A multi-modal large model-based dispatch communication intelligent system

Abstract

The application discloses a kind of scheduling communication intelligent systems based on multimodal big model, it is related to communication technical field, including data acquisition and fusion module, from the modality information and historical scheduling experience data of new arrival data accompanying in communication scheduling environment acquisition, data is updated scheduling strategy by incremental learning algorithm processing, obtain the strategy model after preliminary fusion, experience reservation module, according to the time sequence dependent correlation features in preliminary fusion strategy model, integrate time sequence dependent correlation features with new modality information using federated learning algorithm, determine the updated experience reservation module;The scheduling communication intelligent systems based on multimodal big model improves network resource utilization efficiency while ensuring the stability of data transmission, so as to enhance the intelligent level, adaptability and overall operating performance of communication scheduling system in dynamic complex scene.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and more specifically to an intelligent scheduling communication system based on a multimodal large model. Background Technology

[0002] In the field of communication scheduling, scheduling systems need to rationally allocate and efficiently transmit ever-increasing business data under the condition of limited network resources. Their scheduling capabilities directly affect the overall operating efficiency and service quality of the communication network. As traditional voice communication network services evolve towards intelligence and automation, the original work mode based on manual dialing, call recording, and post-call review for accountability has gradually upgraded to AI-driven voice command control, real-time transcription of the call process, keyword retrieval of the transcribed text, and large-scale event analysis based on massive call records. This expands the business link from "voice recording" to a closed loop of "voice-text-knowledge analysis," significantly increasing the complexity of system processing and management requirements.

[0003] Against this backdrop, the sources and forms of data generated during communication scheduling are constantly enriching. Newly arriving data is often accompanied by multimodal information, such as differences in terminal types, changes in service characteristics, changes in environmental conditions, and different temporal characteristics. These multimodal data often exhibit strong temporal correlations and contextual dependencies with existing historical scheduling experience, posing new challenges to the continuous updating and stable operation of scheduling strategies. However, existing scheduling methods typically lack effective long-term learning and knowledge inheritance mechanisms when dealing with the continuous influx of multimodal data. In current technologies, updating scheduling strategies often relies on centralized or periodic retraining of existing models to absorb the characteristics of newly arriving data. This approach may work in scenarios with small-scale or slowly changing services, but in communication scheduling environments with rapid data growth and frequent changes in service types, it easily leads to high computational resource consumption, long update cycles, and delayed response times. Furthermore, with the continuous introduction of new data and services, scheduling systems are prone to underutilizing historical effective experience, manifesting as a decline in original scheduling capabilities or reduced decision stability—the so-called "knowledge forgetting" phenomenon. Furthermore, multimodal data differ in statistical characteristics and business semantics. When inconsistencies or conflicts arise between new modal information and historical scheduling experience, existing scheduling systems often lack effective conflict identification and adjustment mechanisms, easily leading to unstable scheduling decisions and even local performance degradation. In complex application scenarios, such as industrial park communication environments where video surveillance and a large amount of sensor data services coexist, the reporting frequency and latency requirements of different services vary significantly. If the scheduling system cannot coordinate the needs of multiple services, it may cause delays in critical data transmission, affecting system security and reliability. On the other hand, with the expansion of business scale and the continuous growth of data volume, communication scheduling systems also face the challenge of balancing scheduling performance with network resource constraints. Existing technologies for optimizing scheduling performance mostly focus on single indicators or static scenarios, lacking the ability to continuously evaluate and adaptively optimize performance indicators under dynamic business growth conditions, making it difficult to achieve efficient resource utilization while ensuring transmission stability. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent scheduling and communication system based on a multimodal large model, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent scheduling and communication system based on a multimodal large model, comprising a data acquisition and fusion module, which acquires modal information accompanying newly arriving data and historical scheduling experience data from the communication scheduling environment, processes the data using an incremental learning algorithm to update the scheduling strategy, and obtains a preliminary fused strategy model; an experience retention module, which extracts temporal dependency correlation features from the preliminary fused strategy model, integrates the temporal dependency correlation features with the modal information accompanying newly arriving data using a federated learning algorithm, and determines an updated experience retention module; and a decision construction module, which, if the updated experience retention module detects modal difference conflicts, adjusts the parameters of the conflicting part using a gradient descent algorithm to obtain the conflict... The system comprises the following modules: a reconciled scheduling decision framework; a data sequence optimization module, which extracts diverse reporting patterns from the reconciled scheduling decision framework, performs time-series analysis on these patterns to determine if they meet the requirements for rapid adaptation to changes, and obtains an optimized data processing sequence; a forgetting compensation module, which extracts historical fragments related to knowledge forgetting from the optimized data processing sequence, re-injects these historical fragments into the scheduling decision framework using an incremental learning algorithm, and determines the complete strategy after forgetting compensation; and a scheduling adaptation model construction module, which extracts the latency requirements of sensor data from the industrial park from the complete strategy after forgetting compensation, and integrates them with video surveillance experience through a federated learning algorithm to obtain the final scheduling adaptation model.

[0006] Preferably, the data acquisition and fusion module acquires modal information and historical scheduling experience data accompanying newly arriving data from the communication scheduling environment, processes the data to update the scheduling strategy through an incremental learning algorithm, and obtains a preliminary fused strategy model. This includes acquiring newly arriving multimodal data streams and historical scheduling logs from the communication scheduling environment, processing the newly arriving multimodal data streams and the historical scheduling logs using multidimensional feature mapping technology to obtain a spatiotemporal feature vector sequence; inputting the spatiotemporal feature vector sequence into a deep neural network to generate a high-dimensional state representation matrix; if the high-dimensional state representation matrix satisfies the drift condition, calculating the update gradient of the strategy network weights based on the high-dimensional state representation matrix to obtain the updated strategy parameter set; and using an adaptive weighted aggregation algorithm to fuse the updated strategy parameter set with the global baseline model parameters to obtain the preliminary fused strategy model.

[0007] Preferably, the experience retention module extracts temporal dependency features from the initially fused policy model, integrates the temporal dependency features with new modal information using a federated learning algorithm, and determines the updated experience retention module by: using a long short-term memory network to parse the initially fused policy model to obtain temporal dependency feature vectors; using a multi-head attention mechanism to project the new modal information data stream onto the semantic feature space where the temporal dependency feature vectors are located to generate a multimodal collaborative feature matrix; calculating the federated local gradient update amount based on the multimodal collaborative feature matrix, and performing a safe weighted aggregation calculation on the federated local gradient update amount to generate a globally shared feature increment; mapping the globally shared feature increment to the experience memory network, and adjusting the weights of the experience memory network through an adaptive forgetting gating unit to determine the updated experience retention module.

[0008] Preferably, the decision-making construction module, if the updated experience retention module detects modal discrepancy conflicts, adjusts the parameters of the conflicting part using the gradient descent algorithm to obtain the conflict-resolved scheduling decision framework. This includes acquiring multimodal feature distribution data within the updated experience retention module and calculating the Mahalanobis distance between different modal feature vectors; if the Mahalanobis distance exceeds a preset conflict detection threshold, a modal discrepancy conflict is determined to exist, and a joint loss function is constructed based on the modal discrepancy conflict to calculate the loss value; the loss value is input into the backpropagation computation graph to generate a gradient vector, and the weight parameters are adjusted along the negative direction of the gradient vector using the gradient descent algorithm to obtain a converged corrected parameter set; the internal mapping logic of the experience retention module is reorganized according to the corrected parameter set to generate the conflict-resolved scheduling decision framework.

[0009] Preferably, the data sequence optimization module obtains diverse reporting pattern data from the conflict-resolved scheduling decision framework, performs time-series sequence analysis on the reporting pattern data, determines whether it meets the requirements for rapid adaptation to changes, and obtains an optimized data processing sequence. This includes extracting historical scheduling logs from the conflict-resolved scheduling decision framework to construct an original reporting time sequence set reflecting the dynamic attributes of the business; generating a feature vector matrix for the original reporting time sequence set and calculating an adaptability score based on the feature vector matrix; if the adaptability score is lower than a threshold, identifying the blocking nodes causing response delays; resetting the queue priority based on the blocking nodes and dynamically scaling the processing window according to the queue priority; and rearranging the data flow order according to the processing window to obtain an optimized data processing sequence that meets the requirements for adapting to changes.

[0010] Preferably, the forgetting compensation module extracts historical fragments related to knowledge forgetting based on the optimized data processing sequence, re-injects these historical fragments into the framework using an incremental learning algorithm, and determines the complete strategy after forgetting compensation. This includes obtaining the optimized data processing sequence, labeling forgetting features based on the backtesting results of the data processing sequence on the historical validation set, filtering historical log vectors based on the forgetting features to obtain historical fragments related to knowledge forgetting, generating an incremental model using a historical fragment-driven online incremental learning algorithm, calculating a forgetting compensation factor based on the generalization error of the incremental model, and mapping the forgetting compensation factor to the output layer of the incremental model to determine the complete strategy after forgetting compensation.

[0011] Preferably, the scheduling adaptation model construction module obtains the latency requirement features of industrial park sensor data from the complete strategy after forgetting compensation, and fuses them with video surveillance experience through a federated learning algorithm to obtain the final scheduling adaptation model. This includes parsing the communication protocol messages extracted from the complete strategy after forgetting compensation to generate latency requirement features that quantify the urgency of sensor data transmission; constructing a monitoring experience vector reflecting the sensitivity of video surveillance services to network fluctuations based on the latency requirement features, and inputting the monitoring experience vector into the federated learning framework to aggregate and obtain a global experience model; mapping the cross-regional network congestion prediction value output by the global experience model to the latency requirement features to generate a heterogeneous fusion feature matrix characterizing the correlation between sensor data and video surveillance services; and using the heterogeneous fusion feature matrix to supervise and adjust the pre-set scheduling neural network to obtain the final scheduling adaptation model with multi-service collaborative processing capabilities.

[0012] Preferably, the system also includes a real-time verification and stable output module, which performs real-time verification on the final scheduling adaptation model to determine whether the increase in new data causes a time delay problem. If so, the model parameters are adjusted to obtain a stable transmission scheduling output. Specifically, this includes acquiring the real-time verification data stream generated by sensors in the industrial park, calculating the new data growth rate based on the timestamp difference of the real-time verification data stream, and mapping the new data growth rate to a preset time delay threshold to calculate the transmission delay deviation; inputting the transmission delay deviation into the congestion prediction module to generate a congestion prediction residual; if the congestion prediction residual exceeds the allowable range, generating a gradient update vector to modify the model weight values ​​of the scheduling adaptation model; using the updated model weight values ​​to generate a stable transmission queue that matches the current data scale, and parsing the stable transmission queue to extract the stable transmission scheduling output.

[0013] Preferably, it also includes a resource allocation optimization module, which extracts data transmission performance indicators from the stable transmission scheduling output, uses a gradient descent algorithm to optimize the performance indicators and the constraints of limited network resources, and determines a resource allocation scheme. Specifically, it includes parsing the stable transmission scheduling output to obtain a scheduling log sequence, extracting throughput and packet loss rate values ​​from the scheduling log sequence, and inputting the throughput and packet loss rate values, combined with bandwidth limit boundaries and energy consumption values, into a loss construction module to calculate the loss function value.

[0014] Preferably, the resource allocation optimization module extracts data transmission performance indicators from the stable transmission scheduling output, uses the gradient descent algorithm to optimize the performance indicators and the constraints of limited network resources, and determines the resource allocation scheme by taking the partial derivative of the loss function to obtain the gradient vector value, using the gradient vector value to iteratively correct the resource allocation weight to obtain the optimal resource weight value; determines the physical layer resource block mapping table based on the optimal resource weight value, and determines the power allocation value and spectrum resource block location of each terminal node based on the resource block mapping table to complete the determination of the resource allocation scheme.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This intelligent scheduling and communication system, based on a multimodal large-scale model, can continuously optimize and stably evolve scheduling strategies under complex conditions such as diverse service types, continuously growing data scale, and limited network resources in the communication scheduling environment. By effectively integrating the multimodal information accompanying newly arriving data with historical scheduling experience, scheduling strategy updates can be completed without frequent full retraining, reducing computational resource consumption and system update latency. Simultaneously, by identifying and adjusting conflicts arising from multimodal differences, it effectively alleviates scheduling instability caused by inconsistencies between old and new service characteristics, preventing the rapid forgetting of effective historical experience. Furthermore, the system can adaptively model and coordinate scheduling for significantly different service requirements in different industry scenarios (e.g., different reporting patterns and latency requirements for video surveillance and industrial sensor data). Combined with real-time verification mechanisms and resource allocation optimization processes, it improves network resource utilization efficiency while ensuring data transmission stability, thereby enhancing the intelligence, adaptability, and overall performance of the communication scheduling system in dynamic and complex scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, this invention provides a technical solution: an intelligent scheduling and communication system based on a multimodal large model, comprising a data acquisition and fusion module, which acquires modal information accompanying newly arriving data and historical scheduling experience data from the communication scheduling environment, processes the data using an incremental learning algorithm to update the scheduling strategy, and obtains a preliminary fused strategy model; an experience retention module, which extracts temporal dependency correlation features from the preliminary fused strategy model, integrates the temporal dependency correlation features with the modal information accompanying newly arriving data using a federated learning algorithm, and determines an updated experience retention module; a decision construction module, which, if the updated experience retention module detects modal difference conflicts, adjusts the parameters of the conflicting part using a gradient descent algorithm to obtain a conflict-resolved scheduling decision framework; and a data sequence optimization module, which acquires diverse reporting pattern data of various business types from the conflict-resolved scheduling decision framework, performs time-series sequence analysis on the reporting pattern data, and determines... The system is divided into several modules: A first module determines whether the data processing meets the requirements for rapid adaptation to change and obtains an optimized data processing sequence; a second module, a forgetting compensation module, extracts historical fragments related to knowledge forgetting based on the optimized data processing sequence and re-injects these historical fragments into the framework using an incremental learning algorithm to determine the complete strategy after forgetting compensation; a third module, a scheduling adaptation model construction module, obtains the latency requirements of sensor data in the industrial park from the complete strategy after forgetting compensation, and fuses them with video surveillance experience through a federated learning algorithm to obtain the final scheduling adaptation model; a fourth module, a real-time verification and stable output module, performs real-time verification of the final scheduling adaptation model to determine whether the increase in new data causes time delays, and if so, adjusts the model parameters to obtain a stable transmission scheduling output; and a fifth module, a resource allocation optimization module, extracts data transmission performance indicators from the stable transmission scheduling output, uses a gradient descent algorithm to optimize the performance indicators and the constraints of limited network resources, and determines the resource allocation scheme.

[0019] In the above implementation, the system uses a multimodal large model as its core. By continuously learning and dynamically adjusting the data arriving in the communication scheduling environment, it achieves intelligent and adaptive scheduling communication. The data acquisition and fusion module first collects multimodal information, including text scheduling instructions, sensor values, and video feature vectors, and inputs it along with historical scheduling experience data into the incremental learning algorithm. This allows the model to update its strategy without disrupting its existing knowledge structure, forming a preliminary fused strategy model. Subsequently, the experience retention module extracts temporal dependency features reflecting business evolution trends from this strategy model and performs distributed integration across multiple data sources using a federated learning algorithm. This enhances the model's ability to remember new modal information while protecting the privacy of each data source.

[0020] When conflicts arise between different modalities regarding scheduling priority, latency requirements, or resource usage predictions, the decision-making module detects the conflicting parts and uses the gradient descent algorithm to iteratively correct the relevant parameters, gradually reducing the value of the conflict loss function, ultimately forming a conflict-resolved scheduling decision framework. Based on this, the data sequence optimization module performs time-series analysis on the data reporting patterns reflecting different business types within the framework. By comparing the predicted response speed with a preset rapid adaptation threshold, it filters and reconstructs a better data processing order, obtaining an optimized data processing sequence.

[0021] As the system operates for an extended period, the model may experience knowledge forgetting, leading to a decline in its ability to respond to early scheduling knowledge. The forgetting compensation module analyzes the optimized data processing sequence, identifies historical scheduling segments related to the forgetting phenomenon, and uses an incremental learning algorithm to reinject these historical segments into the current framework, thus forming a complete strategy after forgetting compensation. The scheduling adaptation model construction module further extracts the latency requirement features corresponding to the sensor data in the industrial park from the complete strategy and integrates them with the scheduling experience of video surveillance services through a federated learning algorithm to construct a scheduling adaptation model that can simultaneously meet the needs of multiple types of services. Finally, the real-time verification and stable output module continuously monitors the impact of new data growth on system latency and adjusts model parameters when anomalies are detected to output stable scheduling results. The resource allocation optimization module, based on this, comprehensively considers performance indicators such as throughput, latency, and packet loss rate, and uses the gradient descent algorithm to solve for the optimal resource allocation scheme under the constraint of limited network resources.

