Multi-terminal cooperative remote exhibition viewing load balancing method and system

Abstract

This invention discloses a multi-terminal collaborative remote exhibition load balancing method and system, belonging to the field of multi-terminal collaborative technology. The method includes: obtaining parameter configurations adapted to the scenario based on the requirements of the remote exhibition scenario and the access types of multiple terminals; performing preliminary verification based on the parameter configurations and the data collection requirements of multiple terminals to obtain four-dimensional raw data; performing temporal causal attention fusion processing based on the four-dimensional raw data and the causal logic of the remote exhibition scenario to obtain causal fusion features and user resource demand confidence; and obtaining a multi-terminal collaborative remote exhibition load balancing scheme based on the causal fusion features, user resource demand confidence, and global scheduling objectives. This invention iteratively obtains a multi-terminal collaborative remote exhibition load balancing scheme by completing operations such as decision scheduling, resource management, robust anti-interference, and closed-loop optimization, effectively achieving fair scheduling among users and load balancing of edge nodes.

Description

Technical Field

[0001] This invention relates to the field of multi-terminal collaboration technology, and more specifically to a method and system for load balancing remote exhibition viewing using multi-terminal collaboration. Background Technology

[0002] With the rapid development of digital cultural tourism and online exhibition industries, remote exhibition viewing has become the mainstream exhibition viewing mode. The demand for multi-terminal collaborative access such as VR, PC, and mobile phones is increasing. In order to ensure a smooth exhibition viewing experience for users of different types of terminals, load balancing technology has become the core support for remote exhibition viewing systems. However, remote exhibition viewing scenarios have characteristics such as diverse exhibit types and dynamic changes in user viewing behavior. Existing load balancing methods are difficult to adapt to the special characteristics of this scenario, so existing technologies have shortcomings. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a multi-terminal collaborative remote exhibition load balancing method and system. By completing operations such as decision scheduling, resource management, robust anti-interference, and closed-loop optimization, a decision scheduling scheme, a resource allocation scheme, and an anti-interference scheduling scheme are obtained. Finally, a multi-terminal collaborative remote exhibition load balancing scheme is obtained through iteration, which effectively realizes fair scheduling among users and load balancing of edge nodes.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This invention provides a multi-terminal collaborative remote exhibition load balancing method, comprising:

[0006] Based on the needs of remote exhibition viewing scenarios and the types of multi-terminal access, parameter configurations adapted to the scenarios are obtained;

[0007] Preliminary verification was performed based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data.

[0008] Based on the original four-dimensional data and the causal logic of the remote exhibition scenario, temporal causal attention fusion processing is performed to obtain causal fusion features and user resource demand confidence.

[0009] Based on the causal fusion characteristics, user resource demand confidence, and global scheduling objectives, a multi-terminal collaborative remote exhibition load balancing scheme is obtained.

[0010] As a further improvement of the present invention, the step of obtaining the parameter configuration adapted to the scenario based on the requirements of the remote exhibition scenario and the access types of multiple terminals includes:

[0011] Based on the scale of the remote exhibition scenario and the types of multi-terminal access, cloud center nodes, edge nodes, and multi-terminal devices are deployed, the core responsibilities of each node are clarified, and a hardware deployment plan is obtained.

[0012] Based on the characteristics of various exhibits and the needs of remote exhibition scheduling, the exhibit feature parameters, algorithm running parameters, and user priority parameters are preset to obtain the initial parameter configuration results.

[0013] Based on the initial parameter configuration results and exhibit resource classification standards, an exhibit resource demand profile library is constructed, exhibit resource demand characteristics are classified and archived, and a parameter support system is obtained.

[0014] Based on the hardware deployment plan and parameter support system, the parameter configurations adapted to the scenario are obtained.

[0015] As a further improvement of the present invention, the preliminary verification based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data includes:

[0016] Based on parameter configuration and multi-terminal data collection requirements, collect terminal performance-related data to obtain terminal characteristic data;

[0017] Based on parameter configuration and network monitoring requirements, network status data between terminals and nodes is collected through edge node network probes to obtain network characteristic data;

[0018] Preliminary verification is performed based on terminal feature data and network feature data to remove abnormal data and obtain four-dimensional raw data.

[0019] As a further improvement of the present invention, the step of performing temporal causal attention fusion processing based on four-dimensional raw data and the causal logic of remote exhibition scenarios to obtain causal fusion features and user resource demand confidence includes:

[0020] Based on the original four-dimensional data and the characteristics of the remote viewing scene, a directed causal topology graph is constructed to obtain the scene causal topology structure;

[0021] Based on the scene's causal topology and four-dimensional raw data, the attention weights between each causal node are calculated to obtain the true causal relationship information.

[0022] Based on real causal relationship information and four-dimensional raw data, deep feature fusion processing is performed to obtain causal fusion features;

[0023] Based on the causal fusion characteristics and fluctuation factors, the confidence level and uncertainty of user resource demand are calculated, thus obtaining the causal fusion characteristics and the confidence level of user resource demand.

[0024] As a further improvement of the present invention, the step of obtaining a multi-terminal collaborative remote exhibition load balancing scheme based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives includes:

[0025] Based on the causal fusion characteristics, the confidence level of user resource requirements, and the global scheduling objective, the decision scheduling operation is completed to obtain the decision scheduling scheme.

[0026] Based on the causal fusion characteristics, the confidence level of user resource needs, and the decision scheduling scheme, resource management operations are completed to obtain a resource allocation scheme.

[0027] Based on the causal fusion characteristics, the confidence level of user resource requirements, and the resource allocation scheme, robust anti-disturbance operation is completed to obtain an anti-disturbance scheduling scheme.

[0028] Based on the decision-making scheduling scheme, resource allocation scheme, and anti-interference scheduling scheme, a closed-loop optimization operation is completed to obtain a multi-terminal collaborative remote exhibition load balancing scheme.

[0029] As a further improvement of the present invention, the step of completing the decision scheduling operation based on causal fusion characteristics, user resource demand confidence, and global scheduling objective to obtain a decision scheduling scheme includes:

[0030] Based on the causal fusion characteristics, the confidence level of user resource demand, and the global scheduling objective, user game groups are divided and reward rules are set to obtain user game groups and reward system.

[0031] Based on the user game groups and the reward system, the game equilibrium within the user groups is solved to obtain the basis for fair scheduling among users.

[0032] Based on the fair scheduling criteria among users and the load conditions of edge nodes, the game equilibrium among edge nodes is solved to obtain the node load balancing scheme.

[0033] Based on the node load balancing scheme and the global scheduling objective, and combined with the hierarchical federated learning mechanism, the globally optimal bitrate adjustment instruction and node allocation scheme are obtained as the decision scheduling scheme.

[0034] As a further improvement of the present invention, based on the node load balancing scheme and the global scheduling objective, and combined with the hierarchical federated learning mechanism, a globally optimal bitrate adjustment instruction and node allocation scheme are obtained, which serve as the decision-making scheduling scheme, including:

[0035] Based on the node load balancing scheme and local real-time scheduling requirements, a preliminary scheduling scheme is obtained by making local decisions through the edge node DQN model.

[0036] Based on the preliminary scheduling plan and the global scheduling objectives, the global scheduling strategy is obtained by performing global policy distillation through the cloud center DQN model.

[0037] Based on the global scheduling strategy and the preliminary scheduling scheme, the globally optimal bitrate adjustment instructions and node allocation scheme are integrated to form the decision scheduling scheme.

[0038] As a further improvement of the present invention, based on causal fusion characteristics, user resource demand confidence, and decision scheduling scheme, resource management operations are completed to obtain a resource allocation scheme, including:

[0039] Based on the causal fusion characteristics, the confidence level of user resource demand, and the scheduling scheme, combined with the prediction results of user flow trajectory, user resource quotas are allocated to obtain the user resource quota scheme.

[0040] Based on the user resource quota scheme and the requirements for preventing resource preemption, a user-exclusive resource ownership certificate is generated to obtain the resource ownership system.

[0041] Based on the resource ownership system and user priority, a differentiated and flexible lock-in strategy is implemented to obtain a flexible lock-in solution.

[0042] Based on the flexible lock-in scheme and resource utilization needs, unused resources are released and reused without disturbance, resulting in a resource allocation scheme.

[0043] As a further improvement of the present invention, based on causal fusion characteristics, user resource demand confidence, and resource allocation scheme, robust anti-disturbance operation is completed to obtain an anti-disturbance scheduling scheme, including:

[0044] Based on the causal fusion characteristics, the confidence level of user resource requirements, and the requirements for disturbance detection, uncertainty and its rate of change are monitored in real time to obtain uncertainty monitoring data.

[0045] Based on the uncertainty monitoring data and the disturbance rejection triggering rules, determine whether to trigger the robust disturbance rejection mode and obtain the triggering judgment result;

[0046] Based on the trigger judgment result, the scheduling mode is switched and adapted. Based on the scheduling mode adjustment result, an anti-disturbance scheduling scheme adapted to sudden scenarios is obtained.

[0047] As a further improvement of the present invention, based on the decision scheduling scheme, resource allocation scheme, and anti-interference scheduling scheme, a closed-loop optimization operation is completed to obtain a multi-terminal collaborative remote viewing load balancing scheme, including:

[0048] Based on the execution status of the decision-making scheduling plan, resource allocation plan, and disturbance suppression scheduling plan, the plan operation monitoring data is obtained;

[0049] Based on the monitoring data of the scheme operation and optimization needs, identify the shortcomings of the scheme operation and determine the optimization direction to obtain optimization guidance;

[0050] Based on the optimization guidelines, the operating parameters of each module are dynamically corrected and the relevant models are updated to obtain the optimized parameters and model system.

[0051] Based on the optimized parameters and model system, the load balancing scheme is iteratively optimized to obtain a multi-terminal collaborative remote exhibition load balancing scheme.

