A method, system, and storage medium for dynamic optimization of wireless communication in rail transit

Abstract

This invention discloses a dynamic optimization method, system, and storage medium for wireless communication in rail transit, belonging to the field of rail transit technology. The method includes: receiving scene feature data; matching and calling corresponding scene-specific sub-models based on the scene feature data; issuing the called scene-specific sub-models and training rules to several edge nodes, driving the edge nodes to train the scene-specific sub-models respectively; receiving model training parameters from several edge nodes and dynamically aggregating them to generate a final global scene model; converting the final global scene model into wireless communication policy execution instructions; and issuing policy execution instructions queried through pre-execution to the corresponding edge nodes, instructing the edge nodes to configure parameters for the terminal devices. Based on a federated learning architecture, it uses transmitted model parameters instead of raw data, reducing data transmission and processing latency, avoiding policy lag, and mitigating data privacy issues, significantly improving cross-domain optimization capabilities.

Description

Technical Field

[0001] This invention belongs to the field of rail transit technology, and specifically relates to a dynamic optimization method, system and storage medium for rail transit wireless communication. Background Technology

[0002] Dedicated wireless communication systems for rail transit are core infrastructure for ensuring train dispatching and safe operation, and must meet the differentiated needs of different types of rail transit. Currently, coverage optimization for dedicated wireless communication in rail transit mainly adopts a centralized optimization scheme. Its core logic is "centralized data collection - central analysis and decision-making - network-wide parameter adjustment," primarily through a central server analyzing data from individual lines or equipment from the same manufacturer and generating strategies. This approach has certain limitations. Insufficient real-time performance: Centralized data transmission and analysis latency reaches 50-300ms, which cannot match the coherence time of high-speed rail channels and the dynamic requirements of subway scene switching; Weak cross-domain collaboration: Data from equipment from multiple manufacturers is not interconnected, data from different subway lines or railway sections is not interconnected, and rapid collaborative optimization is not possible in the event of sudden failures. Poor scenario adaptability: The subway or railway uses a unified optimization model and a single scenario model, which cannot distinguish the different needs of subway tunnels and elevated sections, railway mountainous areas and railway tunnels, and multi-line intersection / parallel sections, resulting in low optimization accuracy. Resource waste: Each line / section trains its own basic optimization model, resulting in a duplication rate of over 50% and significant loss of computing power.

[0003] Existing optimization solutions are mostly designed for single-type systems and are difficult to be compatible with complex scenarios across different types and manufacturers. There is an urgent need for a universal wireless communication dynamic optimization framework to achieve collaborative optimization where "data does not leave the local system and parameters flow globally". Summary of the Invention

[0004] To address the above problems, this invention provides a dynamic optimization method for wireless communication in rail transit, the method comprising: Receive scene feature data, match and call the corresponding scene-specific sub-model based on the scene feature data; The scenario-specific sub-models and training rules are sent to several edge nodes to drive the edge nodes to carry out scenario-specific sub-model training respectively. Receive model training parameters from several edge nodes, and dynamically aggregate and generate a final global model of the scene based on the training parameters; The final scenario global model is converted into wireless communication policy execution instructions; The strategy execution instruction obtained through pre-execution query is sent to the corresponding edge node, instructing the edge node to perform parameter configuration on the terminal device.

[0005] Furthermore, the method also includes: Receive feedback data from the terminal device regarding the execution of the parameter configuration; The scene feature threshold is corrected based on the feedback data, and the next iteration is triggered according to the iteration cycle and feedback data anomalies.

[0006] Furthermore, the generation of the scene feature data includes: Determine the rail transit type of the target rail transit system, wherein the rail transit type includes at least one of railway system and metro / light rail system; Based on the determined rail transit type, scene feature collection instructions are issued to several edge nodes; The system receives static and dynamic scene data collected by several edge nodes according to the acquisition instructions, and preprocesses the collected scene data to generate scene feature data.

[0007] Furthermore, matching and invoking the corresponding scene-specific sub-models includes: Cross-validation is performed on the scene feature data sent by several of the edge nodes; The target scene is identified by matching the scene feature data with a subdivided scene library through cross-validation; The corresponding scene-specific sub-model is invoked based on the locked target scene.

[0008] Furthermore, dynamically aggregating and generating the final scene global model based on the training parameters includes: Receive model training parameters from several edge nodes and perform anomaly verification on the model training parameters; Based on the training status labels and scene data quality evaluation indicators reported by each edge node, aggregate weights are dynamically assigned to the model training parameters of each edge node that has passed the anomaly check. The global values ​​of the model training parameters are determined by weighting. Determine the absolute difference between the training parameters of each edge node model and the global value, and take the maximum value of the absolute difference to generate a preliminary global scene model; The initial global model of the scene that meets the iteration termination condition is determined as the final global model of the scene.

[0009] Furthermore, the preliminary global model of the scenario that satisfies the iteration termination condition includes: Determine whether the preliminary global model of the scenario meets the preset iteration termination conditions. The iteration termination conditions include at least the global loss function change reaching the target or the key performance index value reaching the target. When the iteration termination condition is met, the preliminary scene global model is determined as the final scene global model.

[0010] Furthermore, converting the final scenario global model into wireless communication policy execution instructions includes: Obtain the real-time scene feature data of the final scene global model at the current moment; Analyze the scene correlation in real-time scene feature data involving the jurisdiction of at least two edge nodes; Based on the aforementioned scene correlation, the final scene global model is converted into a wireless communication optimization strategy; The wireless communication optimization strategy is subject to multi-dimensional compliance verification, including device hardware compatibility, industry standard compliance, and system initial constraint compliance. Based on a pre-defined protocol adaptation rule base, the verified optimization strategies are converted into private protocol instructions that can be recognized by the target device manufacturer, thereby generating wireless communication strategy execution instructions.

[0011] Furthermore, the pre-execution query of the strategy execution instruction includes: Send a policy pre-execution query request to the wireless switching center, the request carrying at least one wireless communication policy execution instruction to be issued and its metadata; Receive a response from the wireless switching center for the pre-execution query request, the response including a conflict determination result between the wireless communication policy execution instruction and the core network configuration; Based on the received response, if the determination result is no conflict, then an execution permission for the wireless communication policy execution instruction is generated and recorded. The wireless communication policy execution instruction that allows the execution of the permission status identifier is sent to one or more corresponding edge nodes.

[0012] Furthermore, the method also includes: Receive a transitional state report from at least one edge node, the report being generated and sent by the edge node after detecting that the offset of the real-time communication scene characteristics within its jurisdiction relative to the baseline steady-state scene exceeds a preset threshold; Based on the received transition state report, the global transition sub-model is invoked to generate target steady-state parameters that match the scene offset reported in the report; The generated target steady-state parameters are sent to the corresponding edge node that sent the transition state report, so as to trigger the edge node to configure the transition parameters of the device terminals in its jurisdiction based on the target steady-state parameters.

[0013] Furthermore, the method also includes: Receive emergency alarm information containing event type labels from at least one edge node. The information is generated and pushed by the edge node after it continuously monitors the device status and business indicators of the device terminals in the area under its jurisdiction and detects that at least one indicator exceeds a preset emergency threshold. Based on the received emergency alarm information, an emergency policy instruction is generated and sent to the corresponding edge node that sent the emergency alarm information to drive it to suspend non-critical services and prioritize the execution of emergency configuration. Receive emergency response progress reports synchronously reported by the edge nodes at fixed time intervals during the emergency configuration process; The system receives the emergency cancellation notification generated by the edge node after detecting that all emergency indicators have returned to normal range and have remained stable for a preset duration, and performs closed-loop processing on the emergency event.

[0014] The present invention also provides a dynamic optimization system for wireless communication in rail transit, the system comprising a federated server configured to: Receive scene feature data, match and call the corresponding scene-specific sub-model based on the scene feature data; The scenario-specific sub-models and training rules are sent to several edge nodes to drive the edge nodes to carry out scenario-specific sub-model training respectively. Receive model training parameters from several edge nodes, and dynamically aggregate and generate a final global model of the scene based on the training parameters; The final scenario global model is converted into wireless communication policy execution instructions; The strategy execution instruction obtained through pre-execution query is sent to the corresponding edge node, instructing the edge node to perform parameter configuration on the terminal device.

[0015] Another embodiment of the present invention provides a dynamic optimization method for wireless communication in rail transit, the method comprising: The system receives scene-specific sub-models and training rules from the federated server, wherein the scene-specific sub-models are determined by the federated server based on the received scene feature data through matching and invocation. Based on the training rules, scenario-specific sub-models are trained to generate model training parameters, and the model training parameters are sent to the federated server. The system receives a wireless communication policy execution instruction issued by the federation server through a pre-execution query. The wireless communication policy execution instruction is determined by the federation server dynamically aggregating and generating a final scene global model based on the training parameters, and then transforming the final scene global model. The terminal device executes parameter configuration based on the strategy execution instructions.

[0016] Furthermore, the training parameters for generating the model by training scene-specific sub-models include: Collect real-time communication data from terminal devices within the jurisdiction, and preprocess the real-time communication data; Based on the preprocessed real-time communication data, corresponding scene labels are generated to form a scene label dataset. Using the scene label dataset as training input, the parameters of the scene-specific sub-model are trained through a three-step iterative algorithm of prediction-comparison-parameter tuning; Determine whether the scene-specific sub-model has reached a convergence state; When the scenario-specific sub-model reaches the preset convergence condition ahead of schedule, the training parameters of the trained model are encrypted, and the encrypted training parameters are uploaded.

[0017] Furthermore, the method also includes: Monitor the real-time communication scene characteristics of equipment terminals within the jurisdiction. When the deviation of the current scene characteristics from the benchmark steady-state scene exceeds a preset threshold, start the transition scene processing procedure. After initiating the transition scenario processing procedure, a transition status report is sent to the federated server; Receive target steady-state parameters generated by the global transition sub-model based on the transition state report issued by the federated server; Based on the target steady-state parameters, the transition parameter configuration for the equipment terminals within the jurisdiction is initiated.

[0018] Furthermore, the method also includes: Real-time monitoring of scene feature changes resulting from device terminal configuration; During the transition execution process, the proportion of the steady-state characteristics corresponding to the transition target in the total characteristics is continuously monitored; When the percentage continuously exceeds the first threshold and the duration reaches the preset window, the transition is determined to be over, and the execution of the relevant transition strategy is stopped.

[0019] Furthermore, the method also includes: Continuously monitor the equipment status and business indicators of terminals within the jurisdiction; When at least one indicator is detected to exceed a preset emergency threshold, an emergency alarm message containing an event type label is generated and the emergency alarm message is pushed to the federated server. Upon receiving the emergency policy instruction generated by the federal server based on the emergency alarm information, suspend all currently executing non-critical business policies and prioritize the execution of the emergency configuration operation in the emergency policy instruction; During the emergency configuration process, the current emergency response progress report is synchronized to the federal server at fixed time intervals; Continuously monitor key emergency-related indicators. When all emergency indicators have returned to normal range and have remained stable for a preset duration, generate an emergency deactivation notification and push the notification to the federated server.

