Real-time signaling flow intelligent arrangement system based on queue optimization

Abstract

The invention discloses a real-time signaling flow intelligent arrangement system based on queue optimization, and relates to the technical field of communication networks, and the system comprises a multi-dimensional feature perception and fusion module which is used for collecting and outputting various dynamic attribute parameters of signaling in real time; and the self-adaptive scheduling decision module is connected with the multi-dimensional feature perception and fusion module and is used for receiving the dynamic attribute parameters and generating a scheduling instruction. According to the real-time signaling flow intelligent arrangement system based on queue optimization, by introducing a double-loop intelligent architecture of prediction and decision verification, the perspectiveness and reliability of signaling flow scheduling are effectively improved. According to the technical scheme, the system can make decisions by integrating the multi-dimensional dynamic factors, and through double inspection of the rule base and micro-simulation, the reasonability of the scheduling strategy and the system stability are still kept in the face of complex burst traffic, and the end-to-end service quality is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication network, in particular to a real-time signaling flow intelligent scheduling system based on queue optimization. BACKGROUND

[0002] The existing scheduling standards generally present a reactive feature, that is, mainly according to the signaling state of the current arrived queue to make decisions, and cannot make a prediction on the traffic mode change that the system may face in the short term. When complex, sudden and internally related signaling flows appear in the network, this passive coping mechanism appears rigid and slow. The system is difficult to adaptively adjust the weight balance between different scheduling factors according to the real-time load change, the logical relationship between the signaling and the predicted traffic trend, resulting in insufficient flexibility and intelligent level of the overall scheduling strategy, ultimately affecting the stable guarantee of end-to-end service performance. Therefore, the problem to be solved by the present system is how to realize an intelligent scheduling mechanism that can integrate multi-dimensional dynamic factors and make forward-looking scheduling. SUMMARY

[0003] The present application aims to provide a real-time signaling flow intelligent scheduling system based on queue optimization to solve the problems raised in the background. To solve the above technical problems, the present application provides the following technical scheme: a real-time signaling flow intelligent scheduling system based on queue optimization, comprising: A multi-dimensional feature perception and fusion module is used for real-time collection and output of multiple dynamic attribute parameters of signaling; An adaptive scheduling decision module is connected to the multi-dimensional feature perception and fusion module, used for receiving the dynamic attribute parameters and generating a scheduling instruction; A queue execution module is connected to the adaptive scheduling decision module, used for executing the scheduling instruction; The adaptive scheduling decision module internally includes a two-stage decision unit and an internal verification unit; The two-stage decision unit is used for generating a preliminary scheduling scheme based on the currently received dynamic attribute parameters; The internal verification unit includes a rule library comparison subunit and a micro-simulation deduction subunit; The rule library comparison subunit is used for comparing the preliminary scheduling scheme with a reference scheme in a preset rule library, and outputting an optimized scheduling scheme to the micro-simulation deduction subunit according to the comparison result; The micro-simulation deduction subunit is used for making a forward-looking deduction on the optimized scheduling scheme based on a current state snapshot of the system, and feeding back performance index data obtained by deduction to the adaptive scheduling decision module to adjust its decision parameters, and outputting a finally determined scheduling instruction to the queue execution module.

[0004] It is further noted that after the system is started, the multi-dimensional feature perception and fusion module continues to run, and real-time collection of multiple dynamic attribute parameters of each signaling unit in the input signaling stream is performed. The parameters can include but are not limited to the service type to which the signaling belongs, the emergency level identification carried by the signaling itself, the identity level of the signaling source device, and the current depth of each processing queue when the signaling arrives. These parameters are collected and integrated into a set of dynamic data sets representing the current system state and signaling features in real time and in parallel, and output to the adaptive scheduling decision module.

[0005] After receiving the above dynamic data set, the dual-stage decision unit in the adaptive scheduling decision module immediately starts working. The unit first analyzes and weighs the input data based on a set of preset initial strategy model, quickly generates a preliminary scheduling scheme for the current queue, which clearly shows the priority order or resource allocation suggestion of the signaling to be processed. Then, the internal inspection unit starts to work. The rule base comparison subunit first intervenes, which imports the preliminary scheduling scheme into a lightweight simulation environment, which has built-in multiple typical business scenario rules and stability constraints extracted from historical operation and maintenance experience. The comparison subunit quickly runs the preliminary scheme in this simulation environment, monitors and evaluates the risks that may be caused if it is actually executed, such as whether it will cause the processing delay of high-priority signaling to exceed the acceptable range, or whether it will cause the queue length to fluctuate sharply. If the evaluation result indicates that the risk level is below the preset safety threshold, the preliminary scheme can directly enter the next stage; if the risk level exceeds the threshold, the rule base comparison subunit will trigger the adjustment mechanism of the decision model, and modify the preliminary scheme according to the optimization direction provided by the rule base, thereby generating an optimized scheduling scheme.

