Network slice optimization method of 5G private network

Abstract

The invention relates to a network slice optimization method for a 5G private network, and the technical scheme comprises the steps: constructing a time sequence performance model in a slice operation process, and monitoring tiny changes including end-to-end time delay, throughput and packet loss rate; then extracting a short-period fluctuation mode of the index, and comparing the gradient with a historical stability threshold value; judging whether the index is in a performance critical state or not by identifying an index drifting behavior which continuously jitters but does not default yet; non-equal-ratio responses including jitter improvement and delay decline after each time of resource adjustment are recorded; then trend fitting is carried out on the data, and whether a performance platform period or a critical jump point exists or not is judged; delimiting three intervals of resource investment in the model: an inactive area with invariable investment, a high-sensitivity area with rapid performance change and a rising saturation area with stable performance and continuous resources, wherein the three intervals include an inactive area with invariable investment, a high-sensitivity area with rapid performance change and a rising saturation area with stable performance; and identifying the current response area of the slice through the model, determining that the regulation strategy does not take equal proportion input as the core, and taking the optimal response area as a resource allocation interval to make a choice.

Description

Technical Field

[0001] The present invention relates to a network slicing optimization method for a 5G private network. Background Art

[0002] Although certain achievements have been made in network slicing optimization in terms of slice isolation, resource abstraction, and on-demand deployment, there are still many technical bottlenecks and systematic deficiencies in actual deployment and operation. These problems are particularly prominent in private network scenarios such as multi-service concurrency, high reliability, low latency, and resource constraints. First of all, most current network slicing optimization methods are still based on static configuration or quasi-static strategies, lacking a truly dynamic optimization mechanism for the entire life cycle. In the early stages of network slicing creation, the system allocates resources based on preset business templates, service levels, and estimated loads. Although it can meet the initial deployment requirements, it is difficult to cope with the fluctuations in resource demand caused by subsequent dynamic changes in business, especially in manufacturing, transportation, medical, and other scenarios where business is highly sudden and load periodic fluctuations are significant. The original resource configuration often appears to be "over-allocated in the early stage, lagging in the middle stage, and rigid in the later stage", resulting in low resource utilization and inability to adjust intelligently according to business evolution. Secondly, the current mainstream slice optimization mechanism is highly dependent on SLA breach detection as a trigger signal, that is, only when key performance indicators such as latency and packet loss rate exceed the threshold defined by the SLA, the system will start scheduling behaviors such as resource expansion or policy reconstruction. This "passive response" has obvious lags and cannot identify and intervene in critical performance or sub-healthy conditions in advance, which can easily cause short-term fluctuations in service quality or jitter accumulation, and may even lead to irreversible business interruption.

[0003] In addition, existing methods generally lack the ability to continuously model the performance fluctuation trend of slices, ignoring the early drift phenomenon when the indicator has not yet breached the contract but has continuously deviated from the stable state. Although this "sub-drift" state does not cause an immediate alarm, it usually indicates that the resource allocation efficiency is declining or external interference is accumulating. If it cannot be identified and intervened in time, it will cause problems to break out in the later stage. Furthermore, most of the current resource scheduling mechanisms adopt proportional increase and decrease, priority occupation or static weight strategies, which fail to fully explore the nonlinear relationship between resources and performance. For example, after reaching a specific resource threshold, the performance improvement of some services tends to be platform-based, but the system continues to allocate resources, resulting in extremely low marginal benefits or even reverse interference; for example, some resource dimensions (such as cache, queue scheduling priority) have delayed and cross-interference characteristics on performance. Existing strategies often find it difficult to accurately identify "low-sensitivity and high-consumption" resource allocation, which ultimately leads to the coexistence of redundant resources and delayed performance response. In addition, current optimization methods mostly use single slices as scheduling objects, lack the ability to conduct collaborative analysis across slices and indicators, and cannot effectively identify the linkage risks caused by shared links, edge computing nodes or scheduling strategies between multiple slices. In high-density slice deployment scenarios, systemic problems such as "local optimization and global disorder" are prone to occur. What's more serious is that when the performance of multiple business slices degrades at the same time and there is no clear causal path, it is difficult for the existing system to determine whether it is an individual problem or a system structural bottleneck, which leads to errors in scheduling strategy selection or confusion in priority setting. Summary of the invention

[0004] The purpose of the present invention is to provide a network slicing optimization method for 5G private networks, so as to solve some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0006] S1. Use a performance drift perception mechanism based on slice boundary status:

[0007] S1.1. Build a time series performance model during the operation of the slice to monitor small changes in end-to-end latency, throughput, and packet loss rate. Then extract the short-term fluctuation pattern of the indicator and compare the gradient with the historical stability threshold.

[0008] S1.2. By identifying the indicator drift behavior that is continuously jittering but has not yet violated the contract, it is determined whether it is in a critical performance state;

[0009] S2. Dynamic modeling of the nonlinear relationship between slice services and resources:

[0010] S2.1. Record the non-proportional response after each resource adjustment, including jitter improvement and latency reduction. Then perform trend fitting on the data to determine whether there is a performance plateau or critical jump point.

[0011] S2.2. Define three intervals for resource input in the model: an ineffective area with unchanged input, a highly sensitive area with rapidly changing performance, and an ascending saturation area with stable performance but continuous resource input.

[0012] S2.3. Identify the current response area of the slice through the model, and determine that the control strategy does not focus on equal - proportion input, but selects the optimal response area as the resource allocation interval.

[0013] S3. Adopt a resource compression control mechanism based on the performance maintenance area:

[0014] S3.1. Determine the area where each slice is currently located in the non - linear response curve; if it is in the performance saturation area or the plateau period, it means that some resources are redundant at this time; start the progressive resource compression process and slightly recover from the edge resource units.

[0015] S3.2. After each compression, real - time detect whether the performance remains within the set range; if the performance does not decline, continue the next round of compression until the performance boundary is touched, and record the compression limit value as the resource retention point of the slice.