[0020] The present invention also provides a dynamic optimization system for wireless communication in rail transit, the system comprising a plurality of edge nodes, the plurality of edge nodes being configured as follows: The system receives scene-specific sub-models and training rules from the federated server, wherein the scene-specific sub-models are determined by the federated server based on the received scene feature data through matching and invocation. Based on the training rules, scenario-specific sub-models are trained to generate model training parameters, and the model training parameters are sent to the federated server. The system receives a wireless communication policy execution instruction issued by the federation server through a pre-execution query. The wireless communication policy execution instruction is determined by the federation server dynamically aggregating and generating a final scene global model based on the training parameters, and then transforming the final scene global model. The terminal device executes parameter configuration based on the strategy execution instructions.

[0021] Another embodiment of the present invention provides a dynamic optimization method for wireless communication in rail transit, the method comprising: The federated server receives scene feature data from several edge nodes, matches and calls the corresponding scene-specific sub-model based on the scene feature data; The federated server sends scenario-specific sub-models and training rules to several edge nodes, driving the edge nodes to train the scenario-specific sub-models respectively. The edge node receives the scene-specific sub-model and training rules issued by the federation server, performs scene-specific sub-model training based on the training rules to generate model training parameters, and sends the model training parameters to the federation server. The federated server receives model training parameters from several edge nodes and dynamically aggregates them to generate the final global model of the scene based on the training parameters; The federated server converts the final scenario global model into wireless communication policy execution instructions, and sends the policy execution instructions, which are pre-executed through the wireless switching center, to the corresponding edge nodes, instructing the edge nodes to perform parameter configuration on the terminal devices. The edge node receives the wireless communication policy execution instruction issued by the federation server through pre-execution query, and performs parameter configuration on the terminal device based on the policy execution instruction.

[0022] Furthermore, the method also includes an initialization step, which includes: Determine the rail transit type of the target rail transit system, wherein the rail transit type includes at least one of railway system and metro / light rail system; Request the initial parameter constraint range from the wireless switching center based on the determined rail transit type; The federated server stores and encrypts the parameter constraint range locally.

[0023] This invention also provides a dynamic optimization system for wireless communication in rail transit, the system comprising: a federated server, several edge nodes, and a wireless switching center. The federated server is configured to receive scene feature data from several edge nodes, and match and call the corresponding scene-specific sub-model based on the scene feature data; The federated server is also configured to issue scenario-specific sub-models and training rules to several edge nodes, driving the edge nodes to carry out scenario-specific sub-model training respectively. Several edge nodes are connected to the federation server and are configured to receive scene-specific sub-models and training rules issued by the federation server, train scene-specific sub-models based on the training rules to generate model training parameters, and send the model training parameters to the federation server. The federated server is also configured to receive model training parameters from several of the edge nodes and dynamically aggregate them to generate a final global model of the scene based on the training parameters. The federated server is also configured to convert the final scenario global model into wireless communication policy execution instructions; The wireless switching center, which communicates with the federation server, is configured to pre-execute the policy execution instructions queryed; The federated server is also configured to send the policy execution instructions, which are pre-executed through the wireless switching center, to the corresponding edge nodes, instructing the edge nodes to perform parameter configuration on the terminal devices; The edge node is also configured to receive wireless communication policy execution instructions issued by the federation server through pre-execution queries, and to perform parameter configuration on the terminal device based on the policy execution instructions.

[0024] Another embodiment of the present invention also provides a computer storage medium storing one or more instructions that, when executed by one or more computers, cause the one or more computers to perform the method described in the present invention.

[0025] Compared with the prior art, the present invention has the following advantages: The present invention relates to a method, system, and storage medium for dynamic optimization of wireless communication in rail transit. Based on the federated learning architecture, the method of replacing the original data with transmission model parameters significantly reduces the amount of data interaction, reduces the latency of data transmission and processing, enables the optimization strategy to respond in a timely manner to the rapid changes in the channel under high-speed mobile scenarios, avoids policy lag, avoids data privacy issues, and significantly improves cross-domain optimization capabilities.

[0026] By calling dedicated sub-models for different scenarios, the wireless communication parameters of each scenario are accurately matched with the characteristics of the scenario, thereby improving the quality of dedicated wireless communication in different scenarios and solving the problem that a single model cannot adapt to multiple scenarios. By collecting data in real time at edge nodes and iterating the global model at high frequency by federated servers, the response latency of optimization strategies is greatly shortened, which can accurately adapt to the dynamic scenario characteristics of rail transit.

[0027] It has a built-in multi-vendor protocol conversion module, which can be directly adapted to mainstream equipment without the need for additional customized tools, significantly reducing the cost of upgrading existing lines and the difficulty of deploying new lines; it breaks down the data barriers of single lines and single vendors, and achieves cross-line and cross-vendor model parameter collaborative aggregation through federated learning. It can formulate global optimization strategies without relying on direct interaction of raw data, improve cross-regional collaborative response efficiency, and reduce the need for repeated development of basic models for each line by sharing a global model, thereby reducing computing power consumption.

[0028] By using scenario-specific sub-models and transitional collaborative processes, precise parameter matching across all scenarios is achieved, effectively reducing the risk of communication interruptions during transitional scenarios and improving overall coverage stability. Resource configuration is dynamically adjusted based on real-time load to improve resource utilization efficiency. The emergency response process skips unnecessary compliance checks, prioritizing core business operations and shortening fault handling latency. Based on scenario prediction, sub-models are pre-loaded and strategies are smoothly switched, reducing wireless signal fluctuations during scenario switching and lowering the probability of communication interruptions. Furthermore, when adding new scenarios, only the corresponding sub-model needs to be added to complete the adaptation without system reconstruction, significantly improving system scalability.

[0029] The "scene recognition-federated learning-dynamic optimization" architecture constructed in this invention has strong adaptability and can flexibly accommodate different rail transit types, equipment from different manufacturers, and various sub-scenarios. Through modular design, the scene recognition module can expand feature dimensions according to new scenarios, the federated learning framework can adapt to differences in protocols from different manufacturers by loading new sub-models, and the dynamic optimization module can automatically adjust the strategy output form based on scenario characteristics. Without reconstructing the core architecture, it can achieve compatibility and coverage of existing and future new rail transit scenarios and equipment types, forming a generalized technical system of "one-time architecture construction, flexible adaptation to multiple scenarios".

[0030] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0031] 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.

[0032] Figure 1 A schematic diagram of the structure of a dynamic optimization system for wireless communication coverage in rail transit according to an embodiment of the present invention is shown. Figure 2 This illustration shows a flowchart of a closed loop formed by scene adaptation, model training, strategy implementation, and feedback iteration in an embodiment of the present invention. Figure 3 A schematic diagram of the system initialization workflow according to an embodiment of the present invention is shown; Figure 4 A schematic diagram of the dynamic optimization method for wireless communication in rail transit according to Embodiment 1 of the present invention is shown. Figure 5 A schematic diagram of the dynamic optimization method for wireless communication in rail transit according to Embodiment 2 of the present invention is shown; Figure 6 This diagram illustrates the scene recognition and sub-model distribution process according to an embodiment of the present invention. Figure 7 A schematic diagram of the edge training and parameter aggregation process according to an embodiment of the present invention is shown; Figure 8 This diagram illustrates the policy generation and compliance verification process according to an embodiment of the present invention. Figure 9 A schematic diagram of the exchange center collaboration and execution process according to an embodiment of the present invention is shown; Figure 10 A schematic diagram of the feedback loop and iterative process according to an embodiment of the present invention is shown. Detailed Implementation

[0033] 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.

[0034] This invention provides a dynamic optimization system and method for rail transit wireless communication coverage based on federated learning. By constructing a three-tier architecture of "federated server - edge node - device terminal," and employing a full-process approach of "edge local training + federated parameter aggregation + dynamic policy distribution," it systematically addresses the pain points of existing technologies. The aim is to achieve precise optimization across all rail transit scenarios, improving communication quality in each scenario; by transmitting model parameters instead of raw data, optimization latency is reduced, adapting to channel changes under high-speed movement and improving handover success rate; data silos are broken down, solving the problem of excessively long cross-regional data synchronization latency, improving the real-time performance of collaborative optimization, and reducing computational resource waste; policy gaps during scenario switching are eliminated, achieving seamless handover and improving system scalability.

[0035] The three-tiered federated learning architecture of "federated server - edge node - device terminal" in this embodiment of the invention aims at both data privacy protection and dynamic and precise optimization. Through the collaborative operation of the three-tiered nodes, a closed-loop optimization process of "scene recognition - model training - policy implementation - feedback iteration" is formed. The dynamic optimization system for wireless communication coverage in rail transit includes the collaborative operation of key components such as federated servers, edge nodes, and device terminals. The three interact with each other through a hybrid wireless + wired network, breaking through the limitations of traditional centralized architectures and achieving full coverage optimization for different rail transit scenarios such as subways and railways.

[0036] In this embodiment of the invention, the collaborative operation of the three-level nodes refers to the specific content of edge nodes being responsible for local data processing and training, federated servers being responsible for global parameter aggregation and policy generation, and device terminals being responsible for data collection and instruction execution. The following is a detailed description of the specific content of the three-level federated learning architecture: Figure 1 A schematic diagram of the structure of a dynamic optimization system for wireless communication coverage in rail transit according to an embodiment of the present invention is shown. Figure 1The dynamic optimization system architecture includes a federated server deployed at the metro control center / railway bureau dispatching center, specifically at the rail transit wireless communication switching center. This switching center includes TETRA (Terrestrial Trunked Radio) trunking switches, LTE-M (Long Term Evolution for Metro) core networks, GSM-R (Global System for Mobile Communications - Railway) mobile switching centers, and 5G-R core networks. This embodiment does not limit the selection of the wireless switching center. The federated server is also equipped with a display terminal. The dynamic optimization system architecture also includes edge nodes that communicate with the federated server via the rail transit transmission network / data communication network / fiber optic cable. These edge nodes are located in stations / sections and are lightweight distributed computing units adapted to rail transit scenarios, focusing on real-time data processing and lightweight communication. The devices feature a lightweight design and can be embedded in base stations, radios, RRUs (Remote Radio Units), BTS (Base Transceiver Stations), and AAUs (Active Antenna Units) along the rail transit line. In equipment such as units, the power and communication interfaces of base stations, radios, RRUs, BTSs, and AAUs are fully utilized, avoiding the space occupation of separate equipment deployment. This equipment form is suitable for new line construction scenarios. Alternatively, it can be deployed in the cabinets or equipment boxes of key facilities such as base stations, radios, RRUs, and BTSs along the rail transit line, and connected to base stations, RRUs, and other equipment through wired interfaces to read their operating data. This equipment form is suitable for both existing line renovation scenarios and new line construction scenarios. Dynamic optimization of the system structure also includes communication connections with edge nodes. Onboard terminal equipment includes TAU (Train Access Unit) and CIR (Cab Integrated Radio communication equipment).