[0006] After that, the micro-simulation deduction subunit starts to work. It first captures the accurate state snapshot of the system at the current time, including the detailed content of all active queues, the instantaneous load rate of the central processor, the memory occupation situation, etc. Then, it loads the scheduling scheme optimized by the rule base comparison onto the state snapshot, and performs a short-time range, multi-cycle discrete event simulation deduction. This deduction process aims to prospectively observe the influence of the scheduling scheme on the overall behavior of the system in the next few consecutive working periods, and record key performance indicator data, such as the average waiting time of signaling during the deduction period, the change trend of the overall throughput of the system, etc. These performance indicator data obtained by deduction are immediately fed back to the core decision logic of the adaptive scheduling decision module. The decision module uses these data as an immediate effect evaluation signal to fine-tune the parameter weights of its internal strategy model, thereby realizing online learning and optimization of the decision-making ability. At the same time, the final scheduling instruction confirmed by this deduction is sent to the queue execution module.

[0007] The queue execution module reorders the signaling queues in the physical cache or allocates computing resources for specific signaling according to the received final scheduling instruction, thereby completing the scheduling arrangement this time. The entire process forms a continuous and closed-loop automated decision-making chain from data collection, preliminary decision-making, internal rule inspection to micro-simulation deduction and feedback optimization.

[0008] Through the inspection and optimization steps including at least twice different mechanisms: rule base comparison and micro-simulation deduction, the system ensures that the scheduling decision is not only based on the current state, but also fully considers the long-term consequences and potential risks of the decision, thereby realizing high reliability and adaptive intelligent arrangement of real-time signaling flow, effectively overcoming the defects of traditional scheduling methods lacking foresight and adaptive ability.

[0009] Further, the multi-dimensional feature perception and fusion module includes a short-term traffic prediction sub-module; The short-term traffic prediction sub-module is used to analyze historical signaling flow data, predict the arrival trend and correlation of signaling flow within a set time window in the future, and output the prediction results as a set of dynamic weight factors to the adaptive scheduling decision module.

[0010] Further, the short-term traffic prediction sub-module adopts an online learning time series analysis model as the prediction model; The performance index data deduced by the micro-simulation deduction sub-unit is fed back to the short-term traffic prediction sub-module for correcting the prediction bias.

[0011] Further, the logic compared by the rule base comparison sub-unit includes: Running the preliminary scheduling scheme in a simulated environment to evaluate the possible high-priority signaling latency and queue state fluctuation risk; When the evaluation risk exceeds the preset threshold, triggering the strategy adjustment mechanism of the adaptive scheduling decision module to generate the optimized scheduling scheme.

[0012] Further, the process of forward-looking deduction by the micro-simulation deduction sub-unit includes: Building a state snapshot based on the current queue state and resource occupation of the system; Placing the optimized scheduling scheme in the state snapshot, simulating running the subsequent continuous multiple scheduling periods; Recording the signaling average processing latency and system throughput efficiency data generated during the simulation running process as the performance index data.

[0013] Further, after receiving the performance index data, the adaptive scheduling decision module uses it as an immediate reward signal to update the parameters of its internal strategy model, realizing online adaptive optimization of the decision-making model.

[0014] Further, the dynamic attribute parameters collected by the multi-dimensional feature perception and fusion module at least include signaling service type identification, signaling emergency level identification, signaling source identity, and real-time queue depth.

[0015] Further, the core decision model of the adaptive scheduling decision module adopts a policy model constructed based on a reinforcement learning framework, a state space of which is constituted by the dynamic attribute parameters output by the multi-dimensional feature perception and fusion module, and an action space of which is a set of selectable scheduling strategies.

[0016] Further, the queue execution module performs reordering or resource allocation operations on the signaling queue according to the final scheduling instruction.

[0017] Further, the multi-dimensional feature perception and fusion module, the adaptive scheduling decision module, and the queue execution module work cooperatively to form a closed-loop "prediction-decision-verification" process to realize forward-looking intelligent arrangement of real-time signaling flow.

[0018] The application provides a real-time signaling flow intelligent arrangement system based on queue optimization. The real-time signaling flow intelligent arrangement system based on queue optimization effectively improves the forward-looking and reliability of signaling flow scheduling by introducing a double-loop intelligent architecture of prediction and decision verification. The system can make decisions by comprehensively considering multi-dimensional dynamic factors and can be verified by a rule base and micro simulation, thereby maintaining the rationality of scheduling strategies and system stability when facing complex burst traffic, and guaranteeing end-to-end service quality.

[0019] The real-time signaling flow intelligent arrangement system based on queue optimization has online learning and adaptive optimization capabilities and can continuously adjust scheduling strategies according to historical decision effects. This self-evolution mechanism reduces the dependence on artificial preset rules and improves the adaptability of the system under different network conditions, thereby providing effective support for the intelligentization and automation of communication network operation. BRIEF DESCRIPTION OF DRAWINGS