[0016] S3.3. Score each slice, and then predict the available resource trend in the future period according to the current state of the resource pool and the slice weight; then construct the scheduling priority based on the stability entropy between slices.

[0017] Furthermore, the construction of the performance drift perception mechanism based on the slice boundary state includes:

[0018] During the operation of the slice, continuously collect key performance indicators at the second - level or sub - second - level granularity to form time - series data; retain the original data at each moment for analyzing edge perturbations; to characterize the continuity of perturbations, introduce a perturbation intensity function Cv to measure the abnormal fluctuation trend of indicators within a short period:

[0019]

[0020] Where:

[0021] Cv represents the perturbation continuity intensity of a certain performance indicator of the slice in the time period [t, t + ΔT]; ΔT is the time length of the short - period analysis window; X(t) is the value of a specific performance indicator at time t; represents the change rate of the indicator at time t; μ is the expected value of the change rate calculated by the system based on historical stable data; |·| represents the absolute deviation degree of the change.

[0022] Furthermore, the construction of the performance drift perception mechanism based on the slice boundary state includes:

[0023] Build a stable baseline model for each metric; record the reasonable fluctuation range of each metric during stable operation, analyze the coupling relationship between metrics to identify systemic linkage risks; define the perturbation response coupling strength S between metrics ij , which is used to characterize the influence sensitivity of one metric on another metric:

[0024]

[0025] where:

[0026] S ij represents the average response coupling degree of the i-th metric to the j-th metric; N is the number of historical sample points; α ij is the coupling importance weight factor for the i,j pair of metrics; is the local change rate of metric X i relative to X j at the k-th time point, which is used to measure the linkage trend.

[0027] Furthermore, the construction of the performance drift perception mechanism based on the slice boundary state includes:

[0028] Based on the perturbation characteristics in each analysis period, construct a perturbation scoring function D(t) to uniformly evaluate whether the current slice has entered the performance critical state:

[0029]

[0030] where:

[0031] D(t) represents the comprehensive score of the overall perturbation risk of the slice at time t; M is the total number of performance metrics; ω i is the weight factor of the i-th metric; ΔV i (t) is the difference between the maximum and minimum values of the i-th metric within the current window, which measures the fluctuation amplitude; is the maximum allowable fluctuation value of this metric during the stable period; Fv i (t) is the jump frequency, that is, the number of times the metric exceeds the threshold, representing the intensity of fluctuations; γ i is the normalization constant of the jump frequency; Cv i (t) is the perturbation continuity strength of this metric within the current period; θ i is the normalization coefficient of perturbation continuity;

[0032] When the value of D(t) exceeds the critical threshold D set according to historical modeling crit , it indicates that the slice is currently in the performance drift area, that is, although the SLA hard alarm has not been triggered, it has deviated from the stable state and is about to enter the abnormal area.

[0033] Further, the method for constructing a resource compression regulation mechanism based on a performance retention area is as follows:

[0034] Based on the resource allocation and performance feedback data of each network slice at different time periods, construct a response relationship curve between the multi-dimensional resource input and the key performance indicators of the slice. The response relationship curve is used to characterize the non-linear characteristics of the slice's performance with respect to the change in resource input. In the response relationship curve, identify the section where the current resource level is located according to the function:

[0035]

[0036] where R s is the marginal performance response score of the current resource allocation of the slice, represents the instantaneous change rate of the performance indicator P with respect to the resource input R. ∈ is the marginal benefit threshold set by the system, κ(t) is the time-weighted function set for the business importance, and φ(·) is a piecewise response function used to filter out the change rate values below the benefit threshold. If R s is less than the set minimum effective response value, it means that the current slice resource is in the performance saturation area or the platform area, and the resource compression process can be entered.

[0037] Further, the compression process adopts a progressive compression strategy, selects the dimension with the smallest weight in the resource dimension for minor adjustment. After each resource compression, the core performance indicators are collected in real time within a set window, and stability detection is performed based on the performance retention function defined as follows:

[0038]

[0039] where S k represents the perturbation stability of the k-th type of performance indicator within the observation period, and σ k (t) is the real-time value of the k-th performance indicator at time t, is the mean value of this indicator in the baseline stable state, and η k (t) is a dynamic sensitivity weight function constructed based on the historical volatility, used to amplify the abnormal perturbation response.

[0040] Further, in the compression process, when the S after two consecutive rounds of compression k is less than the preset fluctuation tolerance threshold, continue to the next round of compression until the S after a certain round of compression k exceeds the threshold range. Then, the resource configuration of the previous compression in this round is recorded by the system as the resource retention point of the slice. After the system records the retention point, a future resource transferable trend model is constructed based on the service level, historical stability, and resource usage efficiency of each slice, and a transferable resource prediction value is generated based on the following resource scheduling trend function:

[0041]

[0042] Among them, F r represents the overall resource release trend intensity of the slice, ω i is the system weight of the i-th type of resource, represents the release speed per unit time of this type of resource during the compression process, δ i is the sensitivity adjustment factor of this resource, ψ i (t) is a trend function reflecting the stable occupancy rate of the resource, T i is the length of the historical compression observation window of the corresponding resource. Finally, the system comprehensively sorts according to the predicted F r and the resource retention points, safety performance scores, and historical fluctuation entropy indicators of each slice to complete the dynamic adjustment of the resource reallocation priority.

[0043] The 5G private network network slice optimization method proposed by the present invention has significant engineering value and system-level innovation, and its beneficial effects are mainly reflected in the following aspects:

[0044] By constructing a time series performance model and a perturbation intensity scoring mechanism, it can identify the "performance drift" state in advance when the slice has not yet experienced SLA default, thereby effectively avoiding the scheduling lag and service interruption risks caused by delayed judgment in traditional methods; secondly, by establishing a coupling relationship model between performance indicators, the present invention can accurately identify the linkage interference mechanism between cross-indicators and cross-slices, provide a basis for multi-dimensional collaborative regulation for the system, and improve the overall resource coordination and service stability;

[0045] Again, based on the resource value evaluation mechanism of non-linear performance response, it identifies the "performance saturation area" and "marginal failure area" in resource allocation, avoids resource redundancy and waste, and realizes the optimal compression of resources on the premise of performance maintenance through a progressive compression strategy combined with real-time feedback of the perturbation function, enhancing the system elasticity and resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flowchart of the network slice optimization method for the 5G private network of the present invention.