[0052] This invention provides a multi-terminal collaborative remote exhibition load balancing system, comprising:

[0053] The configuration module is used to obtain parameter configurations adapted to the scenario based on the requirements of remote exhibition viewing and the access types of multiple terminals.

[0054] The verification module is used to perform preliminary verification based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data.

[0055] The causal fusion module is used to perform temporal causal attention fusion processing based on the original four-dimensional data and the causal logic of the remote exhibition scenario to obtain causal fusion features and user resource demand confidence.

[0056] The load balancing module is used to obtain a multi-terminal collaborative remote exhibition load balancing solution based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives.

[0057] This invention constructs parameter support adapted to remote exhibition scenarios, then collects and verifies multi-dimensional data to obtain reliable four-dimensional raw data. Next, based on the causal logic of the scenario, it performs temporal causal attention fusion to obtain causal fusion features and user resource demand confidence. Then, by completing operations such as decision scheduling, resource management, robust anti-disturbance, and closed-loop optimization, it obtains decision scheduling schemes, resource allocation schemes, and anti-disturbance scheduling schemes. Finally, it iterates to obtain a multi-terminal collaborative remote exhibition load balancing scheme, which effectively achieves fair scheduling among users and load balancing of edge nodes, avoids resource contention and waste, improves resource utilization, ensures a smooth exhibition experience for users of different terminals, and ensures stable operation of the system in dynamic exhibition scenarios. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the steps of a multi-terminal collaborative remote exhibition load balancing method according to the present invention.

[0059] Figure 2 A schematic diagram illustrating the steps to obtain a resource allocation plan;

[0060] Figure 3 A schematic diagram illustrating the steps to obtain a load balancing solution. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0062] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0063] like Figure 1 As shown in the figure, this application provides a multi-terminal collaborative remote exhibition load balancing method, including:

[0064] Based on the needs of remote exhibition viewing scenarios and the types of multi-terminal access, parameter configurations adapted to the scenarios are obtained;

[0065] Preliminary verification was performed based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data.

[0066] Based on the original four-dimensional data and the causal logic of the remote exhibition scenario, temporal causal attention fusion processing is performed to obtain causal fusion features and user resource demand confidence.

[0067] Based on the causal fusion characteristics, user resource demand confidence, and global scheduling objectives, a multi-terminal collaborative remote exhibition load balancing scheme is obtained.

[0068] Specifically, the parameter configuration for the adapted scenario is a set of customized hardware deployment parameters, algorithm operation parameters, and user priority parameters, taking into account the scale of the remote exhibition, the type of exhibits, and the number of multiple terminals accessing the site. This is also the foundational configuration for the entire load balancing process. Basic information about the remote exhibition scenario needs to be collected beforehand, including the estimated total number of participants, the proportion of different terminal types, and the total number and type distribution of exhibits. For example, the upper limit for access terminals is set at 5,000 for small and medium-sized exhibitions and 20,000 for large exhibitions. Based on the scenario scale and terminal types, the basic parameters for the cloud center nodes, edge nodes, and terminal SDKs are preset. Edge nodes are divided according to city administrative regions. This ultimately yields the parameter configuration for the adapted scenario, providing a unified standard for subsequent data collection and feature processing.

[0069] Next, the system performs the collection and verification of four-dimensional raw data. This four-dimensional raw data is a collection of four core data categories: terminal performance data, network status data, user viewing behavior data, and exhibit feature data. The system initiates the data collection process based on the parameter configuration for the adapted scenario. It collects terminal performance data such as CPU utilization, GPU load, and decoding frame rate through the terminal SDK; network status data such as bandwidth, latency, and packet loss rate through edge node network probes; and simultaneously collects viewing behavior data such as user dwell time and interaction frequency. It also retrieves exhibit resource requirement feature data. After collection, preliminary verification is performed to remove invalid data with abnormal values ​​deviating from the normal range, ultimately obtaining complete and accurate four-dimensional raw data.

[0070] Next, temporal causal attention fusion processing is performed. This processing is based on time series data and combines the causal logic in remote exhibition scenarios—where user behavior determines exhibit demand, exhibit demand determines terminal load, and terminal load affects network status—to filter effective features and eliminate redundant features through an attention mechanism. It requires four-dimensional raw data as input. First, the data is sorted by time series. Then, feature weights are assigned through the attention mechanism, with higher weights given to core features that affect user experience and lower weights for secondary features. Redundant data without causal relationships is eliminated. After fusion, causal fusion features are generated. These causal fusion features are multi-dimensional comprehensive features that, after temporal causal attention fusion processing, eliminate pseudo-related information and retain true causal relationships, enabling precise quantification of user resource needs. Then, user resource requirements are quantified based on causal fusion characteristics, and the stability of requirements is assessed in combination with data fluctuation amplitude to obtain the user resource requirement confidence score. The user resource requirement confidence score is a quantitative assessment result of the degree and stability of the user's demand for resources such as bandwidth, computing power, and bit rate for viewing exhibitions based on causal fusion characteristics. It can reflect the reliability of user resource requirements. The confidence score ranges from 0 to 1. The closer the value is to 1, the more stable the resource requirements are and the more reliable the assessment result is.

[0071] Finally, based on causal fusion characteristics, user resource demand confidence levels, and global scheduling objectives, a final solution is generated. The global scheduling objective is a holistic scheduling approach centered on optimizing the user viewing experience, balancing edge node loads, and maximizing resource utilization. The causal fusion characteristics and user resource demand confidence levels are input into the core scheduling module. Guided by the global scheduling objective, decision-making scheduling, resource management, robust anti-interference, and closed-loop optimization operations are executed sequentially. The results of all sub-operations are integrated to generate a complete load balancing solution covering bitrate adjustment, node allocation, resource allocation, anti-interference guarantees, and iterative optimization. This solution can be directly deployed to the cloud center, edge nodes, and terminals for execution.

[0072] This embodiment constructs parameter support adapted to remote exhibition scenarios, then collects and verifies multi-dimensional data to obtain reliable four-dimensional raw data. Next, based on the causal logic of the scenario, it performs temporal causal attention fusion to obtain causal fusion features and user resource demand confidence. Then, by completing operations such as decision scheduling, resource management, robust anti-interference, and closed-loop optimization, it obtains decision scheduling scheme, resource allocation scheme, and anti-interference scheduling scheme. Finally, it iterates to obtain a multi-terminal collaborative remote exhibition load balancing scheme, which effectively realizes fair scheduling among users and load balancing of edge nodes, avoids resource contention and waste, improves resource utilization, ensures a smooth exhibition experience for users of different terminals, and ensures stable operation of the system in dynamic exhibition scenarios.

[0073] Furthermore, this embodiment provides a step for obtaining parameter configurations adapted to the scenario based on the requirements of remote exhibition scenarios and multiple terminal access types, including:

[0074] Based on the scale of the remote exhibition scenario and the types of multi-terminal access, cloud center nodes, edge nodes, and multi-terminal devices are deployed, the core responsibilities of each node are clarified, and a hardware deployment plan is obtained.

[0075] Based on the characteristics of various exhibits and the needs of remote exhibition scheduling, the exhibit feature parameters, algorithm running parameters, and user priority parameters are preset to obtain the initial parameter configuration results.

[0076] Based on the initial parameter configuration results and exhibit resource classification standards, an exhibit resource demand profile library is constructed, exhibit resource demand characteristics are classified and archived, and a parameter support system is obtained.

[0077] Based on the hardware deployment plan and parameter support system, the parameter configurations adapted to the scenario are obtained.

[0078] Specifically, the hardware deployment architecture is first planned based on the scale of the remote exhibition scenario and the access types of multiple terminals. For example, the cloud center node can deploy 1 main server and 2 backup servers to be responsible for global policy scheduling and data storage. Edge nodes are deployed according to the user gathering areas, with 5-10 deployed for small and medium-sized exhibitions and 15-30 deployed for large exhibitions. The cloud center is responsible for global policy distillation, the edge nodes are responsible for local real-time decision-making, and the terminals are responsible for data collection and instruction execution, forming a standardized hardware deployment scheme.

[0079] The system then presets initial parameters to obtain the initial parameter configuration results. These initial parameter configuration results are a set of preset values ​​for exhibit feature parameters, algorithm execution parameters, and user priority parameters; these are basic parameters that do not require real-time adjustment. The system presets product feature parameters according to exhibit type, allowing for different resource requirement characteristics to be set for VR immersive exhibits, 4K video exhibits, and static display cases. The algorithm execution parameters are preset with 50 federated learning iterations and a DQN model learning rate of 0.01. User priority parameters are set according to viewing behavior, with users who linger for in-depth viewing having the highest priority and those who quickly browse having the lowest. After all parameters are preset, a unified initial parameter configuration result is obtained, ensuring the system can operate normally in the initial state.

[0080] Next, an exhibit resource requirement profile library is constructed to build a parameter support system. This library is a resource requirement feature database categorized and archived by exhibit type, storing the basic resource requirement standards for different exhibits. The parameter support system is a complete parameter support framework formed by integrating the hardware deployment plan, initial parameters, and the exhibit profile library, covering all dimensions of hardware and software configuration. Based on the initial parameter configuration results, the system categorizes exhibits into four main categories according to the exhibit resource classification standards: immersive interactive exhibits, dynamic video exhibits, static display exhibits, and real-time interactive exhibits. Preset exhibit feature parameters are extracted as resource requirement features corresponding to the exhibit type, and each type of exhibit is bound to corresponding resource requirement features. For example, immersive exhibits require 30Mbps bandwidth and a frame rate of 60fps or higher, while static exhibits require 1Mbps bandwidth and a frame rate of 1fps. All exhibit features are categorized and archived to construct the exhibit resource requirement profile library. This library supports real-time retrieval and updates, allowing for rapid adaptation based on exhibit adjustments at the exhibition. The hardware deployment plan, initial parameter configuration results, and exhibit resource requirement profile library are integrated to form a comprehensive parameter support system covering hardware, software, and exhibits.