[0022] The diverse reporting patterns cover various service types within the communication scheduling environment, including scheduling tasks recorded in historical scheduling logs, transmission and processing of sensor data from industrial parks, data streams generated by video surveillance systems, verification data for real-time validation of scheduling adaptation models, and data transmission performance metrics such as throughput and packet loss rate involved in stable transmission output. Through time-series analysis and eigenvector matrix processing, these service types enable optimization of scheduling strategies and resource allocation for multimodal large-scale models, ensuring the system can quickly adapt to service changes and achieve efficient and stable data transmission and scheduling management.

[0023] The data acquisition and fusion module acquires modal information accompanying newly arriving data and historical scheduling experience data from the communication scheduling environment. It processes the data using an incremental learning algorithm to update the scheduling strategy, obtaining a preliminary fused strategy model. This includes acquiring newly arriving multimodal data streams and historical scheduling logs from the communication scheduling environment; processing these data streams and logs using multidimensional feature mapping technology to obtain a spatiotemporal feature vector sequence; inputting this sequence into a deep neural network to generate a high-dimensional state representation matrix; if the high-dimensional state representation matrix satisfies the drift condition, calculating the update gradient of the strategy network weights based on the high-dimensional state representation matrix to obtain the updated strategy parameter set; and finally, using an adaptive weighted aggregation algorithm to fuse the updated strategy parameter set with the global baseline model parameters to obtain the preliminary fused strategy model.

[0024] In this implementation, newly arriving multimodal data streams and historical scheduling logs are first received from the communication scheduling environment, and the two types of data are aligned using the same time base. The time base uses the scheduling period (which is determined by the dequeue period of the communication scheduling platform, and is the smallest feasible period within the stable operating range after statistical analysis of actual link latency jitter and service arrival rate, typically ranging from 10 milliseconds to 1000 milliseconds) as the alignment granularity; during alignment, a sliding time window is used (the length of this time window is determined by the correlation length of the service status in the time dimension, and is determined by statistical analysis of the duration corresponding to the autocorrelation decay of key indicators in the historical scheduling logs to a low correlation level, commonly ranging from 1 second to 120 seconds), and a sliding step size is set (this step size is determined by real-time requirements, set as an integer multiple of the scheduling period, and takes the minimum value that meets the computing power budget). Spatial dimension alignment adopts regional coding (the regional division scale is determined by the distribution density of equipment in the industrial park and the base station coverage unit. First, a basic area is established according to the coverage unit, and then the high-density business area is subdivided to the granularity that meets the positioning error requirements; the positioning error requirements are determined by the scheduling strategy for path loss and congestion location sensitivity, and the sensitivity is obtained through historical congestion event backtracking evaluation). After alignment, the system performs segmentation and semantic fragmentation on text-based scheduling information (segmentation rules are determined by scheduling instruction templates and historical instruction structure statistics), normalization and anomaly suppression on numerical network states (normalization upper and lower bounds are determined by the value range of historical stable periods; the anomaly suppression threshold is explained after its occurrence: this threshold is determined by the mean and standard deviation of statistical stable period indicators, the mean plus 3 times the standard deviation is taken as the alarm boundary, and fine-tuned in combination with the cost of false alarms), event type encoding and time-series sorting on log events (the event type set consists of events in historical logs whose frequency exceeds the minimum frequency threshold; this minimum frequency threshold is jointly determined by storage overhead and coverage requirements, and the coverage requirements are determined after evaluating the contribution of retained events to scheduling decisions to the target level), and frame rate tuning and key segment extraction on video information (the frame rate is determined by the constraints of link bandwidth and inference budget; the length of key segments is determined by the scene change rate, obtained by statistically analyzing the median interval of scene switching intervals and adjusting towards the conservative side to avoid missing sudden events). Missing data is handled using a timeout determination method (the timeout determination threshold is explained after it occurs: this threshold is set as an integer multiple of the scheduling cycle, and if the data is not received after more than 5 scheduling cycles, it is determined to be missing. The selection of 5 is based on the fact that this value covers the vast majority of normal delays and significantly reduces false positives under the historical link jitter distribution). Missing segments are handled using a combination of forward hold and short window interpolation (the interpolation window length is determined by the rate of change of the index, and the rate of change is determined by the statistics of the stable interval of the historical first-order difference).

[0025] After completing the above preprocessing, the system uses multidimensional feature mapping technology to generate a spatiotemporal feature vector sequence. Multidimensional feature mapping includes temporal location mapping, spatial location mapping, business attribute mapping, and modal semantic mapping. Temporal location mapping uses time slot numbers (numbers are generated incrementally from the starting point of the sliding time window, and the range of numbers is determined by the time window length and scheduling cycle). Spatial location mapping uses a regional encoding index (the index table is generated during the regional division phase and remains consistent throughout runtime to avoid encoding drift at the same location in different batches). Business attribute mapping uses a business type dictionary (the dictionary is determined by business classification rules, which are generated based on a combination of business latency level, reliability level, and bandwidth level; the level division boundary is determined by platform service level constraints). Modal semantic mapping generates semantic features for text fragments, log events, and video fragments respectively, and concatenates them with numerical states to form joint features at the same time. When concatenating joint features, modal weights are set (modal weights are explained after their introduction: modal weights are determined by the marginal contribution of each modality to the scheduling objective. The marginal contribution is obtained by observing the degradation of the scheduling objective index after masking a modality one by one on historical data. The greater the degradation, the higher the weight; at the same time, bandwidth occupancy penalties are added to avoid excessive weights for high-cost modalities that lead to a decrease in real-time performance), and an upper limit constraint is applied to the weights (the upper limit constraint threshold is explained after its introduction: this threshold is determined by the real-time inference latency budget. First, the average and peak latency of a single inference are measured, and then the peak latency is controlled within 80% of the latency budget, thereby inversely deriving the upper limit of the allowed feature dimensions and the corresponding upper limit of the weights). After completing the mapping and concatenation, the system outputs a spatiotemporal feature vector sequence in time slot order. The sequence length is consistent with the number of time slots within the time window. Each vector in the sequence contains spatial, service, and modal information within the same time slot, satisfying the input requirements of the subsequent deep neural network for the temporal context.

[0026] The spatiotemporal feature vector sequence is then input into a deep neural network to generate a high-dimensional state representation matrix. The structure of a deep neural network is determined by the input dimension, target complexity, and computational budget: number of network layers (the number of network layers is explained after the introduction of network layers: the number of network layers is determined by comparing the convergence stability and inference latency of different numbers of layers on historical data, and the minimum number of layers that makes the validation set scheduling index improve to the point of saturation is selected while meeting the inference latency budget), width of each layer (the width of each layer is explained after the introduction of each layer width: the width of each layer is searched in the range of 1 to 4 times the input dimension, and the minimum width is selected when the memory usage does not exceed the upper limit and the overfitting risk is controlled; the overfitting risk is judged by the difference between the training period and the validation period index exceeding a threshold, and the threshold is determined based on the generalization error tolerance of historical tasks), activation method (the activation method is determined by the gradient stability and convergence speed evaluation, and the method that converges more smoothly under the same budget is selected), regularization strength (the regularization strength is explained after the introduction of regularization strength: the regularization strength is determined by monitoring the volatility of the strategy output during the validation period. If the volatility exceeds the threshold, the regularization is strengthened. The threshold is determined by the allowable jitter range of the scheduling strategy output, and this range is derived from the business latency level constraint). The generation of the high-dimensional state representation matrix is ​​completed through hierarchical forward propagation: each layer performs linear combination and nonlinear transformation on the input vector to obtain an intermediate representation, which is then fused with time slot index information to preserve temporal positional relationships. When processing sequences, the network introduces context aggregation for the intermediate representations of adjacent time slots, so that the state representation simultaneously includes information at the current moment and recent evolutionary trends. The final output high-dimensional state representation matrix is ​​arranged by time slots, with matrix rows corresponding to time slots and matrix columns corresponding to abstract state feature dimensions. The number of feature dimensions is jointly determined by the inference latency budget and historical policy performance, ensuring sufficient expressive power within real-time constraints.

[0027] After the high-dimensional state representation matrix is ​​generated, the system performs a drift condition determination to decide whether to trigger a policy update. The drift determination is based on comparing the distribution differences between the current window state representation and the baseline window state representation: the system first extracts a statistical summary of the high-dimensional state representation matrix from the current time window, including the mean, variance, and change magnitude of each feature dimension, and then extracts a similar summary from a historically stable period or the most recently confirmed stable window as a baseline summary; subsequently, the difference is calculated, which is synthesized from mean offset, variance offset, and change magnitude offset according to preset weights. The difference weights are explained after their appearance: the weights are determined by business-sensitive items; for latency-sensitive businesses, the change magnitude offset weight is increased; for capacity-sensitive businesses, the mean offset weight is increased; and for stability-sensitive businesses, the variance offset weight is increased; sensitive items are jointly determined by platform business configuration and historical fault review. The drift condition threshold, as explained after its occurrence, is determined by the statistical distribution of historical stable-period differences. First, the mean and standard deviation of these stable-period differences are calculated, then the mean plus three times the standard deviation is used as the initial value. Subsequently, verification is performed on historical change event samples. If the false negative rate exceeds the allowable upper limit, the threshold is lowered; if the false positive rate exceeds the allowable upper limit, the threshold is raised. The allowable upper limit is determined by the assessment of operational response capabilities and the cost of misscheduling. To avoid short-term noise triggering drift, the system sets a consecutive trigger count threshold (as explained after its occurrence, this threshold is counted according to the scheduling cycle, taking a value between 3 and 10 times. Replay verification selects the optimal value between suppressing noise triggering and maintaining response speed; a smaller value is used when business changes are frequent to speed up response, and a larger value is used when link jitter is significant to enhance robustness). When the difference exceeds the drift condition threshold and the consecutive trigger count reaches the set value, the system determines that the high-dimensional state representation matrix meets the drift condition and triggers the policy update process.

[0028] The policy update process calculates the update direction and magnitude of the policy network weights based on a high-dimensional state representation matrix, forming the updated policy parameter set. The system first defines a set of scheduling target indicators, including latency, packet loss, throughput, and fairness, and sets target weights for each indicator (the target weights are explained after their occurrence: target weights are determined by business level and regulatory constraints; higher latency levels result in higher latency weights, and higher reliability levels result in higher packet loss weights; resource stress correction is also introduced, increasing throughput weights and decreasing fairness weights for non-critical services when available bandwidth is below the bandwidth stress threshold; the bandwidth stress threshold is explained after its occurrence: this threshold is determined by historical bandwidth occupancy distribution and the occupancy levels corresponding to congestion events, selecting the occupancy level closest to the congestion inflection point as the threshold, and verifying through replay that the congestion early warning time meets the requirements). In each training batch, the system inputs the high-dimensional state representation matrix within the current window into the policy network to generate a probability distribution or scoring sequence of scheduling actions. A reference signal is then constructed from historical scheduling logs, mapped from the actual executed actions and their resulting metrics. The mapping process first normalizes the associated metric for each action, then synthesizes them into a single evaluation value according to the target weight, thus forming an error metric. The system performs backpropagation layer by layer on the error metric, calculating the contribution of each layer's weight to the error, obtaining the update direction and magnitude of each weight. The update magnitude is controlled by the step size, which is explained after its occurrence: the step size is determined through replay verification, starting with a smaller value and gradually increasing until the maximum value is selected that does not cause severe oscillations in the policy output and meets the upper limit of convergence speed requirements. Policy output oscillations are determined by the change in the action distribution of adjacent batches exceeding a threshold, which is determined by the allowable scheduling jitter range of the business. To prevent over-updates due to a single drift, the system sets an update magnitude cap (the update magnitude cap threshold is explained after its inception: this threshold is determined by the statistical distribution of single-step update magnitudes during training in historical stable periods, taking a safe value near its upper bound, and ensuring that the latency of critical business operations after the update does not exceed the constraint during replay of abrupt change scenarios). To improve real-time performance, the training batch size is determined by the number of samples within the window and the computational budget. The batch size is explained after its inception: the batch size is selected to be the minimum value that ensures the time consumed by a single update does not exceed 20% of the latency budget and that the gradient estimation variance is within a controllable range. Variance controllability is determined by the consistency threshold reached in the update direction of 5 consecutive batches, and the consistency threshold is determined by the policy stability requirements. After completing the above updates, the system summarizes the updated policy parameter set, which contains the latest values ​​of the weights and biases of each layer of the policy network.

[0029] After the updated strategy parameter set is generated, the system uses an adaptive weighted aggregation algorithm to fuse it with the global baseline model parameters to obtain a preliminary fused strategy model. Before fusion, the reliability of this update is calculated. Reliability is jointly determined by data freshness, drift intensity, and verification performance: Data freshness is measured by the latest time span covered by the current window sample. The freshness span is explained after it appears: This span is determined by the rhythm of business changes, which is determined by the statistical analysis of historical change event intervals. The shortest span that can cover one typical change is taken as the lower limit. Drift intensity is measured by the extent to which the difference exceeds the drift condition threshold. The larger the extent to which it exceeds the threshold, the higher the drift intensity. Verification performance is obtained through rapid replay verification. Replay verification uses a mixture of samples from the most recent stable window and the current drift window. The mixing ratio is explained after it appears: The mixing ratio is determined by the trade-off between the need to prevent forgetting and the need for adaptation speed. It is initially set at 1:1, and then the ratio that significantly improves the latency of key businesses and does not degrade performance during the stable period is selected in the historical task replay. The system maps credibility to fusion weights. The fusion weights, once defined, are limited to a range of 0.1 to 0.9 to avoid complete reliance on local updates or baseline parameters. The lower limit of 0.1 ensures sufficient update injection when drift is confirmed, while the upper limit of 0.9 limits the impact of local data deviations on the global picture. These limits are determined through historical replay verification, keeping performance degradation in false alarm scenarios within acceptable limits. The fusion process is performed parameter by parameter: for each parameter in the policy network, the updated parameter and the baseline parameter are first linearly synthesized according to the fusion weights, followed by an offset constraint check. The offset constraint threshold, once defined, is determined by the parameter variation range during the stable period. The upper bound of the parameter variation amplitude after multiple consecutive updates during the stable period is taken as the threshold. If the threshold is exceeded, the offset is pruned back to the threshold boundary to avoid abnormal jumps. After completing the fusion and constraint of all parameters, the system forms a preliminary fused policy model. This model maintains global stability while absorbing new patterns brought about by the current drift.

[0030] The experience retention module extracts temporal dependency features from the initially fused policy model, integrates these features with new modal information using a federated learning algorithm, and determines the updated experience retention module by: using a long short-term memory network to parse the initially fused policy model to obtain temporal dependency feature vectors; employing a multi-head attention mechanism to project the new modal information data stream onto the semantic feature space where the temporal dependency feature vectors reside, generating a multimodal collaborative feature matrix; calculating the federated local gradient update amount based on the multimodal collaborative feature matrix, and performing a safe weighted aggregation calculation on the federated local gradient update amount to generate a globally shared feature increment; mapping the globally shared feature increment to the experience memory network, and adjusting the weights of the experience memory network through an adaptive forgetting gating unit to determine the updated experience retention module.

[0031] In this implementation, the parsing objects are first determined and serialized. The parsing objects include the state representation sequence output by the policy model within a continuous scheduling cycle, the sequence of changes in action preferences over time, and the internal representation sequence related to service categories and link states. The scheduling cycle uses the actual dequeue cycle of the communication scheduling platform as a benchmark, and its value is determined by the minimum stable dequeue interval obtained from historical scheduling log statistics. The serialization time window length, as described after its occurrence, is determined by the duration required for the correlation of key indicators in historical logs to decay to a low correlation level, and a smaller value is chosen to improve response speed while meeting the real-time budget. The sliding step size, as described after its occurrence, is set as an integer multiple of the scheduling cycle, determined by both inference time and update frequency constraints, and selected as the smallest integer multiple without triggering latency exceedances. After completing the above settings, the parsing objects within each time window are arranged chronologically to form an input sequence, serving as a unified entry point for subsequent temporal dependency extraction.

