5G network optimization method and system based on intelligent dynamic multi-band adaptive switching system

Abstract

The invention relates to the technical field of vehicle-mounted communication, and provides a 5G network optimization method and system based on an intelligent dynamic multi-band adaptive switching system, and the method comprises the steps: obtaining network signal quality data, environment perception data and a communication task demand set; dynamic obstacle features are extracted based on environmental perception data, and signal occlusion prediction is output through a double-branch mixed deep learning model; screening qualified candidate links to construct a candidate set, generating a dynamic weight by using a deep Q network, calculating a comprehensive score of the candidate links, determining a main communication link under a hard constraint condition, and generating a communication strategy instruction; and according to the instruction type of the communication strategy instruction, performing dynamic fragmentation and multi-path distribution on the data stream, performing verification, recombination and intelligent delivery by a receiving end, and collecting a feedback data packet for network closed-loop feedback. According to the method, multi-mode sensing and signal shielding prediction are fused, and a self-adaptive switching and cooperative transmission mechanism of multi-band communication is constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle-mounted communication, in particular to a 5G network optimization method and system based on an intelligent dynamic multi-band adaptive switching system. BACKGROUND

[0002] With the in-depth application of 5G technology in the field of intelligent transportation and automatic driving, vehicle remote control has become a core technology to realize high-level automatic driving and unmanned operation. In a complex and changeable electromagnetic environment, how to provide continuous, stable and low-latency communication guarantee for remote control instructions and high-definition video backhaul, effectively making up for the defects of traditional single-path communication scheme such as easy interruption and high delay, has become an important issue to be solved in the process of promoting large-scale commercialization of vehicle remote control technology.

[0003] Chinese patent application with publication number CN120264315A provides a network slice optimization method for 5G private network, which includes: constructing a time series performance model during slice operation, monitoring the micro changes of end-to-end delay, throughput and packet loss rate; then extracting the short-period fluctuation pattern of the index, comparing the gradient with the historical stable threshold; by identifying the continuous jitter but not yet default index drift behavior, it is judged whether it is in a performance critical state; record the non-equivalent response including jitter improvement and delay reduction brought by each resource adjustment; then trend fitting is performed on the data to judge whether there is a performance platform period or a critical jump point; three intervals of resource investment are defined in the model: invalid interval with unchanged investment, high sensitivity interval with rapid performance change, and saturation interval with stable performance but continued resource increase; the current response interval of the slice is identified through the model to determine the control strategy, which is not the core of equal proportion investment, but the optimal response interval as the resource allocation interval, and the selection is made.

[0004] However, the current technology still faces many challenges. The propagation ability of 5G high-frequency signals is limited at the physical layer, especially in urban areas with high-rise buildings or complex terrain areas, their diffraction and penetration ability is weak, and they are easily blocked by buildings or mountains, resulting in frequent interruption of communication links. The existing system adopts a passive link switching mechanism, which usually triggers the backup link only when the current signal quality deteriorates seriously, lacks forward-looking prediction of future channel state, resulting in switching delay or even failure. At the same time, its switching decision logic is simple, and it cannot dynamically and comprehensively weigh the real-time quality of each candidate link and the differentiated needs of businesses such as delay and bandwidth, making it difficult to achieve global optimal scheduling, and thus causing loss of remote control instructions or video backhaul lag, which poses a serious threat to driving safety. SUMMARY

[0005] To achieve the above-mentioned purpose, the present application provides a 5G network optimization method based on an intelligent dynamic multi-band adaptive switching system, and the specific technical solutions are as follows:

[0006] The 5G network optimization method and system based on intelligent dynamic multi-band adaptive switching system, comprising:

[0007] Based on the vehicle integrated multi-band device, multi-source sensor network, positioning system and application sensing identification module, the real-time network signal quality data P sig , environmental perception data D env and communication task demand set R task are synchronously acquired;

[0008] Based on the environmental perception data D env , the dynamic obstacle feature vector set F obs is extracted, and the environmental perception data D env and the dynamic obstacle feature vector set F obs are integrated and input into a parallel double-branch hybrid deep learning model M DL for signal shielding prediction, and a multi-dimensional parameter signal shielding prediction result O pred is output;

[0009] The acquired candidate link is sequentially subjected to signal link feasibility test and network resource availability test, and the qualified candidate link is filtered out to construct a candidate set C set of multi-dimensional performance profile; based on the multi-dimensional performance profile of each candidate link, in combination with a dual-mode decision trigger mechanism, a dynamic weight factor is generated by using a deep Q network when the communication decision flow is activated, and the comprehensive score S score of each candidate link is calculated according to the dynamic weight factor; according to the hard constraint condition of the communication service demand R task , the candidate link sorted in descending order according to the comprehensive score S score is subjected to hard constraint condition test one by one until a candidate link satisfying the hard constraint condition is found as the main communication link, and according to the overall situation of the current network resource, a final communication strategy instruction Start comm is generated by choosing between switching strategy instruction and cooperative strategy instruction;

[0010] According to the communication strategy instruction Start comm , if the communication strategy instruction Start comm is a cooperative strategy, dynamic fragmentation is performed on the data stream to generate data packets P sliced that can be concurrently transmitted; the data packets P sliced are sequentially subjected to cyclic redundancy check, out-of-order reorganization and intelligent delivery at the receiving end to restore the data stream D reass to be complete and in order; after the execution of the communication strategy instruction Start comm , feedback data packets F containing the key performance indicators of the instruction are collected and packaged.back and the feedback data packet F back Iterative training of deep Q-networks to update decision weights, enabling adaptive closed-loop feedback.

[0011] Further, the signal obstruction prediction result O pred The output method comprises:

[0012] The dynamic obstacle data D sensor is processed in multiple stages to generate a dynamic obstacle feature vector set F obs containing shape and motion parameters;

[0013] The static environment data D map , the vehicle's own dynamic data D pos and the dynamic obstacle feature vector set F obs are integrated and input into a parallel dual-branch hybrid deep learning model to extract spatial feature vectors F cnn and temporal feature vectors F rnn respectively, the spatial feature vectors F cnn and the temporal feature vectors F rnn are fused to construct spatio-temporal fusion feature vectors F fused , the spatio-temporal fusion feature vectors F fused are decoded by a prediction decoding head to output a signal obstruction prediction result O k containing predicted affected frequency bands bank k , predicted signal attenuation values A pred , predicted start times of obstruction events and predicted durations of obstruction events.

[0014] Further, the generation of the dynamic obstacle feature vector set F obs containing shape and motion parameters comprises shape feature extraction and trajectory feature extraction;

[0015] Based on the dynamic obstacle data D sensor , a convex hull algorithm is used for shape feature extraction to determine the length l, width w and height h of the dynamic obstacle, further divide the data points into sub-cubes, and perform weighted summation according to the data point density ρ i' of each sub-cube to estimate the total volume V of the obstacle;

[0016] A Kalman filter algorithm is used to extract trajectory features of the centroid of the same obstacle identified in consecutive frames, and by calculating the position change of the centroid between consecutive time steps, the instantaneous speed s and instantaneous motion direction angle θ s of the obstacle are estimated.

[0017] ​​Furthermore, the communication strategy instruction Start comm The generation methods include:

[0018] Obtain the candidate link list and network load rate based on real-time network signal quality data P. Sig Perform a signal link feasibility check on the candidate link list, and perform a network resource availability check on the candidate link list based on the network load rate. After completing the feasibility and availability checks, construct a candidate set C from all remaining qualified candidate links. set And based on the candidate set C set The generation of each candidate link includes signal quality index N sig Network load metric N load and distance index N dis Multidimensional performance profile;

[0019] Based on real-time network signal quality data P Sig And signal occlusion prediction results O pred The system combines a dual-mode decision triggering mechanism to determine whether to activate the communication decision process. If the communication decision process is activated, a deep Q-network is used to generate dynamic weight factors based on the state vector of the current communication environment. The comprehensive score S of each candidate link is then calculated based on the weight factors and the multi-dimensional performance profile of each candidate link. score ;

[0020] Based on communication service requirements R task The hard constraints, based on the comprehensive score S score Candidate set C sorted in descending order set The candidate link ranked first is subjected to hard constraint testing. If the candidate link meets the hard constraints, it is determined as the main communication link; if the candidate link does not meet the hard constraints, the process is automatically iterated to the candidate set C after descending order. set The next candidate link is then subjected to hard constraint checks until a candidate link that meets the hard constraints is found. Based on the overall network resource situation, a choice is made between switching policy instructions and coordination policy instructions to generate the final communication policy instruction, Start. comm ;

[0021] The communication service requirement R task The hard constraint is the communication service demand R task Performance requirements of the critical data stream j of the highest priority data stream task in the middle business Including minimum protection width and maximum transmission delay

[0022] Further, the dual-mode decision trigger is activated when either of the following conditions is met: if the real-time network signal quality data P Sig indicates that the key performance indicator of the currently used link is below the preset minimum availability threshold, indicating that the quality of service has deteriorated unacceptably, the passive response trigger is activated; if the signal obstruction prediction result O pred indicates that the current communication link will face a high-probability, high-attenuation signal obstruction event in the near future, representing that the system anticipates an impending deterioration in service quality, the active prevention trigger is activated.

[0023] Further, the selection method between the handover strategy instruction and the synergy strategy instruction includes:

[0024] When the system determines that the current network environment is relatively tense, i.e., the quality of the remaining candidate links is generally poor in addition to the selected main communication link, or the main purpose of this decision is to avoid the deterministic failure risk of the current link, the system starts the handover strategy instruction generation process;

[0025] When the system determines that the current network environment is abundant in resources, i.e., in addition to the determined main communication link, there are still multiple high-quality links in the candidate set C set , the system starts the synergy strategy instruction generation process;

[0026] The handover strategy instruction includes a strategy instruction type 'HANDOVER', a target band band parameter of the finally selected main communication link, and a trigger reason reason parameter recording this link handover;

[0027] The synergy strategy instruction includes a strategy instruction type 'SYNERGY', a control task task control task , a video task video task , and a telemetry task telemetry .