[0081] Finally, the system integrates the hardware deployment plan and parameter support system, performs overall verification of the parameter support system, checks the matching between hardware deployment and parameter settings, the compatibility between exhibit profiles and algorithm parameters, corrects mismatched parameter items, and finally obtains parameter configurations adapted to the scenario. The parameter configurations can be directly imported into the system for execution without secondary adjustments.

[0082] This embodiment achieves scenario-based and refined parameter configuration through four steps: hardware deployment, parameter preset, profile library construction, and system integration. It solves the problems of generalized parameters and lack of product adaptability in traditional load balancing, and provides stable and reliable hardware and software support for subsequent data collection, feature fusion, and scheduling decisions, thereby improving the feasibility and adaptability of the entire solution.

[0083] Furthermore, this embodiment provides a step for performing preliminary verification based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data, including:

[0084] Based on parameter configuration and multi-terminal data collection requirements, collect terminal performance-related data to obtain terminal characteristic data;

[0085] Based on parameter configuration and network monitoring requirements, network status data between terminals and nodes is collected through edge node network probes to obtain network characteristic data;

[0086] Preliminary verification is performed based on terminal feature data and network feature data to remove abnormal data and obtain four-dimensional raw data.

[0087] Specifically, the process begins by collecting terminal performance data to obtain terminal characteristic data. This data reflects the terminal's hardware performance and operating status, and is used to determine the terminal's decoding and load-bearing capabilities. Based on the parameter configuration for the adapted scenario, the terminal SDK's data collection function is activated. For example, the collection frequency is set to 100ms / time. The collected data includes terminal CPU utilization, GPU load, real-time decoding frame rate, screen resolution, and remaining RAM. For example, the normal CPU utilization range is set to 0%-80%, and values ​​exceeding 80% are considered excessively high terminal load. After collecting the data, the SDK encrypts and uploads it to the corresponding edge node in real time. The edge node temporarily stores the data, and all terminal data is stored categorized by terminal ID to ensure a one-to-one correspondence between data and users. The principle behind setting the collection frequency and value range is to accurately capture dynamic changes in terminal performance while avoiding excessive terminal bandwidth consumption due to excessively high collection frequencies.

[0088] Next, the system collects network status data to obtain network characteristic data. This network characteristic data reflects the quality of network transmission between the terminal and the edge node, and is a core indicator determining the smoothness of the exhibition. The system uses network probes deployed at the edge nodes to collect network characteristic data between the terminal and the edge nodes in real time. For example, the collection frequency is 50ms / time, higher than the terminal data collection frequency. The principle is that network status fluctuates faster, requiring higher frequency monitoring. The collected content includes real-time available bandwidth, network latency, packet loss rate, and network jitter. For example, normal network latency is set to ≤100ms and packet loss rate to ≤3%. Values ​​exceeding these ranges are considered a weak network state. The network probes directly connect to the edge node gateway, and data collection is unnoticed by the terminal, not affecting the user's exhibition operation. After collection, the data is synchronously uploaded to the edge node data processing module.

[0089] Next, user viewing behavior data is collected. This data reflects users' interactive behaviors and dwell times during remote viewing, used to characterize user viewing patterns and assess the urgency of user resource needs. Based on the parameter configuration of the adapted scenario, the user behavior data collection function of the terminal SDK is activated. For example, the collection frequency is set to 100ms / time, synchronized with the terminal feature data collection frequency. The principle is that changes in user behavior and changes in terminal status are strongly correlated, and synchronous collection can ensure the consistency of subsequent timestamp alignment. The collected content includes the duration of stay on exhibits, interaction frequency, exhibit switching frequency, and active resource request behavior. For example, a stay time > 10 seconds is defined as a deep viewing user, and ≥ 3 interactions per minute is defined as a high-engagement user. After the SDK collects the data, it is encrypted and uploaded to the corresponding edge node in real time. The edge node temporarily stores the data. All user behavior data is stored in categories according to user ID, ensuring that the data corresponds one-to-one with the terminal feature data. The data collection is imperceptible to the user and does not affect the user's viewing operation.

[0090] Next, the system retrieves exhibit feature data from the exhibit resource requirement profile database. This data reflects the inherent resource requirements of the exhibits the user is currently viewing, clarifying the basic requirements of different exhibits for bandwidth, computing power, and frame rate. Based on the parameter configuration for the adapted scenario, the system retrieves corresponding data from the preset exhibit resource requirement profile database according to the exhibit ID the user is currently viewing. For example, the retrieval frequency is set to be updated synchronously with the user's viewing behavior data. Retrieval is triggered immediately when the user switches exhibits. The collected content includes exhibit type, bandwidth requirement benchmark, frame rate requirement benchmark, and terminal decoding computing power threshold. For example, immersive interactive exhibits require a bandwidth of 30Mbps and a frame rate of 60fps, while static display exhibits require a bandwidth of 1Mbps and a frame rate of 1fps. After retrieval, the data is synchronously associated with the corresponding user ID and stored in conjunction with the user's viewing behavior data and terminal feature data to ensure that the data is consistent with the user's current viewing status.

[0091] The system then integrates multiple types of data to form unverified four-dimensional raw data. It aligns the terminal feature data, network feature data, user viewing behavior data, and exhibit feature data with timestamps to ensure that the four types of data match one by one at the same time, guaranteeing the consistency of data time sequence and avoiding causal logic errors. After alignment, the four types of data are integrated into a unified data structure to form unverified raw four-dimensional data in a standardized JSON format for easy subsequent processing and transmission.

[0092] Finally, preliminary verification is conducted to obtain reliable four-dimensional raw data. Preliminary verification is a basic processing operation that screens the collected raw data for completeness and accuracy, and removes invalid data. The system initiates the preliminary data verification process. For example, three verification rules can be set: data with more than two missing items is considered invalid, values ​​outside the normal range are considered abnormal, and timestamp errors exceeding 10ms are considered misaligned. All raw four-dimensional data are verified one by one, and data that meets the invalid, abnormal, or misaligned conditions is removed. The remaining data after verification is the reliable four-dimensional raw data, avoiding decision-making errors due to missing data.

[0093] This embodiment solves the problems of messy data collection, invalid data, and disordered time sequence in traditional data collection by collecting data in multiple dimensions and standardizing verification. The resulting four-dimensional raw data is complete, accurate, and time-series consistent, providing high-quality input data for subsequent time-series causal attention fusion processing and improving the accuracy of load balancing decisions from the source.

[0094] Furthermore, this embodiment provides a step for performing temporal causal attention fusion processing based on four-dimensional raw data and the causal logic of a remote viewing scenario to obtain causal fusion features and user resource demand confidence, including:

[0095] Based on the original four-dimensional data and the characteristics of the remote viewing scene, a directed causal topology graph is constructed to obtain the scene causal topology structure;

[0096] Based on the scene's causal topology and four-dimensional raw data, the attention weights between each causal node are calculated to obtain the true causal relationship information.

[0097] Based on real causal relationship information and four-dimensional raw data, deep feature fusion processing is performed to obtain causal fusion features;

[0098] Based on the causal fusion characteristics and fluctuation factors, the confidence level and uncertainty of user resource demand are calculated, thus obtaining the causal fusion characteristics and the confidence level of user resource demand.

[0099] Specifically, a directed causal topology graph is a topology graph constructed using user behavior, exhibit demand, terminal load, and network status as nodes and unidirectional causal relationships as edges. It clearly defines the causal relationships between features. The scene causal topology structure is the overall architecture of the directed causal topology graph, used to reflect the core causal relationships of various features in a remote exhibition scenario. Based on four-dimensional raw data, four types of features are extracted as topology nodes: user exhibition behavior, exhibit resource demand, terminal decoding load, and network transmission status. According to the inherent causal logic of the remote exhibition scenario, a directed causal topology graph with fixed directions is constructed: user exhibition behavior points to exhibit resource demand, exhibit resource demand points to terminal decoding load, and terminal decoding load points to network transmission status. Each node in the topology graph corresponds to a type of feature data, and the edges represent the causal influence relationships between nodes, ultimately forming a stable scene causal topology structure.

[0100] Subsequently, attention weights are calculated to obtain true causal relationship information. This true causal relationship information is the data that retains valid causal relationships after the attention weight calculation, removing spurious correlations. Based on the scene's causal topology, attention weights are calculated on the four-dimensional raw data using a temporal causal attention mechanism. First, according to the pre-defined causal pointing relationships in the topology, only feature influence calculations are allowed along the causal path; reverse or unrelated feature pairs are excluded from the calculation, ensuring that attention calculations are based solely on true causal relationships. For each pair of upstream and downstream features with a causal pointing relationship, an influence score is calculated. For example, based on the acquisition frequency of the four-dimensional raw data, the data from the 10 consecutive time steps prior to the current moment is used as the calculation window to ensure that the data covers the dynamic changes in user viewing. Then, the changing trends of upstream and downstream features within the window are statistically analyzed. For example, for upstream user dwell time, the dwell time increment is calculated every 100ms; for downstream terminal decoding load, the CPU utilization rate change is calculated every 100ms. Next, along the causal direction of the topology, the number of times the downstream feature changes within a lag of 1-2 time steps after the upstream feature changes is statistically analyzed to see if the changes are consistent with expectations. For example, when the user dwells... The length of time increases continuously. The number of times the terminal decoding load increases synchronously within the next 1-2 time steps is counted. If more than 8 times within 10 time steps meet the expected synchronization, it is judged as high synchronization. The score is calculated based on the synchronization matching rate. If the synchronization matching rate is ≥80%, the base score is 0.8. If there is a reverse synchronization situation where the downstream feature changes before the upstream feature, 0.1 points are deducted for each occurrence to ensure that only positive causal influence is included in the score. Finally, the influence score between 0 and 1 is obtained. Then, the influence scores of all feature pairs are globally normalized, that is, the influence score of each feature is divided by the sum of the scores of all features to obtain the initial attention weight. The model is iteratively trained based on historical scheduling data of remote exhibition scenarios. By continuously correcting the judgment rules of synchronization matching and the correction coefficient of change amplitude, the weight allocation is optimized and finally converged to obtain a stable attention weight. The feature association data that is finally retained is the real causal association information. The principle of weight setting is to highlight the influence of core features on decision-making, weaken the interference of secondary features, and improve the effectiveness of features.

