Railway communication cell range intelligent speculation method and system based on artificial intelligence

Abstract

The application discloses a railway communication cell range intelligent speculation method and system based on artificial intelligence, belongs to the field of wireless communication network optimization and deep learning technology, and comprises data collection, data preprocessing, deep learning speculation, GIS display and model optimization. The model input is constructed by arranging and processing the collected data, the deep learning model outputs the cell range speculation result, and the GIS is used to realize visual display. The system is composed of a data collection module, a data preprocessing module, a deep learning speculation module, a GIS display module and a model optimization and updating module, and is used for completing data acquisition, processing, speculation, display and model updating. When the network condition changes or the speculation performance decreases, the system can perform retraining or incremental updating on the model to maintain the accuracy and stability of the speculation. The application realizes the automatic intelligent speculation of the railway communication cell range, and provides support for communication network planning, maintenance and optimization.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication network optimization and deep learning technology, and in particular to an intelligent prediction method and system for railway communication cell range based on artificial intelligence. Background Technology

[0002] In railway communication systems, accurately determining the coverage area of ​​each communication cell is a crucial technical step in network planning, optimization, and troubleshooting. The accuracy of cell coverage not only directly affects communication quality but also relates to the stability and security of critical services such as train control systems and train-to-ground information exchange.

[0003] However, existing technologies generally rely on manual field testing, which has significant limitations:

[0004] On the one hand, the terrain along the railway is complex and varied, including mountains, tunnels, bridges, plains and other environments. In addition, some lines are inconvenient to access, resulting in sparse distribution of measurement points and difficulty in fully covering all areas, which affects the accuracy and comprehensiveness of the judgment results.

[0005] On the other hand, traditional methods fail to fully integrate the mapping relationship between railway-specific kilometer markers and geographical coordinates, and also lack the ability to integrate multi-source signal quality data (such as Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), and Signal-to-Noise Ratio (SINR)) for multi-dimensional analysis, resulting in one-sided or even biased coverage assessment results.

[0006] Furthermore, manual measurement methods are costly in terms of time, manpower, and economic resources, and have long data update cycles, making them difficult to adapt to the needs of frequent base station adjustments and dynamic network structure optimization. When the network environment changes, traditional manual measurement methods often cannot complete the reassessment of coverage in a short period of time, which to some extent limits the real-time optimization and rapid fault location capabilities of railway communication networks.

[0007] Therefore, there is an urgent need for a technical solution that can be automated, efficient, and adaptive, fully combining deep learning with railway scenario characteristics, and utilizing large-scale real-time data during train operation to achieve intelligent prediction of communication cell coverage, thereby significantly improving the operation and maintenance efficiency and network service quality of railway communication systems. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes an intelligent prediction method and system for railway communication cell coverage based on artificial intelligence. By combining data mining, deep learning, and geographic information technology, and through continuous collection of locomotive operation data, training and inference of deep learning models, intelligent prediction and visualization of cell coverage, and continuous optimization and iterative updates of deep learning models, the system achieves accurate prediction and dynamic presentation of railway communication cell coverage.

[0009] The technical solution of this invention not only significantly improves the accuracy and robustness of cell range estimation, but also has good scalability and adaptability, enabling it to adapt to the changing geographical conditions and frequent base station configuration adjustments in the railway communication environment, thus providing strong support for the intelligent operation and maintenance and optimized management of railway communication networks.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] An AI-based intelligent prediction method for railway communication cell range includes the following steps:

[0012] A. Locomotive communication data acquisition: Collect locomotive operation data including location information, kilometer markers, cell IDs, and signal quality parameters; cell IDs are only used as supervised learning labels during the training phase and not as input data for the model;

[0013] B. Data preprocessing: Cleaning, outlier removal, classification, storage and formatting of locomotive operation data to generate input data that can be used by deep learning models;

[0014] C. Cell range prediction through deep learning model training: Input preprocessed locomotive operation data into a deep learning model for training and prediction, output the predicted cell ID and its confidence score, and determine the cell coverage area based on the prediction results.

[0015] D. GIS Map Display: Dynamically visualize the inferred results using a geographic information system;

[0016] E. Model Optimization and Update: The deep learning model is updated based on newly collected locomotive operation data to adapt to changes in base station configuration and improve prediction accuracy.

[0017] Furthermore, the locomotive operation data in this method includes latitude and longitude information, kilometer markers, cell ID, RSRP, RSRQ, and SINR.

[0018] Furthermore, the data preprocessing in this method includes cleaning the locomotive operation data to remove outliers, and then classifying and storing it according to the cell ID before vectorizing it to generate a tensor format that can be accepted by deep learning neural networks.