[0037] The following describes the main functions of the federated server, edge nodes, and vehicle-mounted terminal devices: I. Federation Server: The federated server is the core hub for global decision-making and collaborative optimization. It is responsible for coordinating the local training of edge nodes, aggregating global models, and generating optimization strategies, while solving the problems of cross-domain collaboration and scenario adaptation. The federated server is deployed at the rail transit wireless communication exchange center. It can be deployed at the line level, such as Metro Line 1 or Railway Line XX; or at the network level, such as the Metro Operation Control Center or the Dispatching Center of the Railway Bureau Group Company; or, depending on the current status of the railway, it can be deployed at the China State Railway Group to perform dynamic optimization for all railway lines.

[0038] In this embodiment of the invention, the core function of the federated server is not to directly process raw data, but to achieve global optimization only through model parameter interaction. As the parameter aggregation center of the federated learning framework, it serves two main purposes: first, it connects to the exchange centers of various wireless systems, collaborating with them on parameter constraint synchronization and instruction conflict judgment; second, it connects to the edge nodes set up at various locations in the rail transit system, receiving model parameters uploaded by these edge nodes, aggregating them to generate a globally optimized model, and then distributing sub-models and optimization strategies to the edge nodes, achieving cross-domain collaboration where "data does not leave the local machine, but parameters flow globally." The main functions are as follows: (1) Functional coordination with wireless switching centers ① Real-time synchronization of parameter constraints The wireless switching center proactively pushes parameter constraint ranges to the federation server through a standardized interface, with an adjustable update frequency to ensure the federation server receives the latest constraints. For example, after the core network adjusts the power limit of a base station from 43dBm to 41dBm, it immediately synchronizes this to the federation server, ensuring that the optimization instructions it generates strictly adhere to this range. ② Collaborative verification before instruction execution After generating optimization instructions (such as "adjust power to 40dBm"), the federated server first sends a "pre-execution query" (including instruction parameters and target base station ID) to the wireless switching center via a dedicated link. The wireless switching center returns "execution allowed" or "conflict warning" (such as "this base station is currently performing core network configuration and adjustment is not allowed at this time") within 100ms. The federated server then decides whether to issue the instruction based on the feedback.

[0039] (2) Scene recognition and sub-model management In this embodiment of the invention, the federated server cross-validates and precisely divides the scene into subdivisions based on scene features uploaded by multiple adjacent edge nodes. These subdivisions can be selected as the middle section of a long tunnel, tunnel entrances and exits, etc., to build a subdivision scene library and match a dedicated sub-model for each scene.

[0040] The scenario library includes subdivided scenarios such as subway (tunnels / elevated roads / stations) and railway (plains / mountainous areas / long tunnels / railway hubs / intersecting parallel lines), with each scenario associated with typical signal characteristics (e.g., tunnel multipath delay > 100ns, mountainous area occlusion coefficient > 0.6). The sub-model library contains pre-built, targeted optimized models, such as a tunnel anti-multipath sub-model, a high-speed Doppler compensation sub-model, and a station hall load balancing sub-model. When edge nodes upload new scenario features, the sub-model adaptation mechanism is automatically triggered, calling the sub-model of the most similar scenario and dynamically adjusting parameters.

[0041] Optionally, in this embodiment of the invention, the tunnel anti-multipath sub-model is used to optimize the reflection phase suppression multipath of RIS (Reconfigurable Intelligent Surface), the high-speed Doppler compensation sub-model is used to dynamically adjust the carrier frequency to offset frequency offset, and the station hall load balancing sub-model is used to optimize the resource allocation of PRB (Physical Resource Block).

[0042] (3) Model parameter aggregation and global optimization In this embodiment of the invention, the federated server maintains its own global sub-model for each sub-scenario (such as a subway tunnel or railway viaduct), receives local sub-model parameters uploaded by all edge nodes in that scenario, and aggregates them by dynamically allocating weights to generate a globally optimized model adapted to that scenario. This process only aggregates parameters within the same scenario and does not involve mixing parameters from different scenarios. Specific functions include: ① Dynamic weight allocation In this embodiment of the invention, the federated server allocates aggregation weights based on factors such as the criticality, data quality, and node reliability of edge nodes within the same scenario, avoiding impact on the global model. For example, regarding criticality, edge nodes in the middle section of long tunnels (where multipath interference is most severe, affecting CBTC train control) have higher weights than edge nodes at the entrances and exits of short tunnels in tunnel scenarios; edge nodes in high-speed rail hub sections of elevated railways have higher weights than nodes in ordinary elevated railway sections. In this embodiment, for example, the speed in high-speed rail hub sections of elevated railways is 350 km / h, with high switching requirements, while the speed in ordinary elevated railway sections is 200 km / h. Regarding data quality, nodes with higher data sampling rates and lower noise have higher weights. Regarding node reliability, nodes with better historical optimization performance and more stable online performance within the same scenario have higher weights. In this embodiment of the invention, the aggregation weight coefficient is set to 0.5-1.2, and no specific limitations are provided for this weight coefficient.

[0043] ② Differential privacy protection In this embodiment of the invention, the federated server adds Gaussian noise before aggregating parameters from the same scene to prevent the original data from being inferred from the parameters. Optionally, the noise intensity of the Gaussian noise is dynamically adjusted according to the scene sensitivity.

[0044] ③ Model Iterative Update The federated server in this embodiment of the invention supports "incremental aggregation." For example, the global sub-model for each scenario iterates independently, receiving edge node parameters and updating the global model every 5 minutes (this can be shortened to 1 minute for high-frequency scenarios such as high-speed rail sections), ensuring that the model adapts to the real-time communication status. Model updates stop once the training reaches the preset termination iteration condition, and an executable version of the global model is generated.

[0045] (4) Optimization strategy generation and distribution In this embodiment of the invention, the federated server transforms the global execution model for each scenario into specific strategies executable by edge nodes. Simultaneously, based on a collaborative optimization framework of sub-models for each scenario, it generates specific strategies for cross-scenario coordination, precisely adapting to the control interfaces of base stations from different vendors. It should be noted that this framework does not replace the independence of the sub-models for each scenario, but rather coordinates sub-model strategies to adapt to cross-scenario requirements.

[0046] ① Strategy output format Instructions for adjusting base station parameters include: LTE-M base station power, handover threshold, and leaky cable attenuation; GSM-R base station frequency and handover band length. These instructions are scenario-level general policies (such as power ranges in tunnel / elevated scenarios). Edge nodes need to combine local device characteristics and current load to convert them into equipment vendor-specific instructions.

[0047] Local training instructions for edge nodes include sub-model data acquisition instructions and feature weight adjustment instructions. Sub-model data acquisition instructions include instructions such as "prioritize acquiring signal samples with multipath delay > 80ns for tunnel edge nodes" and "increase the Doppler frequency offset sampling frequency for elevated nodes at speeds > 300km / h". Feature weight adjustment instructions include instructions such as "increase the weight of the user number feature in the load balancing sub-model during peak hours in the station hall".

[0048] ② Compatibility and Transition Handling The federated server in this embodiment of the invention integrates protocol conversion modules from mainstream manufacturers such as Huawei, ZTE, and Nokia, converting general policies into vendor-specific instructions (such as Huawei MML commands and ZTE CLI commands) to ensure that the policies are executable. Cross-scenario collaboration triggering conditions: When an edge node predicts that a train is about to enter the transition zone based on the location data of the vehicle terminal, it reports a scenario switching forecast to the federated server. After receiving consistent forecasts from three or more adjacent edge nodes, the server initiates the collaborative adjustment mechanism. Scene abrupt transition strategy: For example, when a train enters an elevated road from a tunnel, the federated server issues a "coordinated adjustment command" to the edge nodes of the tunnel / elevated road - the tunnel node gradually reduces its power (e.g., power decreases from 38dBm to 36dBm), and the elevated road node synchronously increases its power (e.g., power increases from 35dBm to 38dBm). The switching threshold smoothly transitions from 2dB in the tunnel scene to 5dB in the elevated road scene. It should be noted that this scene abrupt transition strategy is completed in 3 steps within a 100-meter transition zone to avoid signal abrupt changes.

[0049] (5) System monitoring and collaborative management In this embodiment of the invention, the federated server monitors the status of edge nodes, model training results, and policy execution feedback in real time to ensure stable system operation. Specifically: ① Core monitoring indicators Edge node and sub-model state: Basic edge node status: The edge node online rate should be ≥99.9%, data upload latency ≤200ms, and hardware load CPU utilization ≤70%. Sub-model running status: success rate of loading each technical module (≥99.8%, such as whether the anti-multipath sub-model is successfully activated), number of local training iterations (e.g., ≥5 times per hour to ensure model freshness).

[0050] Model performance: Global model optimization accuracy in various scenarios (e.g., handover success rate improvement ≥5%, SINR improvement ≥3dB); Collaborative effect across scenarios (e.g., handover failure rate reduction ≥8% at the tunnel-station junction).

[0051] Strategy execution and scenario linkage: Basic metrics: Strategy delivery success rate (e.g., success rate ≥ 99.5%), anomaly feedback response time (e.g., response time ≤ 10s), etc. Scene transition metrics: such as the smoothness of cross-scene strategy switching (e.g., power adjustment fluctuation ≤2dB / 50m) and the duration of communication interruption in the transition zone (e.g., communication interruption duration ≤50ms).

[0052] ② Emergency Response Node anomaly: When an edge node goes offline, adjacent nodes automatically take over the policy execution of its coverage area (e.g., if a tunnel exit node goes offline, the elevated entrance node temporarily expands its coverage by 50 meters) and loads the corresponding scene sub-model (e.g., temporarily enabling the anti-multipath sub-model to handle residual multipath). Strategy / Sub-model Abnormalities: If the signal quality deteriorates after the strategy is executed (e.g., SINR decreases by >3dB), immediately issue a "rollback strategy" and suspend the corresponding sub-model (e.g., disable the anti-multipath sub-model and enable the basic model); if the sub-model parameters are abnormal (e.g., Doppler compensation value jumps by >1000Hz), automatically load the backup sub-model parameters (based on historical best values) and trigger the federated server to re-aggregate the module parameters.