[0032] After sequence construction is completed, the module uses a Long Short-Term Memory (LSTM) network to perform temporal dependency parsing on the input sequence, generating temporal dependency-related feature vectors. The number of layers in the LSM network is explained below: the number of layers is determined through replay verification, prioritizing the minimum number of layers that stabilizes key scheduling metrics while meeting budget constraints in inference time. The hidden state dimension is explained below: the hidden state dimension is determined based on the input feature dimension and business complexity, initially starting with 1 times the input feature dimension and gradually increasing, selecting the minimum dimension that avoids overfitting during the verification period and ensures inference time does not exceed budget. The parameter initialization range is explained below: the initialization range is determined by forward propagation observation of historical stable sequences to ensure that the gating output does not exhibit long-term saturation, thus reducing the risk of gradient vanishing. The calculation process for each time step is as follows: First, a linear combination is performed on the input features of the current time step and the hidden state of the previous time step. Then, the linear combination result is superimposed with the bias term to form the gating driving quantity. The gating driving quantity is transformed by a compression function to obtain three types of gating coefficients, which are used to filter the current input information, control the retention ratio of historical memory, and generate candidate memory content, respectively. The candidate memory content is obtained by the linear combination of the current input and the previous hidden state and then by the hyperbolic tangent compression function. Subsequently, the historical memory is scaled according to the retention gating coefficient, the candidate memory is scaled according to the input gating coefficient, and the two are superimposed element by element to obtain the memory state of the current time step. Then, the memory state of the current time step is transformed by the hyperbolic tangent compression function and multiplied element by element with the output gating coefficient to obtain the hidden state of the current time step. The above process proceeds step by step within the time window, resulting in a sequence of hidden states arranged by time. The temporal dependency feature vector is obtained by aggregating this hidden state sequence through the time dimension. The aggregation method is explained after the aggregation method appears: the aggregation method is determined based on the business latency sensitivity. For latency-sensitive businesses, a weighted aggregation that is more biased towards the recent time step is adopted. The weight decay rate is obtained by statistically analyzing the duration of the impact of historical events on the current decision.

[0033] After obtaining the temporal dependency feature vector, the module processes the new modal information data stream and performs semantic space alignment to generate a multimodal collaborative feature matrix. The new modal information data stream is first segmented according to the scheduling cycle. For each segment, intramodal feature representations are extracted. The feature representation dimension is explained as follows: the dimension is determined by the effective information density of the modal information and the inference budget. The contribution of different dimensions is evaluated in offline replay, and then the minimum dimension where the contribution tends to stabilize is selected. The missing value threshold is explained as follows: the threshold is set as an integer multiple of the scheduling cycle and is determined by statistically analyzing the upper bound of normal arrival delay under historical link jitter, striking a balance between misjudgment cost and omission cost. Subsequently, a multi-head attention mechanism is entered: the number of attention heads is explained as follows: the number of heads is determined through replay verification, prioritizing the maximum number of heads that tends to stabilize while ensuring the inference time meets the budget. The single-head projection dimension is explained as follows: the dimension is determined based on the hidden state dimension and the head distribution relationship, ensuring that the total dimension after concatenation is consistent with the input of subsequent fusion layers, and taking the minimum feasible value with the goal of reducing redundancy. The computation process for each attention head is as follows: A linear transformation is performed on the temporally dependent feature vector to obtain the query representation; linear transformations are then performed on the new modality segmented features to obtain key and value representations; similarity is calculated between the query representation and each key representation, using a method of summing dimension-wise products followed by scale normalization. The scale normalization factor, as explained below, is determined by the size of the single-head projection dimension and is used to control the numerical range of similarity, preventing excessive concentration of attention weights; the similarity results are then normalized to obtain attention weights. The normalized temperature parameter, as explained below, is determined through playback verification, selecting a value that allows the attention weight distribution to distinguish key segments without exhibiting extreme spikes; finally, the attention weights are used to perform a weighted summation on the value representations to obtain the convergence result for that head. After concatenating the convergence results of each head along the feature dimension, a fusion linear transformation is performed to align them to the semantic scale of the temporally dependent feature vector, forming a multimodal collaborative feature matrix arranged by time slices.

[0034] Based on the multimodal collaborative feature matrix, the module locally calculates the federated local gradient update at each participating node. Local training first determines the composition of the loss metric, which includes a semantic coherence term and a scheduling-related term. The semantic coherence term aims at the similarity and consistency between the multimodal collaborative feature matrix and the temporally dependent feature vectors. The calculation process involves aligning the same time slice, calculating the dimensionally absolute deviation of the differences between the two, and summing them to obtain the coherence error. The scheduling-related term aims at the explanatory power of the features for scheduling results. The calculation process involves inputting the multimodal collaborative feature matrix into the local empirical memory network front end to obtain a predicted representation of changes in key scheduling indicators, and then calculating the difference between this representation and the actual indicator changes in the corresponding time slice of the node's historical logs to obtain the correlation error. The weighting coefficients for the two errors are explained after their appearance: the weighting coefficients are determined by the business level; the higher the latency level, the higher the weight of the scheduling-related term; the more drastic the environmental changes, the higher the weight of the semantic coherence term. The weighting coefficients are adjusted through replay verification to ensure that the latency of key business operations does not exceed constraints and that model output fluctuations are controlled. After obtaining the loss metric, backpropagation is performed layer by layer on the network parameters locally: first, the partial derivatives of the loss metric with respect to the output layer parameters are calculated, and then the propagation is performed layer by layer according to the chain rule to obtain the local gradient of each layer's parameters; the local gradients are pruned, and the pruning threshold is explained after it appears: the pruning threshold is determined by statistically analyzing the gradient magnitude distribution during training in the stable period, taking a safe value near the upper bound of the stable period, and verifying in the abrupt replay scenario that it does not cause convergence stagnation; the update step size is explained after it appears: the update step size is searched step by step from small to large through replay verification, selecting the maximum value that does not cause violent oscillations in the policy output and meets the upper limit requirement for convergence speed. The oscillation criterion is that the change in the action preference distribution between adjacent updates exceeds the threshold, and the change threshold is determined according to the scheduling jitter range allowed by the business. After completing several local updates, the federated local gradient update amount is obtained, and the number of local steps is explained after it appears: the number of local steps is determined by the communication overhead and the model drift speed. The drift speed is obtained by statistically analyzing the frequency of consecutive exceedances of the difference amount. When the drift speed is high, a smaller number of local steps is taken to improve the global synchronization frequency.

[0035] After the local gradient update amount enters the safe weighted aggregation calculation stage, the aggregation weight of each node is generated first. The aggregation weight is synthesized from the data quality score, sample freshness score, and node stability score. The data quality score is calculated by statistically analyzing the missing rate and outlier rate within the local time window. The missing rate is obtained by counting the missing data, and the outlier rate is obtained by counting outlier indicators. The outlier threshold is explained as follows: the threshold is determined by the mean and standard deviation of the indicators during the stable period, with the mean plus three times the standard deviation used as the initial value, and fine-tuned based on the cost of false alarms. The sample freshness score is calculated by calculating the latest time span covered by the local time window. The shorter the span, the higher the freshness. The freshness span lower limit is explained as follows: the lower limit is determined by the pace of business changes, determined by statistically analyzing the intervals of historical change events, ensuring that the time window covers at least one typical change. The node stability score is calculated by statistically analyzing the consistency and amplitude fluctuation of the local gradient direction of the node's recent uploads. Low consistency or high fluctuation lowers the stability score. The fluctuation judgment threshold is explained as follows: the threshold is determined by the upper bound of the fluctuation of stable nodes during the stable period, and the suppression effect is verified in the replay of outlier nodes. After obtaining the aggregated weights, abnormal local update amount suppression is performed: First, the aggregated reference direction of the local gradient update amounts of all nodes is calculated, which is synthesized from the weighted directions; then, the deviation of each node's local update amount from the reference direction is calculated. If the deviation exceeds a deviation threshold, it is explained that this threshold is determined by the upper bound of the deviation amount during the historical stable period and adjusted towards the conservative side in conjunction with the risk assessment of anomalies; the local update amount exceeding the threshold is reduced by a reduction ratio, which is explained after the reduction ratio is determined by replay verification, with the goal of suppressing the impact of anomalies on the global direction while ensuring that the contribution of normal nodes is not excessively diluted. After suppression is completed, the local gradient update amounts of each node are weighted and summed according to the aggregated weights to obtain the global shared feature increment. The upper limit of the magnitude of the global shared feature increment is explained after the magnitude upper limit is determined by the global update magnitude distribution during the stable period, taking its safe upper bound, and verifying that key business indicators do not deteriorate under abrupt change scenarios.

[0036] After the global shared feature increment is generated, the module maps it to the experience memory network and adjusts the weights of the experience memory network through an adaptive forgetting gating unit to complete the update of the experience retention module. Before mapping, the memory slot organization method of the experience memory network is determined. The memory slots form a three-dimensional index structure based on service type, link status level, and time slice index. The service type division boundary is explained after it appears: the boundary is determined by the platform service level configuration and consistency verification is performed in conjunction with the distribution of service indicators in historical scheduling logs. The link status level division threshold is explained after it appears: the threshold is determined by aligning the distribution of link indicators during the stable period with the congestion event trigger point, selecting the indicator level closest to the congestion inflection point as the threshold and performing replay verification. After the index is determined, the global shared feature increment is routed to the parameter subset associated with the corresponding slot according to the index. The suggested update value is calculated for this parameter subset. The suggested value is jointly determined by the increment magnitude and the historical stability of the slot. The historical stability of the slot is explained after it appears: the stability is obtained by statistically analyzing the change magnitude of the slot parameters in multiple consecutive rounds of updates during the stable period. A smaller change magnitude indicates higher stability. The adaptive forgetting gating unit then calculates the retention and forgetting coefficients: first, it calculates the similarity between the new addition and the existing memory representation of the slot. The similarity calculation is obtained by summing the dimension-wise products and then normalizing. The similarity threshold is explained after it appears: the threshold is determined by the similarity distribution during the stable period, taking a safe value near the lower bound of the stable period, and verifying its ability to trigger forgetting enhancement in a timely manner during scenario replay. When the similarity is higher than the similarity threshold, the retention coefficient is increased and the forgetting coefficient is decreased, so that slot parameter updates are mainly based on steady-state inheritance. When the similarity is lower than the similarity threshold, the retention coefficient is decreased and the forgetting coefficient is increased, so that slot parameter updates are mainly based on the absorption of new patterns. The gating output is also constrained by the upper limit of the update rate. The upper limit threshold is explained after it appears: the threshold is determined by the upper bound of the parameter update rate during the stable period, combined with real-time budget verification, to avoid scheduling output fluctuations caused by short-term strong updates. After gating adjustment, the experience memory network performs parameter updates weighted by the gating coefficient on the corresponding parameter subset. The updated network is more stable in extracting scheduling dependencies for the same service type and link state in subsequent time windows.

[0037] The decision-making construction module, if the updated experience retention module detects modal discrepancy conflicts, adjusts the parameters of the conflicting part using the gradient descent algorithm to obtain the scheduling decision framework after conflict resolution. This includes acquiring multimodal feature distribution data within the updated experience retention module and calculating the Mahalanobis distance between different modal feature vectors. If the Mahalanobis distance exceeds a preset conflict detection threshold, a modal discrepancy conflict is determined to exist, and a joint loss function is constructed based on the modal discrepancy conflict to calculate the loss value. The loss value is input into the backpropagation computation graph to generate a gradient vector, and the weight parameters are adjusted along the negative direction of the gradient vector using the gradient descent algorithm to obtain a converged corrected parameter set. The internal mapping logic of the experience retention module is reorganized according to the corrected parameter set to generate the scheduling decision framework after conflict resolution.

[0038] In this implementation, the multimodal feature distribution data related to scheduling decisions is first extracted and organized to form a computable distribution description. During extraction, the continuous running data is segmented using the feature sampling time window length (determined by statistically analyzing the correlation decay of key performance indicators and service arrival sequences over time in historical scheduling logs, taking the duration corresponding to the low correlation state as the lower bound, and then compressing the time window to the minimum value that does not trigger latency exceedances in conjunction with online inference budgets); at the same time, a sampling step size is set (determined by setting the sampling step size as an integer multiple of the scheduling cycle, and selecting the smallest integer multiple that satisfies the budget constraint by online measurement of the time required for a single distribution update). This ensures that the distribution estimation is aligned with the scheduling rhythm. For each modality, a set of feature vector samples is collected within each time window. Anomaly suppression is first performed to improve the stability of distribution estimation. Anomaly suppression threshold (determined by statistically analyzing the norm distribution of the feature vectors of the modality during the stable operating interval, calculating the mean and standard deviation, and then taking the mean plus three times the standard deviation as the initial value, and then fine-tuning it in combination with the cost of misjudgment and the cost of missed judgment) is used to determine whether the sample deviates from the central trend. Samples exceeding the threshold are given less weight in statistical calculation, thereby reducing the pulling effect of abnormal samples on subsequent covariance estimation.

[0039] After sample cleaning, the module calculates the mean vector and covariance matrix for each modality. The mean vector is calculated as follows: the summation of each feature dimension within the time window is divided by the sample size to obtain the mean for that dimension. The mean vector is then formed by combining the mean values ​​in their original order. A lower limit on the sample size (determined based on the feature dimension size and the stability requirements of covariance estimation; first, observe the decreasing trend of covariance element fluctuations as the sample size increases in the replay data; take the minimum sample size when the fluctuation enters the stable range; and verify that the computation time corresponding to this number meets the online budget) is used to avoid unreliable distributions caused by small samples. The covariance matrix is ​​calculated as follows: the mean vector is subtracted from each sample vector dimension by dimension to obtain a centered vector. The products of each dimension of the centered vector are then summed, and normalized according to the sample size and unbiased correction rules to obtain the covariance matrix. To avoid numerical ill-conditioned covariance matrix that could lead to unstable distance calculations, the module performs numerical stabilization on the covariance matrix. The stabilization strength parameter (determined by evaluating the condition number change of the covariance matrix on the stable data and selecting the minimum strength that makes the condition number fall within a controllable range and prevents abnormal fluctuations in the distance calculation results) is used to superimpose small positive values ​​in the diagonal direction, thereby ensuring the matrix has reversibility and suppressing the amplification of small noise.

[0040] After obtaining the mean vector and covariance matrix of each modality, the module begins to calculate the Mahalanobis distance between the feature vectors of different modalities to measure the intensity of modal difference conflicts. Before the distance calculation, alignment filtering is performed, using alignment granularity (the alignment granularity is determined by using the memory slot index of the experience retention module as a benchmark, treating data of the same business type, the same link status level, and the same time slice as the same semantic scenario, and dividing the alignment granularity according to the scheduling cycle to ensure that the difference measurement reflects the modal differences under the same scheduling semantics) to select paired samples. For each pair of modalities, the mean difference vector is calculated first. The calculation process of the mean difference vector is as follows: the mean vectors of the two modalities are subtracted item by item along the dimension to obtain the difference sequence, and then the difference sequence is assembled into a difference vector in dimensional order. Subsequently, a scaling metric baseline matrix is ​​constructed. This matrix is ​​synthesized from the covariance matrices of two modes according to weights. The synthesized weights are modal reliability weights (determined by masking a specific mode in the replay data during the stable period and observing the degradation of key scheduling indicators; the greater the degradation, the higher the contribution of that mode; simultaneously, a penalty is applied based on the sampling missing rate of that mode; the higher the missing rate, the lower the weight; finally, the contribution and penalty are combined and normalized to obtain the final weight). This allows the scaling metric to focus more on the statistical structure of reliable modes. After the scaling metric baseline matrix is ​​formed, it is inverted to obtain the inverse matrix. Before inversion, dimensionality reduction or regularization is performed to reduce the impact of high-dimensional noise. Dimensionality reduction preserves dimensions (determined by statistically analyzing the explanatory power of each main change direction on the total change in the stable period data, and selecting the smallest dimension that fully covers the main change directions and meets the budget for online computation time). The calculation process of Mahalanobis distance then unfolds as follows: First, matrix multiplication is performed between the mean difference vector and the inverse matrix to obtain a standardized difference vector. Then, the standardized difference vector is multiplied by the mean difference vector to obtain a single distance value. The dot product process involves multiplying dimension by dimension and then summing the results to obtain the difference measure of the modality pair. The module then summarizes the distance values ​​obtained from multiple modality pairs according to business weights. The business weights (determined by setting them based on business latency and reliability levels; higher latency levels result in greater weights, and higher reliability levels result in greater weights, with fine-tuning based on resource stress, which is assessed by the ratio of available bandwidth to occupied bandwidth; a lower ratio results in increased weights for critical business operations) are used to improve the sensitivity of critical business conflict identification.