[0019] Furthermore, the data preprocessing in this method includes:

[0020] B1. Raw Data Cleaning: Remove abnormal data where the same locomotive's single displacement exceeds 500 meters; remove low signal values ​​with RSRP less than or equal to -120 dBm that cannot maintain normal communication connections, and abnormally high signal values ​​with RSRP greater than or equal to -50 dBm caused by measurement errors or being too close to the base station; remove low quality signal values ​​with RSRQ less than or equal to -20 dB that cause service interruption, and abnormally high signal values ​​with RSRQ greater than or equal to -3 dB caused by neighboring cell interference or parameter configuration errors; remove low signal-to-noise ratio values ​​with SINR less than or equal to -3 dB that cannot complete signal demodulation, and abnormally high signal-to-noise ratio values ​​with SINR greater than or equal to 30 dB caused by no interference or measurement abnormalities.

[0021] B2. Data Classification and Storage: The cleaned locomotive operation data is classified and stored according to the cell ID, so as to facilitate subsequent modeling at the single cell level;

[0022] B3. Vectorization: Vectorize the data for each category to generate tensor formats that can be accepted by deep learning neural networks, which are then used as input data for the model.

[0023] Furthermore, the step of inferring cell range through deep learning model training in this method includes:

[0024] C1. Divide the preprocessed dataset into training set, validation set, and test set;

[0025] C2. Based on the kilometer marker sequence and signal strength input, as well as the latitude and longitude grid and signal strength input, respectively, they are fed into a dual-path convolutional neural network for feature extraction;

[0026] C3. The feature vectors extracted by the dual-path convolutional neural network are concatenated, and feature fusion is performed through a multi-head self-attention mechanism and a dynamic weighting method of signal intensity to obtain a joint representation of spatial signals.

[0027] C4. Input the fused features into the fully connected layer for high-order feature transformation, and output the probability distribution of each cell through the output layer combined with the Softmax activation function to obtain the target cell ID and its confidence level.

[0028] C5. In the actual prediction stage, the real-time locomotive operation data is processed using the same feature extraction and fusion methods as in the training stage to obtain the cell prediction results and display them visually.

[0029] Furthermore, the GIS map display in this method includes:

[0030] D1. Receive latitude and longitude information or kilometer marker information input by the user;

[0031] D2. Call the trained deep learning model to output the target cell ID and its corresponding confidence score;

[0032] D3. Based on GIS map technology, the predicted geographical boundaries of the community and the current real-time location of the locomotive are dynamically visualized and rendered to realize the identification and management of communication areas.

[0033] Furthermore, the model optimization and update steps in this method include:

[0034] E1. Automatically collect locomotive operation data, including latitude and longitude information, kilometer markers, cell ID, RSRP, RSRQ and SINR, during the daily operation of the locomotive, and store the data along with the collection timestamp and the corresponding base station version number information in the database;

[0035] E2. Real-time monitoring of base station parameter changes. If changes in base station parameters are detected, including new PCI or frequency adjustment hardware configuration changes, the data preprocessing step and deep learning model training step are invoked to perform full model retraining on the changed locomotive operation data and generate a new deep learning model.

[0036] E3. If the base station parameters have not changed, incremental learning is performed only on the newly added locomotive operation data to fine-tune the parameters of the existing deep learning model, obtain the updated deep learning model, and deploy the deep learning model to the production environment to complete the iterative optimization.

[0037] Furthermore, the dual-path convolutional neural network in this method includes a one-dimensional convolutional neural network and a two-dimensional convolutional neural network. The one-dimensional convolutional neural network is used to process kilometer marker sequence data, including four channels: normalized kilometer markers, RSRP, RSRQ, and SINR. It adopts a three-layer one-dimensional convolutional structure with a kernel size of 3 and the number of kernels being 32, 64, and 128 respectively. The two-dimensional convolutional neural network is used to process the signal data after latitude and longitude gridding, with a grid division of 0.01 degrees resolution. The three channels correspond to the grid mean values ​​of RSRP, RSRQ, and SINR, respectively. An improved ResNet18 structure is used to extract two-dimensional spatial features of the coverage area. The feature vectors output by the one-dimensional convolutional neural network and the two-dimensional convolutional neural network are concatenated and input into a multi-head self-attention mechanism. The signal intensity weights are dynamically allocated through the training weight matrix to generate a 256-dimensional joint representation of spatial signals. The joint representation of spatial signals is subjected to high-order feature transformation through a fully connected layer with 512 neurons and using the ReLU activation function. Finally, the probability distribution and confidence score of each cell ID are output through a fully connected output layer combined with the Softmax activation function.

