause rapid channel changes over time; high-density residential areas, due to their dense buildings and complex paths, are prone to channel characteristics changes even when stationary or moving at low speeds.

[0120] Figure 8As shown in the figure, transportation hubs have open structures and relatively simple paths, mainly relying on direct and stable reflection paths. Therefore, the channel response varies little across different frequencies, resulting in high frequency correlation. Low-density commercial areas have sparse building distribution and a limited number of scattering paths, leading to a relatively smooth frequency response and also exhibiting high frequency correlation. Medium-density and low-density residential areas have moderate building layouts and a relatively uniform number and distribution of scatterers, resulting in moderate frequency response variations and thus moderate frequency correlation. High-density commercial and residential areas, due to their complex environments and abundant multipath sources, show significant differences in signal interference modes across different frequencies, resulting in low frequency correlation. It is worth noting that although low-density commercial areas are generally spacious, sudden structural obstructions and dynamic changes in some areas can increase response fluctuations in certain frequency bands, leading to a decrease in frequency correlation.

[0121] In summary, the adaptive classification method for wireless channel scenarios based on electronic maps established in this invention achieves adaptive classification of wireless channel scenarios by combining a SOM neural network to classify different propagation scenarios. This lays an important foundation for subsequent wireless channel modeling and optimization.

[0122] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for wireless channel scenario classification based on adaptive algorithm, characterized in that, The method includes the following steps: Step 1: Determine the format, accuracy, and available physical environment parameters of the electronic map to be used; Step 2: Determine the main coverage area of ​​each cell based on the existing network data to achieve scene segmentation; Step 3: Extract physical environment parameters and merge engineering configuration parameters to construct a scene feature dataset; Step 4: Preprocess the scene feature dataset, including data standardization and dimensionality reduction. Step 5: Train the preprocessed data using the SOM neural network algorithm, traverse SOM networks of different topological sizes, and determine the optimal number of scene classifications. Step 6: Based on the determined optimal number of classifications, use the SOM network to adaptively classify the scene; In step 1, the physical environment parameters include building height, building area ratio, road width, road area ratio, vegetation type and area ratio, and water area ratio. In step 2, the existing network data includes minimum road test (MDT) data containing cell signal strength, reception quality, and coverage boundary information. Based on this data, the signal strength and quality of each cell are analyzed to determine the main coverage area, and the area is segmented based on this. Step 5 specifically includes the following steps: Step 5.1: Select the silhouette coefficient and the Kalski-Hallabus index (CHI) as cluster evaluation indicators. The silhouette coefficient measures intra-cluster compactness and inter-cluster separation, while the CHI measures inter-cluster dispersion and intra-cluster dispersion. Step 5.2: Use the SOM algorithm to classify the data. The training process includes initializing neuron weights, competition phase, weight update and repeated iteration until the network weights converge and form an effective classification result. Step 5.3: Traverse SOM network models of different sizes, select the best mesh size, and determine the most suitable number of scene classifications. The test range is 3 to 15 clusters. Step 5.4: Based on the profile coefficient and CHI score, select the number of categories and grid size with the highest evaluation index score. Finally, determine the 1×8 grid size as the optimal classification structure and the number of categories as 8.

2. The wireless channel scene classification method based on an adaptive algorithm according to claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Based on the geometric calculation function of QGIS, perform corresponding mathematical operations through the attribute calculator to calculate and extract the physical environment parameters of each scene segmentation area; Step 3.2: Associate the extracted physical environment parameters with the corresponding scene segmentation regions to form a dataset containing complete physical environment features; Step 3.3: Obtain the azimuth, downtilt, and coverage type parameters from the engineering parameter configuration table; Step 3.4: Combine the extracted physical environment parameters with the corresponding engineering configuration parameters to construct a comprehensive feature dataset of the scene segmentation region, and export it as a CSV file to provide input data for analysis.

3. The method of claim 1, wherein, Step 4 specifically includes the following steps: Step 4.1: Standardize the scene feature data by mapping all features to the range [0, 1] to eliminate the impact of differences in feature scales. Step 4.2: Use PCA to reduce the dimensionality of the data, retaining principal components with a cumulative variance of 80% to remove redundant information; Step 4.3: Use interpolation to fill in missing values ​​and use anomaly detection methods to remove outliers to ensure data integrity.

4. The method of claim 1, wherein, Step 6 specifically includes the following steps: Step 6.1: Configure the SOM network and train it according to the optimal number of classes and grid size. The network selects the winning neuron through a competition phase and updates the weights until the network weights converge. Step 6.2: After training, the SOM network assigns scene data to different categories based on its weights and topology. Each neuron represents a category, and scene data is classified into the corresponding category based on the similarity between its features and the neurons.

Images