[0041] When the aggregated Mahalanobis distance exceeds the conflict detection threshold, the module determines that a modal difference conflict exists and initiates the conflict resolution process. The conflict detection threshold (determined by establishing a reference distribution of Mahalanobis distance on historical data during the stable period, selecting a safe value near the upper bound of the reference distribution as the initial value; subsequently, verification is performed on historical change events and fault samples. If a missed detection leads to excessive latency in critical business operations, the threshold is lowered; if a false detection leads to frequent unnecessary adjustments, the threshold is raised. The adjustment range is constrained by the acceptable adjustment frequency and performance fluctuation limit of operations and maintenance) is used to distinguish between acceptable differences and conflicting differences. After determining a conflict, the module constructs a joint loss based on the conflict scenario and calculates the loss value. The joint loss consists of a consistency constraint term and a task performance constraint term. The calculation process for the consistency constraint is as follows: The absolute values ​​of the standardized difference vectors of conflicting mode pairs are taken dimension by dimension and summed to obtain the consistency deviation. This deviation is then amplified according to the conflict intensity. The conflict intensity (determined by using the difference between the current Mahalanobis distance and the conflict detection threshold as the over-threshold amplitude; a larger over-threshold amplitude indicates higher conflict intensity, and an upper limit constraint is imposed to prevent excessive amplification; the upper limit is jointly determined by the upper bound of the stable period distance fluctuation and the acceptable adjustment amount for sudden scenarios) reflects the severity of the conflict. The calculation process for the task performance constraint is as follows: The scheduling output under the current parameters is mapped to a sequence of key indicators, including latency, packet loss, throughput, and fairness sequences. Differences are then calculated between these sequences and the target constraint sequence. The difference calculation uses the absolute values ​​of the time-by-time differences, which are then accumulated to obtain the indicator deviation. The target constraint sequence originates from the business service level configuration and the baseline output of the recent stable window. The baseline window length (determined by selecting the most recent continuous time period judged as stable, covering at least three business fluctuation cycles, and satisfying online access overhead constraints) ensures the representativeness of the baseline. The deviations of each indicator are then synthesized according to their respective weights. The indicator weights (determined based on business level: higher latency levels result in higher latency weights, higher reliability levels result in higher packet loss weights, and during periods of bandwidth constraints, throughput weights are increased while fairness weights for non-critical services are decreased; the bandwidth constraint threshold is determined by the occupancy level corresponding to historical congestion inflection points and verified through playback) are used to balance the emphasis on conflict resolution and performance maintenance. Finally, the consistency constraint and task performance constraint are added together according to their composite weights to obtain the loss value. The composite weights (determined monotonically increasing with conflict intensity; higher conflict intensity results in higher consistency constraint weights; upper and lower limits are set for the composite weights to avoid single-item dominance, determined by the performance fluctuation tolerance during the stable period) are used to establish a controllable optimization objective between alignment mode and maintaining task indicators.

[0042] After the loss value is generated, the module constructs a backpropagation calculation path and generates gradient vectors. Then, gradient descent is used to adjust the conflict-related weight parameters to obtain a converged set of corrected parameters. The construction process of the calculation path is as follows: The parameter set involved in conflict modality projection, cross-modal fusion, and gating adjustment in the experience preservation module is identified and marked as updatable parameters. Parameters in non-conflict paths remain unchanged to reduce spillover effects. The generation process of the gradient vector is as follows: Starting from the loss value, the sensitivity of the loss to each updatable parameter is calculated layer by layer. The sensitivity calculation uses a chain-like propagation: first, the impact of the output layer on the loss is calculated, and then the impact is propagated back to the input layer by layer until the gradient component of each parameter is obtained. After the gradient components are generated, gradient pruning is performed. The pruning threshold (determined by statistically analyzing the upper bound of the gradient magnitude during stable training replay, taking a safe value near the upper bound as the initial value, and verifying in the replay of sudden conflict samples that the loss can still decrease continuously after pruning without update stagnation) is used to suppress update jumps caused by abnormal gradients. The gradient descent step size (determined by starting with a small value and gradually increasing it through replay verification, selecting the maximum value when the loss continues to decrease and the scheduling output fluctuation does not exceed the allowable range, which is determined by the upper limit of critical business latency jitter) is used to control the magnitude of each parameter adjustment. Simultaneously, a step size decay coefficient (determined by triggering decay when the loss decreases more slowly after several consecutive iterations, with the decay rate determined through replay verification to make convergence smoother and reduce final-stage oscillations) is used to improve convergence stability. Iteration stopping employs a dual constraint of a convergence threshold and a maximum number of iterations. The convergence threshold (determined by estimating the noise level of loss calculation during stable-period replay, taking the minimum improvement amount above the noise level as the threshold, and determining convergence if the improvement amount is below this threshold for two consecutive iterations) is used to avoid invalid iterations. The maximum number of iterations (determined by working backward from the average time of a single iteration and the online latency budget to ensure that the conflict resolution process is completed within the allowable time window, and reserving 20% ​​of the budget for subsequent reorganization and verification) is used to constrain online latency. During the iteration process, each update adjusts the updatable parameters in the opposite direction of the gradient. After the update is completed, the loss value is recalculated and the next round is entered until the stopping condition is met and the corrected parameter set after convergence is output.

[0043] After obtaining the corrected parameter set, the module reorganizes the internal mapping logic of the module based on the experience retention based on the corrected parameter set, thus forming the scheduling decision framework after conflict resolution. The reorganization process first applies to the conflict mode projection mapping, recalculating the mapping result of the mode in the semantic space according to the corrected projection parameters, so that it is consistent with the central region of the temporal dependency representation; then it applies to the fusion layer, adjusting the fusion weight to the same contribution ratio as the mode reliability weight, so that the reliable mode has a stronger influence on the fusion result, and the influence of the unreliable mode is suppressed; then it applies to the gating adjustment layer, recalculating the retention coefficient and suppression coefficient according to the corrected gating parameters. The retention coefficient is used to enhance the continuity of stable dependency features in the conflict scenario, the suppression coefficient is used to reduce the interference of conflict modes on the selection of scheduling actions, and the upper limit threshold of the gating adjustment amplitude (the upper limit threshold of the gating adjustment amplitude is determined by: statistically analyzing the upper bound of the natural fluctuation of the gating coefficient during the stable period, and verifying in the conflict replay that the upper limit constraint can still reduce the Mahalanobis distance and keep the key indicators from deteriorating) is used to prevent gating mutations from causing output oscillations. After completing the mapping logic reorganization of the three parts of projection, fusion and gating, the scheduling decision framework forms a consistent semantic alignment result when inputting multimodal information. The features after the conflict is resolved are used to drive scheduling action reasoning.

[0044] The data sequence optimization module extracts diverse reporting patterns from the conflict-resolved scheduling decision framework. It performs time-series analysis on this data to determine if it meets the requirements for rapid adaptation to change, resulting in an optimized data processing sequence. This includes extracting historical scheduling logs from the conflict-resolved scheduling decision framework to construct an original reporting time-series set reflecting the dynamic attributes of the business; generating a feature vector matrix from the original reporting time-series set and calculating an adaptability score based on the feature vector matrix; identifying blocking nodes causing response delays if the adaptability score is below a threshold; resetting queue priorities based on the blocking nodes and dynamically scaling the processing window according to the queue priorities; and rearranging the data flow order according to the processing window to obtain an optimized data processing sequence that meets the requirements for adaptability to change.

[0045] In this implementation, the output of the conflict-resolved scheduling decision framework is used as the entry point. Historical scheduling logs are extracted and organized into analyzable time-series data. The truncation time window length for historical scheduling logs is defined as follows: this length is obtained by statistically analyzing the periodicity of service reporting intervals and the duration of sudden fluctuations in historical logs, taking the minimum length that "covers at least three typical fluctuation cycles and the online analysis time does not exceed the latency budget." The log sampling step size is defined as follows: this step size is set as an integer multiple of the scheduling cycle, selected by measuring the online time for single log segmentation, statistics, and scoring, and then choosing the smallest integer multiple that satisfies the online budget constraint. Log field extraction uses a fixed field mapping table, which is determined by the scheduling platform's log structure definition. The fields at least cover the reporting timestamp, service type identifier, enqueue time, dequeue time, processing time, link status summary, and scheduling action identifier. First, the time base of the extracted logs is unified by converting the timestamps generated by different clock sources into the same reference clock. The calibration offset is explained after it appears: the offset is obtained by aligning and comparing the recording time of the same event in multiple log sources, taking the median offset of multiple alignment results as the calibration value, and using progressive correction in subsequent windows to suppress drift.

[0046] After log extraction and alignment, a raw reporting time sequence set reflecting dynamic business attributes is constructed. The set is grouped by business type, with grouping rules explained below: Grouping rules are based on the type code in the business configuration table, determined by the platform's business registration information. If multiple aliases exist for the same business in different subsystems, they are merged based on the business primary key in the logs. For each business group, the reporting sequence is obtained by sorting the reports in ascending order of timestamp. A reporting interval sequence is calculated between adjacent reports, obtained by subtracting the previous timestamp from the next. A queuing time sequence is then calculated using the enqueue and dequeue times, obtained by subtracting the enqueue time from the dequeue time. Simultaneously, a processing time sequence is retained, obtained directly from the log fields and standardized to milliseconds. If a missing field is found, the missing field determination threshold will be explained after it is found: This threshold is set according to the scheduling cycle count. Fields that have not been completed after more than 5 scheduling cycles are marked as missing. The missing data handling strategy adopts "missing fields participate in statistics with reduced weight". The reduction coefficient will be explained after it is found: The reduction coefficient is determined through replay evaluation and is the maximum reduction magnitude that ensures that the fluctuation of the statistical results does not exceed the upper limit of the fluctuation during the stable period.

[0047] After the initial reporting time series set is formed, the module generates a feature vector matrix for this set and calculates the adaptability score accordingly. First, time slices are segmented. The time slice length, as defined after its definition, is constrained by both the rate of service change and the computational budget. Its lower bound is obtained by statistically analyzing the autocorrelation decay time of the service reporting arrival rate, and then the minimum length that satisfies the online budget is selected. The time slice sliding step, as defined after its definition, is either 1 or 1 / 2 times the time slice length, determined by comparing the score refresh frequency and computing power usage, prioritizing the smaller step that meets real-time requirements. For each service, statistical characteristics are calculated within each time slice. These characteristics include reporting arrival rate, reporting interval jitter, queue backlog, average queue time, maximum queue time, average processing time, processing time variance, and number of link state transitions. The calculation process for the reporting arrival rate is as follows: count the number of reported items within the time slice and divide by the time slice length to obtain the arrival rate per unit time; the calculation process for the reporting interval jitter is as follows: calculate the mean of the reporting interval sequence within the time slice, then calculate the absolute value of the deviation of each interval from the mean and take the mean as the jitter degree; the calculation process for the queue backlog is as follows: take the queue length count at the end of the time slice, and the queue length is derived from the cumulative difference between the enqueue count and the dequeue count; the calculation process for the average queuing time is as follows: add up the queuing time sequence within the time slice and divide by the number of items; the maximum queuing time is obtained by traversing the queuing time sequence within the time slice and taking the maximum value; the average processing time is also obtained by adding up the queuing time and dividing by the number of items; the calculation process for the variance of processing time is as follows: first calculate the average processing time, then sum the squares of the differences between each processing time and the mean, and finally divide by the number of items to obtain the variance; the calculation process for the number of link state switching is as follows: traverse the link state sequence in time order, and increment the count by 1 if the current state is different from the previous state. The above features are arranged in a fixed-dimensional order to form the feature vector for each business and each time slice, and then concatenated according to the business dimension and the time dimension to form a feature vector matrix.

[0048] After the feature vector matrix is ​​generated, it undergoes normalization. The normalization upper and lower bounds are defined as follows: These bounds are obtained through statistical analysis of historical data within the stable operating interval. The minimum and maximum observed values ​​of the feature within the stable interval are used as initial values, and then anomaly suppression is applied to robustly correct the upper and lower bounds. The anomaly suppression threshold is defined as follows: The threshold is calculated using the mean and standard deviation of the stable interval. The mean plus three times the standard deviation is used as the upper suppression point, and the mean minus three times the standard deviation is used as the lower suppression point. Observations exceeding the suppression point are truncated before normalization. The normalization calculation process involves subtracting the lower bound from each observation and then dividing by the difference between the upper and lower bounds to ensure that features of different dimensions fall within a uniform range. If the difference between the upper and lower bounds is too small, leading to numerical instability, a minimum difference protection threshold is defined as follows: This threshold is determined by the resolution of the feature and the measurement noise level within the stable interval. The minimum positive value higher than the noise amplitude is used; if the difference is lower than this threshold, the threshold is used to replace the difference.

[0049] The adaptability score is calculated based on the normalized feature vector matrix, and is a weighted composite of three components: response speed, stability, and resource consumption. The response speed calculation process involves monitoring the change in arrival rate between adjacent time slices. When the change exceeds a trigger threshold, a recovery timer is initiated. The trigger threshold is defined as follows: it is determined by the upper bound of the natural fluctuation of arrival rate within a stable interval, taking a safe value near the upper bound, and verifying timely triggering through burst load replay. The recovery timer terminates when the average queuing time falls below the stable upper limit. The stable upper limit is defined as follows: it is determined by the upper bound of the average queuing time within a stable interval, and verified against false triggers during peak business replay. A shorter recovery time results in a higher response speed score. The stability calculation process involves a weighted summation of queuing time jitter and processing time variance. Lower jitter and variance result in a higher stability score. The weights are explained as follows: weights are determined by business latency and reliability classifications. For businesses more sensitive to latency, the weight of queuing time jitter is increased; for businesses more sensitive to reliability, the weight of processing time variance is increased. Classification boundaries are determined by platform service policy configuration. Resource usage is calculated by statistically analyzing the processing load within a time slice. The processing load is obtained by dividing the total processing time within the time slice by the time slice length. Peak load is obtained by the maximum processing load over several consecutive time slices. When the peak load approaches the resource limit, the resource usage score is reduced. The resource limit is explained as follows: this limit is determined by the available computing power, available bandwidth, and number of scheduling threads of the processing node. The maximum stable processing volume per unit time is first measured during stable operation and then converted into the load limit. The composite weights of the three components are explained as follows: Composite weights are determined jointly by service priority and resource stress. Resource stress is assessed by the ratio of available bandwidth to occupied bandwidth. When the ratio is below the stress threshold, the resource occupancy weight is increased and the stability weight is appropriately reduced. The stress threshold is explained as follows: This threshold is determined by the bandwidth occupancy inflection point before historical congestion events, and replay verification confirms that adjustments can be triggered in advance. After completing the calculations for the three components, an adaptive score is obtained. The adaptive score threshold is explained as follows: This threshold is determined by the lower bound of the stable interval score distribution and the maximum allowable response time for critical services. A safe value near the lower bound of the stable interval is initially taken as the initial value, and then adjusted in replays of sudden scenarios to ensure that the response time of critical services meets the constraints and the frequency of false triggers remains within a controllable range.

[0050] When the adaptability score falls below the adaptability score threshold, the module enters the blocking node identification process. First, a topology mapping between queues and processing nodes is constructed, generated by the scheduling platform's queue and processing link configurations. For each queue, the backlog growth rate is calculated by subtracting the queue backlog from the backlog of two adjacent time slices and dividing by the time slice length. For each processing node, the service rate is calculated by counting the number of items processed by the node within a time slice and dividing by the time slice length. The arrival rate is calculated for each service as a reference. A duration threshold is used for blocking candidate determination. The duration threshold is explained after its occurrence: this threshold is determined by the upper bound of the duration of short-term fluctuations within a stable interval, taking a safe value near the upper bound to ensure that short-term jitter does not trigger blocking identification. Queues or nodes that meet the condition of "continuously positive backlog growth rate and continuously lower service rate than arrival rate" are marked as blocking candidates. The congestion contribution is then calculated as follows: The queuing time increment caused by each candidate within a time window is accumulated. This increment is obtained by subtracting the average queuing time of the baseline time slice from the average queuing time of the current time slice. The baseline time slice, as defined after its occurrence, is selected from the most recent time slice where the adaptability score is above a threshold and the load is stable, ensuring the baseline reflects a normal state. The accumulated increment of the candidate is then divided by the sum of the accumulated increments of all candidates to obtain the contribution. Candidates with a contribution exceeding the contribution threshold are identified as congested nodes. The contribution threshold, as defined after its occurrence, is determined through replay verification. Its goal is to control the number of congested nodes without overlooking major bottlenecks, avoiding frequent priority adjustments due to too many nodes.