[0101] Next, deep feature fusion processing is performed to obtain causal fusion features. Deep feature fusion processing is a process that weights and integrates multi-dimensional features to generate a single comprehensive feature vector, which can simplify the decision input dimensions. Based on real causal relationship information, the system deeply fuses multi-dimensional features, summing four types of features according to attention weights to integrate them into a single-dimensional comprehensive feature vector. For example, firstly, the four types of features are bound to weights calculated through the attention mechanism, with a total weight sum of 1. Then, the real-time value of the four-dimensional feature at each time step is multiplied by the attention weight of the corresponding feature to obtain the weighted feature value of that feature. Next, the weighted values ​​of all four-dimensional features within the same time step are summed to obtain the single comprehensive feature value for that time step. The above summation operation is repeated for each time step to generate the single comprehensive feature value corresponding to each time step. These values ​​are then arranged in chronological order to form a comprehensive feature vector with temporal information. The temporal information is preserved during the fusion process to ensure that the features are dynamically updated over time, with the update frequency consistent with the data acquisition frequency. The dimension of the fused feature vector is compressed to 1 / 4 of the original data, reducing the computational pressure on the subsequent decision model. The principle of dimension compression is to improve the model's operating efficiency and meet real-time scheduling requirements while retaining core information.

[0102] Finally, the confidence level of user resource demand is calculated by combining fluctuation factors. Fluctuation factors are variables that affect the stability of user resource demand, including network fluctuations, terminal load fluctuations, and behavioral change fluctuations. It is necessary to quantify the stability of user resource demand based on causal fusion characteristics and the three types of fluctuation factors. For example, first, the time window for calculating fluctuation amplitude is determined. This can be based on the acquisition frequency of the four-dimensional raw data, such as 100ms / time, taking the data from the 10 consecutive time steps before the current moment as the calculation window. Then, the relative fluctuation amplitude of the three types of fluctuation factors is calculated separately. For network fluctuation amplitude, based on the time-series data of network latency and packet loss rate within the window, the average value of each indicator within the window is calculated first, then the difference between the maximum and minimum values ​​of the indicator is calculated, and the difference is divided by the average value to obtain the relative volatility. For example, the network latency within a VR user's window... The average latency is 80ms, the maximum is 85ms, and the minimum is 78ms, with a difference of 7ms and a relative volatility of approximately 8.75%. The average packet loss rate is 2%, the maximum is 2.3%, and the minimum is 1.8%, with a difference of 0.5% and a relative volatility of approximately 25%. Averaging the relative volatility of the two metrics yields a network volatility of approximately 16.88%. For terminal load volatility, based on time-series data of terminal CPU utilization and GPU load within a window, the same calculation logic as for network volatility is used: first, the average value is calculated; then, the difference between the maximum and minimum values ​​is calculated; and finally, the difference is divided by the average value to obtain the relative volatility. For example... If the average CPU usage within the user window is 60%, the maximum is 65%, and the minimum is 58%, with a difference of 7%, the relative volatility is approximately 11.67%. If the average GPU load is 40%, the maximum is 43%, and the minimum is 38%, with a difference of 5%, the relative volatility is approximately 12.5%. Averaging these two values ​​yields a terminal load volatility of approximately 12.08%. For behavioral change volatility, the relative volatility is calculated based on time-series data of user dwell time and interaction frequency within the window. For example, if the average dwell time within the user window is 1.5 seconds, the maximum is 1.8 seconds, and the minimum is 1.2 seconds, the difference is 0.6 seconds. The relative volatility is approximately 40%, the average interaction frequency is 0.2 times / 100ms, the maximum is 0.3 times, the minimum is 0.1 times, the difference is 0.2 times, and the relative volatility is approximately 100%. Taking the average of the two, the behavioral change volatility is approximately 70%. Then, the three types of volatility are weighted to obtain the overall volatility. The smaller the overall volatility, the higher the confidence level. For example, the confidence level is set to 0.9-1.0 when the volatility is ≤5%, 0.7-0.9 when the volatility is 5%-10%, and <0.7 when the volatility is >10%. Finally, the causal fusion feature and the user resource demand confidence level are output synchronously.

[0103] This embodiment achieves accurate transformation from raw data to effective features by constructing a causal topology and calculating confidence levels. It solves the problems of multiple pseudo-correlations and lack of prominence of core features in traditional feature fusion. The resulting causal fusion features and confidence levels truly reflect user resource needs, greatly improving the rationality and accuracy of subsequent scheduling decisions.

[0104] Furthermore, this embodiment provides a step for obtaining a multi-terminal collaborative remote exhibition load balancing scheme based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives, including:

[0105] Based on the causal fusion characteristics, the confidence level of user resource requirements, and the global scheduling objective, the decision scheduling operation is completed to obtain the decision scheduling scheme.

[0106] Based on the causal fusion characteristics, the confidence level of user resource needs, and the decision scheduling scheme, resource management operations are completed to obtain a resource allocation scheme.

[0107] Based on the causal fusion characteristics, the confidence level of user resource requirements, and the resource allocation scheme, robust anti-disturbance operation is completed to obtain an anti-disturbance scheduling scheme.

[0108] Based on the decision-making scheduling scheme, resource allocation scheme, and anti-interference scheduling scheme, a closed-loop optimization operation is completed to obtain a multi-terminal collaborative remote exhibition load balancing scheme.

[0109] Specifically, the first step is to complete the decision-making and scheduling operation to obtain a decision-making and scheduling scheme. This scheme is based on user demand and node load, generating core scheduling instructions for bitrate adjustment and node allocation. The decision-making and scheduling process requires inputs of causal fusion features, user resource demand confidence, and global scheduling objectives. First, the user and edge node game groups are divided, then the game equilibrium is solved. Combined with hierarchical federated learning, bitrate adjustment and node allocation instructions are generated. These instructions cover all online terminals, ensuring that each user corresponds to the optimal bitrate and the optimal node, thus forming a standardized decision-making and scheduling scheme.

[0110] Next, resource management operations are completed to obtain a resource allocation plan. This plan, based on the decision-making and scheduling plan, generates resource quotas, rights confirmation, rights locking, and release instructions. It requires causal fusion characteristics, user resource demand confidence levels, and the decision-making and scheduling plan as inputs to initiate the resource management process. First, resource quotas are allocated to users, generating resource rights confirmation certificates. Then, flexible rights locking is implemented according to user priority, and unused resources are released without disturbance, forming a resource allocation plan. This plan ensures that resources are not preempted or wasted, and that core user resources are prioritized.

[0111] Afterwards, robust disturbance rejection operations are performed to obtain a disturbance rejection scheduling scheme. This scheme is a robust disturbance rejection scheduling instruction that can cope with sudden traffic surges and network fluctuations, ensuring stable system operation. It requires causal fusion characteristics, user resource demand confidence, and resource allocation schemes as inputs to initiate the robust disturbance rejection process. It monitors the uncertainty and rate of change of user resource demand in real time, determines whether to trigger the disturbance rejection mode, switches the scheduling mode to ensure system stability, and forms a disturbance rejection scheduling scheme adapted to sudden scenarios. This scheme can cope with abnormal situations such as sudden traffic surges and network jitter.

[0112] Finally, a closed-loop optimization operation is performed to obtain the final load balancing solution. This closed-loop optimization is a continuous process of adjusting parameters and updating the model based on the solution's execution performance. It requires summarizing the execution data of the decision-making scheduling scheme, resource allocation scheme, and disturbance rejection scheduling scheme; collecting experience metrics, resource metrics, and operational metrics; identifying shortcomings in the scheme; determining optimization directions; adjusting algorithm parameters; updating the scheduling model; and iteratively optimizing to form the final multi-terminal collaborative remote viewing load balancing solution. The solution supports continuous iteration, for example, optimizing every 30 minutes to adapt to dynamic changes in the scenario.

[0113] This embodiment constructs a full-process, closed-loop load balancing execution logic through decision scheduling, resource management, robust anti-interference, and closed-loop optimization. It solves the problems of traditional scheduling being singular, resource wasteful, weak anti-interference capability, and lacking optimization mechanism, and achieves the integration of scheduling, resources, stability, and optimization. This is the core execution link of the entire solution.

[0114] Furthermore, this embodiment provides a step for completing a decision-making scheduling operation and obtaining a decision-making scheduling scheme based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives, including:

[0115] Based on the causal fusion characteristics, the confidence level of user resource demand, and the global scheduling objective, user game groups are divided and reward rules are set to obtain user game groups and reward system.

[0116] Based on the user game groups and the reward system, the game equilibrium within the user groups is solved to obtain the basis for fair scheduling among users.

[0117] Based on the fair scheduling criteria among users and the load conditions of edge nodes, the game equilibrium among edge nodes is solved to obtain the node load balancing scheme.

[0118] Based on the node load balancing scheme and the global scheduling objective, and combined with the hierarchical federated learning mechanism, the globally optimal bitrate adjustment instruction and node allocation scheme are obtained as the decision scheduling scheme.

[0119] Specifically, the process begins by dividing users into game groups to construct a user game group and reward system. This system consists of user groups categorized by exhibits and regions, along with corresponding reward rules, used to solve for a fair scheduling equilibrium. For example, user game groups can be formed based on the exhibit a user is currently viewing and their current region. Users from the same exhibit or region form one group, with each group consisting of 50-200 people to ensure efficient equilibrium solving. Then, reward rules are set for each group, linking user rewards to experience suitability and node rewards to resource utilization. Higher-priority users receive higher reward weights, ultimately forming a complete user game group and reward system.