[0038] Furthermore, this method is implemented through an AI-based intelligent prediction system for railway communication cell range. This system includes: a data acquisition module, a data preprocessing module, a deep learning prediction module, a GIS display module, and a model optimization and update module. The data acquisition module collects multi-dimensional communication data such as latitude and longitude information, kilometer markers, cell IDs, RSRP, RSRQ, and SINR in real time during locomotive operation and uploads the data to the data center via the communication link. The data preprocessing module performs outlier cleaning, classification, storage, and vectorization on the received locomotive operation data, constructing the processed data into a tensor format acceptable to the deep learning model. The deep learning prediction module includes a dual-path convolutional neural network and a multi-head self-attention mechanism, used for feature extraction and feature fusion of the vectorized locomotive operation data, and based on Softmax... The output layer obtains the probability distribution of each cell and outputs the predicted cell ID and its confidence score; the GIS display module receives the predicted cell ID and confidence score output by the deep learning inference module, and performs visualization rendering of the predicted cell coverage and the current position of the locomotive based on GIS map technology; the model optimization and update module detects changes in base station parameters based on newly collected locomotive operation data, performs full model retraining when base station parameters change, and performs incremental learning when no changes occur, in order to update and deploy the deep learning model.

[0039] The beneficial results of this invention are as follows: This invention integrates multi-source data acquisition, deep learning modeling, and geographic information technology, effectively solving problems such as inaccurate cell boundary determination and low coverage assessment efficiency in traditional methods. First, a high-quality training dataset is constructed by systematically collecting multi-dimensional communication data generated during locomotive operation. Second, leveraging the powerful feature extraction capabilities of deep neural networks and combining them with an attention mechanism, accurate modeling of cell ranges in complex coverage scenarios is achieved. Then, GIS visualization technology is used to intuitively display the predicted results, significantly improving the efficiency and operability of network optimization. Furthermore, the model supports a continuous optimization and update mechanism, ensuring that the system can dynamically adjust when base station parameters change or new data appears, always maintaining the effectiveness and foresight of the prediction model.

[0040] In summary, compared with existing technologies, this invention has significant advantages in data utilization, model adaptability, prediction accuracy, and result visualization. Its innovation lies in the deep application of artificial intelligence technology to the field of railway communication network management, constructing a fully intelligent system from data collection to analysis and application, significantly improving the scientific rigor and timeliness of network optimization. Furthermore, the continuous optimization mechanism of this invention provides reliable technical support for the intelligent operation and maintenance of railway communication networks, possessing good engineering practical value and promising prospects for widespread application. Attached Figure Description

[0041] Figure 1 This is the overall flowchart of the present invention;

[0042] Figure 2 This is a flowchart of the data preprocessing process of the present invention;

[0043] Figure 3 This is a flowchart illustrating the cell range estimation process using deep learning model training in this invention.

[0044] Figure 4 This is a flowchart illustrating the practical application of the present invention;

[0045] Figure 5 This is a flowchart illustrating the optimization and update process of the deep learning model in this invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and beneficial effects of this invention clearer and more explicit, the following detailed description of the artificial intelligence-based intelligent prediction method and system for railway communication cell range involved in this invention, in conjunction with the accompanying drawings, is provided. It should be understood that the following embodiments are only used to illustrate the principles and applications of this invention and do not constitute a limitation on the scope of protection of this invention.

[0047] like Figure 1 As shown, the overall process of the intelligent prediction method for railway communication cell range based on artificial intelligence in this invention covers five core steps: locomotive communication data acquisition, data preprocessing, cell range prediction through deep learning model training, GIS map display, and model optimization and updating. Through the organic connection of the above stages, this invention constructs a highly efficient data-driven intelligent prediction system integrating data acquisition, deep learning modeling, visualization, and adaptive model updating. This system can identify the actual coverage area of ​​railway communication cells with high accuracy in a short time, thus providing strong support for the maintenance, optimization, and planning of railway communication networks.

[0048] The method of this invention is implemented through an AI-based intelligent prediction system for railway communication cell range. This system includes: a data acquisition module, a data preprocessing module, a deep learning prediction module, a GIS display module, and a model optimization and update module. The data acquisition module collects multi-dimensional communication data such as latitude and longitude information, kilometer markers, cell IDs, RSRP, RSRQ, and SINR in real time during locomotive operation and uploads the data to a data center via a communication link. The data preprocessing module performs outlier cleaning, classification, storage, and vectorization on the received locomotive operation data, constructing the processed data into a tensor format acceptable to the deep learning model. The deep learning prediction module includes a dual-path convolutional neural network and a multi-head self-attention mechanism, used for feature extraction and feature fusion of the vectorized locomotive operation data, and based on Softmax... The output layer obtains the probability distribution of each cell and outputs the predicted cell ID and its confidence score; the GIS display module receives the predicted cell ID and confidence score output by the deep learning inference module, and performs visualization rendering of the predicted cell coverage and the current position of the locomotive based on GIS map technology; the model optimization and update module detects changes in base station parameters based on newly collected locomotive operation data, performs full model retraining when base station parameters change, and performs incremental learning when no changes occur, in order to update and deploy the deep learning model.

