Flow characteristic adaptive QoS intelligent prediction adjustment method

Abstract

The invention discloses a flow characteristic adaptive QoS intelligent prediction adjustment method, and relates to the field of network flow management, and the method comprises the steps: 1, collecting flow data in real time, and constructing a multi-dimensional characteristic vector based on protocol types, port numbers and user behavior dynamic classification; 2, high-frequency / low-frequency components are separated, and QoS parameter prediction is output through fusion of an LSTM short-term prediction module and a periodic trend analysis module; 3, solving a resource pre-allocation scheme by adopting reinforcement learning by taking minimization of packet delay as a target; and 4, executing traffic identification, speed limiting and priority queue scheduling by using NPU hardware unloading. According to the method, the precision is improved through a high-frequency / low-frequency combined prediction architecture, decision delay is compressed to a large extent through reinforcement learning and NPU cooperation, and meanwhile online model iteration is achieved through a prediction error triggering mechanism.

Description

Technical Field

[0001] The present invention relates to network quality of service (QoS) management technology, and specifically to a QoS dynamic adjustment method based on traffic feature adaptation and intelligent prediction, which is suitable for real-time network scenarios such as intelligent network traffic control networking and high-performance networking. Background Art

[0002] In recent years, QoS technology has shown three major innovation directions:

[0003] 1. Fine-grained control and dynamic adaptation;

[0004] 2. Resource pre-allocation driven by intelligent prediction;

[0005] 3. Integration of protocol layer and hardware acceleration;

[0006] However, existing technologies either focus on static rules (Inspur Cloud) or predictive algorithms (Volcano Engine), and have yet to form a fully closed-loop adaptive system of "feature perception → prediction → decision-making → hardware execution → feedback optimization". In particular, they lack:

[0007] Joint prediction architecture for high / low frequency traffic components;

[0008] Real-time linkage mechanism between reinforcement learning and NPU hardware offloading;

[0009] Online model iteration capability triggered by prediction errors. Summary of the Invention

[0010] Based on this, the purpose of the present invention is to provide a QoS intelligent prediction and adjustment method that is adaptive to traffic characteristics to solve the technical problems raised in the above background technology.

[0011] To achieve the above object, the present invention provides the following technical solutions:

[0012] A QoS intelligent prediction and adjustment method based on traffic characteristics self-adaptation includes the following steps:

[0013] 1. Collect network data streams in real time, dynamically classify and mark traffic based on protocols, port numbers, application characteristics, and user behavior data, and generate multi-dimensional feature vectors;

[0014] 2. Processing the multidimensional feature vector based on a long-term and short-term prediction fusion model: Historical QoS parameters, network status information, and user device behavior data are layered by time series, high-frequency fluctuation components and low-frequency trend components are separated, and the components are fed into the LSTM short-term prediction module and the periodic trend analysis module, respectively, to output QoS prediction parameters for a preset future time period.

[0015] 3. Build a resource optimization model based on QoS prediction parameters, with the goal of minimizing average packet latency. Use reinforcement learning to solve resource pre-allocation schemes in discrete space and dynamically adjust bandwidth allocation and priority strategies.

[0016] 4. Execute resource pre-allocation scheme through hardware acceleration layer: Use NPU to perform hardware offload of traffic identification, rate limiting and priority queue, and output QoS adjustment instructions to network devices in real time.

[0017] As a preferred embodiment of the present technical solution, the dynamic classification and labeling in step 1 includes:

[0018] Extract the protocol type, source / destination port number, and packet size distribution characteristics of traffic data;

[0019] Associate user behavior characteristics, including access frequency, geographic location activity, and historical abnormal behavior tags;

[0020] Based on the symmetric uncertainty algorithm, feature attributes that are strongly related to QoS requirements are screened and an adaptive feature rule base is constructed.

[0021] As a preferred embodiment of the present technical solution, the processing of the long-term and short-term prediction fusion model includes:

[0022] The high-frequency fluctuation component is input into the LSTM network to learn the short-term burst traffic characteristics;

[0023] The low-frequency trend component is analyzed through the kernel function ridge regression model to analyze the periodicity, and the user context features (autonomous system number, service provider, regional topology) are combined to weighted fusion prediction results.

[0024] As a preferred embodiment of this technical solution, the construction of the resource optimization model includes:

[0025] The data packet service process is modeled as an M / M / 1 queuing system, and the constraint relationship between channel quality and QoS flow requirements is defined;

[0026] The resource allocation strategy is trained offline through deep reinforcement learning (DQN), and the bandwidth allocation ratio and packet loss priority threshold are generated according to the real-time traffic prediction value during online deployment.