[0028] Further, the generation process of the synergy strategy instruction includes:

[0029] The highest priority control task control task is assigned to the determined main communication link and written as the first key-value pair in the parameter dictionary;

[0030] Two temporary resource pools, a to-do task pool and an available link pool, are created, where the to-do task pool contains other service streams in addition to the assigned key service in the communication service demand R task , and is sorted in descending order of service priority , and the available link pool contains the candidate set C setThe other high-quality candidate link other than the occupied main communication link is ranked according to the comprehensive score S score from high to low;

[0031] The priority of the to-do task pool The highest service is processed, and the available link pool is traversed in turn to find the best matching link for it.

[0032] Further, the best matching link includes three scenarios:

[0033] Single link meets scenario, if the current to-do service in the available link pool comprehensive score S score The performance of the highest link meets the core needs of the service, then the link is assigned to the service, and the corresponding key-value pair is added in the parameter dictionary, while the service and the link are removed from the respective resource pool, ensuring that resources are not allocated repeatedly;

[0034] Link aggregation scenario, if none of the single links in the available link pool can meet the needs of the current high-bandwidth service, the system further judges whether the bandwidth of multiple links can be aggregated to meet the service demand;

[0035] Link multiplexing scenario, after completing the link allocation of high-priority tasks, the system continues to process lower-priority tasks, in addition to considering the unused link resources in the available link pool, it also intelligently assesses whether there are already allocated but still have remaining capacity links available for multiplexing.

[0036] Further, the implementation method of the adaptive closed-loop feedback includes:

[0037] According to the instruction type of the communication strategy instruction Start comm , if the instruction type is a cooperative strategy, the optimal fragmentation size S size is dynamically calculated based on the real-time available bandwidth B and the target transmission time window T of each link, and the data stream is executed according to the optimal fragmentation size S size Dynamic fragmentation to generate data packets P sliced that can be transmitted in parallel through multiple links; if the instruction type is a switching strategy, the dynamic fragmentation process is skipped;

[0038] Performing cyclic redundancy check on the received data packets P sliced , storing the data packets that pass the check in the ring buffer according to their sequence numbers for out-of-order reorganization, and combining the bidirectional scanning reorganization algorithm to intelligently decode and deliver the data packets in the ring buffer in turn, to restore the received data packets P sliced to a complete and ordered data stream D reass ;

[0039] After executing the communication strategy instruction Start comm , the execution effect of the communication strategy instruction Start comm is evaluated, the key performance indicators collected in the evaluation are packaged as a feedback data packet F back , and the feedback data packet F back is used as a training sample for iterative training of a deep Q network to update decision weights, so as to realize adaptive closed-loop feedback.

[0040] The 5G network optimization system based on the intelligent dynamic multi-band adaptive switching system is used to realize the 5G network optimization method based on the intelligent dynamic multi-band adaptive switching system, and includes a multi-source heterogeneous data acquisition and perception module, an intelligent signal shielding prediction module, a communication strategy generation module, and a closed-loop feedback optimization module.

[0041] The multi-source heterogeneous data acquisition and perception module is used to acquire real-time network signal quality data P sig , environmental perception data D env , and a communication task demand set R task .

[0042] The intelligent signal shielding prediction module extracts a dynamic obstacle feature vector set F obs based on the environmental perception data D env , integrates the environmental perception data D env and the dynamic obstacle feature vector set F obs , inputs them into a parallel double-branch hybrid deep learning model for signal shielding prediction, and outputs a multi-dimensional parameter signal shielding prediction result O pred .

[0043] The communication strategy instruction generation module is used to sequentially perform signal link feasibility verification and network resource availability verification on the acquired candidate links, filter out qualified candidate links to construct a multi-dimensional performance profile candidate set C set ; based on the multi-dimensional performance profile of each candidate link, a deep Q network is used to generate a dynamic weight factor when the communication decision flow is activated, and the comprehensive score S score of each candidate link is calculated according to the dynamic weight factor; according to the hard constraint condition of the communication service demand R task , the candidate links sorted in descending order according to the comprehensive score S score are tested one by one until a candidate link satisfying the hard constraint condition is found as the main communication link, and the final communication strategy instruction Start comm is generated by choosing between the switching strategy instruction and the cooperative communication strategy instruction according to the overall situation of the current network resources.

[0044] the closed-loop feedback optimization module is configured to collect feedback data packets F containing key performance indicators of the instruction according to the communication strategy instruction Start comm perform instruction type judgment, if the communication strategy instruction Start comm is a cooperative strategy, then perform dynamic fragmentation on the data stream to generate data packets P that can be transmitted in parallel sliced ; at the receiving end, the data packets P sliced are sequentially subjected to cyclic redundancy check, out-of-order reorganization and intelligent delivery to restore the data stream D to be complete and ordered reass ; after the communication strategy instruction Start comm is executed, collect and encapsulate feedback data packets F containing key performance indicators of the instruction back , and use the feedback data packets F back for iterative training of the deep Q network to update the decision weight, realizing adaptive closed-loop feedback.

[0045] Compared with the prior art, the beneficial effects of the present application are:

[0046] The present application improves the problem that the traditional decision system only relies on a single network quality indicator and the decision is blind and disconnected from the actual demand due to lack of comprehensive situational awareness by synchronously collecting multi-dimensional heterogeneous information such as network channel, physical environment and upper-layer service demand.

[0047] The present application improves the problem that the traditional system only relies on lagging network indicators and the switching decision lags and the response is passive due to the inability to predict physical obstructions by constructing a digital twin model to make forward-looking predictions of signal obstruction.

[0048] The present application dynamically optimizes the selection strategy of multi-band communication links by taking real-time network quality, environmental perception data, communication service demand and forward-looking obstruction prediction as comprehensive decision factors, avoiding the risk of mismatching or imminent failure of the selected link with critical service demand due to the decision of traditional methods relying on only a single optimal indicator.

[0049] The present application avoids the problem of transmission bottleneck or resource idling due to the inability to aggregate multi-link bandwidth in the traditional single-link transmission mode by dynamically fragmenting and scheduling the data stream to be transmitted according to the communication strategy instruction and the real-time bandwidth of each link. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0051] Figure 1 is a principle flowchart of the 5G network optimization method based on the intelligent dynamic multi-band adaptive switching system of the application;

[0052] Figure 2 is a functional module diagram of the 5G network optimization system based on the intelligent dynamic multi-band adaptive switching system of the application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0054] Embodiment 1

[0055] Please refer to Figure 1 The embodiment provides a 5G network optimization method based on an intelligent dynamic multi-band adaptive switching system, which comprises the following steps:

[0056] S1000, based on a vehicle integrated multi-band device, a multi-source sensor network, a positioning system and an application sensing identification module, synchronously acquiring real-time network signal quality data P sig , environment perception data D env and a communication task demand set R task .

[0057] Specifically, this step aims to provide high-quality data input basis for subsequent prediction and decision-making. By collecting diversified information of the vehicle itself, the surrounding environment and the real-time network channel, and performing standardized processing, the problem of incomplete information of a single data source and inconsistent time and space reference of multi-source data is solved, and this step is the starting point of the entire closed-loop optimization system.

[0058] Further, step S1000 comprises:

[0059] S1100, real-time monitoring of a radio frequency signal of the vehicle integrated multi-band device, pre-processing the radio frequency signal and outputting structured real-time network signal quality data P sig .

[0060] Specifically, this step provides a reliable data basis for the subsequent steps by using the real-time network signal quality data P Sig collected and output by the vehicle integrated multi-band communication device.

[0061] The in-vehicle integrated multi-band communication device integrates millimeter-wave band (24.25GHz-52.6GHz), Sub-6GHz band (3.3GHz-5GHz), and 4G LTE band communication modules. Each communication module is connected to the vehicle's central control system via a high-speed PCIe data bus, ensuring that the large amount of data generated concurrently across multiple bands can be processed without bottlenecks.

[0062] In practice, the highly integrated multi-band communication device within the vehicle terminal uses a built-in signal processing unit to continuously monitor the radio frequency signals of all available communication bands—millimeter-wave, Sub-6GHz, and 4G LTE—at a high sampling frequency of millions of times per second. This high-precision signal monitoring detects the communication link status in real time and accurately measures key signal quality parameters for each band. Simultaneously, to ensure data purity and reliability, the signal processing unit utilizes bandpass filters and adaptive equalizers to preprocess the original signal, effectively filtering out noise interference and compensating for signal attenuation and distortion during transmission.

[0063] After the above measurement and processing procedures, a structured real-time network signal quality data P is finally output. Sig ={Q i |i∈B set}; where P sig Q represents the total set of network quality information collected at a certain moment, encompassing all available communication frequency bands; it is a set, namely, a real-time network signal quality data set. i ={RSSI i SNR i BER i} represents a tuple of quality parameters for frequency band i, containing several key indicators describing the current communication quality of that frequency band, namely the Received Signal Strength (RSSI). i Signal-to-noise ratio (SNR) i and Bit Error Rate (BER) i RSSI i SNR (Signal Response Rate) represents the signal power received by the vehicle-mounted terminal antenna from the base station. It is the most direct and important indicator for assessing network coverage, measured in decibels (dBm). A higher value indicates a stronger signal and better network coverage. i BER (Block Received Power) represents the ratio of the received useful signal power to the background noise power present in the same channel, reflecting the signal clarity. The unit is decibels (dB). A higher BER indicates that the signal is less susceptible to noise interference and that data transmission reliability is higher. idenotes the proportion of the number of error bits decoded by the receiving end within a certain time to the total number of bits, used to measure the accuracy of the data transmission link, the lower the value, the higher the data transmission quality; i denotes an identifier of a specific communication frequency band; B set ={mmWave, Sub-6G, 4G-LTE} denotes a set of all detectable communication frequency bands in the current vehicle environment, including the millimeter wave frequency band mmWave, the Sub-6G frequency band and the 4G LTE frequency band.

[0064] Step S1200, based on the vehicle-mounted multi-source sensor network and the positioning system, collecting static environment data D map , vehicle self-dynamic data D pos and dynamic obstacle data D sensor , based on the static environment data D map , vehicle self-dynamic data D pos and dynamic obstacle data D sensor , constructing environment perception data D env .