[0047] Figure 2 is a flowchart of the construction method of the performance drift perception mechanism for the slice boundary state of the present invention.

[0048] Figure 3 is a flowchart of the construction method of the resource compression regulation mechanism for the performance maintenance area of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0049] The following will give a detailed description of the specific implementation manners of the present invention with reference to the drawings.

[0050] Combined with the attachedFigure 1 As shown, first, a high-resolution time series performance model is constructed around each running network slice. This model conducts continuous and fine-grained data sampling and recording of the key performance indicators during the operation of the slice. The covered indicators include end-to-end transmission delay, service throughput, and packet loss rate, etc. These indicators directly reflect the dynamic state of service quality and slice load in the 5G private network scenario. Different from traditional monitoring methods that only record the periodic average value or extreme value, this model continuously tracks the performance at a frequency of seconds or sub-seconds to ensure that small fluctuations, edge anomalies, and mutation behaviors are not overlooked, providing a sufficient data basis for subsequent trend judgment and risk warning. After obtaining the continuous performance time series, the system divides it into multiple short-cycle windows (such as every 30 seconds) for local analysis. In each analysis window, fluctuation features are extracted, including the difference between the maximum and minimum values of the performance indicators within this period (i.e., the fluctuation amplitude), the frequency of significant jumps, the duration of continuous jitters, and the direction and rate of indicator trend changes. These local perturbation features are constructed into a set of high-dimensional fluctuation mode vectors.

[0051] Next, the system compares and analyzes the fluctuation features extracted in the current observation period with the performance baseline established during the historical stable operation period of this slice. This baseline is obtained from long-term operation data statistics and includes the normal fluctuation range of each indicator in the "stable state", the jitter frequency threshold, and the co-variation pattern between indicators, etc. In this way, the system can identify whether the current fluctuation belongs to an abnormal deviation. Particularly crucial is that during the comparison process, the system not only judges whether it exceeds the hard performance threshold set by the SLA, but further focuses on identifying those "edge states that have not defaulted but have shown continuous drift behaviors", that is, identifying the "critical state" slices near the boundary of the performance stable area. This critical state usually manifests as a slight but continuous deviation of the performance indicators from the stable baseline. For example, the end-to-end delay slowly increases but does not exceed the maximum tolerance value of the SLA, or the packet loss rate remains low and fluctuates but the frequency increases, etc. The system confirms the credibility of these "sub-abnormal" states through the trend slope, jitter continuity, and co-variation pattern between indicators in the time series, and finally forms a judgment on the critical state of the slice performance, and transmits this state as an input signal to the resource scheduling or architecture optimization module to trigger actions such as pre-scheduling, priority adjustment, or architecture reconstruction, so as to achieve the pre-guarantee of slice performance and the improvement of resource utilization efficiency.

[0052] The system records and monitors the response of each resource adjustment behavior, paying special attention to whether the performance feedback brought by the resource adjustment presents a non-proportional and nonlinear relationship. For example, after increasing bandwidth or computing resources, whether there are changes in indicators such as jitter improvement, latency reduction, and packet loss rate compression, and whether these changes are proportional to the amount of resources invested or have marginal change characteristics. These data will be written into the resource-performance response log after each adjustment and form a structured dynamic response data set. Based on this data set, the system performs trend fitting and morphological modeling on the relationship between investment and performance indicators of different resource dimensions, with the aim of analyzing whether the resource investment has entered the so-called "performance plateau" or "critical jump point", that is, whether the performance change has stabilized or is about to reach an inflection point, so as to determine whether further resource investment is still effective. The system sets three types of resource-performance intervals in the response model: the first is the invalid zone, that is, the resource investment increases but the performance is almost unchanged, indicating that the current resource configuration has exceeded the demand;

[0053] The second is the high-sensitivity area, that is, fine-tuning resources will bring about a significant improvement in performance, which is the most critical interval worthy of priority optimization; the third is the saturation area, that is, although the performance remains at a high level, the continuous investment of resources no longer brings actual performance benefits, and may even cause performance regression due to interference or scheduling queues. Through this model, the system can accurately identify the response segment where each slice is currently located, thereby providing strategic suggestions for the scheduler. For example, if the slice is identified as being in the invalid area, stop continuing to invest resources and start the compression mechanism; if it is in the high-sensitivity area, give priority to ensuring the resource elasticity of the slice; if it is in the saturation area, evaluate its marginal benefits and then recycle resources. Overall, this method abandons the traditional resource scheduling ideas based on equal proportions or fixed quotas, and instead identifies the segments based on the resource-performance response curve, defines the "optimal response area" as the decision-making core of resource allocation, and implements a control strategy that maximizes performance benefits under the premise of limited resources.