[0053] II. Edge Nodes In this embodiment of the invention, the core functions of the edge node are as follows: (1) Data collection Real-time acquisition of local data for dedicated wireless communication in rail transit. This data includes communication performance data, scenario-related data, and equipment operating status data. Communication performance data includes reference signal received power, signal-to-interference-plus-noise ratio, uplink and downlink throughput, handover success rate, radio resource block utilization, and multipath delay for LTE-M BBU and RRU equipment in subway scenarios; voice channel quality, call setup success rate, group call delay, and co-channel interference level for TETRA base stations and other equipment; and received level, carrier-to-interference ratio, number of handover failures, data transmission rate, and Doppler frequency offset for GSM-R and 400M wireless train dispatching equipment in railway scenarios. Scenario-related data (used for detailed scenario identification and matching with dedicated sub-models) includes dynamic scenario parameters (such as real-time train location, speed, acceleration, and location within the designated section) and static scenario parameters (base station coverage area type, leaky cable deployment method, base station spacing, and terrain obstruction coefficient). Equipment operating status data (to ensure equipment stability and avoid hardware damage from optimization strategies) includes base station transmit power (current value / maximum value), RF module temperature, fan operating status (normal / alarm), VSWR, power supply voltage, etc. (2) Local preprocessing In this embodiment of the invention, the edge node further includes filtering and normalizing the collected data, extracting "signal-scene" fusion features and scene parameters such as speed and position, and uploading the extracted scene feature labels and key feature values ​​to the federated server for the federated server to identify the current scene, while avoiding the uploading of the original data.

[0054] (3) Sub-model training and execution In this embodiment of the invention, the edge node loads the scenario-specific sub-model (such as the anti-multipath model for subway tunnels and the Doppler compensation model for high-speed railways) issued by the federated server, trains and generates model parameters using local preprocessed data, receives and executes optimization strategies issued by the server, and executes them through the base station control interface; collects the effect data after the strategy execution (such as the change in SINR after adjustment), and feeds it back to the federated server for model iteration.

[0055] (4) Cross-device collaboration and fault response Edge nodes share local information (such as neighboring cell signal strength) with neighboring edge nodes and collaboratively optimize handover bands (such as dynamically adjusting the handover threshold difference between neighboring base stations). When anomalies are detected in a base station, such as a sudden increase in VSWR or interference levels exceeding the threshold, the nodes quickly report to the federated server and initiate local emergency strategies (such as temporarily reducing transmission power to avoid equipment damage).

[0056] Throughout the system's operation, edge nodes receive scenario-specific sub-models from the federated server, train them using locally preprocessed data, generate model parameters, and upload them to the federated server. Simultaneously, edge nodes are also responsible for receiving optimization strategies generated by the federated server based on the global model, translating them into specific device-executable parameters, such as adjusting base station transmit power and optimizing handover thresholds, and executing these strategies in real time to ensure the stable operation of the wireless communication system. Furthermore, edge nodes have the ability to monitor the effectiveness of strategy execution in real time; upon detecting any anomalies, they immediately report back to the federated server for timely adjustments to the optimization strategies.

[0057] III. Terminal Equipment In this embodiment of the invention, the device terminal includes a wireless dynamic network optimization system configuration display terminal and existing device terminals in the urban rail transit dedicated wireless communication system. The system configuration display terminal serves as the operation and maintenance management entry point, and its core support includes network management monitoring, data configuration, and fault tracing. Its main functions include: first, real-time monitoring, visually presenting the status of devices such as federated servers and edge nodes, displaying scenario-specific KPIs and system operation indicators, and synchronizing constraint information from the wireless switching center; second, data configuration, managing basic device information, scenario sub-model parameters, and optimization strategy templates, supporting custom execution cycles and emergency policy issuance; and third, fault and security management, providing tiered alarms and linking emergency handling, recording operation logs, classifying user permissions, ensuring configuration compliance and data security, and connecting the federated server and the switching center to achieve configuration synchronization verification and fault work order dispatch.

[0058] Existing identification terminals include ground-based terminal equipment and vehicle-mounted terminal equipment. Ground-based terminal equipment includes base stations, station radios, RRUs, BTS, etc., while vehicle-mounted terminal equipment includes TAUs, CIRs, etc. These terminals maintain their original communication functions and do not participate in model storage or training processes; they only serve as data sources and policy execution terminals, collaborating through the following methods: (1) Data upload Station and trackside equipment such as base stations / radios / RRUs / BTS: Send communication performance data and equipment status data to edge nodes through wired interfaces. The data is uploaded after local preliminary filtering. Vehicle-mounted terminal: transmits data such as location, speed, and signal quality to edge nodes via vehicle-to-ground wireless links and trackside equipment. (2) Instruction execution Station and trackside equipment such as base stations / radios / RRUs / BTS: Receive commands such as power adjustment, frequency switching, and gain adjustment issued by edge nodes, execute them through vendor proprietary protocols (such as Huawei MML and ZTE CLI), and provide real-time feedback on the execution results; Vehicle-mounted terminal: Receives simple commands from edge nodes, such as antenna switching and receiver gain adjustment, and provides feedback on the execution results through status codes.

[0059] IV. Connection methods between different parts of the system Specifically, in this embodiment of the invention, the connection between the federated server and the edge node is as follows: the federated server and the edge node establish a connection based on the existing rail transit transmission network / data communication network / fiber optic cable. This is a stable, high-bandwidth wired transmission channel that supports data interaction between the two.

[0060] Connection between edge nodes and ground terminal equipment: Edge nodes are directly connected to ground terminal equipment in the same area via wired adapter interfaces, collect raw wireless data from the terminal side in real time, and send edge layer optimization commands to the terminal equipment.

[0061] Connection between edge nodes and vehicle-mounted terminal equipment: Edge nodes connect to vehicle-mounted terminal equipment via air interface and complete information exchange with vehicle-to-ground wireless channel.

[0062] Connection between the federated server and existing terminal devices: Terminal devices do not connect directly to the federated server; all data interactions must be relayed through edge nodes: Terminal → Edge Node → Backbone Network → Federated Server; the reverse is Federated Server → Backbone Network → Edge Node → Terminal Device. This approach ensures data privacy and meets the low-latency, high-reliability communication requirements of rail transit scenarios.

[0063] Figure 2 This diagram illustrates a closed-loop process of scene adaptation, model training, policy implementation, and feedback iteration, according to an embodiment of the present invention. Figure 2 In this process, the closed-loop process is entirely based on a three-tier architecture of "Federated Server - Edge Node - Device Terminal". It is achieved through the initialization of the wireless dynamic network optimization system, scene recognition and sub-model distribution, edge training and parameter aggregation, policy generation and compliance verification, exchange center collaboration and execution, and feedback loop and iteration. Specifically, the federated server leads scene recognition and global policy generation, the edge node undertakes local training and policy execution, and the device terminal provides data input and command implementation, ultimately achieving accurate and safe optimization of wireless communication coverage in rail transit.

[0064] Figure 3 A schematic diagram of the system initialization workflow according to an embodiment of the present invention is shown. Figure 3In the initialization process, after the system equipment is installed, networked, and debugged, the system is powered on and performs a self-test. Once communication between all parts is normal, the system configuration display interface begins initialization, displaying a basic data configuration library selection box. Users can select the appropriate basic data configuration library based on their rail transit type. The system is currently divided into two main categories of basic databases: railway and metro / light rail. The basic databases include a scene feature baseline library, a pre-set initial sub-model library, an equipment parameter template library, and an industry compliance standard library. After the database selection is completed, the system begins basic data configuration and sends a request to the wireless exchange center to synchronize the initial parameter constraint range. The federated server stores and encrypts the constraint parameters locally and pushes them synchronously to the system configuration display terminal monitoring panel, completing the entire initialization process of "equipment integration and debugging → boundary confirmation → monitoring loading." This establishes a safety baseline for subsequent processes. The system initialization is complete, and the system begins operation.

[0065] The following two specific embodiments illustrate the dynamic optimization method for wireless communication in rail transit according to the present invention: Example 1 Figure 4 This diagram illustrates the flow chart of the dynamic optimization method for wireless communication in rail transit according to Embodiment 1 of the present invention. Figure 4 The method includes: receiving scene feature data; matching and calling a corresponding scene-specific sub-model based on the scene feature data; sending the called scene-specific sub-model and training rules to several edge nodes to drive the several edge nodes to perform scene-specific sub-model training respectively; receiving model training parameters from several edge nodes and dynamically aggregating and generating a final scene global model based on the training parameters; converting the final scene global model into a wireless communication policy execution instruction; and sending the policy execution instruction queried through pre-execution to the corresponding edge node, instructing the edge node to perform parameter configuration on the terminal device.

[0066] Specifically, in this embodiment of the invention, the method further includes: receiving feedback data from the terminal device executing the parameter configuration; correcting the scene feature threshold based on the feedback data; and triggering the next iteration according to the iteration cycle and abnormal feedback data.

[0067] In this embodiment of the invention, the generation of scene feature data includes: determining the rail transit type of the target rail transit system, wherein the rail transit type includes at least one of railway system and metro / light rail system; issuing scene feature collection instructions to a number of edge nodes according to the determined rail transit type; receiving scene data containing static and dynamic data collected by the edge nodes according to the collection instructions, and preprocessing the collected scene data to generate scene feature data.

[0068] Specifically, in this embodiment of the invention, matching and calling the corresponding scene-specific sub-model includes: performing cross-validation on scene feature data sent by several edge nodes; locking the target scene based on the cross-validated scene feature data by matching a subdivided scene library; and calling the corresponding scene-specific sub-model according to the locked target scene.

[0069] Figure 6 This diagram illustrates the scene recognition and sub-model distribution process according to an embodiment of the present invention. Figure 6 In the process, the federated server sends scene feature collection instructions to the edge nodes. After receiving the instructions, the edge nodes collect static scene data (such as the deployment method of tunnel leaky cable and the spacing between elevated base stations) and dynamic scene data (such as the real-time speed of trains and multipath delay values) from the terminal devices, and perform noise reduction preprocessing on the data. Then, the standardized scene features are uploaded to the federated server. Next, the federated server cross-validates the data from multiple edge nodes (such as three adjacent nodes all being marked as "tunnel scene"), matches the subdivided scene library to lock the target scene (such as a long subway tunnel or a railway plain section). Then, it calls the scene-specific sub-model (such as a tunnel matching anti-multipath sub-model and an elevated matching Doppler compensation sub-model), and verifies the compliance of the sub-model parameters and adjusts the parameters in combination with the initial constraints. After that, the federated server encrypts and encapsulates the sub-model, attaches "local training rules", and sends it to the corresponding edge nodes through a secure link. After receiving the sub-model and local training rules, the edge nodes report the reception status to the federated server, and at the same time display the "sub-model sending completed" status on the system configuration display terminal.

[0070] Figure 7 This diagram illustrates the edge training and parameter aggregation process according to an embodiment of the present invention. Figure 7 In this process, through a closed-loop design of "local precise training + global collaborative optimization", efficient iteration of the communication model is achieved while ensuring data privacy. Specifically, edge nodes are triggered to carry out model training. The edge nodes determine whether the local model has converged. If the local model is determined to have converged, the trained parameters are encrypted and uploaded to the federated server. The federated server dynamically weights and aggregates to generate a preliminary global model of the scene. The federated server further determines whether the preliminary global model of the scene meets the iteration conditions. If the conditions are met, the iteration is stopped to form an iterative closed loop.