[0020] Fig. 1 Fig. 1 is a module interaction schematic diagram of the real-time signaling flow intelligent arrangement system based on queue optimization of the application; Fig. 2 Fig. 2 is a scheduling logic diagram based on service type and emergency level of the application; Fig. 3 Fig. 3 is a technical flowchart of micro simulation deduction of the application; Fig. 4 Fig. 4 is a decision flowchart of the adaptive scheduling decision module of the application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0022] Please refer to Figs. 1 to 4 The present application provides a technical solution: a real-time signaling flow intelligent arrangement system based on queue optimization, comprising: A multi-dimensional feature perception and fusion module is configured to collect and output multiple dynamic attribute parameters of signaling in real time. An adaptive scheduling decision module is connected to the multi-dimensional feature perception and fusion module and configured to receive the dynamic attribute parameters and generate scheduling instructions. A queue execution module is connected to the adaptive scheduling decision module and configured to execute the scheduling instructions. The adaptive scheduling decision module comprises a two-stage decision unit and an internal verification unit. The two-stage decision unit is configured to generate a preliminary scheduling scheme based on the currently received dynamic attribute parameters. The internal verification unit comprises a rule library comparison subunit and a micro-simulation deduction subunit. The rule library comparison subunit is configured to compare the preliminary scheduling scheme with a reference scheme in a preset rule library, and output an optimized scheduling scheme to the micro-simulation deduction subunit according to the comparison result. The micro-simulation deduction subunit is configured to perform forward-looking deduction on the optimized scheduling scheme based on a current state snapshot of the system, and feed back performance index data obtained by the deduction to the adaptive scheduling decision module to adjust decision parameters thereof, and output a finally determined scheduling instruction to the queue execution module.

[0023] The multi-dimensional feature perception and fusion module comprises a short-term traffic prediction sub-module. The short-term traffic prediction sub-module is configured to analyze historical signaling flow data, predict the arrival trend and correlation of signaling flow within a set time window in the future, and output the prediction result as a set of dynamic weight factors to the adaptive scheduling decision module. It should be further noted that, in the specific implementation process, the short-term traffic prediction sub-module is implemented as follows: the sub-module continuously receives and caches system historical signaling flow data, which are organized in chronological order and attached with signaling type labels. When the prediction process starts, the sub-module first preprocesses the historical data to identify periodic patterns and trend changes therein, such as traffic regularity of a specific service within a fixed time interval or traffic climbing trend caused by a sudden event.

[0024] Subsequently, the time series analysis model of online learning is activated. Based on the processed historical data, the model predicts the signaling flow in a future time window of a set length. The prediction target not only includes the expected arrival amount of different types of signaling, but also analyzes the possible correlation between different signaling flows, such as the centralized arrival of certain control signaling often triggering batch requests of subsequent data signaling. The result of prediction calculation is not a simple traffic value, but is converted into a set of dynamic weight factors. This set of factors quantitatively describes the trend of the relative importance of various scheduling consideration dimensions, including business priority and latency sensitivity, in a short time in the future. For example, when high-priority signaling flow is predicted to surge, the weight factor of the corresponding business type will be enhanced accordingly.

[0025] The set of dynamic weight factors is then output to the adaptive scheduling decision module as one of the key inputs for its preliminary decision, so that the decision basis is expanded from static parameters reflecting the current state to include dynamic and predictive information. In addition, the performance indicator data from the micro-simulation deduction subunit is also fed back to the short-term traffic prediction sub-module. The feedback data includes the difference information between the actual traffic and the predicted traffic in the previous decision period. The prediction sub-module uses these difference information to adjust the internal parameters of the model through its built-in online learning mechanism, so as to correct the accuracy of subsequent prediction, forming a continuously improved prediction closed loop. This implementation ensures that the scheduling system can be arranged based on the traffic prediction that is constantly approaching the real situation, improving the foresight and adaptability of the decision.

[0026] The prediction model used by the short-term traffic prediction sub-module is an online learning time series analysis model; The performance indicator data deduced by the micro-simulation deduction subunit is fed back to the short-term traffic prediction sub-module for correction of prediction bias.

[0027] It needs to be further explained that in the specific implementation process, the adaptive correction mechanism of the prediction model is involved. The online learning time series analysis model used by the short-term traffic prediction sub-module can continuously adjust its prediction logic according to the actual system operation feedback. After the model completes the traffic prediction of a time window, it does not terminate the task, but enters a verification state. When the corresponding scheduling decision is subjected to micro-simulation deduction and actually applied to the system, the system will record the actual arrival of the signaling flow in the time window. The difference between these actual data and the previous prediction data is quantified as a prediction bias signal.

[0028] The prediction bias signal is fed back to the prediction module in real time. An online learning mechanism is then activated, which includes the following steps: if the bias signal indicates that the predicted value is consistently higher or lower than the actual value, the learning mechanism adjusts the underlying trend judgment parameters of the model; if the bias shows irregular fluctuations, the learning mechanism may focus on adjusting the sensitivity of the model to random disturbances. This adjustment process is based on incremental learning algorithms, with small iterative updates to the model parameters rather than complete retraining, ensuring that the model can adapt to gradual changes in traffic patterns while avoiding over-adjustment due to single sudden events.

[0029] By using the actual effect data after micro-simulation as a calibration signal and continuously acting on the prediction model through a closed-loop path, the system makes the traffic prediction ability no longer static and offline, but dynamic and online self-optimization. This implementation detail ensures that the prediction results can gradually approach the real network traffic behavior, providing more and more reliable forward-looking basis for subsequent scheduling decisions, which is the key foundation of the intelligence of the entire system.

[0030] The logic of the rule library comparison sub-unit includes: Run the preliminary scheduling scheme in a simulated environment to evaluate the potential high-priority signaling latency and queue state fluctuation risks; When the evaluation risk exceeds the preset threshold, trigger the strategy adjustment mechanism of the adaptive scheduling decision module to generate an optimized scheduling scheme.