[0054] By constructing a resource-performance non-linear response model, the performance response section where each slice is located under the current resource configuration is identified. This model reflects the real relationship between resource input and performance improvement, and can determine whether the current state is in the highly sensitive area with rapid performance improvement, the plateau area where performance change slows down, or the area where performance has saturated or even regressed. When the system determines that a certain slice is in the plateau area or saturation area, it means that some of its resource configurations have exceeded the effective range of performance gain, that is, resource redundancy. Therefore, the system will trigger a progressive resource compression mechanism, aiming to maximize resource utilization without affecting the business experience. During the compression process, a strategy of giving priority to the edge and recycling low-risk resources first is adopted, that is, resource units with relatively small impact on performance are selected from dimensions such as computing resources, bandwidth, and cache for small-scale decreasing operations. After each round of compression, the system enters a short-cycle performance detection state, and key performance indicators such as end-to-end delay, throughput, and packet loss rate are collected in real time and compared with the reference value before compression. If the performance remains within the fluctuation tolerance range set by the system, it indicates that the compression has not had a substantial impact on the business, and then the system enters the next round of smaller-scale compression operations. This iteration continues until a certain round of compression triggers the performance boundary, that is, the performance significantly drops for the first time or exceeds the stable range. At this time, the system immediately terminates the compression process and records the resource configuration state that did not affect the performance of the slice last time as the "resource retention point", which represents the minimum available resource standard for the slice in the current business scenario. Subsequently, resource dynamic allocation will be based on this. After completing the compression process, the system will comprehensively score each slice according to dimensions such as the business level, resource usage efficiency, and historical performance stability of each slice, and at the same time, combined with the current overall state and resource usage trend of the resource pool, predict the resource release potential of each slice within a future period. This prediction model not only refers to the current resource load, but also includes compression sensitivity, business fluctuation cycle, and policy priority, so as to determine which slices can become resource release sources in the short term. On this basis, the system introduces the concept of "stability entropy" to model the performance fluctuation risk between slices. Stability entropy is used to measure the fluctuation range and uncertainty of performance indicators of a certain slice during long-term operation. The lower the entropy value, the more stable the slice performance, and the more suitable it is to undertake the task of resource elastic adjustment.

[0055] Example 1:

[0056] Combined with the appendix Figure 2As shown in the figure, a manufacturing enterprise has deployed a local 5G private network to control five intelligent production lines. Each production line corresponds to a service slice, which includes different service types such as machine vision recognition, robotic arm control, and automatic quality inspection. Among them, the service slice S1 for controlling the movement of high-speed robotic arms is the most sensitive to latency and stability. The upper limit of the SLA for the end-to-end latency of S1 set by the system is 30 ms. To ensure real-time performance, S1 is usually allocated more resources. The goal of the present invention is to accurately identify the performance critical state, determine whether the resources are redundant, and perform advance scheduling and optimization without affecting performance.

[0057] During the operation of the system, the end-to-end latency data of S1 is collected at a granularity of 100 ms to form a time series X(t). We select an analysis window of ΔT = 5 s, and a total of 50 points are sampled. The following are some of the latency sampling data (unit: ms) in the most recent analysis period: 22.1, 22.4, 22.6, 22.5, 23.0, 23.6, 24.1, 24.3, 24.9, 25.5,..., 28.2. It can be observed that the latency gradually increases, but has not exceeded the SLA threshold. The system first calculates the expected value μ of the index change rate of this service slice in the normal state based on the data in the historical stable period. After statistics, the average change rate in the stable period is μ = 0.12 ms / s. For the current period, the average latency change rate of this period can be calculated through spline fitting or the difference method as This change rate is significantly higher than the steady-state expectation. To characterize this continuous upward behavior, the system calculates the perturbation intensity function Cv and substitutes it into the formula:

[0058]

[0059] For simplicity of calculation, it can be approximately estimated in discrete form:

[0060]

[0061] where Δt = 0.1 s and N = 49. By substituting the data for calculation, it is found that the change rates at most time points are concentrated between 0.4 and 0.5, and the average deviation from the expected value of 0.12 is about 0.34. Therefore

[0062] Cv ≈ 0.34 ms / s

[0063] According to the system experience rule, when Cv exceeds 0.25, it is considered that the slice has a continuous deviation trend, but it has not reached the SLA default state. Therefore, at this time, slice S1 is determined to be in the "performance critical state". According to the policy configuration, the system reports this drift event to the scheduler, indicating to the scheduling system that the current resource investment has not resulted in significant performance improvement, but the latency is gradually approaching the SLA boundary. Two types of scheduling actions can be considered: First, increase the temporary network bandwidth or computing priority for this slice to prevent further drift from leading to default preemptively.

[0064] Second, the system parallelly starts the resource structure analysis mechanism to determine whether there are other low-priority slices preempting the critical link, causing indirect interference. The system discovers in the historical record that S3 (quality inspection task slice) performs high-frequency image uploads during the same time period, and its uplink bandwidth occupancy suddenly increases. Therefore, the system issues a speed limit control instruction to S3 through the slice boundary management logic, reducing its uplink rate from 30 Mbps to 15 Mbps. At the same time, the system assigns fine-tuned edge computing node cache priorities to slice S1, thereby restoring the latency of S1 to about 23.5 ms and stabilizing it. From this example, it can be seen that the perturbation intensity perception mechanism in the present invention not only has the ability to identify potential performance degradation in advance, but also can actively intervene by linking other system policies based on the quantitative calculation results, avoiding service losses caused by misjudgment or delayed scheduling.

[0065] Regarding the value range of the key parameters, ΔT is recommended to be 2 to 10 seconds, which is applicable to different levels of business cycles; μ is dynamically estimated from historical samples, but the system can set a maximum reference offset upper limit not exceeding 0.2 ms / s; the perturbation threshold Cv crit can be configured according to the business importance. Generally, it is recommended to be set to 0.25 to 0.35 ms / s. If it is too high, the warning will be delayed; if it is too low, false alarms may be triggered.

[0066] In the next step of this embodiment, in order to determine whether the performance drift of slice S1 is due to a single indicator anomaly or affected by other systematic factors, the system starts to execute the coupled risk identification method based on the indicator coupling relationship. At this stage, the system constructs a stable baseline model for each performance indicator, records its fluctuation range during the stable period, and analyzes the coupling relationship between each performance indicator, including end-to-end latency (X1), throughput (X2), and packet loss rate (X3).