[0071] In this embodiment of the invention, the federated server dynamically aggregates and generates a final scene global model based on the training parameters, including: receiving model training parameters from several edge nodes and performing anomaly verification on the model training parameters; dynamically assigning aggregation weights to the model training parameters of each edge node that have passed the anomaly verification based on the training status label and scene data quality evaluation index reported by each edge node; determining the global value of the model training parameters using weighted average; determining the absolute difference between the model training parameters of each edge node and the global value, and taking the maximum value of the absolute difference to generate a preliminary scene global model; and determining the preliminary scene global model that meets the iteration termination condition as the final scene global model.

[0072] Optionally, determining the preliminary global model of the scene that meets the iteration termination condition includes: determining whether the preliminary global model of the scene meets the preset iteration termination condition, wherein the iteration termination condition includes at least the global loss function change reaching the standard or the key performance index value reaching the standard; when it is determined that the iteration termination condition is met, the preliminary global model of the scene is determined as the final global model of the scene.

[0073] In this embodiment of the invention, the process of dynamic weight aggregation, iterative judgment, and iterative closure of the federated server is also described in detail: Dynamic weight aggregation: After receiving parameters, the federated server first verifies them, including format compliance, encryption validity, and logical rationality. Abnormal parameters are marked "pending verification" and do not participate in aggregation. After successful verification, weights are dynamically allocated based on "training status label + scene data quality". Subsequently, the global value of each parameter is calculated by weighting, and the absolute difference between the cross-node parameters and the global value is calculated and the maximum value is taken to generate a preliminary global model and an "aggregation report". In this embodiment of the invention, the aggregation report includes the number of participating nodes, weight distribution, and deviation statistics. It should be noted that this embodiment of the invention does not provide specific limitations on the specific content and form of the aggregation report.

[0074] In this embodiment of the invention, the dynamic weight allocation can be as follows: the weight of the "precision convergence + stable data" node is 0.6, the weight of the "performance meets the standard + high fluctuation data" node is 0.3, the weight of the "number of times as a backup" node is 0.05, the weight of the "backup alternative" parameter is 0.05, and the total weight is 1.

[0075] Global model iteration judgment: The federated server judges the initial global model against the preset termination iteration conditions. If the conditions are met, the iteration is terminated; if not, the next round of training process is triggered.

[0076] Specifically, the global model iteration is judged primarily based on core conditions, with auxiliary conditions as a fallback: iteration terminates if any core condition is met and there are no conflicts with auxiliary conditions. Core conditions include changes in the global loss function and KPI values. For example, a global loss function change of <0.003 for three consecutive iterations indicates stable global accuracy; a maximum cross-node parameter deviation of <5% indicates parameter consistency; or a global model simulation KPI ≥ target value (e.g., network-wide switching success rate ≥99.5%). Auxiliary conditions include the global iteration cycle reaching its upper limit (e.g., one week) or optimization time exceeding a preset window (e.g., two hours before the morning rush hour on the subway). If these conditions are triggered, iteration is forcibly terminated. If the core condition is met and no auxiliary condition is triggered, "global convergence, iteration terminated" is determined. If the core condition is not met, or an auxiliary condition is triggered but the core condition is not met, "iteration needs to continue" is determined, and an "iteration adjustment suggestion" is generated. In this embodiment of the invention, the descriptions of core and auxiliary conditions are merely illustrative and can be adapted to specific implementation scenarios.

[0077] Iterative Loop or Deployment Preparation: If the termination conditions are met, the federated server stores the final global model for each scenario in the model library, synchronously pushes it to the system configuration display terminal, marks it as "Model Optimization Complete," and triggers the subsequent "Policy Generation and Compliance Verification" sub-process. If the criteria are not met, iterative iteration continues. Based on the "Iterative Adjustment Suggestions," the "Training Task Rule Package" is updated, such as shortening the upload cycle, increasing the local termination threshold, increasing the data collection frequency of high-weight nodes, issuing a new round of training instructions to all edge nodes with the current global model parameters as initial values, and repeating the "Local Training → Parameter Upload → Federated Aggregation → Iterative Judgment" process until the model meets the criteria.

[0078] In this embodiment of the invention, converting the final scene global model into wireless communication strategy execution instructions includes: acquiring real-time scene feature data of the final scene global model at the current moment; analyzing the scene correlation involving at least two edge node jurisdiction areas in the real-time scene feature data; based on the scene correlation, converting the final scene global model into a wireless communication optimization strategy; performing multi-dimensional compliance verification on the wireless communication optimization strategy, including device hardware compatibility, industry standard compliance, and system initial constraint compliance; and based on a preset protocol adaptation rule base, converting the verified optimization strategy into a private protocol instruction recognizable by the target device manufacturer to generate wireless communication strategy execution instructions.

[0079] Figure 8 This diagram illustrates the policy generation and compliance verification process according to an embodiment of the present invention. Figure 8In the process, the federated server combines real-time scene characteristics and the correlation of edge nodes to transform the global model of each scene into specific optimization strategies, including parameter adjustment items and execution objects (specified edge nodes / base stations). It also synchronously associates quantifiable target key performance indicator values ​​(KPIs, Key Performance Indicators) to form a "strategy-target" binding document. The second step is to initiate a three-layer compliance verification. If any layer fails the verification, the strategy is rolled back to the "edge training and parameter aggregation" sub-process to re-perform edge training and parameter aggregation. After the third step verification is passed, the federated server calls the system's built-in "protocol adaptation module" to convert the optimization strategy into private protocol instructions that can be recognized by various manufacturers' devices, and encapsulates it into "strategy execution instructions" and sends them to the next process. At the same time, the strategy content, target KPIs, protocol conversion results, and execution node information are uploaded to the system configuration display terminal for operation and maintenance monitoring. In this embodiment of the invention, model parameter transmission replaces the original data, reducing interaction latency and bandwidth consumption, and adapting to the multi-device and highly dynamic characteristics of rail transit.

[0080] In this embodiment of the invention, the three-layer compliance verification includes: verifying whether the parameter adjustment item is within the hardware capability threshold range of the target execution device object, whether it conforms to the preset industry technical standards, such as meeting the 3GPP industry standards, and whether it meets the system initial constraints of the radio switching center to which the target execution device object belongs.

[0081] In this embodiment of the invention, the pre-execution query of the policy execution instruction includes: the federated server sending a policy pre-execution query request to the wireless switching center, the request carrying at least one wireless communication policy execution instruction to be issued and its metadata; receiving a response from the wireless switching center for the pre-execution query request, the response including a conflict judgment result between the wireless communication policy execution instruction and the core network configuration; based on the received response, if the judgment result is no conflict, generating and recording an execution permission for the wireless communication policy execution instruction; and issuing the wireless communication policy execution instruction with the execution permission status identifier to one or more corresponding edge nodes.

[0082] Figure 9 This diagram illustrates the exchange center collaboration and execution process according to an embodiment of the present invention. Figure 9In the process, the federated server sends a policy pre-execution query to the wireless switching center, along with the policy identifier, target base station ID, and parameter details. The wireless switching center retrieves the core network configuration (such as temporary emergency base station failures, sudden offline events, and high-priority maintenance tasks issued by the core network consuming equipment resources), determines whether the instruction conflicts with the configuration, and if there is no conflict, it returns "execution allowed" and grants temporary resource locking permission. If there is a conflict, execution is delayed, and the system returns to the policy pre-execution query stage, waiting for the switching center to return no conflict before executing again. After receiving the permission, the federated server sends the converted policy execution instruction to the corresponding edge node. After receiving the instruction, the edge node first decodes and verifies the integrity of the instruction, then locks the target device resources, and then sends the instruction to terminal devices such as base stations and RRUs to drive the devices to execute the parameter configuration. After execution, the device terminal returns an "execution successful / failed" status. The edge node summarizes the status information and initial KPI data, uploads it to the federated server, and simultaneously displays the "policy execution status" and effect data on the terminal in the system configuration display.

[0083] Figure 10 A schematic diagram of the feedback loop and iterative process according to an embodiment of the present invention is shown. Figure 10 In order to ensure that the effectiveness of the strategy implementation can feed back into the model optimization and form a complete closed loop of "execution-feedback-iteration", the feedback closed loop and the iteration sub-process are "with effect evaluation as the core and problem tracing as the guide", linking the federated server, edge nodes and exchange center.

[0084] Specifically, the federated server receives and integrates feedback data, and presents it in a synchronized and visual manner. The federated server receives execution feedback data uploaded by edge nodes, including KPI changes (such as the increase in handover success rate and the decrease in call drop rate), device operating status (such as base station CPU utilization), and synchronously updates the real-time monitoring panel of the system configuration display terminal, generates optimization effect reports, and marks the KPI achievement status (such as "all core indicators have been achieved") and device health status (such as "no abnormal devices"). Feedback data is fed back into the scene recognition stage to optimize scene rules. The feedback data is categorized by scene and fed back to the scene recognition stage to correct scene feature thresholds and update sub-model adaptation rules, ensuring more accurate subsequent scene recognition and that sub-models better meet actual needs. The next iteration is initiated periodically or triggered by an anomaly. The federated server displays the terminal's preset iteration cycle based on system configuration, or automatically triggers the next iteration process when feedback data shows KPI deterioration (such as a handover success rate below 95% or a call drop rate above 0.1%). This process first initiates the scene recognition stage (identifying the current network scene based on optimized rules), and then distributes updated sub-models and training tasks (such as adding training samples related to "sudden interference response") to edge nodes. Subsequent training and policy generation can be seamlessly connected without manual intervention.

[0085] Record end-to-end iteration logs to form a closed loop. Record complete iteration logs on the system configuration display terminal, covering the entire chain of information from "feedback data (such as KPIs / device status) - scenario correction (such as feature thresholds / adaptation rule adjustments) - training start-up (sub-model delivery time / training task parameters) - next round of strategy implementation results"; at the same time, archive the logs to the "iteration optimization database" so that the operations and maintenance team can trace the motivation and effect of each iteration, forming a continuous closed loop of "data-driven scenario optimization, scenario optimization feeding back into the model, and model iteration improving KPIs".

[0086] In Example 1, the federated server is configured to: receive scene feature data, match and call the corresponding scene-specific sub-model based on the scene feature data; send the called scene-specific sub-model and training rules to several edge nodes, driving the several edge nodes to carry out scene-specific sub-model training respectively; receive model training parameters from several edge nodes, and dynamically aggregate and generate a final scene global model based on the training parameters; convert the final scene global model into a wireless communication policy execution instruction; and send the policy execution instruction queried through pre-execution to the corresponding edge node, instructing the edge node to perform parameter configuration on the terminal device.