[0031] It should be further noted that in the specific implementation process, the internal logic judgment and execution process of the rule library comparison sub-unit is implemented as follows: the pre-set rule library in the sub-unit is not a simple static list, but a collection containing multiple constraint conditions and optimization targets. These rules are derived from the refinement of typical scenarios in historical operation data, such as latency constraints to ensure key business links and queue length fluctuation constraints to maintain overall system stability.

[0032] When the preliminary scheduling scheme is generated, the rule base comparison subunit loads it into a simplified simulation environment for deductive analysis. The deduction process simulates the behavior of the system in a short period of time after the scheme is executed. The evaluation logic first focuses on high-priority signaling, calculates the time interval from entering the queue to starting to be processed in the simulation environment, and compares this simulated time delay with the maximum time delay threshold set in the rule base for this type of signaling. At the same time, the evaluation logic monitors the changes in the queue depth during the simulation process, calculates the fluctuation amplitude, and compares it with the stability threshold defined in the rule base. Based on the comparison results, the system makes a logical judgment: if the simulated time delay is not over the limit and the queue fluctuation is within an acceptable range, it is determined that the risk is controllable, and the preliminary scheduling scheme can be directly passed to the next stage. If any evaluation index, such as the simulated time delay of high-priority signaling, exceeds the corresponding threshold, it is determined that there is a risk, and the strategy adjustment mechanism is triggered immediately.

[0033] This adjustment mechanism is not a simple replacement scheme, but a targeted correction based on the optimization strategies preset in the rule base, such as dynamically increasing the scheduling weight of the business type to which the over-limit signaling belongs without violating the basic dependency relationship between signalings, or inserting specific scheduling gaps to smooth the queue fluctuation, thereby generating an optimized scheduling scheme that meets the basic safety requirements of the rule base. This implementation detail provides a preliminary optimized and lower risk input scheme for subsequent more complex micro-simulation deduction through the introduction of rule-based inner loop quick inspection and correction, improving the efficiency and reliability of the entire decision-making process.

[0034] The process of forward-looking deduction by the micro-simulation deduction subunit includes: Constructing a state snapshot based on the current queue state and resource occupation of the system; Placing the optimized scheduling scheme in the state snapshot and simulating the subsequent continuous scheduling periods; Recording the average processing time delay and system throughput efficiency data generated during the simulation process as performance index data.

[0035] It needs to be further explained that in the specific implementation process, the implementation process of forward-looking deduction by the micro-simulation deduction subunit is as follows: after receiving the optimized scheduling scheme from the rule base comparison link, the subunit first triggers a system state snapshot capture mechanism. This mechanism will instantaneously record the key real-time states of the system at the current time, including but not limited to the specific number and attributes of the signaling waiting to be processed in each physical queue, the current load rate of each core of the central processor, the size of the buffer being occupied in the memory, and the instantaneous throughput of the network interface. This state snapshot strives to reflect the real and complete internal situation of the system at the decision moment.

[0036] Next, the micro-simulation deduction engine is activated. The engine takes the optimized scheduling scheme as the input instruction set, and loads it onto the above-mentioned captured system state snapshot, to build a highly realistic but completely controllable simulation scenario. Deduction is not only one step, but also simulates the execution of subsequent continuous multiple complete scheduling periods. In each simulation period, the engine will strictly follow the logic of the scheduling scheme to process the signaling in the queue, while simulating the random arrival of new signaling, the operation time consumption of the processor, and the off-line process after the completion of signaling processing. The whole simulation process is based on the principle of discrete event simulation.

[0037] During the deduction process, the engine will continuously track and record a series of key performance indicators. For example, it will count the average waiting time of all processed signaling in the queue from the beginning to the end of the simulation, and the throughput efficiency embodied by the total number of successfully processed signaling and the time ratio in the whole simulation period. These data are collected and stored in real time. When the preset multiple simulation periods are all executed, the deduction process terminates. At this time, the recorded performance indicators such as average delay and throughput efficiency are summarized to form a predictive evaluation report about the possible effects of the scheduling scheme in the system in the future short term. This evaluation report is the performance indicator data, which will be fed back to the adaptive scheduling decision module to provide important data support for its final decision and self-learning based on forward-looking simulation.

[0038] This implementation detail improves the verification of scheduling decisions from static, single-point evaluation to dynamic, system-level simulation by introducing multi-cycle, system-state-based fine deduction, enhancing the ability to predict the long-term impact of decisions.

[0039] After receiving the performance indicator data, the adaptive scheduling decision module uses it as an immediate reward signal to update its internal strategy model parameters, realizing online adaptive optimization of the decision model. It needs to be further explained that in the specific implementation process, the adaptive scheduling decision module uses the feedback data to implement online optimization: when the performance indicator data generated by the micro-simulation deduction sub-unit is fed back to the decision module, the processing logic inside the module is immediately started. These performance indicator data, such as the average delay and throughput efficiency deduced, are converted into a quantitative immediate reward signal by the evaluation function preset in the decision module. The value of the reward signal reflects the comprehensive performance of the evaluated scheduling scheme in the simulation environment.