[0065] Specifically, this step constructs a high-fidelity digital twin environment that is real-time synchronized with the physical world through a vehicle-mounted multi-source heterogeneous sensor network and a high-precision positioning system.

[0066] In the specific implementation process, key perception data is collected from three dimensions, as follows:

[0067] I. Static environment data D map collection. Load and call the preset high-precision map database. The map database has a centimeter-level spatial resolution and comprehensively covers key elements such as the geometric shape of the road, lane information, traffic sign and marking position, etc. More importantly, the map database records in detail the three-dimensional spatial information of various static facilities that may affect signal propagation along the road section, including their length, width, height and contour characteristics, covering typical buildings such as buildings, bridges and tunnels.

[0068] II. Vehicle self-dynamic data D pos collection: By fusing the global positioning system (GPS) and the inertial navigation system (INS), the real-time perception of the dynamic state information of the vehicle itself is realized. This fusion scheme ensures that in complex environments where GPS signals are weak or interrupted (such as tunnels, urban canyons, etc.), stable and reliable position and attitude information can still be continuously calculated relying on the inertial navigation system. The positioning accuracy of this system can reach sub-millimeter level, the speed measurement error is within ±0.1 m / s, and it has high dynamic response capability and adaptability to complex environments.

[0069] III. Dynamic obstacle data D sensorCollection: A set of environmental perception sensors are deployed on the roof or high points along the base station, continuously collecting surrounding dynamic obstacle information. The sensor set includes a 128-line mechanical laser radar with 360° omnidirectional scanning capability, a sampling frequency of 10 Hz, and a maximum detection distance of 200 meters; and an industrial-grade binocular camera with a baseline distance of 15 cm, supporting the collection of high-quality images with a resolution of 1280x960 at a frame rate of 30 fps.

[0070] Finally, the three types of data collected above are integrated to form a multi-level, structured environmental perception data D env map pos sensor ; wherein D env represents a multi-modal data set collected at a certain time, which comprehensively describes the state of the vehicle itself and its surrounding environment, and is a set, i.e. environmental perception data; D map represents static high-precision map data, which is used to predict signal obstruction caused by certain fixed objects (such as high-rise buildings, mountains), i.e. static environmental data; D pos = {p, v, θ v} represents a vehicle real-time dynamic data tuple, which is used to describe the position and motion trend of the vehicle at the current time, including vehicle geographic location coordinates p, vehicle speed v and vehicle driving direction angle θ v ; vehicle geographic location coordinates p are coordinate pairs containing longitude and latitude, which are used to position the vehicle to the static environmental data D map ; vehicle speed v represents the current instantaneous speed of the vehicle, which is a key parameter for predicting the future position of the vehicle, with a unit of meters per second m / s; vehicle driving direction angle θ v represents the angle between the vehicle's forward direction and the north direction, which is used to determine the vehicle's driving trajectory, i.e. the heading angle, with a unit of degrees °; D sensor represents three-dimensional point cloud data of dynamic obstacles, which is used to predict temporary and sudden signal obstruction caused by surrounding moving objects, i.e. dynamic obstacle data.

[0071] Step S1300, through the vehicle-mounted application perception identification module, the service type of the upper layer data stream is analyzed in real time, and a communication task demand set R is generated, which contains the minimum guaranteed bandwidth task .

[0072] Specifically, this step aims to establish a bridge between the upper layer application demand and the bottom layer network resource allocation, and dynamically map the abstract business demand of the upper layer application to a quantifiable and standardized communication task demand set R task ​​​​​. Ensure not blindly pursue the best signal, but focus on the current vehicle core, the most urgent business needs, avoid high-quality links are occupied by low-priority data, and cause the risk of key control instruction delay.

[0073] In the implementation process, the application knowledge identification module built in the vehicle central control unit runs continuously, and monitors and analyzes the data stream generated by each application program (such as remote driving control module, video stream transmission module, vehicle state monitoring module, etc.) in real time. When a new data stream is detected or the attribute of an existing data stream changes, the identification module will dynamically identify the task type according to the pre-set strategy rule library and label the communication demand. The specific identification process includes the following three typical tasks:

[0074] 1. High bandwidth task identification. When the 4K high-definition video monitoring function is enabled, the identification module will determine it as a large code rate data stream output by the video stream transmission module. According to the strategy rule library, this data stream is labeled with high bandwidth demand, allowable delay and medium priority.

[0075] 2. Low latency task identification. When the emergency braking or obstacle avoidance steering instruction is generated, the identification module will immediately identify its high-risk attribute. The system will assign it the most stringent latency requirement, the highest task priority, and its bandwidth requirement is relatively low.

[0076] 3. Regular task identification. For the vehicle state data (such as battery power, tire pressure, etc.) periodically reported by the vehicle state monitoring module, the system will assign it a relaxed latency and bandwidth requirement, and a lower task priority.

[0077] Through the above method, all non-standardized business instructions are dynamically converted into a standardized communication task demand set Where, R task represents the set of all communication task demands that the vehicle needs to handle at a certain time; j represents the data stream identifier, which is used to distinguish different data streams, for example, j = 1 represents the control signal stream, j = 2 represents the video stream; S stream represents the set of all active data stream identifiers; represents the specific demand tuple of data stream j, including the minimum guaranteed width the maximum transmission delay of data stream j and the business priority of data stream j the minimum guaranteed width of data stream j The minimum data transmission rate necessary to ensure normal operation of the business, usually expressed in Mbps, which varies greatly depending on the type of business, such as 4K high-definition video stream requiring a few hundred kbps for vehicle status data or control instructions; maximum transmission latency of data stream j denotes the maximum time allowed for a data packet to travel from the sender to the receiver, usually in milliseconds ms; service priority of data stream j the main basis for resource scheduling, usually a non-negative integer, the smaller the value, the higher the priority, such as denotes critical safety instructions (such as emergency braking), denotes real-time interactive services (such as high-definition video monitoring) and the like.

[0078] For example, a remote-controlled unmanned truck operating in a mine is assumed to be controlled by an operator at the remote control center at a certain time t, performing the task of loading ore and transporting it to a designated location. At this time, the vehicle central control unit will identify that there are currently four parallel data streams, corresponding to different service requirements, forming the set S of all active data stream identifiers stream ={1,2,3,4}. The system will define the communication task requirements for each data stream according to its service characteristics, and finally form a complete set of communication task requirements R task . Further, as shown in Table 1, Table 1 is an example table for obtaining the current communication task requirements of the present application. As can be seen from the example table, at time t, the communication service quality requirements of the four parallel data streams of the remote-operated unmanned truck. For example, the first data stream (j = 1) is a real-time control instruction from the remote operator, including steering, acceleration, deceleration and braking, and other critical safety operations. Since it is directly related to the safety of the vehicle, the system will assign it the highest service priority and the most stringent transmission latency to ensure instantaneous response to the instruction. At the same time, given the small amount of data, only a low bandwidth requirement is needed to meet the transmission requirements.

[0079] Table 1 Example of obtaining current communication task requirements

[0080]

[0081] S2000, based on environmental perception data D env , extracting a set of dynamic obstacle feature vectors F obs , based on the environmental perception data D env and the set of dynamic obstacle feature vectors F obs , integrating and inputting into a parallel dual-branch hybrid deep learning model M DL for signal occlusion prediction, outputting a multi-dimensional parameter signal occlusion prediction result O pred .

[0082] In particular, this step aims to predict signal occlusion events on the future path of the vehicle in advance based on the environment perception data D env collected in step S1200, providing forward-looking input for subsequent steps.

[0083] Further, step S2000 includes:

[0084] Step S2100, the dynamic obstacle data D sensor is processed in multiple stages to generate a dynamic obstacle feature vector set F obs containing shape and motion parameters.

[0085] In particular, this step aims to serve as a key bridge connecting sensor data and advanced prediction models. Based on the dynamic obstacle data D env in the environment perception data D sensor collected in step S1200, through a series of complex processing procedures, it is transformed into a low-dimensional and highly friendly dynamic obstacle feature vector set F obs for machine learning models, greatly reducing the computational burden of subsequent steps and serving as a technical prerequisite for real-time and high-precision signal occlusion prediction.

[0086] In the specific implementation process, a series of processing is performed based on the dynamic obstacle data D sensor , which can be divided into four core stages as follows:

[0087] I. Spatio-temporal alignment. Since the dynamic obstacle data D sensor is collected by laser radar and binocular camera respectively, and the two are physically separated heterogeneous sensors, their coordinate systems and collection times are biased, so strict spatio-temporal alignment processing is needed. First, a GPS time module is used to provide a unified time reference for the laser radar and binocular camera, ensuring that the timestamp error of each collected data frame is within ±1 ms, achieving time synchronization. Next, Zhang's calibration method is used to calculate the rotation matrix R and translation vector T between the two sensor coordinate systems to establish the spatial transformation relationship between them. Finally, through affine transformation formula, the three-dimensional points in the binocular camera coordinate system are mapped to the laser radar coordinate system, completing the spatial alignment. The specific process formula is as follows:

[0088] P lidar = R·P vision + T;

[0089] where P lidar represents a spatial point after coordinate transformation, the final three-dimensional coordinates in the target coordinate system (i.e. laser radar coordinate system); R represents the rotation matrix, which describes the rotation attitude of the source coordinate system (i.e. binocular camera coordinate system) relative to the target coordinate system in space; P visionrepresents the original three-dimensional coordinates of a spatial point in the source coordinate system; T represents a translation vector, which describes the displacement of the origin of the source coordinate system relative to the origin of the target coordinate system in the physical space.

[0090] II. Denoising and filtering. After the spatio-temporal alignment is completed, the obtained dynamic obstacle data D sensor There can still be outliers and redundant information caused by sensor errors or environmental reflections and other factors. The system uses statistical filtering and voxel filtering methods to filter the dynamic obstacle data D sensor for data cleaning and compression processing.