[0087] Example 2 Figure 5 This diagram illustrates the flow chart of the dynamic optimization method for wireless communication in rail transit according to Embodiment 2 of the present invention. Figure 5 The method includes: receiving a scene-specific sub-model and training rules issued by a federated server, wherein the scene-specific sub-model is determined by the federated server based on matching and calling received scene feature data; training the scene-specific sub-model based on the training rules to generate model training parameters, and sending the model training parameters to the federated server; receiving a wireless communication policy execution instruction issued by the federated server through a pre-execution query, wherein the wireless communication policy execution instruction is determined by the federated server dynamically aggregating and generating a final scene global model based on the training parameters, and transforming the final scene global model; and configuring execution parameters on the terminal device based on the policy execution instruction.

[0088] In this embodiment of the invention, the process of edge nodes training scene-specific sub-models to generate model training parameters includes: collecting real-time communication data from terminal devices within their jurisdiction and preprocessing the real-time communication data; generating corresponding scene labels based on the preprocessed real-time communication data to form a scene label dataset; using the scene label dataset as training input, training scene-specific sub-model parameters through a three-step iterative algorithm of prediction-comparison-parameter tuning; determining whether the scene-specific sub-model has reached convergence; and when the scene-specific sub-model reaches the preset convergence condition ahead of schedule, encrypting the trained model training parameters and uploading the encrypted model training parameters.

[0089] Specifically, in this embodiment of the invention, edge node-triggered model training includes: After receiving the sub-model and local training rules from the federated server, the edge node automatically triggers local training. In addition, the training triggering mechanism also includes automatically triggering local continuation training when the previous round of local training did not meet the termination conditions, and issuing instructions to trigger global iterative training when the federated server judges that the global model has not met the standards, covering the entire iteration cycle.

[0090] The "local training rules" issued by the federated server are loaded synchronously, clearly defining local termination conditions, unified upload rules, and data security specifications. Local termination conditions can be a specific training cycle or number of training iterations, such as stopping training after one week of local training, or specific judgment criteria, such as changes in the loss function or KPI thresholds. Unified upload rules and local termination conditions must be deeply aligned to avoid both invalid operations like uploading before training is complete and outdated parameters caused by delayed uploads after training. Furthermore, the rules should be adapted to the characteristics of idle computing power during off-peak hours and priority given to business during peak hours in rail transit, such as uploading during the nearest maintenance window after training has stopped.

[0091] Edge nodes collect real-time communication data from device terminals, then preprocess the data to filter out outliers caused by temporary device failures (such as sudden signal drops due to momentary disconnection), normalize parameters of different dimensions to a unified range, and finally label the data with scene tags (such as "subway tunnel - morning rush hour") to ensure that the training data fits the sub-model to the scene.

[0092] Edge nodes take preprocessed data as input and iteratively update sub-model parameters through a three-step process: "forward prediction - error calculation - backpropagation". This three-step iteration includes: predicting communication metrics using current parameters, comparing predicted values ​​with actual data and calculating a loss function, and adjusting parameters based on the error using gradient descent. The three-step iteration includes: prediction step: predicting communication performance metrics using current model parameters, generating predicted values ​​through three iterations; loss calculation step: comparing predicted values ​​with actual measured communication performance metrics, calculating model error based on a preset loss function; parameter adjustment step: based on the model error, backpropagating the error using gradient descent and adjusting model parameters. Forward prediction uses current parameters to predict communication metrics; error calculation compares predicted values ​​with actual data and calculates a loss function; backpropagation adjusts parameters based on the error using gradient descent (e.g., an initial learning rate of 0.05, decaying by 10% every 3 rounds).

[0093] Local model training convergence judgment: After each training round, a multi-dimensional termination judgment is performed. Training stops when the termination conditions are met, such as the loss function changing by less than 0.005 for two consecutive rounds (achieving accuracy convergence), the local model simulation KPIs (switching success rate ≥99%, SINR ≥9dB, etc.) meeting the target, or the iteration cycle or number of iterations reaching the upper limit of the task rules. If the termination conditions are not met, the edge nodes automatically adjust the training strategy, including expanding the data window, reducing the learning rate, and introducing regularization terms, and continue local training until the termination conditions are met. If convergence is achieved in 1-3 rounds, the optimal parameters are encrypted and temporarily stored. Scene data is monitored every 10 minutes. When fluctuations exceed the threshold, 2-3 short rounds of retraining are triggered to update the parameters. If the data is stable, the parameters are kept until the unified upload time.

[0094] Training parameters are uploaded using encryption: All edge nodes strictly adhere to the unified upload time stipulated in the task rules (e.g., upload during the window period), completing parameter uploads within this window. Only the changes in model parameters are extracted; the original data is not uploaded. Differential privacy protection is achieved by adding Gaussian noise according to the scenario, followed by uploading via an encrypted link. A digital signature is attached to each node to prevent tampering, and a "training status label" is added, including the convergence type, training epoch, and data quality. If a node fails to upload due to network interruption, the federated server activates a backup plan within a certain period after the window ends. Priority is given to using the node's optimal parameters from the previous round and marking them as "historical reuse." If no historical data is available, a weighted average of parameters from multiple adjacent nodes in the same scenario is used as a substitute, with the substitute parameters marked as "backup." The weight during aggregation is ≤0.1 to minimize impact.

[0095] In Example 2, the edge nodes are configured to: receive scene-specific sub-models and training rules from the federated server, wherein the scene-specific sub-models are determined by the federated server based on the received scene feature data; train the scene-specific sub-models based on the training rules to generate model training parameters, and send the model training parameters to the federated server; receive wireless communication policy execution instructions issued by the federated server through pre-execution queries, wherein the wireless communication policy execution instructions are determined by the federated server dynamically aggregating and generating a final scene global model based on the training parameters, and transforming the final scene global model; and configure execution parameters for the terminal devices based on the policy execution instructions.

[0096] In this embodiment of the invention, the extended application of the core architecture of "Federation Server-Edge Node-Device Terminal" in special scenarios (dynamic switching, sudden failure) is also described. By differentiating the division of labor among the three-level nodes (such as edge node prediction and triggering, federation server policy scheduling, and device terminal execution feedback), the adaptation gap of the conventional process in special scenarios is filled.

[0097] In the scene transition coordination process, the federation server is used to receive transition status reports sent from at least one edge node. The reports are generated and sent by the edge node after detecting that the offset of the real-time communication scene features in its jurisdiction relative to the baseline steady-state scene exceeds a preset threshold. Based on the received transition status reports, the global transition sub-model is invoked to generate target steady-state parameters that match the scene offset reported. The generated target steady-state parameters are then sent to the corresponding edge node that sent the transition status reports to trigger the edge node to configure transition parameters for the device terminals in its jurisdiction based on the target steady-state parameters.

[0098] In the scene transition collaboration process, edge nodes monitor the real-time communication scene characteristics of device terminals within their jurisdiction. When the offset of the current scene characteristics relative to the baseline steady-state scene exceeds a preset threshold, a transition scene processing flow is initiated. After initiating the transition scene processing flow, a transition status report is sent to the federation server. Target steady-state parameters generated by calling the global transition sub-model based on the transition status report are received from the federation server. Based on the target steady-state parameters, transition parameter configuration for device terminals within their jurisdiction is initiated. Optionally, edge nodes are also used to monitor scene characteristic changes generated after device terminal configuration in real time. During transition execution, the proportion of the steady-state characteristics corresponding to the transition target in the total characteristics is continuously monitored. When this proportion continuously exceeds a first threshold and the duration reaches a preset window, the transition is determined to end, and the execution of the relevant transition strategy is stopped.

[0099] In this embodiment of the invention, a smooth parameter transition during scene switching is achieved through "edge prediction - federated rapid decision-making - lightweight exchange collaboration". The core of the dedicated scene transition collaboration process is to solve the parameter adaptation and resource coordination problems in the "dynamic scene switching stage (such as a train entering a tunnel from the ground, exiting a tunnel and entering an elevated road, or entering a city from an open road)", avoiding KPI fluctuations caused by mixed scene features. The process is based on the core logic of "edge node prediction triggering, federated server model support, and lightweight exchange center collaboration", and the specific steps are as follows: The process is initiated by edge nodes: Edge nodes use real-time collected dynamic data (such as train position, signal change slope, and speed fluctuations) combined with locally deployed transition scenario sub-models (specially trained to focus on dynamic features of scenario switching, with KPI fluctuation penalties introduced during training) to immediately determine "transition scenario initiation" when the identified scenario shifts from "steady state A" to "steady state B" by more than 30%. At this time, a "transition warning" (including scenario type and estimated duration) is simultaneously pushed to the federated server, and a "lightweight notification" (such as synchronizing resource usage scope) is sent to the exchange center. No approval is required, only avoiding conflicts with conventional policies.

[0100] The federated server rapidly generates and distributes transition strategies: based on the scenario type of the edge nodes, it calls the global transition sub-model library to generate a "hybrid adaptation strategy" that retains the original steady-state core configuration (such as tunnel basic power), pre-loads target steady-state parameters (such as elevated road interference threshold), and distributes the strategy within 10 seconds after performing only basic legality checks. If the edge nodes have the same type of strategy cached locally, they can be directly called, shortening the latency. In this embodiment of the invention, the global transition sub-model library is trained by aggregating logs from each edge node and fine-tuned quarterly.

[0101] Edge nodes execute transitional strategies and lock resources: Regular strategies are paused, transitional parameter configurations are initiated (such as fine-tuning power and interference thresholds), necessary resources are locked (such as reserved buffer frequency bands), and KPI data is synchronized to the federated server and exchange center every 5 seconds. The exchange center only monitors resources; if high-priority business conflicts occur, it assists edge nodes in fine-tuning resources through "lightweight alerts" without intervening in strategy adjustments.

[0102] After the transition, the normal process continues: when the edge node detects that the proportion of the target steady-state feature exceeds 90%, it determines that the transition has ended, stops the transition strategy, restarts the normal strategy, and pushes an "end report" to the federated server and the exchange center. The federated server updates the sub-model effect labels and archives the logs, and the exchange center releases resources and updates the ledger, completing the process loop.

[0103] The emergency response process is designed to address the issue of rapid handling of sudden failures or extreme events (such as base station downtime, widespread interference, or a sudden increase in call drop rates).

[0104] In this embodiment of the invention, under the emergency response special process, the federated server is used to receive emergency alarm information containing event type labels pushed from at least one edge node. The information is generated and pushed by the edge node after continuously monitoring the device status and business indicators of the device terminals within its jurisdiction and detecting that at least one indicator exceeds a preset emergency threshold. Based on the received emergency alarm information, an emergency policy instruction is generated and sent to the corresponding edge node that sent the emergency alarm information to drive it to suspend non-critical services and prioritize emergency configuration. The server receives emergency handling progress reports synchronously reported by the edge node at fixed time intervals during the emergency configuration process. The server receives an emergency release notification generated by the edge node after detecting that all emergency indicators have returned to normal range and have been stably and continuously maintained for a preset duration, and performs closed-loop processing on the emergency event.