[0051] Once a blocked node is identified, the module resets queue priorities and dynamically scales the processing window based on the blocked node. Queue priorities are expressed using a numerical scale, the range of which is explained after it appears: the range is determined by the number of priority queue layers supported by the scheduling platform, typically an integer range from 1 to 8; the initial priority is generated jointly by business latency classification, reliability classification, and regulatory constraints. The reset process first increases the priority of critical business queues, the increase amount being explained after it appears: the increase amount is determined by business latency constraints and the current risk of exceeding limits. The risk of exceeding limits is calculated as the ratio of the maximum queuing time to the maximum allowed queuing time for that business; the higher the ratio, the greater the increase. The increase amount is subject to an upper limit, which is determined by the priority scale range. Next, queues related to the blocked node's path are sorted according to their contribution to the blockage; the higher the contribution, the greater the priority increase. Non-critical queues with weak association with the blockage are de-prioritized to release processing share. After priority adjustment, the processing window is scaled according to priority. The processing window length is explained below: The processing window length is calculated based on the node's available processing capacity and queue priority distribution. First, the available processing capacity per unit time is calculated, then processing shares are allocated according to the priority weights of each queue, and these processing shares are converted into window length. The priority weights are obtained by monotonically mapping the priority values. The mapping slope is explained below: The slope is determined through replay verification, with the goal of significantly improving queuing time for critical services and controlling degradation for non-critical services by increasing priority. To suppress window jitter, a smoothing coefficient is used to progressively adjust the window length. The smoothing coefficient is explained below: The smoothing coefficient is determined by statistical analysis of load fluctuation amplitude within a stable range. Larger fluctuations result in a smaller smoothing coefficient to enhance stability, while smaller fluctuations result in a larger smoothing coefficient to improve response. The smoothed window length is obtained by a weighted sum of the previous round's window length and the target window length for this round, with the weighting ratio directly determined by the smoothing coefficient.

[0052] After the priority and processing window are updated, the module rearranges the data flow order according to the processing window to form an optimized data processing sequence. The rearrangement generates a processing plan in units of time slices: within each time slice, queues are first ordered from highest to lowest priority; within the same queue, a secondary sort is performed based on urgency. Urgency is determined by the remaining latency budget, calculated by subtracting the consumed latency from the maximum allowed end-to-end latency. The consumed latency is obtained from the cumulative duration from enqueue to the current moment; the smaller the remaining latency, the higher the urgency. When urgency is the same, queues are arranged according to the order of enqueue to maintain fairness. Cross-time slice connection rules are used to avoid share breaks for critical services during time slice switching. Connection rule parameters are determined through replay verification, with the goal of ensuring the maximum queuing time for critical services does not exceed its allowed limit and the number of processing plan switches remains within a controllable range. Under the premise of satisfying the connection rules, high-priority items not processed in the current time slice are carried over to the beginning of the next time slice. After completing the sorting within the time slice and the connection across time slices, a time-progressing flow sequence is formed. This sequence directly reflects the processing queue order, the order of items within the queue, and the processing window allocation results for each time slice, thus obtaining an optimized data processing sequence that meets the requirements of rapid adaptation to changes.

[0053] The forgetting compensation module extracts historical fragments related to knowledge forgetting based on the optimized data processing sequence, re-injects these historical fragments into the framework using an incremental learning algorithm, and determines the complete forgetting compensation strategy. This includes obtaining the optimized data processing sequence, labeling forgetting features based on backtesting results on the historical validation set, filtering historical log vectors based on forgetting features to obtain historical fragments related to knowledge forgetting, generating an incremental model using an online incremental learning algorithm driven by these historical fragments, calculating the forgetting compensation factor based on the generalization error of the incremental model, and mapping the forgetting compensation factor to the output layer of the incremental model to determine the complete forgetting compensation strategy.

[0054] In this implementation, the sequence is first structured and expanded to extract the queue priority order, processing window length, queue item sorting rules, and cross-time-slice connection rules corresponding to each time slice. These are then aligned and archived with the current service type set and link status level set. Alignment uses an alignment granularity determined by using the scheduling cycle as the minimum alignment unit. The time slice boundaries involved in the data processing sequence are converted to integer multiples of the scheduling cycle, ensuring that subsequent backtesting's enqueueing, dequeueing, and window switching occur under the same time base. The sequence coverage duration is determined by statistically analyzing the typical duration of recent service arrival rate changes and the convergence duration of congestion mitigation, taking the larger of the two as the lower bound, and then combining this with online access overhead to select the minimum coverage duration that meets the budget.

[0055] In this implementation, historical fragments related to knowledge forgetting refer to historical data sequences where the system heavily relies on its scheduling strategy during the introduction of new data and services, but which exhibit response delays or decreased processing capacity under the current strategy. Specifically, these historical fragments include service arrival records that triggered important scheduling decisions in the system, the corresponding processing queue status, processing window length, dequeue time, and processing time. By backtesting the historical validation set, time fragments exhibiting performance degradation, decision instability, or response delays under the current strategy can be identified, thereby determining which historical fragments are related to forgetting.

[0056] To quantify the phenomenon of knowledge forgetting, a backtesting indicator system can be established based on the optimized data processing sequence and historical validation set. Specific indicators include: task completion latency, queue waiting time, processing load change rate, cross-time-slice scheduling order deviation, and task completion success rate for historical segments. When a historical segment exhibits these indicators exceeding preset thresholds during backtesting, it indicates that the knowledge in that segment is not being fully utilized in the current strategy, and is therefore identified as related to knowledge forgetting. In this way, those skilled in the art can identify which historical data segments need to be reintroduced into the strategy framework to compensate for the forgetting effect.

[0057] Furthermore, in this embodiment, historical fragments related to knowledge forgetting can also be filtered by aligning them with the current set of service types and link status levels. Specifically, the queue priority order, processing window length, and cross-time-slice connection rules in the historical fragments are compared with the corresponding rules of the current scheduling strategy to identify historical fragments that cannot maintain the original effective strategy under new service conditions. These filtered fragments can then be used as input to the incremental learning algorithm in the forgetting compensation module, thereby generating an incremental model and adjusting the strategy output to effectively retain and compensate for historical experience.

[0058] After obtaining the structured sequence, the module performs backtesting based on the historical validation set and marks forgotten features. The historical validation set is determined by a historical validation set construction window, which is determined as follows: consecutive time periods containing typical peaks, typical troughs, and typical abnormal link states are selected from the historical scheduling logs to ensure coverage of the main forms of business and link fluctuations, while excluding time periods with system outages or severe data loss; the construction window length is the minimum length covering at least three typical fluctuation cycles, and the backtesting time does not exceed the online budget. Backtesting steps are used, and the backtesting steps are determined as follows: set to an integer multiple of the scheduling cycle; the time consumed by measuring the single-step queue evolution is calculated and compared with the latency budget, selecting the smallest integer multiple that meets the budget. During backtesting, the timeline is progressively advanced according to the backtesting steps: At each time step, the historical service arrival records corresponding to that time step are read, and the arrival items are placed into the corresponding queue according to the service type, with the enqueue time recorded; then, the historical link status records corresponding to that time step are read, and the link status level is mapped to constraint parameters of currently available bandwidth, packet loss risk, and latency jitter risk. The constraint parameter mapping table is determined by statistics from the stable operation period, specifically by establishing a correspondence between the link status level and the actual observed bandwidth usage, packet loss, and latency distribution. Subsequently, the queues are processed according to the optimized data processing sequence: first, queues are selected according to queue priority, and then the number of items that can be processed or the amount of processing time that can be consumed in this time step is determined by the processing window length of the queue. The determination of the processing window length here follows the output of the data sequence optimization module; for the processed items, the dequeue time is recorded and the processing time is accumulated. The processing time comes from the processing node capability records in the historical verification set or the processing time field of the historical log. After completing this time step, the queue length, accumulated waiting time, and accumulated processing load are updated, and the next time step is entered.

[0059] Backtesting outputs a sequence of key metrics and labels forgetting features accordingly. Key metrics include end-to-end latency, queuing latency, packet loss, throughput, and fairness. The end-to-end latency is calculated by subtracting the enqueue time from the dequeue time for each dequeued item to obtain the single-item latency, and then averaging the values ​​of all items within the same time step to obtain the time step latency. The queuing latency is calculated by subtracting the enqueue time from the start processing time for each item to obtain the waiting latency, and then averaging the values ​​to obtain the time step waiting latency. Packet loss is calculated by counting the number of items marked as dropped or timed out within the time step and dividing by the number of arriving items to obtain the time step packet loss level. The timeout threshold is explained after it is set: this threshold is jointly determined by the maximum allowable queuing time in the service level and the upper bound of the link latency jitter, with the sum of the two used as the initial value, and the false positives are verified to not exceed the allowable level during historical peak replays. Throughput is calculated by counting the number of successfully dequeued items within the time step and dividing by the time step duration to obtain the time step throughput. Fairness is calculated by performing a dispersion calculation on the processing share obtained by each service type within the time step; the greater the dispersion, the worse the fairness. The backtest metrics are then compared with historical benchmark performance. The historical benchmark performance is determined by a benchmark window, which is selected from consecutive segments in the historical validation set that are considered stable. Stability is determined using a stability threshold, which is calculated by taking a safe value near the upper bound of the natural fluctuations in end-to-end latency and packet loss during the stable period. If the value does not exceed this safe value for several consecutive time steps, the system is considered stable. The benchmark window length is the minimum length covering at least one typical fluctuation cycle. During the comparison calculation, the degradation amount is calculated for each metric. The degradation amount is calculated by subtracting the benchmark metric from the current backtest metric, taking the absolute value of the difference, and then summing it over time steps to obtain the segment-level degradation amount. The forgetting feature labeling threshold is explained after its appearance: This threshold is determined by the distribution of segment-level degradation amounts during the historical stable period. A safe value near the upper bound of the stable period is used as the initial value, and then constrained by the maximum allowable latency degradation and maximum packet loss degradation for critical business operations. If the initial value is lower than the lower constraint limit, the lower constraint limit is used; if the initial value is higher than the upper constraint limit, the upper constraint limit is used. If the fragment-level degradation exceeds the forgetting feature labeling threshold, the fragment is labeled as a forgotten fragment, and a forgetting feature is generated. The forgetting feature includes the service type, link status level, queue status feature summary, occurrence time range, and main degradation index category.

[0060] After completing the forgetting feature labeling, the module filters historical log vectors based on the forgetting features to form historical fragments related to knowledge forgetting. Historical log vectors are generated by historical log vector mapping rules, which are determined by assembling business type, link status level, queue length, arrival rate, waiting latency, processing time, scheduling actions, and resource allocation results into vectors in a fixed-dimensional order, and normalizing each dimension. The upper and lower bounds of normalization are obtained statistically from historical stable periods and anomaly suppression is performed. The filtering process consists of two steps: conditional filtering and sorting / truncation. During conditional filtering, matching vectors are retrieved from the historical log vector library based on the business type and link status level in the forgetting features, with the time range limited to the same seasonal interval or the same production shift interval as the forgotten fragment. The interval matching parameter is explained after its appearance: this parameter is determined by the industrial park's operational rhythm, by statistically analyzing the mean and variance differences in business arrival rates across different shifts. If the difference is significant, shift interval matching is used; if the difference is not significant, a more lenient time interval matching is used. During the sorting and extraction process, the similarity between each candidate vector and the forgotten fragment feature summary is calculated. The similarity calculation process involves subtracting the candidate vector from the forgotten summary vector dimension by dimension, taking the absolute value, and then summing the results to obtain the distance value. The smaller the distance value, the higher the similarity. Then, the top few vectors are selected from high to low similarity to form a historical fragment set. The upper limit for the number of selections is explained after the upper limit is set: this upper limit is derived from the maximum training time budget of online incremental learning to ensure that subsequent incremental updates do not trigger real-time issues. At the same time, a diversity constraint is set. The diversity constraint threshold is explained after the diversity constraint threshold is set: this threshold is determined by the distribution of distance values ​​between candidate vectors. If the distance value is too small, it indicates high repetition. If the repetition value exceeds the upper limit, some duplicate samples are discarded to ensure that the historical fragments cover different sub-forms of forgotten features.

[0061] After the historical fragment set is determined, the module uses historical fragment-driven online incremental learning to generate an incremental model. The incremental model initialization is derived from a parameter snapshot of the current policy model. The acquisition time of the parameter snapshot is consistent with the completion time of forgotten feature labeling to avoid version mismatch. Online incremental learning uses incremental update batch size, which is determined as follows: based on the computation time of a single forward and backward pass, combined with the online budget, the maximum batch size that meets the budget is selected. Simultaneously, the gradient direction is required to remain stable within three consecutive batches. The stability threshold, once established, is determined by the upper bound of gradient amplitude fluctuations during training replay in the stable period. Exceeding the upper bound is considered unstable, and the batch size is reduced. The incremental update step size, once established, is determined by progressively searching from small to large through replay verification, selecting the maximum value that continuously reduces the degradation of the corresponding indicator for the forgotten fragment while ensuring that the fluctuation of the current online key indicator does not exceed the allowable upper limit. The allowable upper limit is jointly determined by the upper limit of critical business latency jitter and the upper limit of packet loss. The upper limit for the number of incremental update rounds is explained as follows: This upper limit is derived from the online budget and is determined by multiplying the average time of a single update round by the number of rounds, which should not exceed 80% of the budget. 20% is reserved for error assessment and compensation mapping. During each update round, historical segments are fed into the model in batches. The model outputs scheduling action preferences and resource allocation tendencies, which are aligned with the best-performing benchmarks in the historical segments. The alignment targets consist of records with the best or near-best performance metrics in the historical segments. The benchmark selection threshold is explained as follows: This threshold is determined by the joint distribution of end-to-end latency and packet loss in the historical segments. Samples with both at low levels are selected as benchmarks. If the sample size is insufficient, the threshold is relaxed while prioritizing critical business operations. Model parameter updates are limited to the output channels related to forgetting features and a small set of upstream fusion parameters. The limited-range parameters are explained as follows: These parameters are determined through sensitivity analysis. Sensitivity analysis uses parameter-by-parameter perturbation to observe changes in the degradation of forgotten segments. Parameters with significant changes are included in the update set, while parameters with insignificant changes are excluded, ensuring that updates are concentrated on forgetting-related paths.

[0062] After the incremental model is generated, the module calculates the forgetting compensation factor based on the generalization error of the incremental model. The generalization error assessment uses a validation subset size, determined by extracting samples with the same business type and link status level as the forgetting characteristics from the historical validation set, adding a small number of non-forgotten samples for stability constraints, and limiting the total size to the maximum value that meets the budget. The generalization error is obtained by weighting multiple indicators. The indicator weights are explained after their appearance: weights are determined by the business service level; critical businesses have increased weights for latency and packet loss, while throughput is increased and non-critical fairness weights are decreased during resource stress. The resource stress threshold is determined by the bandwidth occupancy level corresponding to historical congestion inflection points and is verified through replay. The generalization error calculation process is as follows: the incremental model is run on each item in the validation subset to obtain the predicted scheduling actions and resource allocations, and the indicator sequence is replayed using the same queue evolution method; the indicator sequence is then compared with the baseline indicator sequence of the validation subset to obtain the difference sequence; the absolute value of the difference sequence is taken at each time step and accumulated to obtain the total error; finally, the average error is obtained by normalizing according to the number of samples in the validation subset. The calculation of the forgetting compensation factor uses a compensation mapping interval. The compensation mapping interval is determined as follows: a lower limit is set to prevent insufficient compensation from hindering the reduction of forgotten fragment errors; an upper limit is set to prevent excessive compensation from causing degradation of current online performance. Both limits are determined through historical playback verification, ensuring that key online indicators do not exceed allowable boundaries. The error ratio is obtained by comparing the average error with the average error during historical stable periods. A higher error ratio indicates that the forgetting compensation factor is closer to the upper limit of the compensation mapping interval, while a lower error ratio indicates that the forgetting compensation factor is closer to the lower limit. The mapping slope, once present, is explained as follows: this slope is determined through playback verification, and the goal is to ensure a smooth decrease in the compensation factor as the error decreases, avoiding drastic fluctuations in the compensation factor that could cause output instability.

[0063] After the forgetting compensation factor is determined, the module maps the forgetting compensation factor to the output layer of the incremental model to form a complete strategy after forgetting compensation. Output layer channel selection is based on forgetting features and the binding rules are explained after their occurrence: The business type and link status level in the forgetting features are used as indices to locate the action preference channel and resource allocation channel in the output layer responsible for the combination of these indices. The index table is generated by the channel definition during model training and remains consistent during runtime. The mapping process involves recalibrating the output confidence of the bound channel. The recalibration magnitude is determined by the forgetting compensation factor; a larger forgetting compensation factor results in a more significant increase in the confidence of the channel. Simultaneously, symmetrical suppression is applied to conflicting alternative action channels. The upper limit of the suppression magnitude is explained after its occurrence: This upper limit is determined by the natural fluctuation upper bound of the output channel during the stable period, and the suppression is verified in replays of sudden scenarios to ensure that it does not cause a sudden drop in throughput or queue starvation. After recalibration, the output enters the scheduling decision path to participate in action selection and resource allocation. Action selection uses a consistency verification threshold, which is determined as follows: during the stable period, the consistency of action selection for the same business type under similar link states is statistically analyzed, and the lower bound of the consistency level is taken as the threshold. When recalibration causes the consistency level to fall below the threshold, a slight rollback is triggered, and the forgetting compensation factor is adjusted down by a preset attenuation coefficient. The attenuation coefficient is determined by replay verification to ensure that the error reduction trend is not disrupted. After completing the output layer mapping and consistency verification, a complete strategy after forgetting compensation is formed.