[0091] The statistical outlier filtering method is based on the statistical distribution of the distance between each point and its neighborhood points to identify and remove outliers deviating from the main distribution. Specifically, the number of neighborhood points is set to k = 50, and the average Euclidean distance between each point and its 50 neighboring points is calculated. If the average distance of a point and its neighborhood exceeds the pre-set threshold, the point is determined to be a noise point and is removed.

[0092] The voxel grid filtering method reduces the data density while preserving the contour features of the point cloud. Specifically, the coordinate space in which the entire point cloud is located is divided into a cubic voxel grid with a size of 0.1m x 0.1m x 0.1m. Then, all voxels in the coordinate space are traversed, and for each voxel containing one or more data points, the average value of the three-dimensional coordinates of all data points within it is calculated as the representative point, i.e. the voxel barycenter, which replaces all data points within the voxel. If no data points fall within a voxel, it will be empty after downsampling. In this way, the number of redundant points in local high-density areas can be significantly reduced, and under normal circumstances, data compression of 60% to 80% can be achieved, saving a large amount of computing resources for subsequent steps.

[0093] III. Morphological feature extraction. After denoising and filtering, based on the dynamic obstacle data D sensor The convex hull algorithm (such as Graham scan) is used to calculate the smallest circumscribed cuboid that wraps a single obstacle point cloud cluster, thereby accurately determining its length l, width w, and height h. On this basis, to further obtain the spatial occupancy characteristics of the obstacle, the data points are divided into finer subcubes, and the data point density p i' of each subcube is calculated. Then, weighted summation is performed to estimate the total volume V of the obstacle. The specific process is as follows:

[0094]

[0095] where V represents the total volume of the obstacle, with a unit of cubic meters m 3 ; i' represents the index of each subcube, used to traverse all subcubes constituting the obstacle; n represents the total number of all subcubes of the obstacle; p i'V represents the point density of the i'-th sub-cube, i.e., the number of data points within each sub-cube; i' This represents the volume of each sub-cube, a fixed value by default, in cubic meters (m). 3 .

[0096] IV. Trajectory Feature Extraction. To obtain the motion state information of obstacles, a Kalman filter algorithm is used to track and predict the trajectory of the centroid of the same obstacle identified in consecutive frames. By calculating the position change of the centroid between consecutive time steps, the instantaneous velocity s and instantaneous motion direction angle θ of the obstacle can be accurately estimated. s The specific process formula is as follows:

[0097]

[0098] Where s represents the instantaneous velocity of the obstacle, in meters per second (m / s); Δx and Δy represent the displacement components of the obstacle's center of mass in the X and Y axes respectively during two consecutive time steps, in meters (m); Δt represents the time interval between two consecutive time frames, in seconds (s); θ s The instantaneous direction of the obstacle's motion is represented by the angle between the obstacle's velocity vector and the positive X-axis of the coordinate system, expressed in degrees (°). arctan(·) represents the arctangent function, used to calculate the ratio of the increments of the Y-axis and X-axis in a rectangular coordinate system. Calculate the instantaneous angle θ of the obstacle's motion direction. s .

[0099] The centroid is a single coordinate point obtained by averaging the three-dimensional coordinates of all data points within the obstacle, and is used to represent the tracking anchor point of the entire obstacle.

[0100] Finally, all the extracted morphological and trajectory features are combined to construct a structured set of dynamic obstacle feature vectors. Among them, F obs This represents the total set of feature vectors of all identified dynamic obstacles in the scene at the current moment; j' represents the obstacle identifier, used to distinguish different obstacles, for example, j=1 represents a truck in front, j=2 represents a car on the side; This represents a single characteristic vector corresponding to the j'-th obstacle, a tuple containing obstacle shape and trajectory information, including the obstacle's length l. j' The width of the obstacle j' The height h of the obstacle j' The total volume V of the obstacle j' The instantaneous velocity s of the obstacle j' and the instantaneous direction of motion θ of the obstacle s,j' .

[0101] Step S2200, integrate the static environment data D map , the vehicle's own dynamic data D pos , and the dynamic obstacle feature vector set F obs , and input them into a parallel double-branch hybrid deep learning model M DL , respectively extract spatial feature vectors F cnn and temporal feature vectors F rnn , fuse the spatial feature vectors F cnn and the temporal feature vectors F rnn to construct spatio-temporal fusion feature vectors F fused , decode the spatio-temporal fusion feature vectors F fused using a prediction decoding head, and output a signal obstruction prediction result O pred containing predicted affected frequency bands bank k , predicted signal attenuation values A k , predicted start times of obstruction events , and predicted durations of obstruction events .

[0102] Specifically, this step aims to generate a signal obstruction prediction result O pred with foresight and practicality based on the multi-source environment perception data and dynamic obstacle feature information extracted in all previous steps, to support subsequent active network scheduling and communication link optimization strategies.

[0103] In the specific implementation process, first, the aforementioned data information is collected, including the static environment data D env and the vehicle's own dynamic data D map in the environment perception data D pos collected by step S1200, and the dynamic obstacle feature vector set F obs constructed by step S2100.

[0104] Then, the collected data information is input as an input into a hybrid deep learning model M DL trained with large amounts of historical data. The hybrid deep learning model M DL adopts a parallel double-branch architecture, one branch being a convolutional neural network (CNN) for processing spatial information, and the other branch being a recurrent neural network (RNN) for processing time series information.

[0105] The spatial feature extraction branch is mainly responsible for extracting spatial morphological features that affect signal propagation from static and dynamic physical environments. First, the static environment data D mapand dynamic obstacle feature vector set F obs = {l j' , w j' , h j' , V j' | j' = 1, 2,..., m} are fused and constructed into a structured spatial morphology tensor, denoted as X spatial . Subsequently, the spatial morphology tensor X spatial is input into a convolutional neural network (CNN) branch, which realizes step-by-step abstraction and feature extraction of spatial information through multiple convolution and pooling operations, and finally outputs a compact spatial feature vector F cnn containing all key spatial information. The CNN branch is stacked with multiple convolutional and pooling layers, and the filter parameters in each convolutional layer are optimized through supervised training, which can automatically identify spatial patterns with physical meaning, such as "edge of high wall", "street canyon formed by continuous buildings", "metal block of large vehicle", etc.

[0106] The time series feature extraction branch is mainly responsible for learning and understanding the dynamic evolution law of the entire scene from the historical motion trajectories of vehicles and obstacles. First, the vehicle's own dynamic data D pos and the dynamic obstacle feature vector set F obs = {s j' , θ s,j' | j' = 1, 2,..., m} are integrated in time sequence to construct a structured time sequence state sequence tensor, denoted as X temporal = [State t-N ,..., State t-1 , State t ]; where t represents the index of the current time; N represents the length of the time window constituting the time sequence, i.e. the number of consecutive data frames; Srate t-N represents the first state vector in the time sequence, i.e. the scene state vector N-1 time steps ago; State t-1 represents the scene state vector at the previous time t-1; State t represents the scene state vector at the current time t. Subsequently, the time sequence state sequence X temporalThe input is processed by a recurrent neural network (RNN) branch, which typically uses a long short-term memory (LSTM) network with long-term dependency construction capabilities. The LSTM network can selectively remember key dynamic information from the past (such as an obstacle accelerating) and filter out irrelevant or redundant information for the current task through an internal gating mechanism. By parsing the entire sequence step by step, the LSTM network can capture complex dynamic interaction patterns and achieve prior perception of potential future events. Finally, a time feature vector F rnn .

[0107] The spatial feature vector F cnn output by the two parallel branches and the time feature vector F rnn are concatenated in the feature dimension to form a higher-dimensional spatio-temporal fusion feature vector F fused .

[0108] Finally, the spatio-temporal fusion feature vector F fused is input into the model's prediction decoding head. The prediction decoding head is typically a multilayer perceptron (MLP) composed of several fully connected layers, which is used to decode the high-dimensional spatio-temporal fusion feature vector F fused into understandable, structured signal occlusion prediction results The MLP is trained using a specialized multi-task learning approach and contains multiple parallel output branches corresponding to different types of prediction targets, which are used to answer key questions about future signal states. The functions of each branch are as follows:

[0109] I. Communication band occlusion risk classification, i.e., a classification problem. This branch is used to determine which communication band is most likely to be affected by occlusion. Its output is passed through a Softmax activation function to calculate the probability value for each candidate communication band (e.g.,'mmWave', 'Sub-6G', '4G-LTE'), and the band with the highest probability is selected as the prediction result, i.e., the predicted affected band bank k .

[0110] II. Signal attenuation strength prediction, i.e., a regression problem. This branch is used to predict the signal strength attenuation value caused by the occlusion event. The output is a continuous real number value in decibels (dB), corresponding to the predicted signal attenuation value A k .

[0111] Three, the prediction of the occurrence time of the shielding event, i.e., the regression problem. This branch is used to predict the near risk of the signal terminal, and according to the relative motion relationship between the vehicle and the obstacle, the time countdown is calculated, which is in seconds s, as the predicted start time of the shielding event

[0112] Four, the prediction of the duration of the shielding, i.e., the regression problem. This branch combines the size of the obstacle and the relative speed of the vehicle, i.e., how long the shielding state is expected to last, which corresponds to the predicted duration of the shielding event

[0113] For example, an automatic driving vehicle is driving under a city viaduct and is about to enter the entrance of a long tunnel. In this scenario, the hybrid deep learning model M DL performs parallel analysis on the multi-source input data. Among them, the spatial feature extraction branch is responsible for interpreting the spatial dimension information, and by analyzing the static environment data D map and the dynamic obstacle feature vector set F obs , a large concrete structure with strong electromagnetic shielding characteristics, i.e., the tunnel body, located in front is quickly identified, which poses a significant shielding risk to high-frequency communication signals. At the same time, the time series feature extraction branch processes the time series data, and according to the vehicle's own dynamic data D pos and the dynamic obstacle feature vector set F obs , it captures the motion trend that the current vehicle is gradually approaching the tunnel entrance at a stable speed in real time. The abstract feature vectors output by the two branches respectively represent the spatial feature vector F cnn “there is a tunnel in front” and the time feature vector F rnn “the vehicle is approaching stably”. Then, the two feature vectors are spliced in the feature dimension to construct the spatio-temporal fusion feature vector F fused , which constructs a complete semantic description of the event, i.e., “the vehicle is driving towards a large tunnel at a stable speed”. Then, the spatio-temporal fusion feature vector F fused is input into the prediction decoding head of the model, and the structured shielding risk assessment result is output. The model determines that in this scenario, the millimeter wave mmWave communication frequency band bank k will be affected the most severely, and the predicted signal attenuation value A k will be as high as 50 dB. Based on the condition that the vehicle is currently about 100 meters away from the tunnel entrance and the speed is 20 m / s, it is calculated that the shielding event will occur in about 5.0 seconds . According to the full length of the tunnel in the map, which is 800 meters, it is further predicted that the shielding state will last for about 40.0 seconds Finally, the model outputs a signal shielding prediction result O pred, including at least one key prediction tuple ('mmWave', 50, 5.0, 40.0). The model can also output secondary prediction results for other communication frequency bands, such as ('Sub-6G', 25, 5.0, 40.0). The signal obstruction prediction result O pred contains precise forward-looking information such as specific frequency bands, attenuation levels, start times, and duration, which is transmitted to subsequent steps to ensure that the system takes optimal intervention and adjustment strategies in advance to minimize the impact of obstructions on communication service continuity and reliability.