[0067] During the past 72 hours of stable operation, the system sampled these three indicators once per second, obtaining 259,200 sets of historical sample points. Therefore, N = 259,200. The system calculates the perturbation response coupling strength S ij between each pair of indicators to quantify the dynamic coupling degree between the indicators. Taking the calculation of S 13 as an example, that is, the coupling strength of the end-to-end latency X1 to the packet loss rate X3, and setting the weight factor α13 = 0.6, which represents that in the manufacturing operation, the impact of latency variation on packet loss is moderately strong. In the historical samples, the system samples the corresponding time points per second of These local change rates are calculated from the derivatives fitted by historical time series data. The system analysis finds that in about 65% of the time periods in the samples, the two show positive synchronous fluctuations, and the derivative values are concentrated between 0.3 and 0.5; in about 20% of the time periods, the derivative values are close to zero, indicating no significant correlation; while in the remaining 15% of the time periods, there is an inverse relationship or fluctuating deviation. Substitute these data into the formula:

[0068]

[0069] Calculate to obtain S 13 ≈ 0.21, indicating that the end-to-end latency has a certain degree of positive sensitive linkage relationship with the packet loss rate. To ensure the interpretability of the results, the system also calculates S 12 (X 1 for X 2 ) and the value of S 23 (X2 to X3), and the results are S 12 = -0.05 and S 23 = 0.38 respectively, indicating that there is basically no obvious coupling (negative correlation and very small absolute value) between latency and throughput variation, while there is a strong positive correlation between throughput and packet loss rate.

[0070] Based on these analyses, the system infers that the performance drift of the S1 slice is not an isolated phenomenon, but rather an increase in packet loss caused by the aggravated uplink bandwidth load, which further feeds back to the command response latency of the robotic arm. The system further discovers that this linkage risk also exists in other slices. For example, in the quality inspection service S3, a similar coupling relationship appears during high image upload periods. Therefore, the system immediately includes S1 and S3 in the "coupling-aware scheduling queue" and executes the "coordinated resource isolation" strategy, that is, to allocate a more stringent uplink bandwidth guarantee for S1 with high stability requirements, and implement an upload rate threshold strategy for S3, limiting its maximum upload rate from the original 40 Mbps to 25 Mbps. At the same time, the system adjusts the packet scheduling priority of S1 from the original level 4 to level 2 to enhance its preemption ability. Regarding the parameter selection in the formula, the coupling weight factor α ijThe general setting range is 0.1 to 1.0. The higher the value, the more directly the system considers the influence between indicators. In key business scenarios, the coupling value of, for example, X1 to X3 is recommended to be set between 0.5 and 0.8; when the influence on indicators is weak (such as the influence of latency on throughput), it can be set between 0.1 and 0.3. Through this mechanism, the system can not only identify local problems caused by abnormal single indicators, but also realize cross-indicator and cross-slice linkage risk identification and regulation based on the analysis of the coupling matrix, constructing an intelligent perception ability for system-level dynamic evolution, and improving the controllability, forward-looking and service continuity guarantee ability of 5G private network network slices in a multi-service concurrent environment.

[0071] In this embodiment, to further enhance the system's perception ability and unify the decision-making logic, the system comprehensively scores the perturbation performance of multiple performance indicators in each analysis period, constructs a perturbation scoring function D(t) to quantify the overall stability state and drift risk of the slice, and avoid bias caused by relying on the judgment of single indicators. In the monitoring period of S1, the performance indicators that the system focuses on include end-to-end latency (indicator 1), packet loss rate (indicator 2), and RTT fluctuation (indicator 3), so M = 3. First, the system sets weight factors ω i for these three indicators respectively, where ω1 = 0.5, ω2 = 0.3, ω3 = 0.2, reflecting that latency is the most critical in this slice, followed by the packet loss rate, and the RTT fluctuation has a relatively low weight as an indirect manifestation. In the current analysis period (set as a 5-second window), the system observes that the maximum value of the end-to-end latency is 28.3ms and the minimum value is 22.1ms, so the fluctuation range ΔV1(t) = 6.2mms; the maximum allowable fluctuation value during the stable period indicates that the current fluctuation range exceeds the expectation. At the same time, record the frequency of jumps, that is, the number of times the latency exceeds the local threshold (25ms) within this window is Fv1(t) = 5, set the normalization coefficient γ1 = 10, and the perturbation continuity score has been calculated through the aforementioned Cv as Cv1(t) = 0.34, and the normalization coefficient θ1 = 0.5. Substitute the above data into the scoring formula, and the scoring part of indicator 1 is calculated as follows:

[0072]

[0073] For the packet loss rate indicator 2, the current fluctuation range ΔV2(t) = 1.8%, the maximum fluctuation during the stable period is V2 max = 2.5%, the number of jumps Fv2(t) = 2, γ2 = 5, the perturbation intensity Cv2(t) = 0.12, θ2 = 0.3, substitute into the formula to get:

[0074]

[0075] For the RTT fluctuation indicator 3, ΔV3(t) = 4.5ms, Fv3(t) = 1, γ3 = 5, Cv3(t) = 0.08, θ3 = 0.2. Substituting these values gives:

[0076]

[0077] Adding the three scores together gives:

[0078] D(t) = 1.365 + 0.456 + 0.27 = 2.091

[0079] According to the critical threshold D set by the system based on historical modeling crit = 1.85, it can be seen that the current score D(t) = 2.091 > D crit , indicating that the slice S1 has currently entered the obvious performance drift area, that is, although there is no SLA default, it has substantially deviated from the stable state and there is a risk of further deterioration.