[0120] Subsequently, the game equilibrium within the user group is solved to obtain the basis for fair scheduling among users. The result of the game equilibrium within the user group is the core basis for ensuring the fairness of resource allocation among users. Specifically, an evolutionary game theory algorithm can be used to solve the game equilibrium within the user group. By replicating dynamic rules, user resource request strategies are adjusted. For example, first, the game participants and initial settings are defined. The game participants are user groups in the same exhibit and the same area. For example, in the VR immersive exhibit scenario, the size of the user group in the same city area is controlled between 50 and 200 people. Each user's initial resource request strategy is divided into three levels: high request strategy, requesting the highest resource requirement of the exhibit, such as 30Mbps bandwidth for VR exhibits; medium request strategy, requesting the basic resource requirement of the exhibit, such as 15Mbps bandwidth; and low request strategy, only meeting the basic loading requirements, such as 5Mbps bandwidth. The initial strategy can be preset according to user priority. High-priority users take the high request strategy first, and low-priority users take the low request strategy first. Then, user benefit is defined as user experience benefit - resource occupation cost. Among them, the user experience benefit is directly proportional to the actual amount of resources obtained. The calculation formula is the actual bandwidth obtained / the bandwidth required by the exhibit. For example, if the VR exhibit requires 30Mbps and the user actually obtains 24Mbps, the experience benefit is 0.8. The resource occupation cost is related to the proportion of users with the same strategy within the group. If the proportion of users adopting the high request strategy exceeds 60%, the bandwidth competition intensifies, and the resource occupation cost coefficient for each user is set to 0.3. If the proportion is less than 40%, the cost coefficient is set to 0.1, so as to guide users to avoid excessive competition for high resource requests. Then, user strategies are adjusted by replicating dynamic rules. The core logic of replicating dynamic rules is that users dynamically adjust their strategies based on the difference between their own strategy's payoff and the group's average payoff. Specifically, in each iteration, the average payoff for each strategy within the group is calculated. If a user's current strategy payoff is lower than the group's average payoff, and the payoff difference exceeds 0.2, then the user has a 50% probability of switching to a higher-payoff strategy. If the payoff difference is less than 0.2, the switching probability drops to 10% to avoid frequent strategy fluctuations. The iteration frequency is set to 1 second / time to match the dynamic change rhythm of users' viewing behavior. Finally, the game equilibrium state is determined. When the change rate of the proportion of users using the three strategies is less than 5% in three consecutive iterations, and the standard deviation of the payoffs of all users within the group is less than 0.1, it is determined that the payoffs within the group have reached equilibrium, and the iteration stops. If the VR exhibit user group is in a balanced state, the proportion of users with high request strategy is stable at 30%, users with medium request strategy is 50%, and users with low request strategy is 20%. The revenue of all users is stable between 0.7 and 0.8, and no user's revenue deviates from the group average. At this time, the balance result ensures fair resource allocation among users in the same group, and no user over-occupies resources. It provides a basis for fair scheduling among users, which can be directly used for resource allocation to ensure the basic viewing experience of all users.

[0121] Next, the game equilibrium among edge nodes is solved to obtain the node load balancing scheme, which is the result of the game equilibrium among edge nodes. The game equilibrium among edge nodes can be solved using the fair scheduling criteria among users and the real-time load of edge nodes as inputs. The node revenue is the local experience revenue minus the scheduling energy cost, guiding nodes to spontaneously balance the load. Specifically, firstly, the resource demand level and priority of each user are obtained from the fair scheduling criteria among users. Simultaneously, real-time load data of each edge node is collected, including CPU utilization and bandwidth utilization, to calculate the load rate of each node. For example, the node load rate threshold is set to 80%; exceeding this value is considered excessive node load, requiring user migration. Next, node policies and revenue functions are defined. Edge node policies are divided into three categories: prioritizing local users, migrating low-priority users to adjacent nodes, and accepting overflow users from adjacent nodes. The node revenue function is: Net node revenue = Local user experience compliance rate - Scheduling cost. The local user experience compliance rate is the percentage of users within a node whose resource needs are met. For example, out of 100 users within a node, 95 users receive resources that meet their needs, resulting in a compliance rate of 0.95. Scheduling costs include bandwidth transmission costs for user migration and communication costs between nodes. For each low-priority user migrated, the scheduling cost coefficient is set to 0.01; migrating 10 users results in a cost of 0.1. If a node receives overflow users from neighboring nodes, it also needs to calculate the computational cost due to the increased load rate. For every 5% increase in load rate, the cost coefficient increases by 0.05. Then, each node adjusts its strategy based on its current net benefit compared to the benefits of neighboring nodes. For example, if node A's current load rate is 82% and its net benefit is 0.9, while neighboring node B's load rate is 70% and its net benefit is 0.98, node A will prioritize migrating low-priority users to node B. If node B receives the user and its load rate rises to 73%, but its net benefit is still higher than its current benefit, then it agrees to accept the user. Nodes synchronize load data every 2 seconds to update user allocation strategies and avoid network fluctuations caused by frequent migrations. Finally, the game equilibrium state between nodes is determined. When the load rate of all edge nodes stabilizes between 70% and 80%, the number of user migrations between nodes is 0 for three consecutive iterations, and the standard deviation of the net revenue of all nodes is less than 0.05, the game equilibrium between nodes is considered to have been reached, and the iteration stops. After the equilibrium is achieved, a node load balancing scheme is obtained, which clarifies the rules for user migration between nodes and the resource allocation ratio. For example, low-priority users are prioritized to migrate to nodes with a load below 75%, while high-priority users are prioritized to stay on their local nodes, ensuring that the load of each node is evenly distributed and that no node is overloaded or idle. All values ​​in this embodiment are merely examples and are not intended to limit the scope of this embodiment. Those skilled in the art can set them according to actual conditions.

[0122] Finally, a decision-making and scheduling scheme is obtained by combining a hierarchical federated learning mechanism. This mechanism is a distributed learning approach where edge nodes train locally and the cloud center performs global distillation, balancing real-time performance with global optimization. Specifically, the hierarchical federated learning mechanism is adopted, with edge nodes responsible for local real-time decision-making and the cloud center responsible for global policy integration. By combining a node load balancing scheme and a global scheduling objective, a globally optimal bitrate adjustment command and node allocation scheme are generated. Node allocation prioritizes local edge nodes to reduce latency. Ultimately, the commands are integrated into a decision-making and scheduling scheme that precisely adapts to each terminal and user.

[0123] This embodiment solves the problems of resource contention among users, uneven node load, and local optimum in traditional scheduling by dividing the user group into game theory groups and federated learning decision-making. It achieves the dual goals of fair user scheduling and balanced node load, thereby improving the fairness and global optimum of the scheduling scheme.

[0124] Furthermore, this embodiment provides a step for obtaining the globally optimal rate adjustment instruction and node allocation scheme as a decision-making scheduling scheme based on the node load balancing scheme and the global scheduling objective, combined with a hierarchical federated learning mechanism. The steps include:

[0125] Based on the node load balancing scheme and local real-time scheduling requirements, a preliminary scheduling scheme is obtained by making local decisions through the edge node DQN model.

[0126] Based on the preliminary scheduling plan and the global scheduling objectives, the global scheduling strategy is obtained by performing global policy distillation through the cloud center DQN model.

[0127] Based on the global scheduling strategy and the preliminary scheduling scheme, the globally optimal bitrate adjustment instructions and node allocation scheme are integrated to form the decision scheduling scheme.

[0128] Specifically, firstly, edge DQN models and cloud-center DQN models are constructed and trained. The edge DQN model is a deep Q-network model deployed on edge nodes, responsible for local real-time scheduling decisions. The cloud-center DQN model is a deep Q-network model deployed in the cloud center, responsible for global policy distillation and optimization. The edge DQN model uses causal fusion features, node load, and user resource demand confidence as its state space, and bitrate adjustment and node switching as its action space, constructing a lightweight four-layer neural network DQN model. It is trained using an experience replay mechanism with a batch size of 32 and a discount factor of 0.95. The training data is local historical scheduling data, and training stops when the loss function is below 0.01 to ensure model accuracy. The cloud-center DQN model constructs a DQN model with the same structure as the edge nodes, receiving model gradients from all edge nodes but not receiving raw data to protect data privacy.

[0129] Subsequently, the edge nodes make local decisions using the edge DQN model to obtain a preliminary scheduling scheme. This preliminary scheduling scheme is the locally optimal scheduling instruction generated by the edge DQN model and is only suitable for users in the local area. The edge nodes input real-time state data into the locally trained DQN model, which generates a preliminary scheduling scheme adapted to local users. The scheme includes local bitrate adjustment and local node allocation, without involving cross-region scheduling, thus fully ensuring the real-time performance of local decisions.

[0130] Next, the cloud center performs global policy distillation using its DQN model to obtain a global scheduling policy. This global scheduling policy is the globally optimal policy generated by integrating the policies of all edge nodes. Specifically, each edge node, after completing a round of local DQN model training, only uploads the gradient vector of the model parameters to the cloud center, without transmitting any raw user data to ensure data privacy and security. During upload, node status information is simultaneously included, such as the total number of users, real-time load rate, and coverage area identifier. The cloud center assigns a fusion weight to the gradient of each node based on its user scale and load rate: Node weight = (Number of users / Total number of users at the exhibition) × User scale weight + (Node load rate / 100) × Load rate weight. All node weights are normalized to a sum of 1. The user scale weight and load rate weight are determined based on the core needs of the exhibition scenario. For example, user scale directly determines the representativeness of node scheduling experience, so a higher weight, such as 0.6, is assigned to prioritize learning the scheduling patterns of hotspot areas. The load rate reflects the node's operational bottleneck, so a second-highest weight, such as 0.4, is assigned to balance system stability. The cloud center then uses the fused global gradient to update the parameters of the global DQN model, setting the learning rate to 0.005 and the update frequency to be synchronized with the edge nodes every second. The state space and action space of the global DQN model are fully aligned with the edge nodes. The state includes causal fusion features, user resource requirement confidence, and node load rate. The actions include four bitrate adjustments (30 / 24 / 18 / 12Mbps) and node allocation options to ensure compatibility of subsequent policies with the edge nodes. After the cloud center model update is completed, three core global gradients are generated for all areas and exhibit types across the entire exhibition. Scheduling rules include cross-regional user migration rules, such as prioritizing allocation to adjacent nodes with a load rate <75% when users switch to popular exhibits in other regions; popular exhibit resource allocation rules, such as setting the maximum bitrate for VR popular exhibits to 24Mbps to avoid single-node overload; and node load correction rules, such as forcing low-priority users to migrate to adjacent nodes when the load rate of a node is >80%. For example, the global policy for a certain VR popular exhibit is: all users watching this exhibit have a maximum bitrate of 24Mbps and are prioritized for allocation to nodes with a load rate <75%, and are forcibly migrated when the local node's load rate is >80%.