[0114] S3000, the acquired candidate links are sequentially subjected to signal link feasibility verification and network resource availability verification to filter out qualified candidate links to construct a candidate set C of multi-dimensional performance profiles set ; based on the multi-dimensional performance profiles of the candidate links, a deep Q network is used to generate a dynamic weight factor when the communication decision-making process is activated, and the comprehensive score S of each candidate link is calculated based on the dynamic weight factor score ; based on the hard constraint conditions of the communication service requirements R task , the candidate links sorted in descending order of the comprehensive score S score are subjected to hard constraint condition verification one by one until a candidate link that meets the hard constraint condition is found as the main communication link. According to the overall situation of the current network resources, a choice is made between the handover strategy instruction and the coordination strategy instruction to generate the final communication strategy instruction Start comm .

[0115] Specifically, this step aims to model and comprehensively analyze the real-time network signal quality data P Sig collected in step S1100, the environment perception data D env collected in step S1200, the communication service requirements R task obtained in step S1300, and the signal obstruction prediction result O pred generated in step S2200, and finally generate a unique and optimal communication strategy instruction Start comm in the form of an instruction to the execution layer.

[0116] Further, step S3000 includes:

[0117] Step S3100, obtain the candidate link list and network load rate, based on real-time network signal quality data P Sigperforming a signal link feasibility check on the candidate link list, performing a network resource availability check on the candidate link list based on the network load rate, and after the feasibility and availability checks are completed, all remaining qualified candidate links are collectively constructed into a candidate set C set , and based on each candidate link in the candidate set C set , a multi-dimensional performance profile is generated, which includes a signal quality indicator N sig , a network load indicator N load , and a distance indicator N dis .

[0118] In a specific implementation process, first, the vehicle-mounted system initiates a request to the current serving base station and obtains a neighbor relation list (Neighbor Relation List). The neighbor relation list not only contains the identity and frequency information of the candidate link list of the neighboring cells (Neighboring Cells) or available backup communication frequency bands, but also contains corresponding geographic location information and other static data, which provides necessary support for subsequent evaluation. After obtaining the candidate link list, a strict two-stage pre-qualification and pruning (Pre-qualification and Pruning) process is performed on each candidate link, as follows:

[0119] I. Signal link feasibility check. Based on real-time network signal quality data P Sig , the signal strength of each candidate link is evaluated. Any candidate link with a received signal strength RSSI lower than -110 dBm will be directly excluded from the candidate list because the signal coverage is too weak and does not have the basic conditions to establish a stable physical connection.

[0120] II. Network resource availability check. If a strong signal link is already saturated in resources, it cannot provide quality service. Therefore, the network load rate obtained from the system information (System Information Blocks, SIBs) or proprietary signaling broadcast by the analysis base station is further analyzed to determine the current load status of the candidate link. Any overloaded link with a network load rate exceeding 90% will be pruned because it cannot guarantee service quality.

[0121] After the above feasibility and availability screening, all remaining qualified candidate links collectively constitute the final candidate set C set . To compare these qualified candidate links in a unified mathematical framework, a multi-dimensional performance profile is further constructed for each link, and the dimensions are normalized to eliminate the influence of dimensions. The multi-dimensional performance profile covers multiple key performance indicators, which are derived from different data dimensions: a signal quality indicator N , a network load indicator N , and a distance indicator N .sig is calculated according to the signal-to-noise ratio SNR in real-time network signal quality data P Sig , which is used to reflect the physical transmission capability of the link; the network load indicator N load is extracted by analyzing the system information (System Information Blocks, SIBs) or dedicated signaling broadcast by the base station, which is used to measure the resource usage of the candidate link; the distance indicator N dis is calculated by real-time Euclidean distance between the real-time GPS coordinates of the vehicle itself dynamic data D env in the environment perception data D pos and the geographic coordinates of the candidate base station, which is used to estimate the potential propagation delay and interference risk of the link. To realize the unified comparison between different performance indicators, the maximum-minimum normalization paradigm is adopted to map all indicators to the closed interval [0, 1]. Taking the signal quality SNR as an example, the specific process of calculating its score N sig is as follows:

[0122]

[0123] where N sig represents the normalized signal quality indicator, which is a dimensionless floating point number between 0 and 1, and the higher the value, the closer the signal quality to the ideal state; SNR current represents the current real-time signal-to-noise ratio of the candidate link, which is used to measure the core indicator of signal clarity, with the unit of decibel dB; SNR min and SNR max represent the lower limit of acceptable signal-to-noise ratio and the upper limit of ideal signal-to-noise ratio respectively, which are reference parameters set according to engineering experience.

[0124] Step S3200, based on real-time network signal quality data P Sig and signal shielding prediction results O pred , a dual-mode decision trigger mechanism is used to determine whether to activate the communication decision process. If the communication decision process is activated, a deep Q network is used to generate a dynamic weight factor according to the state vector of the current communication environment, and the dynamic weight factor and the multi-dimensional performance profile of each candidate link are used to calculate the comprehensive score S score of each candidate link.

[0125] In the specific implementation process, in order to reduce the calculation burden and improve the response efficiency, a dual-mode decision trigger mechanism (Dual-mode Decision Trigger) is designed to ensure that the complete communication decision process is activated only when necessary. This mechanism is activated when any of the following conditions is met: first, the passive response trigger mechanism. When the real-time network signal quality data P SigThe key performance indicator (such as received signal strength indication, RSSI) of the current link in use is shown to be lower than the preset minimum availability threshold of -110 dBm, indicating that the quality of service has deteriorated unacceptably; secondly, the active prevention trigger mechanism. When the signal blocking prediction result P pred The current communication link is shown to be facing a high-probability, high-attenuation signal blocking event in the near future, indicating that the system anticipates the impending deterioration of the quality of service.

[0126] When the communication decision-making process is activated, the system enters the dynamic weighting and comprehensive scoring stage to evaluate the suitability of all candidate communication links. Based on a deep Q network (DQN) reinforcement learning model, a set of optimal dynamic weight factors (W sig ,W load ,W dis ) is dynamically generated in combination with the current communication environment state vector. The state vector includes the communication service demand R task , the candidate set C set generated based on step S3100, and the static environment data D env in the environment perception data D map . Subsequently, each qualified candidate link in the candidate set C set is traversed in turn, and the signal quality indicator N sig , network load indicator N load , and distance indicator N dis of each candidate link are calculated in combination with the dynamic weight factors (W sig ,W load ,W dis ) and step S3100. A comprehensive, quantitative comprehensive score S score is calculated for each candidate link through a multi-objective optimization formula, and the specific process formula is as follows:

[0127] S score =W sig ·N sig +W load ·(1-N sig )+W dis ·(1-N dis );

[0128] wherein S score represents the final comprehensive score of the candidate link; W sig ,W load ,W dis represent the dynamic weight factors of signal quality, network load, and distance, respectively, and the sum of the three is 1, dynamically generated by the DQN model according to the real-time scene; (1-N sig ) and (1-N disThis is used to convert negative metrics, such as network load and distance metrics, where smaller values ​​are considered better, into positive metrics.

[0129] Step S3300, based on communication service requirements R task The hard constraints, based on the comprehensive score S score Candidate set C sorted in descending order set The candidate link ranked first is subjected to hard constraint testing. If the candidate link meets the hard constraints, it is determined as the main communication link; if the candidate link does not meet the hard constraints, the process is automatically iterated to the candidate set C after descending order. set The next candidate link is then checked against the hard constraints until a candidate link that meets the hard constraints is found. Based on the overall network resource situation, a choice is made between the handover policy instruction and the coordination policy instruction to generate the final communication policy instruction, Start. comm .

[0130] Specifically, this step aims to act as a bridge between abstract analysis results and concrete actions, by combining the comprehensive score S calculated using S3200. score Candidate set C generated in step S3100 set Communication service requirements R obtained from S1300 task Generate a well-structured communication strategy instruction Start that can be directly parsed and executed. comm .

[0131] In the specific implementation process, firstly, the current communication service requirement R is analyzed. task Extract the data stream task with the highest business priority (i.e., priority) from it. Performance requirements of the key data stream j (the item with the smallest numerical value) And identify its core performance requirements, including minimum protection width. and maximum transmission delay As a hard constraint for link allocation, it must be met unconditionally. Subsequently, based on the comprehensive score S... score For candidate set C set The candidate links are sorted in descending order, and the top-ranked candidate link is selected as the initial optimal target. A constraint satisfaction check is then performed on this candidate link to determine whether it simultaneously satisfies the aforementioned minimum guaranteed width. and maximum transmission delay The system considers the hard constraints. If a candidate link meets all hard constraints, it is confirmed as the primary link capable of carrying critical tasks, and the system proceeds to the next step of strategy formulation. If a candidate link does not meet any hard constraints, the system discards the candidate link and automatically iterates to the descendingly sorted candidate set C.set the next candidate link (i.e., the comprehensive score S score of the second candidate link), the same constraint satisfaction check is performed. This iteration process proceeds in order according to the comprehensive score S score ranking until the first candidate link that satisfies all the critical service requirements is located.