[0080] At this time, the system triggers two types of linkage actions: one is to implement "performance locking protection" for the S1 slice, that is, lock the current resource configuration and suspend the occupation of its shared link by other non-core services; the other is to mark S3 (quality inspection service) as a "coupling interference source", the scheduling module actively reduces the resource scheduling priority of S3, and compresses its available uplink bandwidth by 20%; at the same time, the system pushes the Dt score of this time to the prediction module for constructing the resource scheduling weight reference for the next cycle to avoid repeated triggering of performance fluctuations due to similar risks. In terms of parameters, the weight ω in the perturbation scoring function i is recommended to be set in the range of 0.1 - 0.6, and the sum is 1; the jump normalization coefficient γ i can be set between 5 - 15 according to the index sensitivity; the perturbation intensity normalization coefficient θ i is suitable to be set between 0.2 - 1.0 to keep the scores at the same order of magnitude. This example shows that the present invention organically unifies the fluctuation degree, jump frequency and continuous perturbation intensity of multiple performance indicators by constructing the perturbation scoring function Dt, forms a quantitative evaluation model for the overall operation state of the slice, enables the system to perceive abnormal trends in advance, and realizes dynamic decision-making behaviors with basis, weight and priority, comprehensively improving the stability guarantee ability of key service slices in 5G private networks and the intelligent level of resource scheduling.

[0081] Example 2:

[0082] Combined with the attached Figure 3 As shown, based on Example 1, further evaluate the cost performance of the current resource configuration of this slice to determine whether there is resource redundancy. In 5G private networks, the resources allocated to different service slices may be over-allocated due to inaccurate estimation or security redundancy strategies, especially in control services that require low latency but have non-constant service loads, which is more common.

[0083] Therefore, the system performs retrospective analysis on the data of the S1 slice in the past two weeks, extracts the corresponding performance feedback at different resource levels, and forms a resource-performance mapping dataset. The resource dimension R includes the number of allocated computing cores, bandwidth throughput capacity (Mbps), and slice-level cache capacity. The performance metric P selects the end-to-end delay (unit: ms) as the main observation metric. After the system performs fitting processing on the time series data, it is found that: when the resource R changes between 60 and 80 units, the performance P shows an obvious downward trend; but when R exceeds 80, the decline rate of the P value slows down and gradually converges to about 22 ms. When R is further increased to more than 100, the change of P tends to be stable, and even a slight rebound occurs due to effects such as scheduling queues, and the delay rises to 23 ms. It can be seen that the performance of the S1 slice shows typical non-linear characteristics with the resource input, and there are obvious performance plateau periods and intervals of diminishing marginal benefits of resources. To quantitatively judge whether the current resource level is in a redundant state, the system uses the marginal performance response function:

[0084]

[0085] During this analysis period (T = 20 s), according to the current data fitting derivative The marginal benefit threshold ∈ set by the system is -0.1, which means that at least 0.1 ms of delay improvement per unit of resource input is considered an "effective input"; the function φ(x) is set to take its absolute value when the input x < 0, and if x ≥ 0, the output is 0, that is, when the decline rate of performance is less than expected or instead rises, the function output is 0; κ(t) is set to 1.0 (constant weight, because this service is in a critical control period); after actual calculation, is obviously less than 0, so φ(x) = 0, and the result is:

[0086]

[0087] Therefore, R s = 0, which is much lower than the minimum effective response value threshold R min = 0.2. Thus, it can be determined that the current slice S1 is in the performance plateau area, and continuing to maintain the current resource configuration no longer has effective performance benefits. The system then enters the resource compression control process, and selects to perform progressive compression operations on the non-core resource dimensions - cache capacity and uplink bandwidth. First, the bandwidth is reduced from 100 Mbps to 90 Mbps, and the performance metric is recalculated within a 5-second monitoring window. It is found that the end-to-end delay remains at 22.7 ms; then it is further reduced to 85 Mbps, and the delay slightly rises to 23.1 ms, still lower than the SLA-set threshold of 30 ms, and the perturbation score Dt remains below 1.5, far lower than the warning threshold of 1.85.

[0088] Finally, when the bandwidth is compressed to 75 Mbps, the latency rises to 24.8 ms, and the frequency hopping times increase significantly. The system triggers the threshold and determines that the "compression limit has been reached", and records the configuration before the previous compression (bandwidth 80 Mbps) as the performance retention point of S1. Subsequently, when the system performs resource scheduling, it will use this performance retention point as the reference minimum configuration boundary to avoid over-allocation again. In terms of the parameter range, the marginal benefit threshold ∈ can be set between -0.2 and -0.05 ms / resource unit according to business sensitivity. κ(t) can be a fixed value of 1, or it can be increased to 1.2 - 1.5 in the critical control cycle weight to enhance the response sensitivity, R min The recommended value of R is between 0.1 and 0.3 to ensure that the resource utilization judgment has the ability to filter out small disturbances. In summary, this mechanism realizes the resource value judgment based on the real business response effect, avoids static quotas and blind redundancy, and enables the resource scheduling to move from the optimization level of "whether it is sufficient" to "whether it is necessary".

[0089] After the system determines that the current resources of S1 (the robotic arm control service slice) are in the performance plateau area, it has entered the resource compression regulation process, and through the marginal performance response function R s = 0, it determines that the resource input no longer brings effective performance improvement, and then decides to start the progressive resource compression strategy to release redundant resources. In the specific implementation, the system first analyzes the resource allocation dimensions of this slice and finds that the currently used resources include: the number of CPU processing cores (4 cores), the uplink bandwidth (100 Mbps), and the cache space (500 MB). According to the system's preset resource weight table, the scheduling priority of the cache resource is the lowest, with a weight of 0.2, the bandwidth is 0.4, and the computing resource is 0.4. Therefore, the system preferentially selects the cache dimension for resource compression operations and adopts a strategy of reducing 5% each time for progressive regulation. In the first compression, the cache is reduced from 500 MB to 475 MB, and the system immediately enters a 5-second observation window to collect the main performance indicators of this slice in real time, including the end-to-end delay (indicator 1), jitter (indicator 2), and RTT (indicator 3), and performs stability detection on each indicator based on the "performance retention function". Taking indicator 1 (delay) as an example, the sampled data (unit: ms) is as follows: 22.5, 22.6, 22.7, 22.5, 22.8, 22.9, 23.0, 22.9, 23.1, 23.0. Set the mean value of this indicator in the historical stable period as Use the following performance retention function:

[0090]

[0091] The discrete calculation is approximately:

[0092]

[0093] The system sets η1(t i ) = 1.0 (no amplification) according to historical jitter data, and calculates that the deviations are 0.1, 0.0, 0.1, 0.1, 0.2, 0.3, 0.4, 0.3, 0.5, 0.4 ms respectively, and the sum is 2.4 ms. Therefore, S1 = 2.4 / 10 = 0.24. Set the perturbation stability threshold of this index to S 1,crit = 0.35. Since S1 < S 1,crit , it indicates that this compression has not caused substantial interference to the delay, and the system allows the next round of compression to continue. The cache is further compressed to 450 MB, and a new round of performance data is collected again.