[0064] The scheduling adaptation model construction module extracts the latency requirement features of industrial park sensor data from the complete strategy after forgetting compensation. It then fuses these features with video surveillance experience through a federated learning algorithm to obtain the final scheduling adaptation model. This includes parsing communication protocol messages extracted from the complete strategy after forgetting compensation to generate latency requirement features that quantify the urgency of sensor data transmission; constructing a monitoring experience vector reflecting the sensitivity of video surveillance services to network fluctuations based on the latency requirement features; inputting this monitoring experience vector into the federated learning framework to aggregate it into a global experience model; mapping the cross-regional network congestion prediction values ​​output by the global experience model to the latency requirement features to generate a heterogeneous fusion feature matrix characterizing the correlation between sensor data and video surveillance services; and using the heterogeneous fusion feature matrix to supervise and adjust a pre-set scheduling neural network to obtain the final scheduling adaptation model with multi-service collaborative processing capabilities.

[0065] In this implementation, the extraction, alignment, and parsing of communication protocol messages are first completed to form latency requirement characteristics for quantifying the urgency of sensor data transmission. Message extraction employs a message sampling time window length (determined by statistically analyzing the correlation between service arrival rate and end-to-end latency over time in the historical logs of sensor services in the industrial park, taking the duration corresponding to the low correlation state as the lower bound, and then combining this with the online parsing time budget to select the minimum length that meets the budget), and sets a sampling step size (determined by setting the sampling step size as an integer multiple of the scheduling cycle, and selecting the minimum integer multiple where the peak time does not exceed the latency budget by measuring the average and peak time of a single message extraction and field segmentation). After the packets are extracted, time base unification is performed first. The time calibration offset (determined by aligning and comparing the recording times of the same packet at multiple collection points, taking the median offset of the multiple alignment results as the calibration value, and using progressive correction in subsequent time windows to suppress clock drift) is used to correct timestamp differences between different links and nodes. Field parsing is performed based on the protocol field boundary table. The field boundary table (determined by establishing a fixed offset and length mapping according to the protocol implementation specifications of the campus communication system, and verifying that the offset of each field remains consistent across different device models through packet capture playback before going online) is used to decompose the packet into service type identifier, priority identifier, generation timestamp, sequence number, acknowledgment mechanism identifier, acknowledgment timeout parameter, historical retransmission count, payload length, and path identifier.After decomposition, latency requirement features are assembled according to the same business semantics, including: business latency limit (determined by: determining the business service level configuration, and verifying the consistency of the over-limit alarm threshold setting corresponding to the limit in historical peak logs), consumed latency (calculated by: subtracting the message generation timestamp from the current time to obtain the end-to-end consumed time), remaining available latency (calculated by: subtracting consumed latency from the business latency limit to obtain the remaining time; if a negative value occurs, it is treated as 0 to indicate an over-limit status), queuing waiting time (calculated by: subtracting the enqueue time from the current time to obtain the queue waiting time; the enqueue time is recorded by the queue system), and retransmission risk level (determined by: counting the number of acknowledgment timeouts and retransmissions for this business within the most recent consecutive time window). The confirmation timeout threshold is explained as follows: This threshold is jointly determined by the confirmation timeout parameter in the protocol implementation and the upper limit of link latency jitter. The confirmation timeout parameter is used as the initial value, and then the jitter upper limit is added to form the operating threshold. The retransmission risk classification threshold is explained as follows: The threshold is determined by the dividing point between the upper limit of the distribution of retransmissions during the stable period and the lower limit of the distribution of retransmissions during the fault period. The false alarm rate and the missed alarm rate are verified to be within a controllable range during playback. The data freshness level is determined as follows: The data freshness level is divided into segments based on the ratio of consumed latency to the upper limit of the service latency. The segment boundary is explained as follows: The boundary is determined by the maximum information staleness allowed by the service level. The target staleness upper limit is given first based on the sensitivity of the service to real-time performance, and then the control effect and false trigger frequency under this boundary are verified in historical data. The above elements are combined to form the time delay requirement characteristics, and dimensional unification and truncation protection are implemented for each component. The truncation upper limit is determined by: statistically analyzing the upper limit of the natural fluctuation of each component during the stable operation period, taking a safe value near the upper limit, and avoiding extreme values ​​from having a dominant effect on subsequent fusion.

[0066] After defining latency requirements, the module constructs a monitoring experience vector around the video surveillance service to characterize the video service's sensitivity to network fluctuations. This vector is then fed into a federated learning framework to form a global experience model. The monitoring experience vector consists of real-time bitrate requirements, keyframe intervals, buffer occupancy levels, frame drop tolerance, and stuttering sensitivity. The calculation process for real-time bitrate requirements is as follows: the average bitrate is obtained by accumulating the number of bytes transmitted within the monitoring stream's statistical time window and dividing it by the time window length. The statistical time window length is determined by combining the frequency of video encoding parameter changes and network fluctuation frequency, taking the minimum length that covers at least three typical bitrate fluctuations while satisfying online statistical overhead constraints. The calculation process for keyframe intervals is as follows: the video frame type sequence within the time window is traversed, the timestamp difference between adjacent keyframes is recorded to obtain the interval sequence, and the mean and maximum values ​​of the interval sequence are calculated as sensitivity inputs. The calculation process for buffer occupancy levels is as follows: the buffer depth monitoring value of the edge node or playback end is read, and the mean and minimum values ​​are calculated according to the time window. The buffer alarm threshold is defined after the occurrence of a buffer alarm threshold. This threshold is determined by the service... The maximum allowed stutter probability in the service level is derived by working backwards. First, a correspondence between buffer depth and stutter events is established in the playback, and then the buffer depth corresponding to the stutter inflection point is taken as the threshold. The method for determining frame drop tolerance is: based on the service level configuration, and combined with the correlation between "frame drop rate and subjective experience degradation" in historical experience logs, a correction is made, with an upper limit for the correction range (determined by the upper bound of the natural fluctuation of experience indicators during the stable period, to avoid excessive relaxation of tolerance). The method for determining stutter sensitivity is: statistically analyzing the mapping relationship between the duration of stutter events and alarm levels in historical monitoring experience logs to form a sensitivity level table, with a level boundary threshold. After the level boundary threshold appears, it is explained that the threshold is determined by the alarm policy requirements and its consistency is verified in multi-scenario playback. After completing the construction of monitoring experience vectors, each regional node generates a monitoring experience vector sequence locally and enters the federated learning framework for aggregation. The methods for determining the federation round period (based on the speed of network state changes, firstly, the average rise time and average relief time of link congestion events are statistically analyzed, and half of the smaller of the two is taken as the lower bound of the round period, and then the minimum period that meets the budget is selected in combination with the cross-node communication overhead), local update steps (based on the trade-off between upload overhead and model convergence speed, comparing the global model convergence curves under different synchronization numbers in the replay, and selecting the minimum number of steps that significantly improves the convergence speed under the premise of controllable communication frequency), and update step size (based on the method of gradually increasing the update step size from a small value, selecting the maximum value that is selected when the global indicators continue to improve and local updates do not cause output oscillations, and the oscillation judgment threshold is explained after the oscillation judgment threshold appears: the threshold is determined by the upper bound of the output fluctuation of the empirical vector during the stable period monitoring).Each node locally generates a predicted output using a monitoring experience vector as input and aligns it with local congestion observation labels. The congestion observation labels are constructed by extracting bandwidth usage, queue backlog, packet loss, and latency jitter from local link statistics, summarizing them by time slice, and marking time slices exceeding the congestion judgment threshold as congested. The congestion judgment threshold is defined as follows: the threshold is determined by the inflection point level of historical congestion events. First, the inflection point of bandwidth usage and queue backlog before the congestion event occurs is located; then, the level corresponding to the inflection point is taken as the threshold, and the early warning capability is verified through replay. Based on this, the node calculates the local error and generates a local gradient update. The local gradient update is pruned before uploading. The pruning threshold is determined by: statistically analyzing the upper bound of the gradient magnitude during stable period replay, taking a safe value near the upper bound as the initial value, and then verifying convergence after pruning during the replay of sudden congestion samples. The aggregation phase employs a secure weighted aggregation method. The aggregation weight is determined by a combination of node data freshness, regional coverage, and node stability. Freshness is determined by the latest coverage duration within the time window; coverage is determined by the topological centrality and traffic volume share of the node's region; and stability is determined by the volatility of historical uploaded updates. A volatility threshold is defined after a threshold is reached: the threshold is the upper bound of the volatility of stable nodes, and the suppression effect is verified through abnormal node replay. Updates that significantly deviate from the group's direction are reduced. A deviation threshold is determined by statistically analyzing the upper bound of deviation during the stable period, taking a safe value near the upper bound, and verifying its ability to suppress abnormal directions through abnormal data injection replay. After aggregation, a global empirical model is formed to output cross-regional network congestion prediction values.

[0067] In this implementation, to ensure that the video surveillance experience vector can be effectively fused with the latency requirements of the industrial park's sensor data, the dimensions of the vector are first standardized and unified. Specifically, for each dimension, such as real-time bitrate requirements, keyframe intervals, buffer usage levels, frame drop tolerance, and stuttering sensitivity, linear normalization or standard deviation normalization methods are used to map their values ​​to a standard range of 0 to 1. This ensures that features with different physical meanings and dimensions are comparable in subsequent fusion calculations and avoids a dominant effect of an excessively large value in any one dimension on the fusion result.

[0068] When performing federated learning algorithm fusion, a heterogeneous fusion feature matrix is ​​constructed, aligning the standardized monitoring experience vector with the latency requirement features according to the correspondence between business type and time slice. The fusion method includes: firstly, element-wise weighted combination of the monitoring experience vector of each time slice and the corresponding sensor latency requirement feature. The weights are determined based on historical performance playback results, with higher weights indicating a greater impact of that dimension on network fluctuation sensitivity; subsequently, principal component analysis (PCA) or normalized covariance matrix processing is performed on the weighted feature matrix to extract the main influencing directions, and the processing results are fed into the scheduling neural network for supervised adjustment and policy optimization.

[0069] Furthermore, to maintain the accuracy of cross-regional network predictions, the local monitoring experience vectors uploaded by each node are securely weighted and aggregated during the federated learning process. The aggregation weights comprehensively consider the freshness of node data, service coverage, and node stability. Vectors that are obviously abnormal or deviate are pruned or reduced to suppress the interference of abnormal data on the global model. Through these methods, it can be ensured that the fused feature matrix reflects both the sensitivity of video surveillance services to network fluctuations and the latency requirements of sensor data in industrial parks, thereby generating a final scheduling adaptation model with multi-service collaborative processing capabilities.

[0070] After generating cross-regional network congestion predictions, the module maps them to latency requirement features to obtain a heterogeneous fusion feature matrix characterizing the correlation between sensor data and video surveillance services. The regional division granularity (determined by establishing basic regions based on the coverage units of campus base stations and the convergence boundaries of switching nodes, then merging or subdividing these basic regions according to congestion propagation correlation; correlation is obtained through statistical analysis of the propagation delay and probability of historical congestion events in adjacent regions, aiming for stronger consistency of congestion characteristics within the same region) is used to define the index set for cross-regional predictions. The prediction lead time (determined by the shortest executable time from detection to adjustment in the scheduling system, obtained from the cumulative time consumed by online inference, distribution, and queue activation; the peak value of the cumulative time is taken as the lower bound, and then one scheduling cycle is added as a safety margin) is used to align the predictions to subsequent scheduling windows. The mapping process first performs time alignment, projecting the congestion prediction value of each region onto the corresponding scheduling time slice according to the prediction lead. Then, it performs service alignment, concatenating the sensor latency requirement features within the same region and time slice with the sensitivity components of the video surveillance experience vector to form a joint feature row. The dimensional order of the joint feature row adopts a fixed dimension table, which is determined by generating it based on the input structure definition of the subsequent scheduling neural network to ensure consistency in the meaning of each dimension during training and deployment. This is used to avoid dimensional mismatch. Subsequently, the joint features are normalized. The upper and lower bounds of normalization (determined by using the minimum and maximum observed values ​​of each feature component as initial values ​​during the stable period, and applying upper and lower bound suppression to suppress abnormal fluctuations; the suppression thresholds are the mean plus three standard deviations and the mean minus three standard deviations, verified through playback) are used to suppress the impact of extreme values. After alignment and normalization, the matrix rows are arranged by time slice, the matrix columns are arranged by feature dimension, and the matrix blocks are formed by region index, thus forming a heterogeneous fusion feature matrix. This matrix carries sensor urgency, video sensitivity, and congestion prior in the same representation space, realizing cross-service and cross-regional correlation expression.

[0071] After the heterogeneous fusion feature matrix is ​​formed, the module uses this matrix to supervise and adjust the pre-set scheduling neural network, resulting in a final scheduling adaptation model with multi-service collaborative processing capabilities. The input of the pre-set scheduling neural network receives the heterogeneous fusion feature matrix, and the output generates scheduling action selection and resource allocation suggestions. The supervision signal is composed of historical excellent scheduling records and service level targets. The excellent scheduling record screening threshold is defined as follows: this threshold is determined by the joint conditions of critical service latency not exceeding limits, packet loss not exceeding limits, low frequency of video stuttering events, and resource consumption not exceeding the safety line. The safety line is determined by the upper bound of resource load during the stable period. During screening, separate databases are built for different regions to avoid label offsetting caused by differences in capabilities between regions. The training batch size (determined by working backward from the online training time budget, selecting the largest batch that meets the budget, while requiring the loss reduction direction of three adjacent batches to be consistent to a stable requirement, the consistency judgment threshold is determined by playback during the stable period), training step size (determined by starting with a smaller value and gradually increasing, selecting the maximum value when the loss continues to decrease and the fluctuation of key online indicators does not exceed the allowable upper limit, the allowable upper limit is jointly determined by the upper limit of latency jitter of key sensor services and the upper limit of video stuttering tolerance), and maximum number of iterations (determined by working backward from the average time of a single iteration and the online budget, the budget usage ratio does not exceed 80%, and 20% is reserved for verification and rollback) are used to constrain the stability and timeliness of online updates. The calculation process for each iteration is as follows: the heterogeneous fusion feature matrix is ​​input into the network in batches to obtain action suggestions and resource suggestions; the action suggestions are compared with the good scheduling records to generate action deviations; the resource suggestions are mapped to the predicted results of latency, throughput, packet loss and video stuttering risk, and compared with the service level target to generate performance deviations; then the action deviations and performance deviations are combined into a supervision error according to weights. The weights are determined according to the business level, with increased weights for latency-related services for critical sensor services, increased weights for stuttering risk-related services for video surveillance services, increased weights for throughput-related services and decreased weights for non-critical fairness services when resources are scarce, and the threshold for determining resource scarcity is determined by the historical congestion inflection point level and verified by playback. The supervision error is used to generate the update direction of network parameters. During the update process, gradient pruning is performed. The pruning threshold (determined by statistically analyzing the upper bound of the gradient magnitude in the stable period replay and taking a safe value, and then verifying that the pruning still maintains a downward trend in the sudden congestion scenario replay) is used to suppress abnormal update jumps. At the same time, an early stopping threshold is set. The early stopping threshold (determined by estimating the computational noise level of the supervision error in the stable period, taking the minimum improvement amount higher than the noise level as the threshold, and stopping when the improvement amount is lower than the threshold for two consecutive iterations) is used to avoid invalid iterations.After completing the supervised adjustment, the final scheduling adaptation model is obtained. When the input includes a fusion matrix containing congestion prediction and business urgency, the model outputs scheduling decisions and resource allocation suggestions that can simultaneously take into account the low latency of sensors and the anti-fluctuation requirements of video surveillance.

[0072] The real-time verification and stable output module performs real-time verification on the final scheduling adaptation model, determining whether the increase in new data causes time delay issues. If so, it adjusts the model parameters to obtain a stable transmission scheduling output. This includes acquiring the real-time verification data stream generated by sensors in the industrial park, calculating the new data growth rate based on the timestamp difference of the real-time verification data stream, and mapping the new data growth rate to a preset time delay threshold to calculate the transmission delay deviation. The transmission delay deviation is input into the congestion prediction module to generate a congestion prediction residual. If the congestion prediction residual exceeds the allowable range, a gradient update vector is generated to modify the model weight values ​​of the scheduling adaptation model. The updated model weight values ​​are used to generate a stable transmission queue that matches the current data scale, and the stable transmission queue is parsed to extract the stable transmission scheduling output.