[0131] Finally, the edge nodes integrate the global scheduling strategy and the preliminary scheduling scheme to obtain the globally optimal bitrate adjustment instruction and node allocation scheme, which serves as the decision scheduling scheme. First, after receiving the global scheduling strategy from the cloud center, the edge nodes align the global strategy with the local preliminary scheduling scheme to ensure that the user ID, terminal ID, exhibit ID, and timestamp are completely matched. Then, the edge nodes apply differentiated fusion rules according to the user's scenario type. For ordinary local scenarios, such as users watching non-hotspot exhibits in their local area, the local preliminary scheduling scheme takes precedence. For example, the fusion weight is 70% for the local scheme and 30% for the global strategy, prioritizing local adaptation while considering global balance. For instance, if user Y is watching static exhibits in their local area, the local... The proposed bitrate is 1Mbps, with a global policy recommendation of 1.2Mbps. The resulting bitrate after fusion is 1Mbps, based on the local solution. For cross-region / hotspot scenarios, such as users viewing cross-region or hotspot exhibits, the global scheduling policy takes precedence. For example, the fusion weight is 70% global policy and 30% local policy, prioritizing global cross-region rules and supplementing local corrections. For instance, if user X is viewing cross-region VR hotspot exhibits, the local solution bitrate is 30Mbps, the global policy recommendation is 24Mbps, and the resulting bitrate after fusion is 24Mbps. For conflict correction, if the two instructions conflict, such as opposite bitrate adjustment directions or node allocation conflicts, the global policy instruction is executed first, and the reason for the conflict is recorded and uploaded to the cloud center for subsequent model optimization. Subsequently, for cross-regional scheduling and hot exhibit scheduling scenarios, the edge nodes make two additional corrections. First, cross-regional bitrate correction is performed. Considering the network latency of cross-regional transmission, the merged bitrate is reduced by one level. For example, if the merged bitrate for user X is 24Mbps, it is corrected to 20Mbps during cross-regional transmission to reserve cross-regional bandwidth redundancy and avoid lag. Second, hot exhibit priority correction is performed. Users who watch hot exhibits are given one level higher priority to ensure their resource allocation and prevent them from being preempted by low-priority users. Finally, a final decision scheduling scheme is generated. After the edge nodes complete the fusion and correction, a conflict-free and complete final decision scheduling scheme is generated. The scheme includes each user's bitrate adjustment instruction, node allocation instruction, and priority level. For example, user X's final instruction is: bitrate 20Mbps, node allocation as B, highest priority; user Y's instruction is: bitrate 1Mbps, node allocation as A, medium priority. After the scheme is generated, it is sent to the corresponding terminal and node for execution every 200ms, consistent with the frequency of the scheduling instructions sent above, in order to balance real-time performance and system load, and can be directly used for subsequent resource allocation.

[0132] This embodiment addresses the problems of poor real-time performance, insufficient global optimization, and data privacy leakage in traditional scheduling by employing a hierarchical DQN mechanism that combines local edge training and global distillation in the cloud. It achieves a combination of local real-time decision-making and global optimal scheduling, thereby improving decision-making efficiency and accuracy.

[0133] Furthermore, this embodiment provides a step for completing resource management operations and obtaining a resource allocation plan based on causal fusion characteristics, user resource demand confidence, and decision scheduling schemes, including:

[0134] Based on the causal fusion characteristics, the confidence level of user resource demand, and the scheduling scheme, combined with the prediction results of user flow trajectory, user resource quotas are allocated to obtain the user resource quota scheme.

[0135] Based on the user resource quota scheme and the requirements for preventing resource preemption, a user-exclusive resource ownership certificate is generated to obtain the resource ownership system.

[0136] Based on the resource ownership system and user priority, a differentiated and flexible lock-in strategy is implemented to obtain a flexible lock-in solution.

[0137] Based on the flexible lock-in scheme and resource utilization needs, unused resources are released and reused without disturbance, resulting in a resource allocation scheme.

[0138] Specifically, such as Figure 2As shown, user resource quotas are first allocated to form a user resource quota scheme. This scheme represents the upper limit of exhibition resources allocated to each user, preventing excessive resource consumption. Then, based on causal fusion features, user resource demand confidence, and a decision-making scheduling scheme, the LSTM user flow trajectory prediction results are combined. For example, a user flow trajectory prediction LSTM model is first constructed. The model input is a time series of user exhibition behavior, such as the dwell time in the last 10 time steps, the type and resource demand of the currently viewed exhibit, causal fusion features, and user resource demand confidence. The model uses a two-layer LSTM structure with 64 hidden units. The output is the predicted probability distribution of the exhibits the user will view in the next 30 seconds, including the probability of the current exhibit, adjacent exhibits, and popular exhibits. The model is trained using historical exhibition user viewing sequence data, optimized with a cross-entropy loss function. Training stops when the validation set prediction accuracy reaches 85% or higher. After training, the model can predict a user's viewing trajectory based on their current state. The principle behind setting a 30-second prediction window is that the average interval between users switching exhibits is 20-30 seconds. Predicting exhibits to be viewed 30 seconds in advance allows sufficient time to pre-allocate resources for users, avoiding loading lag when switching exhibits. For each online user, the system calls the LSTM model every second, inputting the user's current viewing data for prediction. For example, if a user is currently watching a VR immersive exhibit, and their dwell time is stable and interaction frequency is high in the past second, the model predicts that the probability of them continuing to watch this VR exhibit in the next 30 seconds is 70%, the probability of switching to an adjacent 4K video exhibit is 25%, and the probability of switching to a static exhibit is 5%. The model automatically selects the exhibit with the highest probability as the prediction result, meaning the user will still watch the VR exhibit in the next 30 seconds. Based on the predicted exhibit type, the corresponding resource requirement benchmark is retrieved from the exhibit resource requirement profile database, and a matching resource quota is allocated to the user. For example, the resource requirement for VR exhibits is 30Mbps bandwidth and 60fps frame rate, so a 30Mbps bandwidth quota and a CPU utilization threshold of 70% are allocated to this user. If the model predicts that the user will switch to 4K video exhibits, the quota is adjusted in advance to 15Mbps bandwidth and a CPU utilization threshold of 50% to avoid resource waste. At the same time, the quota is adjusted based on the user's resource requirement confidence level. If the confidence level is <0.7, the quota is reduced by 10% to reserve redundant resources to cope with fluctuations. The quota allocation results of all users are integrated to form a standardized user resource quota scheme.

[0139] Next, resource ownership certificates are generated to establish a resource ownership system. These certificates serve as unique authorization credentials for user resource quotas. Nodes allocate resources only to users holding valid certificates. The resource ownership system is a resource allocation and authorization control system built upon these certificates. Each user's resource quota corresponds to a unique, encrypted resource ownership certificate. This certificate is bidirectionally bound to the user ID and terminal ID to prevent impersonation and theft. Before allocating resources, edge nodes verify the validity of the user's resource ownership certificate, allocating resources only to users with valid certificates. This constructs a complete resource ownership system, preventing resource preemption and unauthorized allocation at the authorization level.

[0140] Next, an elastic locking strategy is implemented to obtain an elastic locking solution. This strategy differentiates resource quotas based on user priority, ensuring that core user resources are not preempted. For example, firstly, based on the priority of user viewing behavior, the locking strategy is divided into three levels, corresponding to the user priority definition mentioned earlier. For instance, hard locking corresponds to high-priority users. For users who linger for extended viewing (e.g., dwell time > 10 seconds, interaction frequency ≥ 3 times / minute), their resource quota is completely locked, usable only by that user, and cannot be preempted by other users. The locking level field is marked as Level 1 in the resource ownership certificate. When edge nodes allocate resources, they prioritize guaranteeing the quota for this type of user, and their resources are not reclaimed even if the node load rate increases. For example, for users deeply viewing VR immersive exhibits, the 30Mbps bandwidth quota is hard locked, ensuring it will not be preempted by other users under any circumstances. Medium-priority users are assigned a lock-in level. For users waiting to load content, such as those loading high-bitrate exhibits with a dwell time > 5 seconds, resource quotas can be preempted by high-priority users with minimal disruption. The preemption ratio is capped at 30% of the quota, and the lock-in level is marked as Level 2. High-priority users can only preempt a portion of the resources from medium-priority users when their resource demand exceeds their own quota, and the basic loading bandwidth of medium-priority users must be guaranteed to be no less than 70% of their demand. For example, for users waiting to load 4K video, with a 15Mbps bandwidth quota under medium-priority lock-in, high-priority VR users can preempt up to 4.5Mbps, leaving 10.5Mbps to ensure normal video loading. Soft-priority users are assigned a lock-in level. For users browsing quickly, such as those with a dwell time < 3 seconds and no interactive behavior, resource quotas can be reasonably preempted by high and medium-priority users, with no upper limit on the preemption ratio. The lock-in level is marked as Level 3. When node resources are scarce, all of their quotas can be reclaimed and allocated to core users. Low-priority users can only obtain the minimum guaranteed bandwidth, such as 50% of the 1Mbps bandwidth for static exhibits, i.e., 0.5Mbps. For example, a user quickly browsing static exhibits has a 1Mbps bandwidth quota that is soft-locked, meaning it can be completely preempted by other users without affecting the overall viewing experience. When a high-priority user's resource demand exceeds their hard-locked quota, the system preempts resources according to the following rules: First, it prioritizes preempting resources from users with medium-locked quotas within the same area, executing at the maximum preemption ratio to ensure that the basic loading of users with medium-locked quotas is not affected; if resources from users with medium-locked quotas are insufficient, then it preempts resources from users with soft-locked quotas, reclaiming all of their quotas; each preemption operation only adjusts the quota of one user to avoid experience fluctuations caused by preempting multiple users simultaneously; after preemption, the lock level and quota information in the resource ownership certificates of all relevant users are updated to ensure consistent allocation permissions.The system updates priority every 5 seconds based on user viewing behavior, and adjusts the access control level accordingly. For example, if a user changes from quickly browsing to staying on view, the priority rises from low to high, the access control level changes from level 3 to level 1, and the resource quota is automatically hard-locked, preventing other users from preempting it. If the user switches back to quickly browsing, the access control level changes from level 1 to level 3, and the quota becomes soft-locked, allowing it to be preempted. This access control strategy ensures a good viewing experience for core users while improving overall resource utilization.