[0049] Specifically, the execution process of this inference method is as follows: First, a communication monitoring module is deployed on the locomotive equipment to continuously collect key real-time communication data during train operation, including latitude and longitude information, kilometer markers, cell IDs, RSRP (Reference Signal Receiving Power), RSRQ (Reference Signal Receiving Quality), SINR (Signal to Interference plus Noise Ratio), and other core signal quality parameters. This collected data is uploaded to the data center in real time via a stable data link. After receiving the raw data, the data center performs systematic preprocessing on the dataset and stores it according to cell IDs. Next, the preprocessed data is fed into a deep learning model for training and predictive inference. Using the trained and optimized deep learning model, the coverage area of ​​railway communication cells is intelligently inferred, and predicted cell IDs and their confidence scores are generated. The input cell ID refers to the raw data required for model training; the output cell ID refers to the result output during inference using the pre-trained model, which also includes a confidence score. The confidence score represents the probability of the input location belonging to a cell during inference, such as 0.98 or 0.95. This result is dynamically visualized using Geographic Information System (GIS) technology, allowing for an intuitive representation of the relationship between cell boundaries and the current location. A GIS map is a digital intelligent map that combines geospatial data (such as location) with attribute data to achieve the collection, storage, analysis, and visualization of geographic information. During operation, the system continuously performs automatic data collection and executes model optimization and updates according to a predetermined process, thereby continuously improving the accuracy and adaptability of cell prediction. When operators adjust base station parameters (such as new PCI physical cell identifiers or frequency changes), the system automatically initiates a new data collection task and executes it completely. Figure 1The entire process shown ensures that the deep learning model reflects the latest network status in real time, guaranteeing the accuracy and timeliness of the prediction results. In this invention, the intelligent cell range prediction process based on the deep learning model can be divided into two stages: a training stage and an actual prediction stage. In the training stage, the system uses a large amount of historical operating data collected by the locomotive during daily operation to construct training samples. Each sample includes at least: latitude and longitude information, kilometer markers, RSRP, RSRQ, SINR, etc., as input features, and the target cell ID actually accessed by the locomotive at that sampling time as a supervised learning label. The deep learning model iteratively updates the model parameters by minimizing the loss function between the predicted cell ID and the labeled cell ID, thereby learning the mapping between input features and cell coverage relationships. In the actual prediction stage, the system only needs to obtain the latitude and longitude or kilometer marker information of the current locomotive position and the corresponding signal quality parameters, without needing to input the currently accessed cell ID. The deep learning model outputs the predicted cell ID and its confidence score based on the above input features. Furthermore, to clarify the technical logic of the training and inference processes, this invention explicitly distinguishes the roles of "cell IDs (labels) in the training phase" and "predicted cell IDs (model outputs) in the inference phase." In the training phase, the cell IDs serve only as supervised learning labels. The model input includes only latitude and longitude information, kilometer markers, and signal quality parameters (RSRP, RSRQ, SINR) to learn the statistical relationship between different locations and cell coverage. In the inference phase, cell IDs are no longer input; only location and signal quality data are input. The trained model automatically outputs the predicted cell IDs and their probability values. The cell IDs in the training phase serve only as supervised learning labels, used to establish the mapping from location and signal quality features to cell coverage relationships. In the inference phase, cell IDs are no longer input; instead, the model automatically outputs the predicted cell IDs and confidence scores based on the input features. This invention establishes a mathematical mapping relationship of "point-cell coverage probability" through a deep learning model, rather than a fixed correspondence between serving cells and covered cells. Furthermore, at the data organization level, this invention treats each sampling point as a five-dimensional vector (Lat, Lon, Km, RSRP, RSRQ, SINR) of "latitude / longitude / kilometer marker + signal quality features," where Lat represents latitude coordinates, Lon represents longitude coordinates, and Km represents railway kilometer markers. Each cell ID corresponds to a set of points obtained from a large number of real-world operational samples. The geometric coverage area of ​​the cell can be approximately constructed through the distribution of this point set. In practice, a cell ID typically covers a circular or near-circular area, and its coverage boundary can be fitted with high accuracy through a spatial distribution model of a large number of points. This invention utilizes a deep learning model to learn this spatial coverage pattern from a large number of point samples, achieving probabilistic inference from points to cells.