[0040] The decision module is a policy model based on a reinforcement learning framework, which receives the instant reward signal as the key evaluation basis for the effectiveness of this decision cycle. The updating process of the model parameters follows the following logic: if the reward signal is positive, indicating that the optimized scheduling scheme performs well in the deduction, such as reduced latency and stable throughput, the model will tend to strengthen the decision logic based on which the scheme is generated, i.e. fine-tune the connection weights inside the neural network, so that the probability of making similar scheduling decisions in the future when encountering similar system states and signaling characteristics increases. Conversely, if the reward signal is negative, indicating that the scheme has deficiencies, the model will weaken the decision path that leads to this result.

[0041] This parameter updating is a continuous and incremental process, rather than a periodic batch adjustment. Each cycle from decision generation to micro-simulation verification to feedback updating constitutes a complete learning period. Through a large number of such periods, the model gradually accumulates experience of which scheduling strategy is more effective under different system states and signaling loads. This implementation makes the scheduling strategy not fixed and not simply a pile of rules, but can continuously evolve with the evolution of the actual running environment of the system, thereby dynamically improving its adaptability and effectiveness under different working conditions, realizing a fundamental change from static configuration to dynamic self-learning, and constituting a continuous guarantee for the intelligent arrangement capability of the system.

[0042] The dynamic attribute parameters collected by the multi-dimensional feature perception and fusion module at least include the signaling service type identifier, the signaling urgency level identifier, the signaling source identity identifier, and the real-time queue depth. It needs to be further explained that in the specific implementation process, the multi-dimensional feature perception and fusion module collects dynamic attribute parameters in the following manner: the module continuously monitors the signaling flow through multiple detection points arranged inside the system, and whenever a new signaling arrives, the module synchronously analyzes the signaling header and payload content to extract the key attributes.

[0043] The signaling service type identifier is obtained by analyzing the preset protocol type field or service class code of the signaling, which is used to distinguish the business category to which the signaling belongs, such as session control, mobility management or emergency alarm. The signaling urgency level identifier is determined by reading the specific priority flag bit or service quality level label carried in the signaling, which directly reflects the urgency requirement for processing the signaling. The signaling source identity identifier is confirmed by verifying the digital signature of the source address of the signaling or the certificate of the access network device, which is used to judge the credibility and importance level of the signaling source.

[0044] The real-time queue depth is directly obtained by the module timing query of the buffer counter of each processing queue, representing the current instantaneous load pressure of the system. After these parameters are collected in parallel, the fusion logic performs standardization and normalization processing to form a unified and structured dynamic data set, providing a comprehensive and real-time state perception basis for subsequent intelligent decision-making.

[0045] The core decision model of the adaptive scheduling decision module adopts a policy model based on a reinforcement learning framework, and its state space is composed of dynamic attribute parameters output by the multi-dimensional feature perception and fusion module. The action space is a set of selectable scheduling strategies. It needs to be further explained that the implementation of the internal decision mechanism of the policy model based on the reinforcement learning framework is as follows: the model constructs a comprehensive state vector after the dynamic attribute parameter set output by the multi-dimensional feature perception and fusion module in real time is normalized. The vector fully describes the current signaling load condition, resource pressure, and inherent characteristics of the signaling flow of the system.

[0046] The decision process of the model begins with the analysis of the current state vector. The policy model, as a neural network pre-trained and with online learning ability, receives the state vector as input. The internal calculation layer performs nonlinear transformation and feature extraction on the input information, and finally generates a probability distribution corresponding to all selectable scheduling strategies in the output layer. The action space is composed of these specific scheduling strategies, such as "prioritize processing of signaling of business type A", "allocate additional resources for signaling source level B", or "use weighted round robin to balance queue load", and combinations of basic scheduling actions. The model determines the scheduling scheme to be adopted in the current period according to the output probability distribution by sampling or selecting the action with the highest probability.

[0047] The entire decision-making process is a learning cycle that continuously interacts with the system environment. After the model executes the selected action, i.e., generates a preliminary scheduling scheme, the scheme is subjected to inspection and evaluation through rule library comparison and micro-simulation deduction. The performance indicators generated by the deduction are converted into reward signals, such as positive rewards for lower latency and negative rewards for queue overflow. This reward signal, together with the new system state, constitutes a complete experience data. The model uses these experience data to adjust its neural network weights through optimization algorithms such as gradient descent, aiming to increase the probability of selecting actions that can obtain higher cumulative rewards in a specific state. Through massive interactive learning, the model gradually masters how to balance multi-dimensional factors under different complex working conditions, thereby autonomously generating efficient and robust scheduling strategies, realizing a fundamental change from relying on fixed rules to having adaptive optimization capabilities in scheduling decision-making.

[0048] The queue execution module reorders or allocates resources to the signaling queue according to the final scheduling instruction. It needs to be further explained that in the specific implementation process, the implementation mode of the queue execution module according to the final scheduling instruction is as follows: the module continuously monitors the output instruction channel of the adaptive scheduling decision module, and immediately starts the execution process when receiving the final scheduling instruction after double verification and optimization. The execution process first parses the instruction to identify the specific operation intention implied in the instruction, such as the need to adjust the order of specific signaling in the queue or the need to reserve processing resources for a certain type of signaling.