[0105] Under the emergency response process, edge nodes continuously monitor the device status and business metrics of terminals within their jurisdiction. When at least one metric exceeds a preset emergency threshold, an emergency alarm message containing an event type label is generated and pushed to the federated server. The edge nodes receive emergency policy instructions generated by the federated server based on the emergency alarm message, suspend all currently executing non-critical business policies, and prioritize the execution of emergency configuration operations within the emergency policy instructions. During emergency configuration, the edge nodes synchronize current emergency response progress reports to the federated server at fixed time intervals. They continuously monitor core emergency-related metrics, and when all emergency metrics have returned to normal ranges and stably reached a preset duration, an emergency deactivation notification is generated and pushed to the federated server.

[0106] Optionally, in this embodiment of the invention, the emergency response process follows the logic of "edge alarm triggering, federated priority scheduling, and on-demand exchange coordination," prioritizing the stability of core services (such as train control signals and emergency communications). The specific process is as follows: The process is initiated by either an edge node or the switching center: Edge nodes monitor equipment status (e.g., base station CPU utilization, RRU operating temperature) and service metrics (e.g., call drop rate, interference intensity). When an "emergency threshold exceeding the limit" is detected (e.g., call drop rate ≥ 5%, interference intensity ≥ -60dBm), or the switching center detects cross-node resource anomalies (e.g., backbone link interruption), an "emergency alarm" is immediately pushed to the federated server, clearly indicating the event type, scope of impact (involving edge node IDs, coverage areas), and core service risk level (e.g., "high risk: affecting train control"). Simultaneously, edge nodes do not notify the switching center of "minor local emergencies" (e.g., weak signal from a single node), but simultaneously send an "emergency resource demand forecast" to the switching center for "cross-node / core emergencies" (e.g., multiple base stations offline), preparing for subsequent collaboration.

[0107] The federated server rapidly dispatches emergency strategies: Upon receiving an alarm, the federated server prioritizes retrieving the "emergency strategy library" (which pre-stores standardized handling schemes for various faults, such as neighbor cell power compensation schemes when a base station is offline and frequency band switching schemes for interference scenarios). It then generates a "customized emergency strategy" based on the alarm details. For example, when a single base station is offline, it instructs neighboring base stations to increase their power from 25dBm to 28dBm while simultaneously reducing the coverage radius to avoid interference. In cases of multi-node interference, it uniformly schedules affected nodes to switch to backup frequency bands. After strategy generation, it skips routine compliance checks and resource approvals, only verifying "whether it affects core services," ensuring that the strategy is distributed to the corresponding edge nodes within 30 seconds. If it is a "cross-node emergency," it simultaneously pushes an "emergency resource scheduling request" to the switching center, specifying the cross-node resources that need to be coordinated (such as backup link bandwidth and shared power capacity). The switching center must respond to resource availability within one minute to avoid delays in handling.

[0108] Edge nodes execute emergency strategies and seize resources: Upon receiving the strategy, edge nodes immediately suspend non-critical service strategies and prioritize emergency configurations, such as adjusting power, switching frequency bands, and activating backup equipment. Simultaneously, they lock necessary emergency resources (e.g., occupying backup frequency bands) and synchronize "emergency response progress" (e.g., "neighboring cell power has been adjusted, coverage gap reduced by 50%", "call drop rate reduced from 6% to 3.2%)" with the federated server and switching center every 10 seconds. The switching center assumes the responsibility of "resource coordination" in "cross-node emergency response": If resources are available, it immediately locks the corresponding resources and prevents other services from occupying them; if resources are scarce, it prioritizes core service-related nodes (e.g., prioritizing backup power allocation for train control-related base stations) and simultaneously pushes "emergency work orders" to the operations and maintenance team, indicating the fault location and handling suggestions.

[0109] The closed-loop process after emergency response is lifted is as follows: When the edge node detects that "emergency indicators have returned to normal" (e.g., call drop rate ≤ 0.1%, base station online again), or the operations and maintenance team reports that "the fault has been repaired," it immediately pushes an "emergency lift notification" to the federation server and the switching center. The edge node stops the emergency strategy, restarts the regular business strategy, and releases the redundant resources occupied in the emergency. After receiving the notification, the federation server issues a "parameter recovery instruction" to the edge node (e.g., adjusting the power of neighboring base stations back to normal values), and archives the "emergency handling log" (including event type, strategy content, and KPI recovery data) to the "emergency optimization library" for subsequent updates to the emergency strategy library (e.g., optimizing power compensation parameters). The switching center releases the coordinated cross-node resources, updates the resource ledger, and synchronizes the emergency handling records to the operations and maintenance audit system, thus completing the closed loop of the emergency response process.

[0110] This invention also provides a method for dynamic optimization of wireless communication in rail transit, the method comprising: The federated server receives scene feature data from several edge nodes, matches and calls the corresponding scene-specific sub-model based on the scene feature data; The federated server sends scenario-specific sub-models and training rules to several edge nodes, driving the edge nodes to train the scenario-specific sub-models respectively. The edge node receives the scene-specific sub-model and training rules issued by the federation server, performs scene-specific sub-model training based on the training rules to generate model training parameters, and sends the model training parameters to the federation server. The federated server receives model training parameters from several edge nodes and dynamically aggregates them to generate the final global model of the scene based on the training parameters; The federated server converts the final scenario global model into wireless communication policy execution instructions, and sends the policy execution instructions, which are pre-executed through the wireless switching center, to the corresponding edge nodes, instructing the edge nodes to perform parameter configuration on the terminal devices. The edge node receives the wireless communication policy execution instruction issued by the federation server through pre-execution query, and performs parameter configuration on the terminal device based on the policy execution instruction.

[0111] Specifically, the method further includes an initialization step, which includes: determining the rail transit type of the target rail transit system, wherein the rail transit type includes at least one of railway system and metro / light rail system; requesting an initial parameter constraint range from the wireless switching center according to the determined rail transit type; and the federated server storing and encrypting the parameter constraint range locally.

[0112] In this embodiment of the invention, a computer storage medium is also provided, which stores one or more instructions, which, when executed by one or more computers, cause the one or more computers to implement the dynamic optimization method for rail transit wireless communication described in this invention.

[0113] The present invention relates to a method, system, and storage medium for dynamic optimization of wireless communication in rail transit. Based on the federated learning architecture, the method of replacing the original data with transmission model parameters significantly reduces the amount of data interaction, reduces the latency of data transmission and processing, enables the optimization strategy to respond in a timely manner to the rapid changes in the channel under high-speed mobile scenarios, avoids policy lag, avoids data privacy issues, and significantly improves cross-domain optimization capabilities.

[0114] By calling dedicated sub-models for different scenarios, the wireless communication parameters of each scenario are accurately matched with the characteristics of the scenario, thereby improving the quality of dedicated wireless communication in different scenarios and solving the problem that a single model cannot adapt to multiple scenarios. By collecting data in real time at edge nodes and iterating the global model at high frequency by federated servers, the response latency of optimization strategies is greatly shortened, which can accurately adapt to the dynamic scenario characteristics of rail transit.

[0115] It has a built-in multi-vendor protocol conversion module, which can be directly adapted to mainstream equipment without the need for additional customized tools, significantly reducing the cost of upgrading existing lines and the difficulty of deploying new lines; it breaks down the data barriers of single lines and single vendors, and achieves cross-line and cross-vendor model parameter collaborative aggregation through federated learning. It can formulate global optimization strategies without relying on direct interaction of raw data, improve cross-regional collaborative response efficiency, and reduce the need for repeated development of basic models for each line by sharing a global model, thereby reducing computing power consumption.

[0116] By using scenario-specific sub-models and transitional collaborative processes, precise parameter matching across all scenarios is achieved, effectively reducing the risk of communication interruptions during transitional scenarios and improving overall coverage stability. Resource configuration is dynamically adjusted based on real-time load to improve resource utilization efficiency. The emergency response process skips unnecessary compliance checks, prioritizing core business operations and shortening fault handling latency. Based on scenario prediction, sub-models are pre-loaded and strategies are smoothly switched, reducing wireless signal fluctuations during scenario switching and lowering the probability of communication interruptions. Furthermore, when adding new scenarios, only the corresponding sub-model needs to be added to complete the adaptation without system reconstruction, significantly improving system scalability.

[0117] The "scene recognition-federated learning-dynamic optimization" architecture constructed in this invention has strong adaptability and can flexibly accommodate different rail transit types, equipment from different manufacturers, and various sub-scenarios. Through modular design, the scene recognition module can expand feature dimensions according to new scenarios, the federated learning framework can adapt to differences in protocols from different manufacturers by loading new sub-models, and the dynamic optimization module can automatically adjust the strategy output form based on scenario characteristics. Without reconstructing the core architecture, it can achieve compatibility and coverage of existing and future new rail transit scenarios and equipment types, forming a generalized technical system of "one-time architecture construction, flexible adaptation to multiple scenarios".

[0118] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic optimization method for wireless communication in rail transit, characterized in that, The method includes: Receive scene feature data, match and call the corresponding scene-specific sub-model based on the scene feature data; The scenario-specific sub-models and training rules are sent to several edge nodes to drive the edge nodes to carry out scenario-specific sub-model training respectively. Receive model training parameters from several edge nodes, and dynamically aggregate and generate a final global model of the scene based on the training parameters; The final scenario global model is converted into wireless communication policy execution instructions; The strategy execution instruction obtained through pre-execution query is sent to the corresponding edge node, instructing the edge node to perform parameter configuration on the terminal device.

2. The dynamic optimization method for wireless communication in rail transit according to claim 1, characterized in that, The method further includes: Receive feedback data from the terminal device regarding the execution of the parameter configuration; The scene feature threshold is corrected based on the feedback data, and the next iteration is triggered according to the iteration cycle and feedback data anomalies.

3. The dynamic optimization method for wireless communication in rail transit according to claim 1 or 2, characterized in that, The generation of the scene feature data includes: Determine the rail transit type of the target rail transit system, wherein the rail transit type includes at least one of railway system and metro / light rail system; Based on the determined rail transit type, scene feature collection instructions are issued to several edge nodes; The system receives static and dynamic scene data collected by several edge nodes according to the acquisition instructions, and preprocesses the collected scene data to generate scene feature data.

4. The dynamic optimization method for wireless communication in rail transit according to claim 3, characterized in that, Matching and invoking the corresponding scene-specific sub-models includes: Cross-validation is performed on the scene feature data sent by several of the edge nodes; The target scene is identified by matching the scene feature data with a subdivided scene library through cross-validation; The corresponding scene-specific sub-model is invoked based on the locked target scene.