[0050] like Figure 2As shown, the data preprocessing flow of this invention mainly includes three steps: raw data cleaning, data classification and storage, and model input construction. In the raw data cleaning stage, the system removes erroneous or abnormal records from the collected data. Specifically, this includes: removing abnormal location data with a single displacement exceeding 500 meters; removing records with signal values ​​below the threshold for maintaining normal communication (e.g., RSRP ≤ -120 dBm), and abnormal signal values ​​caused by measurement errors or interference due to proximity to the base station (e.g., RSRP ≥ -50 dBm); removing low-quality signal values ​​that may cause service interruption (e.g., RSRQ ≤ -20 dB) and high abnormal values ​​caused by neighboring cell interference or configuration abnormalities (e.g., RSRQ ≥ -3 dB); and removing low signal-to-noise ratio values ​​that cannot be demodulated (e.g., SINR ≤ -3 dB) and abnormally high values ​​caused by the absence of interference or measurement abnormalities (e.g., SINR ≥ 30 dB). The cleaned data is classified and stored according to cell ID for processing at the individual cell level in subsequent modeling. Subsequently, the system vectorizes the data, constructing it into a tensor format acceptable to neural networks, which serves as input data for the deep learning model. Further, this invention strictly represents each data sample as: x = (lat, lon, km, RSRP, RSRQ, SINR), i.e., an independent spatial sampling point. Multiple sampling points belonging to the same cell ID form the point set S_cell for that cell. The model needs to learn the spatial distribution and signal strength characteristics of the point set S_cell in multidimensional space. Through statistical learning of a large number of points, the model can infer the probability that any new input point will fall into which S_cell, thereby obtaining the predicted cell ID.

[0051] like Figure 3 The diagram shows a flowchart of the intelligent prediction of cell range through deep learning model training according to the present invention. The process covers two stages: model training and actual prediction. The whole process includes four core links: model input construction, feature extraction, feature fusion and prediction output.

[0052] During the model training phase, the preprocessed dataset is first divided into training, validation, and test sets in an 8:1:1 ratio. Then, based on two spatial information sources, input formats of "kilometer marker sequence + signal strength" and "latitude / longitude grid + signal strength" are constructed respectively, and fed into a one-dimensional convolutional neural network (1D CNN) and a two-dimensional convolutional neural network (2D CNN). Feature extraction is performed using a dual-path convolutional neural network of 1D CNN and 2D CNN. The resulting two feature vectors are then concatenated and fused into a 256-dimensional feature vector using a multi-head self-attention mechanism and dynamic signal strength weighting. Finally, this fused feature is processed through a fully connected layer and an output layer to generate probability values ​​for each cell, achieving accurate cell range estimation based on multimodal signal features.

[0053] The essential mathematical form of the model is: p = Softmax(W·Φ(x)), where x is the multi-dimensional feature vector of the input, Φ(x) is the high-dimensional representation vector after fusion of a dual-path convolutional neural network and attention, W is the trainable weight parameter matrix in the model, and Softmax is the probability normalization function. The Softmax output p = (p1, p2, ..., p_N) represents the probability that the input point belongs to different cell IDs. The predicted cell ID of the model is argmax(p i ), confidence score = max(p i This probabilistic output is clear and feasible. In the actual prediction phase, the input can come from real-time collected operational data or user-inputted kilometer markers / latitude and longitude information. After the same feature extraction and fusion processing as in the training phase, the predicted cell ID and its corresponding confidence score are directly output, and can be visualized using a GIS map. For example, if a user inputs longitude 110.439995 and latitude 39.029295, after processing by the pre-trained model, the output cell ID is 0BEC611, and the confidence score is 0.98.