[0141] Finally, unused resources are released without disturbance to obtain a resource allocation plan. Undisturbed release is a resource reclamation method that releases unused resources along a smooth curve, without affecting system operation and user experience. The system monitors resource quota usage in real time. For example, unused pre-allocated resources can be released gradually along an exponential smooth curve to avoid system load fluctuations caused by sudden large-scale releases. Released resources are automatically merged into the elastic resource pool for reuse by other users in need. The entire process of quota allocation, authorization, locking, and release is integrated to obtain a complete resource allocation plan.

[0142] This embodiment solves the problems of preemption, waste, and release disturbance in traditional resource management by allocating quotas, confirming rights, locking rights, and releasing them without disturbance, thereby achieving precise resource allocation, efficient reuse, and secure control, and reducing resource waste rate.

[0143] Furthermore, this embodiment provides a step for completing robust anti-interference operations and obtaining an anti-interference scheduling scheme based on causal fusion characteristics, user resource demand confidence, and resource allocation scheme, including:

[0144] Based on the causal fusion characteristics, the confidence level of user resource requirements, and the requirements for disturbance detection, uncertainty and its rate of change are monitored in real time to obtain uncertainty monitoring data.

[0145] Based on the uncertainty monitoring data and the disturbance rejection triggering rules, determine whether to trigger the robust disturbance rejection mode and obtain the triggering judgment result;

[0146] Based on the trigger judgment result, the scheduling mode is switched and adapted. Based on the scheduling mode adjustment result, an anti-disturbance scheduling scheme adapted to sudden scenarios is obtained.

[0147] Specifically, the system first performs real-time monitoring to obtain uncertainty monitoring data. This data represents the fluctuation range and rate of change of user resource demand confidence levels, reflecting the system's operational stability. Then, based on causal fusion characteristics and user resource demand confidence levels, an anti-interference monitoring process is initiated. For example, the monitoring frequency can be set to 50ms / time, continuously collecting the uncertainty and rate of change of user resource demand confidence levels. Uncertainty reflects the fluctuation range of resource demand, and the rate of change reflects the speed of fluctuation. The monitoring data is uploaded in real-time to the edge node anti-interference module, providing accurate data support for subsequent anti-interference judgments. For example, the uncertainty is a standardized value quantified from the time-series data of user resource demand confidence levels. The calculation first uses the user resource demand confidence level collection frequency as a basis, taking the confidence level data from the previous 10 consecutive time steps as the calculation window. The average of the 10 confidence level data within the window is calculated, then the difference between the maximum and minimum confidence levels is calculated. The difference is divided by the average to obtain the relative volatility. The relative volatility is then normalized to the 0-1 range to obtain the uncertainty value. The rate of change of uncertainty is an indicator that reflects how fast uncertainty changes over time. First, determine the time interval between two consecutive uncertainty calculations, obtain the uncertainty value at the current moment, and subtract it from the uncertainty value at the previous monitoring moment to obtain the change. Divide the change of uncertainty by the time interval to obtain the rate of change of uncertainty per unit time.

[0148] Subsequently, a judgment is made based on the anti-disturbance triggering rules to obtain the triggering result. The anti-disturbance triggering rules are the standard conditions for determining whether to enter the robust anti-disturbance mode. The robust anti-disturbance mode is a stable scheduling mode to deal with sudden scenarios, which can reduce frequent adjustments and ensure the smooth operation of the system. For example, when the uncertainty exceeds 0.3 and the rate of change exceeds 0.1 / second, the current scenario is determined to be a sudden abnormal scenario, and the robust anti-disturbance mode needs to be triggered. When this condition is not met, it is determined to be a normal viewing scenario, and the fine scheduling mode is maintained.

[0149] Finally, based on the trigger judgment result, the scheduling mode is switched and adapted to obtain an anti-interference scheduling scheme adapted to the sudden scenario. The scheduling mode adaptation adjustment is the parameter adjustment operation when switching between anti-interference mode and fine scheduling mode. If the trigger judgment result is to enter robust anti-interference mode, frequent bitrate adjustment and node switching operations are immediately stopped, only the parameter adjustment of the large level is retained, the resource quota of core users such as stationary users is locked, the pre-allocation of resources for non-core users is turned off, and the stable operation of the system is prioritized. If robust anti-interference mode is not triggered, the system maintains fine scheduling mode, continuously optimizes scheduling parameters, improves user experience, and generates an anti-interference scheduling scheme adapted to the current scenario based on the mode switching and adjustment results.

[0150] This embodiment addresses the issues of traditional systems being prone to lag and instability when dealing with sudden traffic surges and network fluctuations by using uncertainty monitoring, rule-based judgment, and mode switching. It improves the system's robustness and anti-interference capabilities, ensuring uninterrupted viewing for users in high-concurrency and weak-network scenarios.

[0151] Furthermore, this embodiment provides a step-by-step process for completing closed-loop optimization based on a decision-making scheduling scheme, a resource allocation scheme, and an anti-interference scheduling scheme to obtain a multi-terminal collaborative remote viewing load balancing scheme, including:

[0152] Based on the execution status of the decision-making scheduling plan, resource allocation plan, and disturbance suppression scheduling plan, the plan operation monitoring data is obtained;

[0153] Based on the monitoring data of the scheme operation and optimization needs, identify the shortcomings of the scheme operation and determine the optimization direction to obtain optimization guidance;

[0154] Based on the optimization guidelines, the operating parameters of each module are dynamically corrected and the relevant models are updated to obtain the optimized parameters and model system.

[0155] Based on the optimized parameters and model system, the load balancing scheme is iteratively optimized to obtain a multi-terminal collaborative remote exhibition load balancing scheme.

[0156] Specifically, the system first aggregates execution data to obtain solution operation monitoring data. This data encompasses all dimensions of experience, resources, and operation after the load balancing solution is implemented, comprehensively reflecting the solution's effectiveness. The system aggregates execution data for decision-making scheduling, resource allocation, and anti-interference scheduling at fixed intervals, such as every minute. The collected data includes core indicators such as user viewing lag rate, edge node load rate, resource utilization, and system oscillation index. This data covers user experience, resource usage, and system operation dimensions, forming complete solution operation monitoring data.

[0157] like Figure 3As shown, the monitoring data is then analyzed to obtain optimization guidance, which corresponds to the optimization directions identified in the monitoring data. The monitoring data of the solution's operation is compared with preset target values. First, the solution's operation monitoring data is collected, summarizing three categories of core indicators every minute: user experience indicators such as lag rate, loading success rate, and experience compliance rate; node operation indicators such as node load rate, load balancing, and scheduling energy consumption; and resource utilization indicators such as bandwidth utilization and resource waste rate. All data is stored categorized by user, node, and exhibit type for targeted comparison. Second, preset target values ​​are defined, setting thresholds for each type of indicator, balancing user experience, node stability, and resource efficiency. For example, lag rate ≤ 2%, node load rate ≤ 80%, bandwidth utilization ≥ 85%, and resource waste rate ≤ 10%, forming a standardized comparison benchmark. Specific values ​​can be fine-tuned according to the exhibition scale and exhibit type. Next, the comparison is performed item by item, comparing the actual monitoring data with the target values ​​to identify abnormal indicators exceeding the thresholds. For example, monitoring data for a VR exhibit area showed a 5% stuttering rate, 90% load on node A, 55% load on node B, 70% bandwidth utilization, and 15% resource waste. Comparison revealed three types of problems: excessively high stuttering rate, uneven node load, and resource waste. Finally, the root cause was located, and the cause of the anomaly was traced through the scheduling process. If an excessively high stuttering rate was accompanied by an excessively high node load, the root cause was node overload leading to scheduling delays. If uneven node load was accompanied by an excessively high load balancing, the root cause was an unreasonable user game equilibrium rule. If an excessively high resource waste rate was accompanied by insufficient bandwidth utilization, the root cause was the failure of the elastic locking strategy to reclaim idle resources in a timely manner. For each problem, corresponding parameter adjustments, model updates, and logic optimization directions were determined, forming clear and actionable optimization guidelines.

[0158] Following the optimization guidelines, parameters are corrected and models are updated to obtain optimized parameters and model systems. These optimized parameters and model systems are a set of corrected parameters and updated models, capable of adapting to dynamic changes in the scenario. Next, according to the optimization guidelines, the feature weights, game payoff parameters, and robust disturbance rejection trigger thresholds of the temporal causal attention fusion are dynamically corrected. Simultaneously, the temporal causal graph attention model and the DQN scheduling model are updated. For example, the update cycle is set to 30 minutes to ensure that the parameters and models always conform to the dynamic changes of the remote exhibition scenario, forming the optimized parameter and model system.