[0049] For queue reordering operation, the execution module will directly operate the data pointer in the physical queue or logical queue according to the priority adjustment rule specified in the instruction, changing the processing sequence of the signaling unit. For resource allocation operation, the execution module will interact with the system's resource manager to adjust the quota proportion of computing cores, memory buffers, or network bandwidth between different signaling processing threads according to the instruction requirements. All operations are completed in an atomic transaction manner to ensure the instantaneous integrity and consistency of queue state switching, avoiding signaling loss or state confusion during adjustment. After execution is completed, the module will feedback the operation result state to the system, thus completing the closed loop from intelligent decision to physical execution, ensuring that the arrangement strategy can be accurately implemented.

[0050] The multi-dimensional feature perception and fusion module, the adaptive scheduling decision module, and the queue execution module work together to form a closed-loop "prediction-decision-verification" process to realize the forward-looking intelligent arrangement of real-time signaling flow. It needs to be further explained that in the specific implementation process, the closed-loop process formed by the whole system working together is reflected in the continuous data flow and feedback interaction between modules. The multi-dimensional feature perception and fusion module, as the starting point of the process, continuously collects and outputs dynamic data reflecting the real-time state of the system. These data are received by the adaptive scheduling decision module to start its double verification decision process including preliminary generation, rule comparison, and micro-simulation deduction.

[0051] In this process, the preliminary decision first accepts rapid risk screening and correction based on the rule base, and then the optimized scheme is subjected to multi-cycle forward-looking deduction evaluation based on the system state snapshot. The performance data generated by the deduction is not only used for the final scheduling instruction, but more importantly, this part of data is immediately fed back to the core strategy model of the adaptive scheduling decision module for online fine-tuning of its parameters, and also fed back to the short-term traffic prediction sub-module for correction of its prediction algorithm. Finally, the confirmed scheduling instruction is implemented by the queue execution module, thereby affecting the actual running state of the system.

[0052] This process, from perception to decision, from verification to execution, and from execution effect feedback to perception and decision starting point, runs repeatedly. It makes the system no longer a static pipeline that passively responds to the signaling flow, but turns into an organic whole that can learn from history and real-time interaction, can predict the short-term future state, and dynamically optimize its scheduling strategy accordingly. This closed-loop, learning and forecasting-capable "prediction-decision-verification" process is the core mechanism to realize the forward-looking intelligent arrangement of real-time signaling flow, ensuring the adaptability and robustness of the system in the face of complex and variable network environment.

[0053] It should be further pointed out that in the implementation process of the real-time signaling flow intelligent arrangement system based on queue optimization described in the present application, it is realized through a series of closely coordinated steps. The operation of the whole system starts from the continuous perception of the multi-dimensional characteristics of the signaling flow. The perception module built in the system will analyze each input signaling in real time and extract its inherent attribute information, which includes but is not limited to the type of service represented by the signaling, the emergency level marked by the signaling itself, the identity of the signaling source device, and the current length of each processing queue in the system at the time of signaling arrival. After all these parameters are collected and standardized, a comprehensive dynamic data set is formed, which fully describes the real-time load condition and the characteristics of the signaling flow faced by the system.

[0054] Next, the core decision-making mechanism of the system begins to operate: an adaptive scheduling decision maker based on online learning algorithm receives the above dynamic data set. The decision maker first uses its internal strategy model to quickly generate a preliminary signaling processing sequence scheme based on the current data. However, this preliminary scheme is not directly adopted. The system designs a key double verification process to improve the robustness and foresight of the decision. The first double verification is to compare the preliminary scheme with a preset rule base. The rule base defines a series of basic rules that the system needs to follow to run stably, such as ensuring the upper limit of the time delay of high-priority signaling and preventing excessive fluctuation of queue length. The comparison process is carried out in a simplified simulation environment to evaluate whether the preliminary scheme may violate these core rules. If potential risks are found in the evaluation, the system will automatically adjust the preliminary scheme according to the guidance of the rule base to generate an optimized alternative scheme.

[0055] The subsequent second-stage verification is more refined. The system captures a precise snapshot of the current state of all key components, including queue contents, processor load, and memory usage, constructing a highly realistic simulation environment. The scheduling scheme optimized through the first-stage verification is placed in this environment, and its execution effect over several consecutive work cycles is simulated. This micro-simulation process can proactively predict the impact of the scheme on the overall system performance, such as the changing trends of average signaling processing latency and system throughput efficiency. The performance prediction data generated after the simulation is immediately fed back to the adaptive scheduling decision-maker. The decision-maker uses this data as a learning signal to fine-tune the parameters of its internal policy model, enabling subsequent decisions to better optimize the overall system performance. Simultaneously, the scheduling instructions deemed feasible in this simulation are finally confirmed.

[0056] Finally, the confirmed scheduling instructions are sent to the queue execution unit. Based on the specific content of the instructions, this unit reorders the signaling in the physical queue or adjusts the resource allocation strategy, thereby translating intelligent decision-making into actual scheduling actions. The entire process, from feature perception, preliminary decision-making, rule verification, simulation deduction to learning optimization and instruction execution, forms a complete, closed-loop automated intelligent chain. Through this integration of multi-level verification and online learning mechanisms, the system achieves a transformation from passive response to proactive prediction and optimization of real-time signaling streams, effectively improving the intelligence level and reliability of signaling scheduling in complex network environments.