[0073] In this implementation, the real-time verification data stream generated by sensors in the industrial park is first acquired, and the data stream is then time-aligned, deduplicated, and its field integrity is checked to ensure that the subsequent calculations of growth rate and latency deviation have consistent statistical standards. The real-time verification data stream includes a generation timestamp, enqueue timestamp, dequeue timestamp, service type identifier, region identifier, load size, link status summary, and queue identifier. The sampling time window length is explained after its introduction: this length is determined jointly based on the data growth rate and the online computing budget. First, the typical duration required for historical data growth mutations to stabilize is statistically analyzed, and half of this duration is taken as the lower bound. Then, the minimum length that meets the budget is selected in conjunction with the online computing time. The sampling step size is explained after its introduction: the sampling step size is set as an integer multiple of the scheduling cycle. By measuring the peak time of a single statistical analysis, residual calculation, and queue reconstruction, the minimum integer multiple that does not trigger latency budget overruns is selected. Time alignment is based on the scheduling cycle, mapping the timestamp of each data item to the corresponding time slice. The mapping rule is to round down the timestamp to the nearest scheduling cycle boundary. If the same data arrives repeatedly, the deduplication key is specified after it appears. The deduplication key is determined by the combination of the message sequence number, the generation timestamp, and the device identifier. The combination method is determined based on the consistency check result of the protocol message fields, ensuring that the same data can be identified as the same event at different collection points.

[0074] After completing the data stream processing, the module calculates the new data growth rate based on the timestamp difference of the real-time verification data stream and uses the growth rate for subsequent latency threshold mapping. The growth rate statistics use an arrival count metric. The arrival count metric is defined as follows: it is based on enqueued events to avoid miscounting growth due to duplicate generation times caused by retransmissions; the counting object is limited to data entries verified through field integrity checks. The growth rate calculation process is as follows: The current arrival quantity is obtained by counting the number of enqueued entries within the current sampling time window; the previous arrival quantity is obtained by counting the number of enqueued entries within the previous sampling time window; the difference between the two is the arrival increment; and the arrival increment is divided by the sampling time window length to obtain the increase in arrivals per unit time, which is used as the new data growth rate. To suppress the amplification effect of short-term spikes on the growth rate, the growth rate smoothing coefficient is defined as follows: the smoothing coefficient is determined by the upper bound of the natural fluctuation of the growth rate during the stable period; the larger the fluctuation, the smaller the smoothing coefficient to enhance stability; the smaller the fluctuation, the larger the smoothing coefficient to improve sensitivity. The smoothed growth rate is obtained by weighting the previous growth rate and the current growth rate according to the smoothing coefficient. If the load size also needs to be considered, the load conversion weight is explained after it appears: This weight is determined by comparing the processing time per unit byte with the processing time per unit entry. It is used to convert the number of bytes into the equivalent growth of the number of entries to maintain consistency with the threshold table.

[0075] After obtaining the growth rate, the module maps the new data growth rate to a preset time delay threshold and calculates the transmission delay deviation. The time delay threshold table has a segmented mapping structure. The segment boundaries are explained after their occurrence: the segment boundaries are obtained by statistically analyzing the achievable delay upper bounds under different growth levels during historical stable periods. First, the growth rate is divided into intervals, then the corresponding delay upper bound is extracted for each interval as a candidate threshold, and the final boundary is formed by combining it with the maximum allowable delay in the critical business service level. The threshold table update cycle is explained after its occurrence: the update cycle is determined by the frequency of campus network topology changes and equipment online frequency. When changes are frequent, the cycle is shortened to keep the threshold aligned with the current network capability; when changes are stable, the cycle is extended to reduce maintenance overhead. The mapping process is as follows: the current growth rate is located to the corresponding interval in the threshold table, and the allowable delay upper limit corresponding to that interval is read as the current time delay threshold. Transmission latency adopts a queue transmission latency caliber, which is explained after its introduction: This caliber is defined by subtracting the enqueue timestamp from the dequeue timestamp to avoid deviations caused by fluctuations in the time generated on the end side. When it is necessary to evaluate the end-to-end experience, the switching conditions between the end-to-end caliber and the queue caliber are explained after the switching conditions are introduced: The switching conditions are determined by the service type and alarm level. The end-to-end caliber is used for critical alarm-type sensor data, while the queue caliber is used for others. The queue transmission latency is calculated by subtracting the enqueue timestamp from the dequeue timestamp for each dequeued item within the current sampling time window to obtain the single queue latency, then averaging the values ​​to obtain the window average queue latency, and simultaneously calculating the maximum queue latency for risk identification. The transmission delay deviation is calculated by subtracting the current time delay threshold from the average queue delay of the window to obtain the deviation value. A deviation value greater than 0 indicates an over-limit risk, while a deviation value less than or equal to 0 indicates that it is within the safe range. To avoid a single deviation triggering an erroneous update, the number of deviation persistence determinations is specified after the occurrence of the deviation persistence determination count: This count is determined by statistically analyzing the short-term rebound duration of the deviation during the stable period, taking a safe value near the upper limit of the rebound duration, and requiring this count to be reached consecutively before entering the residual verification process.

[0076] After the transmission delay deviation is formed, the module inputs it into the congestion prediction module to generate congestion prediction residuals, and determines whether to trigger model parameter updates based on the residuals. The input of the congestion prediction module consists of three parts: transmission delay deviation, queue backlog characteristics, and link occupancy characteristics. The calculation process for queue backlog characteristics is as follows: at the end of the sampling time window, the length of each queue is read. The queue length is obtained by the cumulative difference between the enqueue count and the dequeue count, and the increment of the queue length compared to the previous time window is calculated. The calculation process for link occupancy characteristics is as follows: the link bandwidth occupancy, packet loss count, and delay jitter count are read, and the mean and maximum values ​​are calculated according to the sampling time window. The congestion prediction prior comes from the congestion output channel of the final scheduling adaptation model. The prior output time alignment advance is explained after the occurrence of the prior output time alignment advance: this advance is determined by the sum of the peak values ​​of online inference time, scheduling deployment time, and queue activation time, plus one scheduling cycle as a safety margin, used to align the prior prediction to the current verification window. The calculation process for congestion prediction residuals is as follows: The residual value is obtained by subtracting the real-time congestion estimate from the prior congestion prediction. The real-time congestion estimate is synthesized from queue backlog characteristics and link occupancy characteristics according to weights. The synthesis weights are explained as follows: The weights are determined through historical congestion event replay to ensure that the contribution ratios of queue backlog changes and link occupancy changes to the estimation before congestion occur conform to the actual sequential relationship. The prior congestion prediction value is taken from the model output and scale-aligned. The scale alignment factor is explained as follows: This factor is determined by comparing the mean difference between the "model prediction level" and the "real-time estimation level" during the stable period, used to eliminate system bias caused by inconsistencies in their dimensions and ranges. The allowable residual range is explained as follows: The allowable range is determined by the upper bound of the residual fluctuation during the stable period, and is also constrained by the upper limit of critical business latency jitter. The stricter of the two is taken as the boundary to avoid tolerating large residuals even when latency risks increase. If the residual value exceeds the allowable residual range and meets the deviation persistence judgment number, the module generates a gradient update vector and uses it to modify the model weight values ​​of the scheduling adaptation model.

[0077] The generation and application of gradient update vectors employ an online small-step update strategy to reduce disturbances to overall scheduling stability. The set of updatable parameters, once defined, is determined through sensitivity analysis. This analysis involves observing the residual decrease magnitude by perturbing each parameter; parameters showing significant decreases are included, while those not showing significant decreases are excluded. Priority is given to covering congested output channels, queue allocation channels, and a small number of upstream fusion parameters. The update step size, once defined, is determined through a progressive search from small to large using replay verification. The maximum value is selected when the residual continuously decreases and scheduling output fluctuations do not exceed the allowable upper limit. The allowable upper limit is jointly determined by the latency jitter limit for critical sensor services and the stuttering tolerance limit for video services. The gradient pruning threshold, once defined, is determined by the upper bound of the gradient magnitude in online update replays during the stable period. A safe value near the upper bound is selected, and verification is performed in replays of sudden growth scenarios to ensure that pruning still reduces the residual without causing convergence stagnation. The upper limit for the number of update iterations is explained after it appears: This upper limit is derived from the online budget and is determined by multiplying the peak time of a single iteration by the number of iterations, which does not exceed 80% of the budget. 20% is reserved for queue reconstruction and output parsing. During the update process, in each iteration, the verification error, centered on the residual, is calculated first. The calculation process involves taking the absolute value of the residual and adding a penalty term for transmission delay deviation. The penalty weight is explained after its occurrence: this weight is determined by the business level; for critical sensor businesses, the penalty weight is increased to prioritize delay control. Then, the verification error is passed along the model calculation path with reverse sensitivity propagation to obtain the gradient component of each parameter in the updatable parameter set. After pruning the gradient components, the parameters are adjusted in the reverse direction according to the update step size. After the adjustment, the residual is recalculated and it is determined whether the stopping condition has been met. The stopping condition adopts a dual constraint of residual fallback threshold and minimum improvement threshold. The residual fallback threshold is explained after its occurrence: this threshold is taken as a more stringent safety value within the allowable range of residuals to ensure that there is a margin after fallback. The minimum improvement threshold is explained after its occurrence: this threshold is determined by the noise level of the residual calculation during the stable period; if the improvement amount is lower than this threshold for two consecutive iterations, the process stops to avoid invalid iterations.

[0078] After the model weights are updated, the module uses the updated model weight values ​​to generate a stable transmission queue that matches the current data scale, and parses the queue to obtain the stable transmission scheduling output. The construction of the stable transmission queue first determines the queue hierarchy. The number of hierarchical levels is explained after the hierarchy is defined: the number of hierarchical levels is determined by the priority levels supported by the scheduling platform, and configured in conjunction with the number of business types and the number of critical businesses to ensure that critical businesses occupy independent levels to avoid being squeezed out. The queue capacity limit is explained after the queue capacity limit is defined: the limit is jointly constrained by available memory, the maximum allowed queuing time, and the maximum growth rate. The required capacity is obtained by multiplying the number of arrivals per unit time under the maximum growth rate by the allowed queuing time, and then taking the smaller value between this and the memory capacity. Subsequently, based on the updated scheduling adaptation model, the priority and processing share of the output queue are calculated. The calculation process for the processing share is as follows: The amount of data that can be processed per unit time is allocated according to the current growth rate and the urgency of each business's latency. Urgency is calculated from the remaining available latency; the smaller the remaining latency, the higher the urgency. The allocation ratio is constrained by the minimum guaranteed share. The minimum guaranteed share, once it is determined, is explained as follows: This share is determined by service level requirements to ensure that non-critical businesses do not experience prolonged starvation while not affecting the achievement of critical business targets. The queue enqueue rule is to allocate entries to the corresponding queues according to business type and regional index. The sorting rule within the queue is to first sort by urgency, then by enqueue order to maintain stability. The processing window length is dynamically adjusted according to the current data scale. The window adjustment step size, once it is determined, is explained as follows: The step size is determined by the growth rate change and online stability requirements; the larger the change, the larger the step size, but it is constrained by the maximum step size upper limit, which is determined by the upper bound of the natural fluctuation of the window during the stable period, to avoid scheduling oscillations caused by sudden window changes.

[0079] After a stable transmission queue is formed, the module parses it and extracts the scheduling output. The scheduling output includes the queue service order, the processing window length of each queue, the resource allocation instructions for each queue, and the action sequence within the scheduling period. The parsing process is as follows: At the beginning of each scheduling period, the priority and current backlog of each queue are read, and a service order is generated according to priority from high to low. Then, a processing window is allocated to each queue, and a dequeue plan for the current period is generated. The number of dequeue plan entries is calculated by converting the available processing time of the processing window into the average processing time of the entries. The average processing time is obtained by statistical analysis of the most recent consecutive time windows. Subsequently, the dequeue plan is converted into resource allocation instructions. The resource allocation instructions include bandwidth share, time slot share, or scheduling period range. The instruction generation rules are defined by the underlying scheduling execution interface and correspond one-to-one with the queue identifier. Finally, the service order, dequeue plan, and resource allocation instructions are combined and output as the scheduling output for stable transmission.

[0080] The resource allocation optimization module extracts data transmission performance metrics from the stable transmission scheduling output. It then uses a gradient descent algorithm to optimize these performance metrics against the constraints of limited network resources. The process for determining the resource allocation scheme includes parsing the stable transmission scheduling output to obtain a scheduling log sequence, extracting throughput and packet loss rate values ​​from the log sequence, and combining these values ​​with bandwidth limitations and energy consumption values ​​as input to a loss function construction module. The module calculates the loss function by taking the partial derivative of the loss function to obtain the gradient vector, and iteratively adjusts the resource allocation weights using this gradient vector to obtain the optimal resource weight values. Finally, based on the optimal resource weight values, it determines the physical layer resource block mapping table and, according to this table, determines the power allocation value and spectrum resource block location for each terminal node to complete the resource allocation scheme determination.

[0081] In this implementation, the scheduling output is first structured and parsed to form a scheduling log sequence, ensuring consistent statistical standards for throughput and packet loss rate extraction. The scheduling log sequence includes a scheduling cycle identifier, terminal node identifier, service type identifier, queue service order, allocated bandwidth share or time slot share, actual number of bytes sent, number of bytes successfully delivered, number of attempted packets, number of successfully delivered packets, number of retransmissions, number of timeouts and packet drops, link quality summary, and transmit power record. The log truncation time window length is explained after its appearance: this length is determined based on the link state change rate and the online optimization time budget. First, the typical duration of historical link quality deterioration to recovery is statistically analyzed as a lower bound, and then the minimum length that satisfies the optimization calculation without triggering the latency budget is selected. The truncation step size is explained after its appearance: the step size is set as an integer multiple of the scheduling cycle, determined by measuring the peak time of single-round metric statistics, loss calculation, and weight update, selecting the minimum integer multiple where the peak time does not exceed the budget. First, perform an integrity check on the fields in the log. The missing data determination threshold is explained as follows: The threshold is set according to the scheduling cycle count. If the missing data is not filled in for more than 5 scheduling cycles, it is considered missing. Missing records will have their weight reduced in the statistics. The weight reduction coefficient is determined by the upper limit of the statistical fluctuation during the stable period to ensure that the indicator sequence does not have non-physical jumps after the weight reduction.

[0082] After the scheduling log sequence is established, throughput and packet loss rate values ​​are extracted from it. Throughput uses the effective throughput metric, which is explained after it appears: This metric is based on the number of successfully delivered payload bytes, eliminating retransmitted and duplicate bytes to avoid artificially inflating throughput; The throughput is calculated by accumulating the number of successfully delivered bytes of all terminal nodes in each scheduling cycle and then dividing it by the duration of that scheduling cycle to obtain the cycle throughput; When it is necessary to decompose by terminal, the number of successfully delivered bytes for each terminal is accumulated and divided by the cycle duration to obtain the terminal throughput. Packet loss rate uses a delivery-based loss metric, which is explained below: This metric uses the number of attempted packets as the denominator and the number of unsuccessfully delivered packets as the numerator. The number of unsuccessfully delivered packets is obtained by subtracting the number of successfully delivered packets from the number of attempted packets, and timeout-out packets are included in the unsuccessful deliveries. The packet loss rate is calculated by counting the number of attempted packets and successfully delivered packets for each scheduling cycle, calculating the number of unsuccessfully delivered packets, and then dividing by the number of attempted packets to obtain the cycle-wide packet loss rate. To avoid amplifying packet loss rate fluctuations due to a small number of packets, a minimum sample threshold is defined below: This threshold is determined by the lower bound of the number of packets sent per cycle during the stable period. Cycles below this threshold are statistically combined using adjacent cycles to improve stability. After throughput and packet loss rate are established, the module performs smoothing processing on the indicator sequence. The smoothing coefficient is explained below: The smoothing coefficient is determined by the upper bound of the natural fluctuation of the indicator during the stable period. The larger the fluctuation, the smaller the smoothing coefficient to enhance stability; the smaller the fluctuation, the larger the smoothing coefficient to enhance sensitivity. The smoothed indicator is obtained by weighting the previous round's indicator and the current indicator according to the smoothing coefficient.