[0094] In this round of sampling, it is found that the maximum deviation rises to 0.8 ms, and η1(t i ) automatically amplifies to 1.2 for the abnormal deviation range. After calculation, S1 = 0.39 > 0.35. At the same time, the jitter index (index 2) S2 reaches 0.42 ms, which also exceeds its tolerance threshold of 0.3 ms. The system determines that the compression touches the performance boundary, that is, continuing to compress the current resources will cause a substantial decline in the service experience. Therefore, the system rolls back to the state before the previous compression, that is, the cache capacity is set to 475 MB, which is marked as the "resource retention point" of this slice. In the parameter setting, Δt is recommended to be set to 3 - 10 seconds depending on the real-time nature of the service; η k (t) ranges from 1.0 to 1.5, where the higher the value, the more sensitive it is to the index fluctuation. Usually, when the deviation exceeds the mean by more than 5% continuously three times, η k (t) is dynamically increased; S k,crit is recommended to be configured in the range of 0.2 - 0.5 according to the importance of the index. In this example, the change of performance perturbation is quantified by the formula, so that the compression operation has a clear feedback closed-loop, and at the same time, the high-sensitivity recognition of fine-grained performance risks is achieved by combining dynamic weights, effectively avoiding SLA default caused by excessive compression.

[0095] This embodiment slightly reclaims the cache resource dimension and evaluates the perturbation degree of key performance indicators after each round of compression based on the performance retention function S k . In the previous steps, the system has completed two rounds of cache compression, compressing the cache resources of S1 from 500 MB to 475 MB and 450 MB respectively, and continuously calculating the corresponding S1 values of the end-to-end delay as 0.24 and 0.27 respectively in these two cycles, both of which are lower than the system-set threshold S 1,crit= 0.35, thus meeting the judgment condition of "stable disturbance values in two consecutive rounds", and entering the third round of compression attempt. When the cache is further compressed to 425MB, the delay jitter rises rapidly, the fluctuation amplitude of σ1(t) increases, and the fluctuations are concentrated in the range of 0.5 - 0.9ms deviating from the mean. Substituting into the formula for calculation, we get S1 = 0.43 > 0.35. The system determines that the current compression has reached the performance stability boundary, immediately rolls back the configuration, and takes the previous round, i.e., 450MB, as the "resource retention point" of slice S1.

[0096] On this basis, to achieve the macroscopic optimization of the network-level resource pool, the system needs to further determine whether S1 has the potential to continue releasing resources within a certain period in the future, and perform a priority ranking of resource reallocation for multiple slices in the entire resource pool. Therefore, the system uses the "resource scheduling trend function" to calculate the predicted value F of the available resources for each slice r , and takes S1 as an example for analysis. S1 includes three types of resource dimensions: computing resources (R1), bandwidth resources (R2), and cache resources (R3), and their current release rates per unit time are (CPU uncompressed), (bandwidth uncompressed), Considering the resource priority settings in the system, ω1 = 0.4, ω2 = 0.4, ω3 = 0.2; the sensitivity adjustment factor δ3 of the cache resource is 1.2, which is used to amplify the evaluation of the impact of compression on future behavior; the historical compression observation window T3 = 60s. Within this window, the system monitors the stable occupancy trend function ψ3(t) of the cache resource. According to data analysis, the cache occupancy remains at about 80% on average without obvious fluctuations. Therefore:

[0097]

[0098] Therefore, the predicted value of the available resources of slice S1 is currently 0.04, which is a slice with "weak release potential". The system then performs the same process on other slices such as S3 and S4 and obtains their respective F r values, and finds that S4 has released a total of 10% of the cache, bandwidth, and CPU resources in the past 3 rounds. Its bandwidth release rate is -1.2Mbps / s, and the stable occupancy has dropped to 65%. After calculation, F r= 0.17, which is significantly higher than S1. The system sorts all slices in descending order of the Fr value, and at the same time integrates the resource retention point position of the slice (such as whether it has approached the boundary), the stability entropy within the past 7 days (the lower the entropy value, the more stable), and the criticality level (such as whether it is a Class A control service) to form a multi-dimensional scoring model, and transmits the scoring result to the resource reallocation scheduling engine. Finally, the system dynamically adjusts the resource reallocation priority, preferentially schedules the slice resources with great release potential, stable performance, and medium service level, and sets high-level sensitive slices such as S1 to the "reserved state" and does not participate in cross-slice resource migration. In terms of coefficient setting, the system recommends: ω i The sum is 1.0, and each type of resource is set between 0.1 and 0.5; δ i Is set to 1.0 to 1.5 according to resource volatility and service sensitivity, ψ i (t) ranges from 0 to 1, reflecting the trend of resource occupancy rate; T i Can be 60 to 600 seconds, specifically configured according to the business cycle length. It can be seen that through the quantitative calculation of the stability recognition and trend prediction function F r of the resource compression process, the present invention realizes the transformation of resource scheduling from passive response to forward-looking planning, and provides a more intelligent, stable and resource-efficient optimization and control mechanism for the 5G private network multi-service, multi-slice, high-density deployment environment.

[0099] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. 5G private network network slice optimization method, characterized in that: S1. Adopt a performance drift perception mechanism based on slice boundary status: S1.

1. During the operation of the slice, construct a time series performance model to monitor minute changes including end-to-end delay, throughput, and packet loss rate; then extract the short-period fluctuation patterns of the metrics, and compare the normalized gradient with the historical stability threshold; S1.