[0027] As a preferred embodiment of this technical solution, the execution of the hardware acceleration layer includes:

[0028] Map traffic classification marking results to sub-queues of the multi-queue FIFO chip. Each sub-queue is independently configured with storage depth and almost full / empty offset values.

[0029] The microcode ratio control algorithm is used to schedule packets in descending order of QoS priority, and delay-sensitive data flows are allocated to low-latency queues.

[0030] As a preferred embodiment of the present technical solution, the updating of the adaptive feature rule base includes:

[0031] Periodically collect malicious traffic feature samples and generate new rules using the random forest model;

[0032] The similarity of redundant feature attributes is calculated. If the weight difference is lower than the threshold, the features are merged to dynamically compress the size of the rule base.

[0033] As a preferred embodiment of this technical solution, the encoding of the user context feature includes:

[0034] One-hot encode the user ID, service ID, and region ID and map them into dense vectors;

[0035] The relationship weights between context features are calculated through the self-attention mechanism to generate a normalized interaction feature matrix.

[0036] As a preferred embodiment of the present technical solution, the online optimization of the resource allocation strategy includes:

[0037] When real-time traffic exceeds the bandwidth limit, the packet loss ratio is calculated in ascending order according to QoS priority, and redundant traffic is released;

[0038] A microcode proportional control algorithm is used for critical state traffic, and the remainder is compensated in sections according to the ratio of (sent traffic s: discarded traffic d).

[0039] As a preferred embodiment of this technical solution, an adaptive feedback closed loop is also included:

[0040] Monitor network congestion indicators and user equipment delay data after QoS execution;

[0041] When the prediction error exceeds the threshold, the feature rule base is retrained and the long-term and short-term prediction model parameters are updated.

[0042] In summary, the present invention mainly has the following beneficial effects:

[0043] 1. Achieve full closed-loop adaptation: By building a full-link system from feature perception → prediction → decision-making → hardware execution → feedback optimization, this system overcomes the limitations of static rules and pure algorithmic solutions, effectively reducing dynamic response latency.

[0044] 2. Accurate multi-frequency traffic prediction: This approach uses a high-frequency / low-frequency joint prediction architecture, combined with wavelet decomposition and LSTM dual-channel processing, to reduce prediction errors.

[0045] 3. Hardware Collaborative Acceleration: Based on the real-time linkage between the reinforcement learning decision engine and NPU hardware offload, QoS policy execution latency is optimized from a software perspective.

[0046] 4. Online iterative optimization: Trigger model updates through prediction errors, quickly complete parameter adjustments, and support maximum SLA guarantees under burst traffic. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the method of the present invention;

[0048] Figure 2 This is a flow chart of the dynamic classification and marking module of the present invention;

[0049] Figure 3 This is a diagram of the prediction fusion processing module architecture of the present invention;

[0050] Figure 4 This is a diagram of the resource optimization decision module architecture of the present invention;

[0051] Figure 5 This is a diagram of the hardware acceleration execution module architecture of the present invention;

[0052] Figure 6 This is a diagram of the system layered architecture of the present invention;

[0053] Figure 7 This is a diagram of the priority scheduling module architecture of the present invention;

[0054] Figure 8 This is a diagram of the feedback closed-loop module architecture of the present invention; DETAILED DESCRIPTION

[0055] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be understood as limiting the present invention.

[0056] The following describes an embodiment of the present invention based on its overall structure.

[0057] like Figure 1-8 As shown, this embodiment provides a traffic feature adaptive QoS intelligent prediction and adjustment method, including the following steps:

[0058] Step 1. Collect network data flows in real time, dynamically classify and label the traffic based on protocols, port numbers, application characteristics, and user behavior data, and generate multi-dimensional feature vectors.

[0059] Step 2. Process the multidimensional feature vector based on the long-term and short-term prediction fusion model: Stratify historical QoS parameters, network status information, and user device behavior data by time series, separate high-frequency fluctuation components and low-frequency trend components, and input them into the LSTM short-term prediction module and the periodic trend analysis module, respectively, to output QoS prediction parameters for a preset future time period;

[0060] Step 3. Build a resource optimization model based on QoS prediction parameters. Minimize the average packet delay. Use reinforcement learning to solve the resource pre-allocation scheme in discrete space and dynamically adjust bandwidth allocation and priority strategies.