[0083] After obtaining the throughput and packet loss rate values, the module introduces bandwidth limitation boundaries and energy consumption values ​​to construct and calculate the loss. The bandwidth limitation boundary is determined by the number of available spectrum resource blocks, the available bandwidth per resource block, and the available time slots within a scheduling period. The bandwidth limitation boundary is determined by reading the available spectrum configuration and time slot configuration of the base station or switching node, deducting the reserved control signaling resources to obtain the available resource set, and then converting the available resource set into the maximum effective bandwidth that can be allocated within a unit scheduling period. The energy consumption value is obtained by statistically analyzing the transmit power and continuous transmit duration of the terminal node. The energy statistics time window is explained after its appearance: this time window is consistent with the log truncation time window to avoid inconsistencies. The energy consumption calculation process involves reading the transmit power record and actual occupied time slot duration for each terminal in each scheduling period, multiplying the power by the duration to obtain the energy consumption for that period, and then summing the periodic energy consumption within the window to obtain the window energy consumption. To avoid abnormal power records affecting energy consumption estimation, the power anomaly threshold is explained after its occurrence: the threshold is determined jointly by the terminal's rated power limit and the upper bound of natural power fluctuations during the stable period; power records exceeding the threshold are truncated. The process then proceeds to the loss construction module, where the loss is composed of throughput penalties, packet loss penalties, energy consumption penalties, and bandwidth overrun penalties. The throughput penalty is calculated by comparing throughput with the target throughput. The target throughput is determined jointly by the service level target and the upper bound of historical stable period throughput, taking the minimum throughput that meets critical business needs as the target; a penalty is incurred when the actual throughput is lower than the target throughput, with higher penalties for larger discrepancies. The packet loss penalty is calculated by comparing the packet loss rate with the allowable packet loss upper limit. The allowable packet loss upper limit is determined by the service reliability level and combined with the upper bound of natural packet loss fluctuations during the stable period, taking the more stringent one; the larger the packet loss rate exceeds the upper limit, the higher the penalty. The calculation process for the energy consumption penalty involves comparing the window energy consumption with the energy budget. The energy budget, as explained after its appearance, is determined by the terminal's battery capacity, desired battery life target, and temperature rise safety boundary, and is issued in the device management policy. The closer the window energy consumption is to or exceeds the budget, the higher the penalty. The calculation process for the bandwidth overrun penalty involves calculating the total bandwidth demand requested by the current resource allocation. This bandwidth demand is calculated by multiplying the number of resource blocks allocated to each terminal by the bandwidth of a single block, summing the results, and then comparing this to the bandwidth limit boundary. Any excess is penalized according to the excess amount. The penalty strength coefficient, as explained after its appearance, is determined through replay verification. The goal is to quickly return to within the boundary once an overrun occurs without causing a sudden drop in throughput.The weights of each penalty item are explained after their appearance: Weights are determined jointly by the service level and resource stress. Critical services have increased throughput and packet loss weights, while power consumption weights are increased when terminal battery power is low. Resource stress is assessed by the ratio of available bandwidth to required bandwidth; when the ratio is below a stress threshold, the out-of-bounds penalty weight is increased. The stress threshold is determined by the occupancy level corresponding to historical congestion inflection points and verified through replay. After calculating each penalty, they are added together according to their weights to obtain the total loss.

[0084] After obtaining the loss value, the partial derivative of the loss is taken to obtain the gradient vector value. The gradient vector is then used to iteratively correct the resource allocation weights to obtain the optimal resource weight values. The resource allocation weights are defined by terminal nodes or by service queues. The initial weight values ​​are explained after their occurrence: the initial values ​​are generated jointly by the queue priority in the stable transmission scheduling output and the historical average resource occupancy ratio. Objects with higher priority have larger initial values, while also satisfying the normalization constraint that the sum of the weights is 1. The gradient vector generation process is as follows: the resource allocation weights are used as input parameters for the loss. The magnitude of the change in the loss value when the weight of a certain object changes slightly is calculated, and the magnitude of the change is used as the gradient component corresponding to that object. The magnitude of the change is calculated using a finite difference approach. The difference step size is explained after its occurrence: the difference step size is determined by the lower bound of the natural fluctuation of the weight change in the stable option. If it is too small, it will be overwhelmed by numerical noise; if it is too large, it will destroy the local approximation. The smallest step size is selected to make the gradient estimation stable and the optimization converge. After obtaining the gradient vector, gradient pruning is performed. The pruning threshold, as explained after its occurrence, is determined by the upper bound of the gradient magnitude during the stable period. This threshold is verified in burst load replays to ensure that pruning still reduces loss and does not lead to convergence stagnation. Next, weights are updated. The update step size, as explained after its occurrence, is determined by progressively searching from small to large through replay verification, selecting the maximum value where loss continues to decrease and throughput and packet loss fluctuations do not exceed the allowable upper limit. The allowable upper limit is jointly constrained by the critical business latency jitter limit and the packet loss limit. The update rule is to adjust the weights in the opposite direction of the gradient to reduce loss. After the update, constrained projection is performed. The projection process involves truncating negative weights to 0 and then renormalizing all weights to sum to 1. A single weight change upper limit is set, as explained after its occurrence. This upper limit is determined by the upper bound of the natural fluctuation of weights during the stable period to avoid drastic changes in resource allocation caused by a single update. The iteration stopping condition is jointly constrained by a convergence threshold and a maximum number of iterations. The convergence threshold, once reached, is determined by calculating the noise level through the steady-state loss; iterations stop if the loss improvement is below this threshold for two consecutive iterations. The maximum number of iterations, once reached, is determined by working backward from the online budget, ensuring that the peak time of a single iteration multiplied by the number of iterations does not exceed 80% of the budget, with 20% reserved for resource block mapping and power calculation. The optimal resource weight value is output when the stopping condition is met.

[0085] After obtaining the optimized resource weight values, the module determines the physical layer resource block mapping table accordingly, and further determines the power allocation value and spectrum resource block location for each terminal node to complete the resource allocation scheme. The construction of the resource block mapping table first defines the set of available resource blocks. The set of available resource blocks is determined by reading the spectrum grid configuration and time slot configuration of the physical layer scheduler, removing control signaling reserved blocks and guard band blocks to obtain the remaining resource block list, and numbering them according to spectrum location and time location. Then, the number of resource blocks is allocated according to the resource weights. The calculation process for the allocation number is to multiply the total number of available resource blocks by the weight ratio to obtain the target number of blocks for each terminal, and then perform rounding on the target number of blocks. The rounding strategy is explained after the rounding strategy is introduced: the rounding strategy first allocates the integer part and then pads the remaining decimals from largest to smallest, ensuring that the total number of blocks does not exceed the number of available blocks and that the allocation error is minimized. Resource block locations are selected based on link quality summaries. The link quality ranking parameters are explained after their appearance: these parameters are determined by bit error and packet loss statistics for different frequency bands during the stable period. Frequency bands with low bit error and packet loss are prioritized for allocation to terminals with high weight and high reliability requirements. After completing the allocation of resource block quantity and location, a resource block mapping table is formed.

[0086] Power allocation is performed after the resource block mapping table is determined. The goal of power allocation is to meet throughput and packet loss constraints while controlling energy consumption. The power allocation value for each terminal is determined by the target transmission success rate and link quality. The target transmission success rate is explained after it is determined: this indicator is derived from the allowable packet loss upper limit; the lower the allowable packet loss, the higher the target success rate. Link quality is obtained by combining real-time measured channel quality indicators and historical bit error statistics. The power calculation process is as follows: first, the target modulation and coding level of the terminal on its allocated resource block is determined. The modulation and coding level selection rule is explained after it is determined: the rule is determined based on the link quality classification table, and the classification boundary is determined by the trade-off between stable period bit error and throughput, and is verified to meet reliability requirements during playback. After determining the modulation and coding level, the lower power limit required to achieve the target success rate at this level is estimated based on the link quality. Then, the lower power limit is compared with the terminal power upper limit, which is determined by the device rated upper limit and temperature rise safety boundary. If the lower power limit exceeds the power upper limit, the modulation and coding level is reduced and the lower power limit is re-estimated until the power upper limit constraint is met. To suppress energy consumption fluctuations caused by power fluctuations, the power smoothing coefficient is explained as follows: the coefficient is determined by the upper bound of the natural power fluctuation during the stable period, and the smoothed power is obtained by weighting the power from the previous round and the power calculated in the current round according to the smoothing coefficient. Finally, the power allocation value and spectrum resource block location for each terminal are output, and the resource block mapping table, power allocation, and scheduling cycle instructions are combined to form a resource allocation scheme.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A scheduling and communication intelligent system based on a multimodal large model, characterized in that, include: The data acquisition and fusion module acquires modal information and historical scheduling experience data accompanying newly arriving data from the communication scheduling environment, processes the data through an incremental learning algorithm to update the scheduling strategy, and obtains a preliminary fused strategy model. The experience retention module extracts temporal dependency features from the initially fused strategy model, and uses a federated learning algorithm to integrate the temporal dependency features with the modal information accompanying the newly arrived data to determine the updated experience retention module. If the updated experience retention module detects a modal difference conflict, the decision-making construction module adjusts the parameters of the conflicting part through the gradient descent algorithm to obtain the scheduling decision framework after conflict resolution. The data sequence optimization module obtains diverse reporting patterns from the scheduling decision framework after conflict resolution, performs time-series sequence analysis on the reporting patterns, determines whether it meets the requirements for rapid adaptation to changes, and obtains the optimized data processing sequence. The forgetting compensation module extracts historical fragments related to knowledge forgetting based on the optimized data processing sequence, and re-injects these historical fragments into the scheduling decision framework using an incremental learning algorithm to determine the complete strategy after forgetting compensation. The scheduling adaptation model construction module obtains the latency requirement characteristics of industrial park sensor data from the complete strategy after forgetting compensation, and then integrates them with video surveillance experience through federated learning algorithm to obtain the final scheduling adaptation model.

2. The intelligent scheduling and communication system based on a multimodal large model according to claim 1, characterized in that: The data acquisition and fusion module acquires modal information and historical scheduling experience data accompanying newly arriving data from the communication scheduling environment, processes the data using an incremental learning algorithm to update the scheduling strategy, and obtains a preliminary fused strategy model, including: Newly arriving multimodal data streams and historical scheduling logs are obtained from the communication scheduling environment. Multidimensional feature mapping technology is used to process the newly arriving multimodal data streams and the historical scheduling logs to obtain a spatiotemporal feature vector sequence. The spatiotemporal feature vector sequence is input into a deep neural network to generate a high-dimensional state representation matrix; If the high-dimensional state representation matrix satisfies the drift condition, then the update gradient of the policy network weights is calculated based on the high-dimensional state representation matrix to obtain the updated policy parameter set. An adaptive weighted aggregation algorithm is used to fuse the updated policy parameter set with the global baseline model parameters to obtain a preliminary fused policy model.

3. The intelligent scheduling and communication system based on a multimodal large model according to claim 1, characterized in that: The experience retention module extracts temporal dependency features from the initially fused policy model, integrates these features with new modality information using a federated learning algorithm, and determines the updated experience retention module as follows: The temporal-dependent correlation feature vector is obtained by analyzing the preliminarily fused policy model using a long short-term memory network. A multi-head attention mechanism is used to project the new modal information data stream onto the semantic feature space where the temporally dependent associated feature vectors are located, generating a multimodal collaborative feature matrix; The federated local gradient update amount is calculated based on the multimodal collaborative feature matrix, and a safe weighted aggregation calculation is performed on the federated local gradient update amount to generate a globally shared feature increment. Globally shared feature increments are mapped to the experience memory network, and the weights of the experience memory network are adjusted through an adaptive forgetting gating unit to determine the updated experience retention module.

4. The intelligent scheduling and communication system based on a multimodal large model according to claim 1, characterized in that: If the updated experience retention module detects a modal difference conflict, the decision-making construction module adjusts the parameters of the conflicting part using the gradient descent algorithm to obtain the scheduling decision framework after conflict resolution, including: Obtain the updated multimodal feature distribution data within the experience retention module and calculate the Mahalanobis distance between feature vectors of different modalities; If the Mahalanobis distance exceeds the preset collision detection threshold, a modal difference collision is determined to exist, and a joint loss function is constructed based on the modal difference collision to calculate the loss value; The loss value is input into the backpropagation computation graph to generate a gradient vector. The gradient descent algorithm is used to adjust the weight parameters along the negative direction of the gradient vector to obtain the converged corrected parameter set. Based on the internal mapping logic of the module retained through the reorganization of the modified parameter set, a scheduling decision framework after conflict resolution is generated.

5. The intelligent scheduling and communication system based on a multimodal large model according to claim 1, characterized in that: The data sequence optimization module obtains diverse reporting patterns from the conflict-resolved scheduling decision framework, performs time-series analysis on the reporting patterns, determines whether they meet the requirements for rapid adaptation to changes, and obtains the optimized data processing sequence, including: Historical scheduling logs are extracted from the scheduling decision framework after conflict resolution to construct a set of original reporting time sequences that reflect the dynamic attributes of the business. A feature vector matrix is ​​generated for the original reported time series set, and an adaptability score is calculated based on the feature vector matrix; If the fitness score is below the threshold, then identify the blocking node that is causing the response delay; Reset queue priority based on blocked nodes, and dynamically scale the processing window according to queue priority; The data flow order is rearranged according to the processing window to obtain an optimized data processing sequence that meets the requirements of adapting to changes.

6. The intelligent scheduling and communication system based on a multimodal large model according to claim 1, characterized in that: The forgetting compensation module extracts historical fragments related to knowledge forgetting based on the optimized data processing sequence, re-injects these historical fragments into the framework using an incremental learning algorithm, and determines the complete strategy after forgetting compensation, including: Obtain the optimized data processing sequence, and label the forgetting features based on the backtesting results of the data processing sequence on the historical validation set; By filtering historical log vectors based on forgetting characteristics, historical fragments related to the phenomenon of knowledge forgetting can be obtained; An incremental model is generated using a historical fragment-driven online incremental learning algorithm, and a forgetting compensation factor is calculated based on the generalization error of the incremental model. The forgetting compensation factor is mapped to the output layer of the incremental model to determine the complete strategy after forgetting compensation.

7. The intelligent scheduling and communication system based on a multimodal large model according to claim 1, characterized in that: The scheduling adaptation model construction module obtains the latency requirement characteristics of industrial park sensor data from the complete strategy after forgetting compensation, and fuses them with video surveillance experience through a federated learning algorithm to obtain the final scheduling adaptation model, which includes: The communication protocol messages extracted from the complete strategy after forgetting compensation are analyzed to generate latency requirement characteristics that quantify the urgency of sensor data transmission. Based on the latency requirements, a monitoring experience vector reflecting the sensitivity of video surveillance services to network fluctuations is constructed, and the monitoring experience vector is input into a federated learning framework to aggregate and obtain a global experience model. The cross-regional network congestion predictions output by the global empirical model are mapped to latency requirement features to generate a heterogeneous fusion feature matrix characterizing the correlation between sensor data and video surveillance services. A heterogeneous fusion feature matrix is ​​used to supervise and adjust the pre-set scheduling neural network, resulting in a final scheduling adaptation model with multi-service collaborative processing capabilities.

8. The intelligent scheduling and communication system based on a multimodal large model according to claim 1, characterized in that, It also includes a real-time verification and stable output module, which performs real-time verification on the final scheduling adaptation model to determine whether the increase in new data causes time delays. If so, it adjusts the model parameters to obtain a stable transmission scheduling output. Specifically, it includes: Acquire the real-time verification data stream generated by sensors in the industrial park, calculate the new data growth rate based on the timestamp difference of the real-time verification data stream, and map the new data growth rate to a preset time delay threshold to calculate the transmission delay deviation. The transmission delay deviation is input into the congestion prediction module to generate congestion prediction residuals. If the congestion prediction residuals exceed the allowable range, a gradient update vector is generated to modify the model weight values ​​of the scheduling adaptation model. The updated model weight values ​​are used to generate a stable transmission queue that matches the current data scale, and the stable transmission queue is parsed to extract the scheduling output of stable transmission.

9. The intelligent scheduling and communication system based on a multimodal large model according to claim 8, characterized in that, It also includes a resource allocation optimization module, which extracts data transmission performance metrics from the stable transmission scheduling output, uses a gradient descent algorithm to optimize the performance metrics and the constraints of limited network resources, and determines the resource allocation scheme, specifically including: Parse the scheduling output of stable transmission to obtain the scheduling log sequence, and extract the throughput and packet loss rate values ​​from the scheduling log sequence; The throughput and packet loss rate values ​​are combined with the bandwidth limit boundary and energy consumption values ​​and input into the loss construction module to calculate the loss function value.

10. A scheduling and communication intelligent system based on a multimodal large model according to claim 9, characterized in that: The resource allocation optimization module extracts data transmission performance metrics from the stable transmission scheduling output, uses a gradient descent algorithm to optimize the performance metrics against the constraints of limited network resources, and determines the resource allocation scheme by further including: The partial derivative of the loss function is used to obtain the gradient vector value. The gradient vector value is then used to iteratively correct the resource allocation weights to obtain the optimal resource weight values. The physical layer resource block mapping table is determined based on the optimized resource weight values, and the power allocation value and spectrum resource block location of each terminal node are determined based on the resource block mapping table to complete the determination of the resource allocation scheme.

Images