2. By identifying the metric drift behavior that is continuously jittering but has not defaulted, determine whether it is in a performance critical state; S2. Adopt dynamic modeling of the non-linear relationship between slice services and resources: S2.

1. Record the non-proportional responses including jitter improvement and delay reduction brought about after each resource adjustment; then perform trend fitting on the data to determine whether there is a performance plateau or a critical jump point; S2.

2. Define three intervals for resource investment in the model: an ineffective area where the investment remains unchanged, a highly sensitive area where the performance changes rapidly, and an ascending saturation area where the performance is stable but the resources continue to increase; S2.

3. Identify the response area where the slice is currently located through the model, and determine that the control strategy does not take proportional investment as the core, but selects the optimal response area as the resource allocation interval; S3. Adopt a resource compression control mechanism based on the performance retention area: S3.

1. Determine the area where each slice is currently located on the non-linear response curve; if it is in the performance saturation area or the plateau period, it means that some of the resources are redundant at this time; start the progressive resource compression process and slightly recover from the edge resource units; S3.

2. After each compression, real-time detect whether the performance remains within the set range; If the performance does not decrease, continue the next round of compression until the performance boundary is touched, and record the compression limit value as the resource retention point of the slice at this time; S3.

3. Score each slice, and then predict the available resource trend in the future period according to the current state of the resource pool and the slice weight; then construct a scheduling priority based on the stability entropy between slices.

2. The method for optimizing network slices of a 5G private network according to claim 1, wherein The construction of the performance drift perception mechanism based on slice boundary status includes: During the operation of the slice, continuously collect key performance indicators at the second-level or sub-second-level granularity to form time series data; retain the original data at each moment for analyzing edge perturbations; to characterize the continuity of the perturbations, introduce a perturbation intensity function Cv to measure the abnormal fluctuation trend of the metrics in the short period: Wherein: $C_v$ represents the perturbation continuity strength of a certain performance indicator of the slice over the time period $[t, t + \Delta T]$; $\Delta T$ is the time length of the short-term analysis window; $X(t)$ is the value of a specific performance indicator at time $t$; represents the rate of change of the indicator at time $t$; $\mu$ is the expected value of the rate of change calculated by the system based on historical stable data; $|\cdot|$ represents the absolute deviation degree of the change.

3. The network slice optimization method for the 5G private network according to claim 2, wherein The construction of the performance drift perception mechanism based on slice boundary status includes: constructing a stable baseline model for each metric; recording the fluctuation range of each metric during stable operation, analyzing the coupling relationship between the metrics to identify systematic linkage risks; defining the perturbation response coupling intensity between the metrics to characterize the influence sensitivity of one metric to another metric.

4. The method for optimizing network slices of a 5G private network according to claim 3, wherein The construction of the performance drift perception mechanism based on slice boundary status includes: Based on the perturbation characteristics in each analysis period, construct a perturbation scoring function D(t) to uniformly evaluate whether the current slice enters the performance critical state: Wherein: D(t) represents the comprehensive score of the overall perturbation risk of the slice at time t; M is the total number of performance indicators; ω i is the weight factor of the i-th indicator; ΔV i (t) is the difference between the maximum and minimum values of the i-th indicator within the current window, measuring the fluctuation range; is the maximum allowable fluctuation value of the indicator during the stable period; Fv i (t) is the hopping frequency, that is, the number of times the indicator exceeds the threshold, representing the intensity of fluctuations; γ i is the normalization constant of the hopping frequency; Cv i (t) is the perturbation continuity intensity of the indicator within the current period; θ i is the normalization coefficient of the perturbation continuity; When the value of D(t) exceeds the critical threshold D set according to historical modeling crit it indicates that the slice is currently in the performance drift region, that is, although the SLA hard alarm has not been triggered, it has deviated from the stable state and is about to enter the abnormal region.

5. The network slice optimization method for the 5G private network according to claim 1, wherein The construction method of the resource compression control mechanism based on the performance retention area: Based on the resource allocation and performance feedback data of each network slice at different time periods, a response relationship curve between the multi-dimensional resource input and the key performance indicators of the slice is constructed, and the response relationship curve is used to characterize the non-linear characteristics of the slice's performance as the resource input changes.

6. The method for optimizing network slices of a 5G private network according to claim 5, wherein The compression process adopts a progressive compression strategy, selects the dimension with the smallest weight in the resource dimension for minor adjustment. After each resource compression, the core performance indicators are collected in real time within a set window, and stability detection is performed based on the performance retention function defined as follows: Among them, S k represents the perturbation stability of the k-th type of performance index within the observation period. σk(t) is the real-time value of the k-th performance index at time t, is the mean value of this index in the baseline stable state, and ηk(t) is a dynamic sensitivity weight function constructed based on historical volatility, which is used to amplify the abnormal perturbation response.

7. The method for optimizing network slices of a 5G private network according to claim 6, wherein In the compression process, when S after two consecutive rounds of compression k is less than the preset fluctuation tolerance threshold, the next round of compression is continued until S k exceeds the threshold range. Then, the compression configuration of the previous round is recorded by the system as the resource retention point of this slice. After recording the retention point, the system constructs a future resource transferable trend model based on the service level, historical stability, and resource utilization efficiency of each slice, and generates a predicted value of transferable resources based on the following resource scheduling trend function: Among them, F r represents the overall resource release trend intensity of the slice, ω i is the system weight of the i-th type of resource, represents the release speed per unit time of this type of resource during the compression process, δ i is the sensitivity adjustment factor of this resource, ψ i (t) is a trend function reflecting the stable occupancy rate of the resource, T i is the historical compression observation window length of the corresponding resource. Finally, the system comprehensively sorts according to the predicted F r and the resource retention points, safety performance scores, and historical fluctuation entropy indicators of each slice to complete the dynamic adjustment of the resource reallocation priority.

Images