[0132] Finally, after a main communication link that satisfies the constraint condition is selected, the system generates a communication strategy instruction Start comm according to the overall situation of the current network resources, between two structurally and logically different communication strategies, as follows:

[0133] I. Generation of the Handover strategy instruction. When the system determines that the current network environment is relatively tense (i.e., the quality of the remaining candidate links is generally poor in addition to the selected main communication link), or the main purpose of this decision is to avoid the risk of deterministic failure of the current link, the system starts a simple handover strategy instruction generation process. This process aims to construct a strategy instruction tuple in the form of ('HANDOVER', {params}); where 'HANDOVER' represents the type of strategy instruction, used to instruct the network switching of the execution layer; {params} represents the parameter dictionary, including the target band parameter and the reason parameter. The target band parameter is used to explicitly indicate the frequency band to which the finally selected main communication link belongs, and its value is determined by the final result of the above-mentioned candidate link constraint satisfaction check. For example, if the system finally selects a Sub-6G frequency band link as the main communication link for the current task, the value of target band is 'Sub-6G'. The reason parameter is used to support subsequent backtracking and model optimization, and records the trigger reason for this link switching. For example, if the decision is activated by the passive response trigger mechanism in step S3200, such as the received signal strength indication RSSI being lower than the minimum available threshold, then reason is set to 'ReactiveThreshold Cross'; if the decision is activated by the proactive response trigger mechanism in step S3200, such as the signal blockage prediction result O pred predicting blockage, then reason is set to 'Proactive PredictedBlockage'.

[0134] II. Generation process of the Synergy strategy instruction. When the system determines that the current network environment is abundant in resources (i.e., in addition to the determined main communication link, the candidate set C setThere are also other high-quality links, and the system will start a higher-level synergy strategy instruction generation process. This process aims to build a strategy instruction tuple in the form of ('SYNERGY', {params'}); where 'SYNERGY' represents the strategy instruction type, which is used for multi-concurrent transmission by the instruction execution layer; {params'} represents the parameter dictionary of the dynamic routing table, which describes the mapping relationship between the current multi-service flow and the multi-communication link, including control task parameters, video task parameters and telemetry task parameters. Among them, control task parameters represent control tasks, that is, real-time control class data streams of vehicles, which belong to the most critical and highest priority communication tasks; video task parameters represent video tasks, that is, video data streams collected and transmitted by vehicle-mounted cameras; telemetry task parameters represent telemetry tasks, that is, telemetry data streams of the running state of the vehicle. The specific process is as follows:

[0135] I. Assign key services. First, assign the control task control task with the highest priority to the determined main communication link, and write it as the first key-value pair in the parameter dictionary. For example, if the system finally selects a mmWave frequency band link as the main communication link for the current task, the parameter dictionary is initialized as {'control task ':'mmWave'}.

[0136] II. Build a pool of to-be-allocated resources. Then, create two temporary resource pools: a to-do task pool and an available link pool. The to-do task pool contains other service streams (such as video, telemetry, etc.) in addition to the allocated key service in the communication service demand R task , and is sorted in descending order of service priority ; the available link pool contains other high-quality candidate links in addition to the occupied main communication link in the candidate set C set , sorted in descending order of comprehensive score S score .

[0137] III. Iterative matching and allocation. Next, process the highest priority service from the to-do task pool and iterate through the available link pool to find the best matching link. The matching process involves complex intelligent decision-making logic, including the following three typical scenarios:

[0138] I. Single-link satisfaction scenario. For the current to-do service (e.g., video task video task ), first evaluate the comprehensive score Sscore The highest link, whether its single-link performance meets the core requirements of the service (e.g. minimum guaranteed bandwidth ). If the condition is met, the link is assigned to the service, and the corresponding key-value pair is added to the parameter dictionary, such as {'video task ': 'Sub-6G'}. At the same time, the service and the link are removed from the respective resource pools, ensuring that resources are not duplicated.

[0139] II. Link aggregation scenario. If none of the single links in the available link pool can meet the current high-bandwidth service (such as 4K video) requirements, the system will further determine whether the bandwidth of multiple links can be aggregated to meet the service requirements. For example, the total bandwidth of the top two links in the available link pool is calculated score The sum of the bandwidths of the top two links, if the sum exceeds the minimum guaranteed bandwidth The system will decide to aggregate the bandwidths of the two links to jointly carry the service. At this time, the key-value pair added to the parameter dictionary is in the form of a list, such as {'video task ': ['Sub-6G', '4G-LTE']}.

[0140] III. Link multiplexing scenario. After completing the link allocation of high-priority tasks (such as video task video task ), the system continues to process lower-priority tasks (such as telemetry task telemetry task ), not only considering the unused link resources in the available link pool, but also intelligently assessing whether there are already allocated but still have remaining capacity of the link can be multiplexed. For example, taking the telemetry task telemetry task as an example, due to its extremely low bandwidth requirement, the system will prefer to load it onto a link that has been activated for other tasks (such as video task video task ) (e.g. 4G-LTE), rather than allocating a new link resource. This strategy can significantly improve the overall link resource utilization and reduce system power consumption while ensuring communication quality. At this time, the parameter dictionary may appear {'telemetry_task': '4G-LTE'}, sharing the same link with the video task video task .

[0141] The above iterative matching and allocation process is repeated until all services in the pending task pool are successfully allocated, or all resources in the available link pool are exhausted. Finally, the parameter dictionary output by the system constitutes an optimal communication strategy instruction Start comm tailored to the current network environment and service demand.

[0142] S4000, according to the communication strategy instruction Start comm Perform instruction type determination; if the communication strategy instruction Start... comm For a collaborative strategy, dynamic fragmentation is performed on the data stream to generate data packets P that can be transmitted concurrently. sliced At the receiving end, the data packet P sliced Cyclic redundancy check, out-of-order reordering, and intelligent delivery are performed sequentially to restore the complete and ordered data stream D. reass ; in the communication strategy instruction Start comm After execution, a feedback data packet F containing the key performance indicators of the instructions is collected and encapsulated. back and the feedback data packet F back Iterative training of deep Q-networks is used to update decision weights and achieve adaptive closed-loop feedback.

[0143] Specifically, this step aims to implement the communication policy instruction Start generated in step S3000. comm The process is executed and feedback is given to the decision-making level, forming a closed loop.

[0144] Further, step S4000 includes:

[0145] Step S4100, according to the communication strategy instruction Start comm If the instruction type is a cooperative strategy, then the optimal fragment size S is dynamically calculated based on the real-time available bandwidth B of each link and the target transmission time window T. size And based on the optimal fragment size S size Dynamically fragment the data stream to generate data packets P that can be transmitted in parallel across multiple links. sliced If the instruction type is a switching strategy, then the dynamic sharding process is skipped.

[0146] Specifically, this step aims to start the communication policy instruction. comm This translates into executable behavior at the physical network layer, acting as a bridge and scheduler between the decision-making and execution layers. The core of this step lies in using dynamic data stream fragmentation and multi-link multiplexing to precisely translate upper-layer strategies into specific data transmission operations, ensuring the system's high efficiency and real-time performance.

[0147] In the specific implementation process, the system follows the communication strategy instruction Start comm Different instruction types execute different processing logic:

[0148] If the instruction type is 'HANDOVER', the system needs to seamlessly migrate a complete and unchanged data stream from the current link to another target link. This process mainly involves underlying signaling interaction and network path redirection, without the need to split or reconstruct the data stream itself. Therefore, the data slicing logic of this step is directly skipped.

[0149] However, when the instruction type is 'SYNERGY', the core logic of this step is activated. At this time, the system needs to intelligently divide one or more large-volume data streams (such as high-definition video streams) and schedule them in parallel to multiple communication links. At this time, dynamic slicing becomes a necessary and critical operation. This process aims to achieve the best trade-off between transmission efficiency and network robustness: if the slice is too large, a single packet loss will result in large data retransmission overhead and is not conducive to delay control; if the slice is too small, it will introduce excessive packet header overhead, reducing overall bandwidth utilization. Therefore, the system needs to dynamically calculate the optimal slice size S size according to the real-time carrying capacity of each link and the delay requirement of the application. The specific process formula is as follows:

[0150]

[0151] where S size represents the optimal slice size, i.e. the effective data amount that each data packet can carry, with the unit of bytes Bytes; B represents the current real-time available bandwidth of the target link, which is a dynamically changing value obtained by network measurement, reflecting the instantaneous carrying capacity of the channel, with the unit of bits per second bps; T represents the target transmission time window, which is used to limit the expected transmission time of each slice, balancing transmission delay and protocol overhead, and the value is determined according to the scene and application requirements, with the unit of seconds s; 8 is a unit conversion factor, used to convert the data stream in bits to bytes, i.e. 1 Byte = 8 bits.

[0152] After calculating the S size of each link participating in synergy for each link participating in synergy, the slicing engine cuts out data blocks of the corresponding size from the original data stream in sequence. Each sliced data packet P sliced is attached with key metadata, at least including a unique 64-bit sequence number to ensure that the receiving end can correctly reassemble the data order, and a synchronized millisecond timestamp to ensure timing consistency. Finally, these processed and labeled data packets will be accurately distributed to the designated communication links according to the dynamic routing table defined by Start comm .