5. The dynamic optimization method for wireless communication in rail transit according to claim 1 or 2, characterized in that, The final global model of the scene is generated by dynamically aggregating the training parameters, including: Receive model training parameters from several edge nodes and perform anomaly verification on the model training parameters; Based on the training status labels and scene data quality evaluation indicators reported by each edge node, aggregate weights are dynamically assigned to the model training parameters of each edge node that has passed the anomaly check. The global values ​​of the model training parameters are determined by weighting. Determine the absolute difference between the training parameters of each edge node model and the global value, and take the maximum value of the absolute difference to generate a preliminary global scene model; The initial global model of the scene that meets the iteration termination condition is determined as the final global model of the scene.

6. The dynamic optimization method for wireless communication in rail transit according to claim 5, characterized in that, The initial global model of the scenario that satisfies the iteration termination condition includes: Determine whether the preliminary global model of the scenario meets the preset iteration termination conditions. The iteration termination conditions include at least the global loss function change reaching the target or the key performance index value reaching the target. When the iteration termination condition is met, the preliminary scene global model is determined as the final scene global model.

7. The dynamic optimization method for wireless communication in rail transit according to claim 1 or 2, characterized in that, Converting the final scenario global model into wireless communication policy execution instructions includes: Obtain the real-time scene feature data of the final scene global model at the current moment; Analyze the scene correlation in real-time scene feature data involving the jurisdiction of at least two edge nodes; Based on the aforementioned scene correlation, the final scene global model is converted into a wireless communication optimization strategy; The wireless communication optimization strategy is subject to multi-dimensional compliance verification, including device hardware compatibility, industry standard compliance, and system initial constraint compliance. Based on a pre-defined protocol adaptation rule base, the verified optimization strategies are converted into private protocol instructions that can be recognized by the target device manufacturer, thereby generating wireless communication strategy execution instructions.

8. The dynamic optimization method for wireless communication in rail transit according to claim 1 or 2, characterized in that, The pre-execution query of the strategy execution instruction includes: Send a policy pre-execution query request to the wireless switching center, the request carrying at least one wireless communication policy execution instruction to be issued and its metadata; Receive a response from the wireless switching center for the pre-execution query request, the response including a conflict determination result between the wireless communication policy execution instruction and the core network configuration; Based on the received response, if the determination result is no conflict, then an execution permission for the wireless communication policy execution instruction is generated and recorded. The wireless communication policy execution instruction that allows the execution of the permission status identifier is sent to one or more corresponding edge nodes.

9. The dynamic optimization method for wireless communication in rail transit according to claim 1 or 2, characterized in that, The method further includes: Receive a transitional state report from at least one edge node, the report being generated and sent by the edge node after detecting that the offset of the real-time communication scene characteristics within its jurisdiction relative to the baseline steady-state scene exceeds a preset threshold; Based on the received transition state report, the global transition sub-model is invoked to generate target steady-state parameters that match the scene offset reported in the report; The generated target steady-state parameters are sent to the corresponding edge node that sent the transition state report, so as to trigger the edge node to configure the transition parameters of the device terminals in its jurisdiction based on the target steady-state parameters.

10. The dynamic optimization method for wireless communication in rail transit according to claim 1 or 2, characterized in that, The method further includes: Receive emergency alarm information containing event type labels from at least one edge node. The information is generated and pushed by the edge node after it continuously monitors the device status and business indicators of the device terminals in the area under its jurisdiction and detects that at least one indicator exceeds a preset emergency threshold. Based on the received emergency alarm information, an emergency policy instruction is generated and sent to the corresponding edge node that sent the emergency alarm information to drive it to suspend non-critical services and prioritize the execution of emergency configuration. Receive emergency response progress reports synchronously reported by the edge nodes at fixed time intervals during the emergency configuration process; The system receives the emergency cancellation notification generated by the edge node after detecting that all emergency indicators have returned to normal range and have remained stable for a preset duration, and performs closed-loop processing on the emergency event.

11. A dynamic optimization system for wireless communication in rail transit, characterized in that, The system includes a federated server, which is configured as follows: Receive scene feature data, match and call the corresponding scene-specific sub-model based on the scene feature data; The scenario-specific sub-models and training rules are sent to several edge nodes to drive the edge nodes to carry out scenario-specific sub-model training respectively. Receive model training parameters from several edge nodes, and dynamically aggregate and generate a final global model of the scene based on the training parameters; The final scenario global model is converted into wireless communication policy execution instructions; The strategy execution instruction obtained through pre-execution query is sent to the corresponding edge node, instructing the edge node to perform parameter configuration on the terminal device.

12. A dynamic optimization method for wireless communication in rail transit, characterized in that, The method includes: The system receives scene-specific sub-models and training rules from the federated server, wherein the scene-specific sub-models are determined by the federated server based on the received scene feature data through matching and invocation. Based on the training rules, scenario-specific sub-models are trained to generate model training parameters, and the model training parameters are sent to the federated server. The system receives a wireless communication policy execution instruction issued by the federation server through a pre-execution query. The wireless communication policy execution instruction is determined by the federation server dynamically aggregating and generating a final scene global model based on the training parameters, and then transforming the final scene global model. The terminal device executes parameter configuration based on the strategy execution instructions.

13. The dynamic optimization method for wireless communication in rail transit according to claim 12, characterized in that, The training parameters for generating the model by training a scenario-specific sub-model include: Collect real-time communication data from terminal devices within the jurisdiction, and preprocess the real-time communication data; Based on the preprocessed real-time communication data, corresponding scene labels are generated to form a scene label dataset. Using the scene label dataset as training input, the parameters of the scene-specific sub-model are trained through a three-step iterative algorithm of prediction-comparison-parameter tuning; Determine whether the scene-specific sub-model has reached a convergence state; When the scenario-specific sub-model reaches the preset convergence condition ahead of schedule, the training parameters of the trained model are encrypted, and the encrypted training parameters are uploaded.

14. The dynamic optimization method for wireless communication in rail transit according to claim 12, characterized in that, The method further includes: Monitor the real-time communication scene characteristics of equipment terminals within the jurisdiction. When the deviation of the current scene characteristics from the benchmark steady-state scene exceeds a preset threshold, start the transition scene processing procedure. After initiating the transition scenario processing procedure, a transition status report is sent to the federated server; Receive target steady-state parameters generated by the global transition sub-model based on the transition state report issued by the federated server; Based on the target steady-state parameters, the transition parameter configuration for the equipment terminals within the jurisdiction is initiated.

15. The dynamic optimization method for wireless communication in rail transit according to claim 14, characterized in that, The method further includes: Real-time monitoring of scene feature changes resulting from device terminal configuration; During the transition execution process, the proportion of the steady-state characteristics corresponding to the transition target in the total characteristics is continuously monitored; When the percentage continuously exceeds the first threshold and the duration reaches the preset window, the transition is determined to be over, and the execution of the relevant transition strategy is stopped.

16. The dynamic optimization method for wireless communication in rail transit according to claim 12, characterized in that, The method further includes: Continuously monitor the equipment status and business indicators of terminals within the jurisdiction; When at least one indicator is detected to exceed a preset emergency threshold, an emergency alarm message containing an event type label is generated and the emergency alarm message is pushed to the federated server. Upon receiving the emergency policy instruction generated by the federal server based on the emergency alarm information, suspend all currently executing non-critical business policies and prioritize the execution of the emergency configuration operation in the emergency policy instruction; During the emergency configuration process, the current emergency response progress report is synchronized to the federal server at fixed time intervals; Continuously monitor key emergency-related indicators. When all emergency indicators have returned to normal range and have remained stable for a preset duration, generate an emergency deactivation notification and push the notification to the federated server.

17. A dynamic optimization system for wireless communication in rail transit, characterized in that, The system includes several edge nodes, which are configured as follows: The system receives scene-specific sub-models and training rules from the federated server, wherein the scene-specific sub-models are determined by the federated server based on the received scene feature data through matching and invocation. Based on the training rules, scenario-specific sub-models are trained to generate model training parameters, and the model training parameters are sent to the federated server. The system receives a wireless communication policy execution instruction issued by the federation server through a pre-execution query. The wireless communication policy execution instruction is determined by the federation server dynamically aggregating and generating a final scene global model based on the training parameters, and then transforming the final scene global model. The terminal device executes parameter configuration based on the strategy execution instructions.

18. A dynamic optimization method for wireless communication in rail transit, characterized in that, The method includes: The federated server receives scene feature data from several edge nodes, matches and calls the corresponding scene-specific sub-model based on the scene feature data; The federated server sends scenario-specific sub-models and training rules to several edge nodes, driving the edge nodes to train the scenario-specific sub-models respectively. The edge node receives the scene-specific sub-model and training rules issued by the federation server, performs scene-specific sub-model training based on the training rules to generate model training parameters, and sends the model training parameters to the federation server. The federated server receives model training parameters from several edge nodes and dynamically aggregates them to generate the final global model of the scene based on the training parameters; The federated server converts the final scenario global model into wireless communication policy execution instructions, and sends the policy execution instructions, which are pre-executed through the wireless switching center, to the corresponding edge nodes, instructing the edge nodes to perform parameter configuration on the terminal devices. The edge node receives the wireless communication policy execution instruction issued by the federation server through pre-execution query, and performs parameter configuration on the terminal device based on the policy execution instruction.

19. The dynamic optimization method for wireless communication in rail transit according to claim 18, characterized in that, The method further includes an initialization step, the initialization including: Determine the rail transit type of the target rail transit system, wherein the rail transit type includes at least one of railway system and metro / light rail system; Request the initial parameter constraint range from the wireless switching center based on the determined rail transit type; The federated server stores and encrypts the parameter constraint range locally.

20. A dynamic optimization system for wireless communication in rail transit, characterized in that, The system includes: a federated server, several edge nodes, and a wireless switching center. The federated server is configured to receive scene feature data from several edge nodes, and match and call the corresponding scene-specific sub-model based on the scene feature data; The federated server is also configured to issue scenario-specific sub-models and training rules to several edge nodes, driving the edge nodes to carry out scenario-specific sub-model training respectively. Several edge nodes are connected to the federation server and are configured to receive scene-specific sub-models and training rules issued by the federation server, train scene-specific sub-models based on the training rules to generate model training parameters, and send the model training parameters to the federation server. The federated server is also configured to receive model training parameters from several of the edge nodes and dynamically aggregate them to generate a final global model of the scene based on the training parameters. The federated server is also configured to convert the final scenario global model into wireless communication policy execution instructions; The wireless switching center, which communicates with the federation server, is configured to pre-execute the policy execution instructions queryed; The federated server is also configured to send the policy execution instructions, which are pre-executed through the wireless switching center, to the corresponding edge nodes, instructing the edge nodes to perform parameter configuration on the terminal devices; The edge node is also configured to receive wireless communication policy execution instructions issued by the federation server through pre-execution queries, and to perform parameter configuration on the terminal device based on the policy execution instructions.

21. A computer storage medium, characterized in that, The computer storage medium stores one or more instructions, which, when executed by one or more computers, cause the one or more computers to perform the method of any one of claims 1-10.

Images