[0054] In terms of model architecture, a 1D CNN is used to process kilometer marker sequence data, including four channels: normalized kilometer markers, RSRP, RSRQ, and SINR. It employs a three-layer one-dimensional convolutional structure (kernel_size=3, filters=32 / 64 / 128). Here, kernel_size represents the kernel length of the one-dimensional convolution (1D CNN, One-Dimensional Convolutional Neural Network), and kernel_size=3 indicates that the kernel covers three consecutive sequence sampling points. filters represents the number of convolutional kernels in the convolutional layer (i.e., the number of output channels). filters=32 / 64 / 128 means that the number of convolutional kernels in the three one-dimensional convolutional layers is 32, 64, and 128 respectively, corresponding to a progressively increasing number of output feature channels to extract richer sequence features. This is used to capture signal propagation features along the trajectory direction. RSRP, RSRQ, and SINR are typical metrics for wireless channel quality. A 2D CNN is used to process the latitude and longitude gridded signal data (divided into grids with a resolution of 0.01°, with the three channels corresponding to the grid mean values ​​of RSRP, RSRQ, and SINR, respectively), and an improved ResNet-18 structure is used to extract the two-dimensional spatial features of the coverage area. After dual-path feature extraction, the feature vectors are concatenated and input into a multi-head self-attention mechanism, typically with four heads. Signal intensity weights are dynamically assigned through a trainable weight matrix to generate a 256-dimensional space-signal joint representation. This representation undergoes high-order feature transformation through a fully connected layer with 512 neurons using the ReLU activation function. Finally, the probability distribution of each cell ID is output through a fully connected output layer (the number of neurons equals the total number of candidate cells) combined with the Softmax activation function, realizing end-to-end inference from the original multimodal signal to the cell prediction result. The probability value corresponding to each cell ID serves as the confidence score for that cell, used to quantify the likelihood that the input location belongs to each candidate cell. Each neuron in the Softmax output layer corresponds to a candidate cell ID, and the output value represents the predicted probability that the current input sample belongs to that cell. The system selects the cell ID corresponding to the category with the highest probability value as the predicted cell ID, and uses this highest probability value as the confidence score for the prediction result. The confidence score ranges from 0 to 1, with a higher value indicating a higher degree of reliability for the predicted cell ID. Simultaneously, to improve the engineering interpretability of the prediction results, this invention constructs a "circular or near-circular" display boundary for cell coverage based on the spatial distribution of the point set corresponding to each cell ID, using density analysis and the minimum coverage circle fitting method. This ensures intuitive consistency between the model's prediction results and the GIS map display, allowing maintenance personnel to directly observe the spatial outline of the predicted cell range.

[0055] likeFigure 4 As shown, this invention proposes a system process for predicting cell locations based on user input information and visualizing the positioning results on a GIS map. The system first requires the user to input the necessary information on a web page. The web interface adopts a front-end / back-end separation architecture: the front-end interface is based on JavaScript and developed using the Vue framework within the Node.js runtime environment. Leveraging Vue's rich component ecosystem and flexible two-way data binding mechanism, a dynamic, responsive, and user-friendly interface is constructed. The back-end service is built using Java and the Spring Boot framework. Utilizing Spring Boot's efficient dependency management, embedded server support, and modular configuration capabilities, the system's development efficiency and maintainability are significantly improved.

[0056] When a user inputs the locomotive's current location's latitude, longitude, or kilometer marker information into the management interface, the system immediately initiates the model prediction process. It calls upon a pre-trained and optimized deep learning model to rapidly analyze and infer the input data. The inference result includes the cell ID and its corresponding confidence score, used to measure the reliability of the prediction. The cell ID and corresponding confidence score in the inference result refer to the cell to which the user-input latitude, longitude, or kilometer marker information belongs and the probability of belonging to that cell. The model's output includes the target cell's unique identifier cell ID and its corresponding confidence score, used to measure the reliability of the prediction. Simultaneously, the system represents the model's final layer output as a vector of length N, p = (p1, p2, ..., p_N), where each p... i This represents the predicted probability that the current input sample belongs to the i-th candidate cell, satisfying p i ≥ 0 and Σp i = 1, the system selects the cell ID corresponding to the category with the highest probability as the "predicted cell ID", and uses this maximum probability value as the confidence score of the prediction result, i.e., Predicted Cell = argmax(p i ), Confidence Score = max(p i Predicted Cell is the predicted cell ID, i.e., the predicted cell identifier. Confidence Score is the confidence score, i.e., the maximum value of the predicted probability. The larger the value, the higher the confidence level of the model in the prediction result.

[0057] After the prediction results are generated, the system will use Geographic Information System (GIS) technology to dynamically overlay and render the geographical boundary of the target cell with the locomotive's current real-time location, and present it to the user intuitively through map visualization. This visualization not only clearly shows the coverage area of ​​the communication cell, but also assists maintenance personnel in accurately identifying and managing railway communication areas, thus providing strong data support for subsequent communication maintenance, optimization, and decision-making.

[0058] like Figure 5 The diagram illustrates the process for optimizing and updating the deep learning model in this invention. The system first collects a new dataset and determines whether the base station parameters have changed. If the base station parameters have changed, such as new PCI physical cell identifiers or frequency adjustments, a full model retraining is performed to generate a new model. If the base station parameters have not changed, incremental learning is used to fine-tune the existing model, thereby improving its adaptability and accuracy while maintaining its stability. Finally, the updated model is deployed to the production environment to ensure that the intelligent prediction system for railway communication cell ranges continues to maintain optimal performance.