[0057] The core of this invention lies in the working mechanism of its adaptive scheduling decision module: the decision-making basis of this module is a comprehensive state vector, which is constructed from real-time data output by the multi-dimensional feature perception and fusion module. State vector S t It can be represented as: S t =(T1, T2, ..., T m P1, P2, ..., P n ;Q depth Res util ); where S t The system's overall state vector at time t, where t represents the time node of the current scheduling decision, T1 to T2. m The number of arrivals of m different service types of signaling within a specific time window in the past; P1~P n Q: The number of signaling entries of n different priority levels in the current system queue; depth : The current depth of the system's critical business queue, Res util : Current utilization of the system's core processing resources, which include CPU and bandwidth, with values ​​ranging from [0, 1]; State vector S t The construction integrates dynamic traffic characteristics with static system states.

[0058] A parameterized policy function π(a|S t ; θ) is adopted inside the decision module, where π(·) is a parameterized scheduling policy function; a is a selectable scheduling action belonging to the action space A, such as “prioritize processing signaling with priority X” “allocate resource Z for service Y” and the like; S t is the current system state vector, i.e. the input variable; θ is the parameter set of the policy model, the core parameters to be optimized, such as neural network weights, decision tree node thresholds and the like; the meaning of the policy function is: under the condition of the given current state S t and model parameters θ, the probability distribution of selecting each selectable scheduling action a is output. The action space A includes basic scheduling instructions such as “prioritize processing signaling with priority X” “allocate additional resource Z for service type Y” and combinations thereof.

[0059] The decision-making process introduces an online learning mechanism based on a reward signal: after a scheduling action a t is executed, i.e. after subsequent inspection and implementation, the system observes the new system state S t+1 at the next decision point t+1, and calculates an immediate reward R t . The design of this reward function R t takes into account the scheduling effect, and its formula is: R t = w1·(-ΔDelay high ) + w2·(ΔThroughput) - w3·(ΔQ variance ); where R t : the immediate reward value obtained after the action a t is executed at time t, the larger the value, the better the action effect; w1, w2, w3: preset weight coefficients, non-negative real numbers, used to balance the importance of different performance indicators, satisfying w1+w2+w3=1 or adjusted according to business needs; ΔDelay high : the average delay change of high-priority signaling from t to t+1, where Δ<0 indicates a decrease in delay, corresponding to a positive reward contribution; ΔThroughput: the change in system throughput from t to t+1, where Δ>0 indicates an increase in throughput, corresponding to a positive reward contribution; ΔQ variance : the change in queue length fluctuation from t to t+1, Δ<0 indicates that the queue is more stable, corresponding to a positive reward contribution; the quantitative value of the reward R t directly reflects the good or bad of the scheduling action a t under the state S t .

[0060] The update of policy model parameter θ is the core of system implementation of self-adaptation. The system adopts a policy gradient method for optimization, and a basic form of its update rule is: θ ← θ + α · ∇ θ logπ(a t |S t ;θ) · G t ; wherein, θ ← θ +...: parameter update symbol, indicating "assign the calculation result on the right side to θ on the left side"; α: learning rate, positive real number, controlling the step size of parameter update, taking value range [10 -5 , 10 -2 ], avoiding oscillation caused by too large step size or slow convergence caused by too small step size; ∇ θ logπ(a t |S t ;θ): the logarithm of the gradient of the policy function with respect to the parameter θ, indicating how to adjust θ to increase the probability of selecting action a t in state S t ; G t is the cumulative future discounted reward calculated from time t, representing the long-term value of the current decision; a t : the scheduling action actually executed at time t; by continuously updating the parameter θ according to the above rule, the policy model gradually learns what scheduling action should be taken in what system state to achieve long-term optimal performance.

[0061] In addition, the implementation details of the short-term traffic prediction sub-module can be further specified. This module can use time series prediction methods such as autoregressive integrated moving average model. The basic idea is to use historical data Y t , Y t-1 , Y t-2 ,... to predict future values Y t+h ; wherein, Y t-k represents the signaling flow observation value at time t-k; prediction output: Y t+h : the signaling flow value at time t+h predicted based on historical data, h is the prediction step.

[0062] The model makes predictions by identifying trends and seasonal patterns in the data. The prediction results, such as the expected arrival intensity I forecast of a certain type of signaling in the future, are normalized to a dynamic weight factor W forecast between zero and one, and integrated into the construction of the state vector S t , or directly used as additional input to the policy function π, so that the decision has foresight; wherein, dynamic weight conversion: W forecast = Norm(I forecast ), wherein: I forecast: future expected arrival intensity of a certain class of signaling; Norm(·): a normalization function, such as Min-Max normalization, to make W forecast ∈ [0, 1]; W forecast : final fusion state vector S t or the prediction weight factor of policy function π.

[0063] The operation of the micro-simulation reasoning sub-unit relies on the discrete event simulation technique. It maintains a list of future events, and accurately calculates the performance index output of the system under a specific scheduling scheme by simulating the advancement of events such as signaling arrival, scheduler decision, signaling processing completion and departure on the virtual timeline. This event-based simulation can effectively capture the randomness and nonlinear characteristics of queue behavior in signaling flow.

[0064] By deeply integrating reinforcement learning, time series prediction and discrete event simulation techniques, an intelligent orchestration system with self-optimization and forward-looking decision-making capability is constructed.