[0159] Finally, based on the optimized parameters and model system, iterative optimization is performed to obtain the final load balancing solution. Iterative optimization is the operation of continuously updating the load balancing solution based on the optimized system. The optimized parameters and model are imported into the scheduling core module to comprehensively update the load balancing decision, resource, and anti-interference logic. This allows the solution to continuously adapt to changes in exhibition traffic, exhibit popularity, terminal type updates, and other scenario changes, ensuring that the system always maintains optimal operating status, ultimately resulting in a stable and efficient multi-terminal collaborative remote exhibition load balancing solution.

[0160] This embodiment constructs a closed-loop mechanism for continuous self-optimization through data monitoring and iterative optimization, solving the problem of traditional scheduling schemes being rigid and unable to adapt to dynamic scenarios, achieving long-term stable and efficient operation of the scheme, and improving the adaptability and life cycle of the entire system.

[0161] Furthermore, embodiments of this application provide a multi-terminal collaborative remote exhibition load balancing system, including:

[0162] The configuration module is used to obtain parameter configurations adapted to the scenario based on the requirements of remote exhibition viewing and the access types of multiple terminals.

[0163] The verification module is used to perform preliminary verification based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data.

[0164] The causal fusion module is used to perform temporal causal attention fusion processing based on the original four-dimensional data and the causal logic of the remote exhibition scenario to obtain causal fusion features and user resource demand confidence.

[0165] The load balancing module is used to obtain a multi-terminal collaborative remote exhibition load balancing solution based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives.

[0166] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-terminal collaborative remote exhibition load balancing method, characterized in that, include: Based on the needs of remote exhibition viewing scenarios and the types of multi-terminal access, parameter configurations adapted to the scenarios are obtained; Preliminary verification was performed based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data. Based on the original four-dimensional data and the causal logic of the remote exhibition scenario, temporal causal attention fusion processing is performed to obtain causal fusion features and user resource demand confidence. Based on the causal fusion characteristics, user resource demand confidence, and global scheduling objectives, a multi-terminal collaborative remote exhibition load balancing scheme is obtained.

2. The multi-terminal collaborative remote exhibition load balancing method according to claim 1, characterized in that, The parameter configuration adapted to the scenario, based on the requirements of remote exhibition viewing and the access types of multiple terminals, includes: Based on the scale of the remote exhibition scenario and the types of multi-terminal access, cloud center nodes, edge nodes, and multi-terminal devices are deployed, the core responsibilities of each node are clarified, and a hardware deployment plan is obtained. Based on the characteristics of various exhibits and the needs of remote exhibition scheduling, the exhibit feature parameters, algorithm running parameters, and user priority parameters are preset to obtain the initial parameter configuration results. Based on the initial parameter configuration results and exhibit resource classification standards, an exhibit resource demand profile library is constructed, exhibit resource demand characteristics are classified and archived, and a parameter support system is obtained. Based on the hardware deployment plan and parameter support system, the parameter configurations adapted to the scenario are obtained.

3. The multi-terminal collaborative remote exhibition load balancing method according to claim 1, characterized in that, The preliminary verification based on parameter configuration and multi-terminal data collection requirements yields four-dimensional raw data, including: Based on parameter configuration and multi-terminal data collection requirements, collect terminal performance-related data to obtain terminal characteristic data; Based on parameter configuration and network monitoring requirements, network status data between terminals and nodes is collected through edge node network probes to obtain network characteristic data; Preliminary verification is performed based on terminal feature data and network feature data to remove abnormal data and obtain four-dimensional raw data.

4. The multi-terminal collaborative remote exhibition load balancing method according to claim 1, characterized in that, The process involves performing temporal causal attention fusion processing based on the original four-dimensional data and the causal logic of the remote exhibition scenario to obtain causal fusion features and user resource demand confidence levels, including: Based on the original four-dimensional data and the characteristics of the remote exhibition scene, a directed causal topology graph is constructed to obtain the scene causal topology structure; Based on the scene's causal topology and four-dimensional raw data, the attention weights between each causal node are calculated to obtain the true causal relationship information. Based on real causal relationship information and four-dimensional raw data, deep feature fusion processing is performed to obtain causal fusion features; Based on the causal fusion characteristics and fluctuation factors, the confidence level and uncertainty of user resource demand are calculated, thus obtaining the causal fusion characteristics and the confidence level of user resource demand.

5. The multi-terminal collaborative remote exhibition load balancing method according to claim 1, characterized in that, The multi-terminal collaborative remote exhibition load balancing scheme, based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives, includes: Based on the causal fusion characteristics, the confidence level of user resource requirements, and the global scheduling objective, the decision scheduling operation is completed to obtain the decision scheduling scheme. Based on the causal fusion characteristics, the confidence level of user resource needs, and the decision scheduling scheme, resource management operations are completed to obtain a resource allocation scheme. Based on the causal fusion characteristics, the confidence level of user resource requirements, and the resource allocation scheme, robust anti-disturbance operation is completed to obtain an anti-disturbance scheduling scheme. Based on the decision-making scheduling scheme, resource allocation scheme, and anti-interference scheduling scheme, a closed-loop optimization operation is completed to obtain a multi-terminal collaborative remote exhibition load balancing scheme.

6. The multi-terminal collaborative remote exhibition load balancing method according to claim 5, characterized in that, The step of performing decision-making scheduling operations based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives to obtain a decision-making scheduling scheme includes: Based on the causal fusion characteristics, the confidence level of user resource demand, and the global scheduling objective, user game groups are divided and reward rules are set to obtain user game groups and reward system. Based on the user game groups and the reward system, the game equilibrium within the user groups is solved to obtain the basis for fair scheduling among users. Based on the fair scheduling criteria among users and the load conditions of edge nodes, the game equilibrium among edge nodes is solved to obtain the node load balancing scheme. Based on the node load balancing scheme and the global scheduling objective, and combined with the hierarchical federated learning mechanism, the globally optimal bitrate adjustment instruction and node allocation scheme are obtained as the decision scheduling scheme.

7. The multi-terminal collaborative remote exhibition load balancing method according to claim 6, characterized in that, Based on the node load balancing scheme and the global scheduling objective, and combined with the hierarchical federated learning mechanism, the globally optimal rate adjustment instructions and node allocation scheme are obtained, which serve as the decision-making scheduling scheme, including: Based on the node load balancing scheme and local real-time scheduling requirements, a preliminary scheduling scheme is obtained by making local decisions through the edge node DQN model. Based on the preliminary scheduling plan and the global scheduling objectives, the global scheduling strategy is obtained by performing global policy distillation through the cloud center DQN model. Based on the global scheduling strategy and the preliminary scheduling scheme, the globally optimal bitrate adjustment instructions and node allocation scheme are integrated to form the decision scheduling scheme.

8. A multi-terminal collaborative remote exhibition load balancing method according to claim 5, characterized in that, Based on causal fusion characteristics, user resource demand confidence, and decision scheduling schemes, resource management operations are completed to obtain resource allocation schemes, including: Based on the causal fusion characteristics, the confidence level of user resource demand, and the scheduling scheme, combined with the prediction results of user flow trajectory, user resource quotas are allocated to obtain the user resource quota scheme. Based on the user resource quota scheme and the requirements for preventing resource preemption, a user-exclusive resource ownership certificate is generated to obtain the resource ownership system. Based on the resource ownership system and user priority, a differentiated and flexible lock-in strategy is implemented to obtain a flexible lock-in solution. Based on the flexible lock-in scheme and resource utilization needs, unused resources are released and reused without disturbance, resulting in a resource allocation scheme.

9. A multi-terminal collaborative remote exhibition load balancing method according to claim 5, characterized in that, Based on causal fusion characteristics, user resource demand confidence, and resource allocation schemes, robust disturbance rejection operations are performed, resulting in a disturbance rejection scheduling scheme, including: Based on the causal fusion characteristics, the confidence level of user resource requirements, and the requirements for disturbance detection, uncertainty and its rate of change are monitored in real time to obtain uncertainty monitoring data. Based on the uncertainty monitoring data and the disturbance rejection triggering rules, determine whether to trigger the robust disturbance rejection mode and obtain the triggering judgment result; Based on the trigger judgment result, the scheduling mode is switched and adapted. Based on the scheduling mode adjustment result, an anti-disturbance scheduling scheme adapted to sudden scenarios is obtained.

10. A multi-terminal collaborative remote exhibition load balancing method according to claim 6, characterized in that, Based on the decision-making scheduling scheme, resource allocation scheme, and anti-interference scheduling scheme, a closed-loop optimization operation is completed to obtain a multi-terminal collaborative remote viewing load balancing scheme, including: Based on the execution status of the decision-making scheduling plan, resource allocation plan, and disturbance suppression scheduling plan, the plan operation monitoring data is obtained; Based on the monitoring data of the scheme operation and optimization needs, identify the shortcomings of the scheme operation and determine the optimization direction to obtain optimization guidance; Based on the optimization guidelines, the operating parameters of each module are dynamically corrected and the relevant models are updated to obtain the optimized parameters and model system. Based on the optimized parameters and model system, the load balancing scheme is iteratively optimized to obtain a multi-terminal collaborative remote exhibition load balancing scheme.

11. A multi-terminal collaborative remote exhibition load balancing system, used to implement the multi-terminal collaborative remote exhibition load balancing method as described in any one of claims 1-10, characterized in that, include: The configuration module is used to obtain parameter configurations adapted to the scenario based on the requirements of remote exhibition viewing and the access types of multiple terminals. The verification module is used to perform preliminary verification based on parameter configuration and multi-terminal data acquisition requirements to obtain four-dimensional raw data. The causal fusion module is used to perform temporal causal attention fusion processing based on the original four-dimensional data and the causal logic of the remote exhibition scenario to obtain causal fusion features and user resource demand confidence. The load balancing module is used to obtain a multi-terminal collaborative remote exhibition load balancing solution based on causal fusion characteristics, user resource demand confidence, and global scheduling objectives.

Images