[0059] During the daily operation of the locomotive, the system automatically collects multi-dimensional data, including latitude and longitude, kilometer markers, cell IDs, and signal strength such as RSRP, RSRQ, and SINR, and stores it in the database. It also records the collection timestamp and the corresponding base station version number. The system monitors changes in base station parameters in real time. When adjustments are detected, it automatically triggers the data preprocessing and model training modules to model and train the changed data, ensuring the accuracy and timeliness of the prediction results. If no parameter changes occur, an incremental fine-tuning strategy is implemented to ensure the model maintains stable and accurate predictive capabilities in constantly changing real-world environments. Finally, the updated deep learning model is deployed to the production environment, completing the optimization and iteration of the intelligent cell-range prediction method.

Claims

1. A method for intelligently predicting the range of railway communication cells based on artificial intelligence, characterized in that, Includes the following steps: A. Locomotive communication data acquisition: Acquire locomotive operation data including location information, kilometer markers, cell IDs, and signal quality parameters; the signal quality parameters include RSRP, RSRQ, and SINR, and the cell ID is used only as a supervised learning label during the training phase and not as model input data; B. Data Preprocessing: The locomotive operation data is cleaned, outlier removed, classified, stored, and formatted to generate input data that can be used by deep learning models. The data preprocessing includes cleaning the locomotive operation data to remove outliers, classifying and storing it according to cell ID, and then vectorizing it to generate a tensor format that can be accepted by deep learning neural networks. C. Cell Range Inference Through Deep Learning Model Training: Preprocessed locomotive operation data is input into a deep learning model for training and inference. The model outputs predicted cell IDs and their confidence scores, and the coverage area of ​​the cells is determined based on the prediction results. The deep learning model includes a dual-path convolutional neural network, which includes a one-dimensional convolutional neural network and a two-dimensional convolutional neural network. The one-dimensional convolutional neural network is used to extract features from the first input, which consists of kilometer marker sequence data and RSRP, RSRQ, and SINR signal quality parameters corresponding to the kilometer marker sequence. The two-dimensional convolutional neural network is used to extract features from the second input, which consists of latitude and longitude gridded spatial data and RSRP, RSRQ, and SINR signal quality parameters corresponding to the latitude and longitude grid. The feature vectors output by the two convolutional neural networks are concatenated, and feature fusion is performed through a multi-head self-attention mechanism and a dynamic signal strength weighting method to obtain a joint representation of spatial signals. The fused features are input into a fully connected layer, and the probability distribution of each cell ID is output through the output layer combined with a probability normalization function. The cell ID corresponding to the category with the highest probability value is selected as the target cell ID, and the highest probability value is used as the confidence score. D. GIS Map Display: Dynamically visualize the inferred results using a geographic information system; E. Model Optimization and Update: The deep learning model is updated based on newly collected locomotive operation data to adapt to changes in base station configuration and improve prediction accuracy.

2. The intelligent estimation method for railway communication cell range based on artificial intelligence according to claim 1, characterized in that, The data preprocessing steps include: B1. Raw Data Cleaning: Remove abnormal data where the same locomotive's single displacement exceeds 500 meters; remove low signal values ​​with RSRP less than or equal to -120 dBm that cannot maintain normal communication connections, and abnormally high signal values ​​with RSRP greater than or equal to -50 dBm caused by measurement errors or being too close to the base station; remove low-quality signal values ​​with RSRQ less than or equal to -20 dB that cause service interruption, and abnormally high signal values ​​with RSRQ greater than or equal to -3 dB caused by neighboring cell interference or parameter configuration errors; remove low signal-to-noise ratio values ​​with SINR less than or equal to -3 dB that cannot complete signal demodulation, and abnormally high signal-to-noise ratio values ​​with SINR greater than or equal to 30 dB caused by no interference or measurement abnormalities. B2. Data Classification and Storage: The cleaned locomotive operation data is classified and stored according to the cell ID, so as to facilitate subsequent modeling at the single cell level; B3. Vectorization: Vectorize the data for each category to generate tensor formats that can be accepted by deep learning neural networks, which are then used as input data for the model.

3. The intelligent estimation method for railway communication cell range based on artificial intelligence according to claim 1, characterized in that, The steps for predicting cell range through deep learning model training include: C1. Divide the preprocessed dataset into training set, validation set, and test set; C2. Based on the kilometer marker sequence and signal strength input, as well as the latitude and longitude grid and signal strength input, respectively, they are fed into a dual-path convolutional neural network for feature extraction; C3. The feature vectors extracted by the dual-path convolutional neural network are concatenated, and feature fusion is performed through a multi-head self-attention mechanism and a dynamic weighting method of signal intensity to obtain a joint representation of spatial signals. C4. Input the fused features into the fully connected layer for high-order feature transformation, and output the probability distribution of each cell through the output layer combined with the Softmax activation function to obtain the target cell ID and its confidence level. C5. In the actual prediction stage, the real-time locomotive operation data is processed using the same feature extraction and fusion methods as in the training stage to obtain the cell prediction results and display them visually.