[0065] By introducing a double-loop intelligent architecture of prediction and decision verification, the forward-looking and reliability of signaling flow scheduling are effectively improved. The system can make decisions by integrating multi-dimensional dynamic factors and undergo double verification by rule library and micro-simulation, so that it can maintain the rationality of scheduling strategy and system stability when facing complex burst traffic, and guarantee the end-to-end service quality.

[0066] In addition, the system has online learning and self-adaptive optimization capability, and can continuously adjust the scheduling strategy according to the historical decision effect. This self-evolution mechanism reduces the dependence on artificial preset rules and improves the adaptability of the system under different network conditions, providing effective support for the intelligentization and automation of communication network operation.

[0067] It should be noted that in this paper, relationship terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0068] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A real-time signaling flow intelligent orchestration system based on queue optimization, characterized in that, The application relates to a multi-dimensional feature perception and fusion module for collecting and outputting multiple dynamic attribute parameters of signaling in real time. An adaptive scheduling decision module is connected to the multi-dimensional feature perception and fusion module and is used for receiving the dynamic attribute parameters and generating scheduling instructions. A queue execution module is connected to the adaptive scheduling decision module and is used for executing the scheduling instructions. The adaptive scheduling decision module comprises a two-stage decision unit and an internal inspection unit. The two-stage decision unit is used for generating a preliminary scheduling scheme based on the currently received dynamic attribute parameters. The internal inspection unit comprises a rule library comparison subunit and a micro-simulation deduction subunit. The rule library comparison subunit is used for comparing the preliminary scheduling scheme with a reference scheme in a preset rule library, and outputting an optimized scheduling scheme to the micro-simulation deduction subunit according to a comparison result. The micro-simulation deduction subunit is used for performing forward-looking deduction on the optimized scheduling scheme based on a current state snapshot of the system, feeding back performance index data obtained through the deduction to the adaptive scheduling decision module to adjust decision parameters of the adaptive scheduling decision module, and outputting a finally determined scheduling instruction to the queue execution module. The multi-dimensional feature perception and fusion module comprises a short-term traffic prediction sub-module.

2. The real-time signaling flow intelligent arrangement system based on queue optimization according to claim 1, characterized in that: The short-term traffic prediction sub-module is used for analyzing historical signaling flow data, predicting an arrival trend and a correlation of signaling flow in a future set time window, and outputting a prediction result as a set of dynamic weight factors to the adaptive scheduling decision module. The prediction model used by the short-term traffic prediction sub-module is an online learning time series analysis model.

3. The real-time signaling flow intelligent arrangement system based on queue optimization according to claim 2, characterized in that: The performance index data obtained through the deduction of the micro-simulation deduction subunit is synchronously fed back to the short-term traffic prediction sub-module to correct a prediction deviation. The logic of the rule library comparison subunit for comparison comprises the following steps:

4. The real-time signaling flow intelligent arrangement system based on queue optimization according to claim 1, characterized in that: The preliminary scheduling scheme is run in a simulation environment, and high-priority signaling time delay and queue state fluctuation risks possibly caused by the preliminary scheduling scheme are evaluated. When the evaluation risks exceed a preset threshold, a strategy adjustment mechanism of the adaptive scheduling decision module is triggered to generate the optimized scheduling scheme. The process of forward-looking deduction performed by the micro-simulation deduction subunit comprises the following steps:

5. The real-time signaling flow intelligent orchestration system based on queue optimization according to claim 1, characterized in that: A state snapshot is constructed based on a current queue state and resource occupation condition of the system. The optimized scheduling scheme is placed in the state snapshot, and a subsequent continuous multiple scheduling periods are simulated and run. Average signaling processing time delay and system throughput efficiency data generated in the simulation and running process are recorded as the performance index data. After receiving the performance index data, the adaptive scheduling decision module takes the performance index data as an instant reward signal to update strategy model parameters in the adaptive scheduling decision module, so that online adaptive optimization of a decision model is realized.

6. The real-time signaling flow intelligent orchestration system based on queue optimization according to claim 5, characterized in that: The dynamic attribute parameters collected by the multi-dimensional feature perception and fusion module at least comprise a service type identifier of signaling, an emergency degree identifier of signaling, an identity identifier of a signaling source and a real-time queue depth.

7. The real-time signaling flow intelligent orchestration system based on queue optimization according to claim 1, characterized in that: ​ 8. The real-time signaling flow intelligent arrangement system based on queue optimization according to claim 1, characterized in that: The core decision model of the adaptive scheduling decision module adopts a policy model constructed based on a reinforcement learning framework, a state space of the policy model is constituted by dynamic attribute parameters output by the multi-dimensional feature perception and fusion module, and an action space is a scheduling policy set.

9. The real-time signaling flow intelligent arrangement system based on queue optimization according to claim 1, characterized in that: The queue execution module performs reordering or resource allocation operations on the signaling queue according to the final scheduling instruction. 10.The real-time signaling flow intelligent arrangement system based on queue optimization according to any one of claims 1 to 9, characterized in that: The multi-dimensional feature perception and fusion module, the adaptive scheduling decision module and the queue execution module work cooperatively to form a closed-loop "prediction-decision-verification" process.

Images