[0153] For example, to transmit a 4K high-definition video stream, the communication strategy instruction Startcomm {'video task ['Sub-6G','4G-LTE']}} indicates that the system needs to simultaneously distribute the high-definition video stream to both Sub-6G and 4G-LTE links for parallel transmission. Assuming that the system measures the available bandwidth B sub6g of the Sub-6G link to be 80 Mbps, the available bandwidth B 4glte of the 4G-LTE link to be 20 Mbps, and the currently set target transmission time window T to be 0.2 seconds. According to the above parameters, the system calculates the optimal fragment size for each link respectively: Sub-6G link: 4G-LTE link: Sub-6G link: 2.0 MB 4G-LTE link: 0.5 MB Subsequently, the fragment engine proportionally splits the original 4K high-definition video data stream according to the above results, generating two data fragments in each 0.2-second time window: one with a size of 2.0 MB scheduled to the Sub-6G link, and the other with a size of 0.5 MB allocated to the 4G-LTE link. Finally, the system completes a total data throughput of 2.5 MB in each transmission period, corresponding to an equivalent total bandwidth of about 100 Mbps (2.5 MB x 8 ÷ 0.2 s), achieving the maximum utilization of multi-link aggregation and effectively guaranteeing the continuity and stability of 4K high-definition video in complex network environments.

[0154] Step S4200, the received data packet P sliced is subjected to cyclic redundancy check, and the data packets that pass the check are stored in a ring buffer according to their sequence numbers for out-of-order reorganization. A bidirectional scanning reorganization algorithm is used to sequentially perform intelligent decoding and delivery on the data packets in the ring buffer, so as to restore the received data packet P sliced to a complete and ordered data stream D reass .

[0155] Specifically, this step aims to verify and reorganize the data packets P sliced of step S4100 one by one, ensuring that they can be accurately restored to the original, ordered, error-free complete data stream at the millisecond level after experiencing parallel, multi-link transmission, thereby providing reliable and highly consistent data support for upper-layer applications such as video decoding and control instruction parsing.

[0156] In the specific implementation process, all data packets P sliced arriving at the receiving end will enter a three-stage automated processing flow in turn, as follows:

[0157] I. Integrity Check via CRC. Each data packet P slicedAfter arriving at the receiving end, it is necessary to pass through the cyclic redundancy check (CRC) to verify whether it has occurred error in the transmission process. In order to balance the calculation burden and reliability requirements, the system adopts hierarchical checking mechanism: for key data packets, such as packets marked in the header of the video I frame or P frame header information, since its content correctness is directly related to the decoding accuracy of subsequent image sequence, the system preferentially adopts CRC-32 check with stronger error detection capability. Although compared with the lightweight check method, it increases the check overhead of 4 bytes, but greatly guarantees the integrity of the key structure data in the transmission process; for non-key data packets, such as packets carrying video B frame content, since its importance is lower than that of I frame or P frame, even if individual loss, the decoder can still tolerate reconstruction by referring to other frames, the system then preferentially adopts CRC-16 check algorithm with lower calculation overhead in such scenarios, to ensure the smoothness of the basic integrity and reduce the processing burden.

[0158] Any data packet that does not pass the cyclic redundancy check CRC will be considered as damaged in the transmission process, the system will immediately discard the data packet and initiate a retransmission request to the upstream sending end, to ensure the reliability and integrity of the overall data transmission.

[0159] II. Reordering and reassembly based on sequence number. Under the cooperative strategy, data packets from multiple links will usually arrive in disordered form. To solve this problem, the receiving end maintains a circular buffer with sufficient capacity for each session. When the data packet passing through the cyclic redundancy check CRC enters this stage, the system reads its unique 64-bit sequence number in the packet header, and deposits the payload of the data packet into the corresponding slot in the buffer according to the sequence number. As data packets continue to arrive, the slots in the buffer are gradually filled, and the disordered data is reordered in space to restore the correct transmission order.

[0160] III. Intelligent Decoding and Delivery. To avoid the large initial delay caused by waiting for the buffer to be completely filled, the system adopts an efficient bi-directional scanning reassembly algorithm. This algorithm continuously scans the ring buffer, starting from the last successfully delivered position, and identifies and submits one or more complete and independently decodable video frames, especially the key I frame as the decoding starting point. Once it is confirmed that all the data packets corresponding to these complete frames are in place, the bi-directional scanning reassembly algorithm immediately extracts the complete data block from the buffer and delivers it to the upper-layer application (such as a video decoder) without waiting for subsequent data packets with further sequence numbers to arrive. This active and continuous delivery mechanism effectively shortens the reassembly delay caused by network jitter and out-of-order.

[0161] Finally, after the complete verification, sorting and reassembly process, the originally scattered data packets are accurately restored to complete and ordered data stream D reass , which is delivered to the upper-layer application.

[0162] Step S4300, after executing the communication strategy instruction Start comm , the execution effect of the communication strategy instruction Start comm is evaluated, the key performance indicators collected in the evaluation are packaged as feedback data packet F back , and the feedback data packet F back is used as a training sample for iterative training of the deep Q network to update the decision weight, realizing adaptive closed-loop feedback.

[0163] In specific implementation, the execution process of this step mainly includes two stages of strategy execution and effect evaluation, as follows:

[0164] I. Strategy execution. In this stage, the system processes the communication strategy instruction Start comm with the instruction type 'HANDOVER' from step S3000. When the handover instruction is received, the system issues it to the underlying communication module, triggering the standard Radio Resource Control (RRC) connection reconfiguration process. This process is a strict signaling interaction mechanism designed to achieve coordinated handover between the current base station and the target base station. By dynamically negotiating the data transmission link between the source path and the target path, the interruption time during the handover process is minimized, ensuring the continuity of data transmission and the stability of communication services.

[0165] II. Effect evaluation and feedback packaging. After each communication strategy instruction Start commAfter the action is completed, whether it is a 'HANDOVER' switching strategy or a 'SYNERGY' synergy strategy, the monitoring module of the system will immediately evaluate its execution effect. The evaluation process will collect a series of key performance indicators, and package the collected information into a structured feedback data packet F back . The feedback data packet F back includes four key fields: the timestamp 'timestamp' indicates the timestamp of this strategy execution and its evaluation, which records the specific time of the event occurrence, providing a time reference for subsequent training; the action 'action taken ' indicates the complete communication strategy instruction Start comm of this execution, such as {'SYNERGY', {'video task ': ['Sub-6G', '4G-LTE']}}; the action result 'outcome' is used to mark the result state of the strategy execution, such as 'SUCCESS' for success, 'FAILURE' for failure, and 'TIMEOUT' for timeout; and the quantitative performance indicators'metrics' indicate the quantitative performance indicator dictionary of the network state after the strategy execution, such as {'latency': 85, 'post action-rssi ': -89, 'post action-snr ': 25}. Among them, 'latency' represents the end-to-end communication delay after the strategy execution, in ms; 'post action-rssi ' represents the received signal strength of the new link after the strategy execution, in dBm; and 'post action-snr ' represents the signal-to-noise ratio of the new link after the strategy execution, in dB.

[0166] Finally, the feedback data packet F back is transmitted to the deep Q network (DQN) learning system deployed in step S3200, and is included in the model update process as a new training sample. By continuously introducing feedback data packets F back derived from real environments for iterative training, the decision weight factor of the DQN model can be continuously and dynamically updated. With this closed-loop feedback mechanism, the system can achieve self-evolution and continuous optimization in complex and variable communication environments, enhancing the overall intelligent decision-making ability and environmental adaptability.

[0167] Embodiment 2

[0168] This embodiment is based on embodiment 1 and provides a 5G network optimization system based on an intelligent dynamic multi-band adaptive switching system, as shown in Figure 2 , which includes:

[0169] A multi-source heterogeneous data acquisition and sensing module is used to acquire real-time network signal quality data P. sig Environmental perception data D env and the communication task requirement set R task ;

[0170] The intelligent signal obstruction prediction module is based on environmental perception data D. env Extract the set of dynamic obstacle feature vectors F obs Based on the environmental perception data D env and the set of dynamic obstacle feature vectors F obs The data is integrated and fed into a parallel dual-branch hybrid deep learning model for signal occlusion prediction, outputting a multi-dimensional parameter signal occlusion prediction result. pred ;

[0171] The communication strategy instruction generation module is used to sequentially perform signal link feasibility checks and network resource availability checks on the acquired candidate links, and select qualified candidate links to construct a candidate set C of multi-dimensional performance profiles. set Based on the multidimensional performance profiles of each candidate link, and combined with a dual-mode decision triggering mechanism, a deep Q-network is used to generate dynamic weight factors when the communication decision process is activated. The comprehensive score S of each candidate link is then calculated based on these dynamic weight factors. score Based on communication service requirements R task The hard constraints, based on the comprehensive score S score Candidate links, sorted in descending order, are checked against hard constraints one by one until a candidate link that meets the hard constraints is found to be the primary communication link. Based on the overall network resource situation, a choice is made between switching policy instructions and cooperative communication policy instructions to generate the final communication policy instruction, Start. comm ;

[0172] The closed-loop feedback optimization module is used to optimize the communication strategy based on the Start instruction. comm Perform instruction type determination; if the communication strategy instruction Start... comm When using a collaborative strategy, dynamic fragmentation is performed on the data stream to generate data packets P that can be transmitted concurrently. sliced At the receiving end, the data packet P sliced Cyclic redundancy check, out-of-order reordering, and intelligent delivery are performed sequentially to restore the complete and ordered data stream D. reass ; in the communication strategy instruction Start comm After execution, a feedback data packet F containing the key performance indicators of the instructions is collected and encapsulated. back and the feedback data packet F back Iterative training of deep Q-networks is used to update decision weights and achieve adaptive closed-loop feedback.