4. The intelligent prediction method for railway communication cell range based on artificial intelligence according to claim 1, characterized in that, The GIS map display includes: D1. Receive latitude and longitude information or kilometer marker information input by the user; D2. Call the trained deep learning model to output the target cell ID and its corresponding confidence score; D3. Based on GIS map technology, the predicted geographical boundaries of the community and the current real-time location of the locomotive are dynamically visualized and rendered to realize the identification and management of communication areas.

5. The intelligent estimation method for railway communication cell range based on artificial intelligence according to claim 1, characterized in that, The steps for model optimization and updating include: E1. Automatically collect locomotive operation data, including latitude and longitude information, kilometer markers, cell ID, RSRP, RSRQ and SINR, during the daily operation of the locomotive, and store the data along with the collection timestamp and the corresponding base station version number information in the database; E2. Real-time monitoring of base station parameter changes. If changes in base station parameters are detected, including new PCI or frequency adjustment hardware configuration changes, the data preprocessing step and deep learning model training step are invoked to perform full model retraining on the changed locomotive operation data and generate a new deep learning model. E3. If the base station parameters have not changed, incremental learning is performed only on the newly added locomotive operation data to fine-tune the parameters of the existing deep learning model, obtain the updated deep learning model, and deploy the deep learning model to the production environment to complete the iterative optimization.

6. The intelligent estimation method for railway communication cell range based on artificial intelligence according to claim 3, characterized in that, The dual-path convolutional neural network includes a one-dimensional convolutional neural network and a two-dimensional convolutional neural network. The one-dimensional convolutional neural network is used to process kilometer marker sequence data, including four channels: normalized kilometer markers, RSRP, RSRQ, and SINR. It adopts a three-layer one-dimensional convolutional structure with a kernel size of 3 and the number of kernels being 32, 64, and 128 respectively. The two-dimensional convolutional neural network is used to process signal data after latitude and longitude gridding. The grid is divided with a resolution of 0.01 degrees, and the three channels correspond to the grid mean values ​​of RSRP, RSRQ, and SINR, respectively. An improved ResNet18 structure is used to extract two-dimensional spatial features of the coverage area. The feature vectors output by the one-dimensional convolutional neural network and the two-dimensional convolutional neural network are concatenated and input into a multi-head self-attention mechanism. The signal intensity weights are dynamically allocated through a training weight matrix to generate a 256-dimensional spatial signal joint representation. The spatial signal joint representation undergoes high-order feature transformation through a fully connected layer with 512 neurons and using the ReLU activation function. Finally, the probability distribution and confidence score of each cell ID are output through a fully connected output layer combined with the Softmax activation function.

7. An intelligent prediction system for railway communication cell range based on artificial intelligence, characterized in that, The system includes: a data acquisition module, a data preprocessing module, a deep learning inference module, a GIS display module, and a model optimization and update module. The data acquisition module collects latitude and longitude information, kilometer markers, cell IDs, RSRP, RSRQ, and SINR multi-dimensional communication data in real time during locomotive operation and uploads the data to the data center via a communication link. The data preprocessing module performs outlier cleaning, classification, storage, and vectorization on the received locomotive operation data, constructing the processed data into a tensor format acceptable to the deep learning model. The deep learning inference module includes a dual-path convolutional neural network and a multi-head self-attention mechanism, used for feature extraction and feature fusion on the vectorized locomotive operation data, and obtains the probability distribution of each cell based on a probability normalization function, outputting the predicted cell ID and its confidence score. The dual-path convolutional neural network includes a one-dimensional convolutional neural network and a two-dimensional convolutional neural network. The one-dimensional convolutional neural network extracts features from the kilometer marker sequence data and the first input consisting of RSRP, RSRQ, and SINR signal quality parameters corresponding to the kilometer marker sequence. The two-dimensional convolutional neural network extracts features from the latitude and longitude gridded spatial data. The system extracts features from the data and the second input consisting of RSRP, RSRQ, and SINR signal quality parameters corresponding to the latitude and longitude grid. The deep learning inference module concatenates the feature vectors output by the two convolutional neural networks and fuses the features through a multi-head self-attention mechanism and a dynamic weighting method for signal strength to obtain a joint representation of the spatial signal. Based on the joint representation of the spatial signal, it outputs the predicted cell ID and its confidence score. The GIS display module receives the predicted cell ID and confidence score output by the deep learning inference module and performs visualization rendering of the predicted cell coverage and the current location of the locomotive based on GIS map technology. The model optimization and update module detects changes in base station parameters based on newly collected locomotive operation data. When the base station parameters change, it performs full model retraining; when no changes occur, it performs incremental learning to update and deploy the deep learning model.

Images