[0173] The parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0174] The specific embodiments described above further illustrate the objects, technical solutions, and beneficial effects of the present application. It should be understood that the above description is merely a specific embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A 5G network optimization method based on intelligent dynamic multi-band adaptive switching system, characterized in that, The method comprises the following steps: acquiring real-time network signal quality data, environmental perception data, and a set of communication task requirements; based on the environmental perception data, extracting a set of dynamic obstacle feature vectors, integrating the environmental perception data and the set of dynamic obstacle feature vectors, inputting into a parallel double-branch hybrid deep learning model for signal obstruction prediction, and outputting a multi-dimensional parameter signal obstruction prediction result; performing signal link feasibility verification and network resource availability verification on the acquired candidate links in sequence, screening out qualified candidate links to construct a candidate set of multi-dimensional performance profiles; based on the multi-dimensional performance profiles of the candidate links, combining a dual-mode decision trigger mechanism, generating a dynamic weight factor using a deep Q network when the communication decision flow is activated, and calculating the comprehensive score of each candidate link according to the dynamic weight factor; according to the hard constraint conditions of the communication service requirements, performing hard constraint condition verification on the candidate links sorted in descending order of the comprehensive score one by one until a candidate link satisfying the hard constraint condition is found as the main communication link, and choosing between the switching strategy instruction and the cooperative strategy instruction according to the overall situation of the current network resources to generate the final communication strategy instruction; according to the communication strategy instruction, performing instruction type judgment, if the communication strategy instruction is a cooperative strategy, performing dynamic fragmentation on the data stream to generate data packets that can be transmitted concurrently, performing cyclic redundancy check, out-of-order reorganization, and intelligent delivery on the data packets in sequence at the receiving end to restore the data stream to be complete and in order, and after the communication strategy instruction is executed, collecting and packaging feedback data packets containing key performance indicators of the instruction, and using the feedback data packets for iterative training of the deep Q network to update the decision weight, thereby realizing adaptive closed-loop feedback. 2.The 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to claim 1, wherein, The output method of the signal obstruction prediction result comprises: performing multi-stage processing on dynamic obstacle data to generate a set of dynamic obstacle feature vectors containing shape and motion parameters; integrating static environmental data, vehicle self-dynamic data, and the set of dynamic obstacle feature vectors, inputting into a parallel double-branch hybrid deep learning model, extracting spatial feature vectors and temporal feature vectors respectively, fusing the spatial feature vectors and the temporal feature vectors to construct spatio-temporal fusion feature vectors, decoding the spatio-temporal fusion feature vectors using a prediction decoding head, and outputting a signal obstruction prediction result containing a predicted affected frequency band, a predicted signal attenuation value, a predicted start time of an obstruction event, and a predicted duration of the obstruction event. 3.The 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to claim 2, characterized in that, The generation of the set of dynamic obstacle feature vectors containing shape and motion parameters comprises shape feature extraction and trajectory feature extraction; based on dynamic obstacle data, performing shape feature extraction using a convex hull algorithm to determine the length, width, and height of the dynamic obstacle, further dividing the data points into sub-cubes, and performing weighted summation according to the data point density of each sub-cube to estimate the total volume of the obstacle; using a Kalman filter algorithm to perform trajectory feature extraction on the center of mass of the same obstacle identified in consecutive frames, and estimating the instantaneous speed and instantaneous motion direction angle of the obstacle by calculating the position change of the center of mass between consecutive time steps. 4.The 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to claim 1, wherein, The generation method of the communication strategy instruction comprises: Obtaining a candidate link list and a network load rate, performing signal link feasibility verification on the candidate link list based on real-time network signal quality data, performing network resource availability verification on the candidate link list based on the network load rate, after the feasibility and availability verification are completed, jointly constructing a candidate set from all remaining qualified candidate links, and generating a multi-dimensional performance profile including a signal quality index, a network load index and a distance index based on each candidate link in the candidate set; Based on real-time network signal quality data and signal blocking prediction results, combined with a dual-mode decision trigger mechanism, it is judged whether to activate a communication decision process, if the communication decision process is activated, a dynamic weight factor is generated according to the state vector of the current communication environment by using a deep Q network, and the dynamic weight factor and the multi-dimensional performance profile of each candidate link are used to calculate the comprehensive score of each candidate link; According to the hard constraint condition of the communication service demand, the candidate link ranked first in the candidate set sorted in descending order of comprehensive score is subjected to hard constraint condition verification, if the candidate link meets the hard constraint condition, the candidate link is determined as the main communication link, if the candidate link does not meet the hard constraint condition, the next candidate link in the candidate set sorted in descending order of comprehensive score is automatically iterated, and the hard constraint condition verification is continued, until a candidate link meeting the hard constraint condition is found; and according to the overall situation of the current network resources, a choice is made between the handover strategy instruction and the cooperative strategy instruction, and a final communication strategy instruction is generated; The hard constraint condition of the communication service demand is the performance requirement of the key data flow of the data flow task with the highest service priority in the communication service demand, including the minimum guaranteed width and the maximum transmission delay. 5.The 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to claim 4, characterized in that, The dual-mode decision trigger mechanism is activated when any of the following conditions is met, if the real-time network signal quality data shows that the key performance indicators of the currently used link are lower than the preset minimum availability threshold, indicating that the quality of service has deteriorated unacceptably, the passive response trigger mechanism is activated; if the signal blocking prediction result shows that the current communication link will face a high-probability, high-attenuation signal blocking event in a short time in the future, representing that the system predicts that the quality of service will deteriorate soon, the active prevention trigger mechanism is activated. 6.The 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to claim 4, wherein, The choice method between the handover strategy instruction and the cooperative strategy instruction comprises: When the system judges that the current network environment is relatively tense, that is, the quality of the remaining candidate links except the selected main communication link is generally poor, or the main purpose of this decision is to avoid the deterministic failure risk of the current link, the system will start the handover strategy instruction generation process; When the system judges that the current network environment resources are abundant, that is, in addition to the determined main communication link, there are still other multiple high-quality links in the candidate set, the system will start the cooperative strategy instruction generation process; The handover strategy instruction includes a strategy instruction type 'HANDOVER', the frequency band parameters of the finally selected main communication link, and the recording of the trigger reason of this link handover. The cooperative strategy instruction includes a strategy instruction type 'SYNERGY', a control task, a video task, and a remote sensing task. 7.The 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to claim 6, wherein, The generation process of the cooperative strategy instruction includes: allocating the control task with the highest priority to the determined main communication link and writing it as the first key-value pair in the parameter dictionary; creating two temporary resource pools, a to-do task pool and an available link pool, wherein the to-do task pool contains other service streams in the communication service demand except the allocated key service and is sorted from high to low according to the service priority, and the available link pool contains other high-quality candidate links in the candidate set except the occupied main communication link and is sorted from high to low according to the comprehensive score; processing the service with the highest priority in the to-do task pool and sequentially traversing the available link pool to find the best matching link for it. 8.The 5G network optimization method based on the intelligent dynamic multi-band adaptive switching system according to claim 7, wherein, The best matching link includes three scenarios: a single-link satisfaction scenario, if the performance of the link with the highest comprehensive score in the available link pool satisfies the core demand of the current to-do service, the link is allocated to the service, a corresponding key-value pair is added in the parameter dictionary, and the service and the link are removed from the respective resource pools to ensure that the resources are not repeatedly allocated; a link aggregation scenario, if none of the single links in the available link pool can satisfy the demand of the current high-bandwidth service, the system further judges whether the bandwidth of multiple links can be aggregated to satisfy the service demand; a link multiplexing scenario, after completing the link allocation of high-priority tasks, the system continues to process lower-priority tasks, and intelligently evaluates whether the allocated link with remaining capacity can be multiplexed in addition to considering the unused link resources in the available link pool. 9.The 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to claim 1, wherein, The implementation method of the adaptive closed-loop feedback includes: according to the instruction type of the communication strategy instruction, if the instruction type is cooperative strategy, the optimal fragmentation size is dynamically calculated based on the real-time available bandwidth of each link and the target transmission time window, and the data stream is executed dynamic fragmentation according to the optimal fragmentation size to generate data packets that can be transmitted in parallel through multiple links; if the instruction type is switching strategy, the dynamic fragmentation process is skipped; performing cyclic redundancy check on the received data packets, storing the data packets that pass the check in a ring buffer according to their sequence numbers for out-of-order reorganization, and using a bidirectional scanning reorganization algorithm to sequentially intelligently decode and deliver the data packets in the ring buffer to restore the received data packets to a complete and ordered data stream; after executing the communication strategy instruction, the execution effect of the communication strategy instruction is evaluated, the key performance indicators collected in the evaluation are packaged as feedback data packets, and the feedback data packets are used as training samples for iterative training of the deep Q network to update the decision weight, realizing adaptive closed-loop feedback.

10. The 5G network optimization system based on intelligent dynamic multi-band adaptive switching system, for implementing the 5G network optimization method based on intelligent dynamic multi-band adaptive switching system according to any one of claims 1-9, characterized in that, It includes: a multi-source heterogeneous data acquisition and perception module, an intelligent signal shielding prediction module, a communication strategy instruction generation module, and a closed-loop feedback optimization module; the multi-source heterogeneous data acquisition and perception module is used to acquire real-time network signal quality data, environmental perception data, and a communication task demand set; The intelligent signal shielding prediction module extracts a dynamic obstacle feature vector set based on the environment perception data, integrates the environment perception data and the dynamic obstacle feature vector set, inputs them into a parallel double-branch hybrid deep learning model for signal shielding prediction, and outputs a multi-dimensional parameter signal shielding prediction result. The communication strategy instruction generation module performs signal link feasibility verification and network resource availability verification on the obtained candidate links in sequence, filters out qualified candidate links, and constructs a candidate set of multi-dimensional performance profiles; Based on the multi-dimensional performance profiles of the candidate links, a deep Q network is used to generate a dynamic weight factor when the communication decision-making process is activated, and the comprehensive scores of the candidate links are calculated according to the dynamic weight factor; According to the hard constraint conditions of the communication service requirements, the candidate links sorted in descending order according to the comprehensive scores are tested one by one until a candidate link that meets the hard constraint conditions is found as the main communication link, and the overall situation of the current network resources is used to make a choice between the switching strategy instruction and the cooperative strategy instruction to generate the final communication strategy instruction; The closed-loop feedback optimization module is used to determine the type of the communication strategy instruction, and if the communication strategy instruction is a cooperative strategy, dynamic fragmentation is performed on the data stream to generate data packets that can be transmitted concurrently; the data packets are sequentially subjected to cyclic redundancy check, out-of-order reorganization and intelligent delivery at the receiving end to restore the data stream to be complete and in order; after the execution of the communication strategy instruction, feedback data packets containing key performance indicators of the instruction are collected and packaged, and the feedback data packets are used for iterative training of the deep Q network to update the decision weight, thereby realizing adaptive closed-loop feedback.

Images