[0061] Step 4. Execute the resource pre-allocation plan through the hardware acceleration layer: Use the NPU to perform hardware offload of traffic identification, rate limiting, and priority queues, and output QoS adjustment instructions to network devices in real time.

[0062] The dynamic classification and labeling in step 1 include:

[0063] Extract the protocol type, source / destination port number, and packet size distribution characteristics of traffic data;

[0064] Associate user behavior characteristics, including access frequency, geographic location activity, and historical abnormal behavior tags;

[0065] Based on the symmetric uncertainty algorithm, feature attributes that are strongly related to QoS requirements are screened and an adaptive feature rule base is constructed.

[0066] In this embodiment, the traffic dynamic classification and feature extraction method is as follows:

[0067] 1. Data collection:

[0068] In intelligent network traffic control or high-performance network architectures, for example, lightweight probes (based on the DPDK framework for zero-copy packet capture) are deployed on network base station edge servers to capture the following four-dimensional traffic characteristics in real time:

[0069] 1.1. Protocol layer features: Parse the IP packet header to obtain the protocol type (TCP / UDP / ICMP / QUIC) and extract the transport layer source / destination port number combination (e.g., 443-5001 identifies HTTPS video stream);

[0070] 1.2. Packet statistics: Calculate the packet size variance within a sliding window (window size = 100 packets, variance > 8000 bytes identifies a video stream), and calculate the jitter of packet arrival intervals per unit time (standard deviation > 5ms triggers a burst flag).

[0071] 1.3. User behavior characteristics: Linking to the user history database, loading the geographic location activity index (e.g., a movement speed > 30 km / h triggers an in-vehicle scene tag), and injecting historical abnormal behavior tags (e.g., matching the DDoS attack source IP database);

[0072] 1.4. Network environment characteristics: Collect wireless channel quality indicators (CQI) and base station load rates (>80% triggers congestion warning);

[0073] 2. Adaptive rule base construction:

[0074] Use the Symmetric Uncertainty algorithm for feature screening:

[0075] SU(X,Y)=2×I(X;Y)H(X)+H(Y)SU(X,Y)=2×H(X)+H(Y)I(X;Y)

[0076] Where I(X; Y) is the mutual information between feature X and QoS requirement Y, and H(X) is the feature information entropy.

[0077] 3. Screening mechanism:

[0078] Set SU threshold = 0.65 (determined by ROC curve analysis)

[0079] Examples of highly correlated features:

[0080] Video streaming scenario: packet size variance (SU = 0.78)

[0081] Industrial control scenario: Message inter-arrival jitter (SU = 0.82)

[0082] 4. Rule base update:

[0083] New threat samples (such as blockchain node encrypted computing interaction traffic generated by Spirent TestCenter) are injected every 2 hours, and a new rule set is generated through incremental training of the random forest model (tree depth = 10, feature bag sampling rate = 0.8).

[0084] The processing of the long-term and short-term prediction fusion model includes:

[0085] The high-frequency fluctuation component is input into the LSTM network to learn the short-term burst traffic characteristics;

[0086] The low-frequency trend component is analyzed through the kernel function ridge regression model to analyze the periodicity, and the user context features (autonomous system number, service provider, regional topology) are combined to weighted fusion prediction results.

[0087] The encoding of the user context feature includes:

[0088] One-hot encode the user ID, service ID, and region ID and map them into dense vectors;

[0089] The relationship weights between context features are calculated through the self-attention mechanism to generate a normalized interaction feature matrix.

[0090] In this embodiment, the long-term and short-term prediction fusion model is as follows:

[0091] 1. Spectrum separation processing:

[0092] Perform wavelet decomposition on the input traffic time series $T(t)$:

[0093] T(t)=Alow(t)+Dhigh(t)T(t)=Alow(t)+Dhigh(t)

[0094] 1.1 Low-frequency component Alow(t): Input cycle trend analysis module

[0095] Kernel Ridge Regression is used to predict periodic patterns:

[0096] y=K(X,Xtrain)(K(Xtrain,Xtrain)+λI)-1ytrainy=K(X,Xtrain)(K(Xtrain,Xtrain)+λI)-1ytrain

[0097] The Gaussian kernel σ=0.5.

[0098] 1.2 High-frequency component D high (t): Input LSTM network structure: 3-layer LSTM (128 units) + Dropout (0.2), input window = 50 steps.

[0099] 2. Contextual feature fusion:

[0100] 2.1. Feature encoding and weighting mechanism:

[0101] Feature vectorization preprocessing: One-hot encoding is performed on discrete features such as user ID, service ID, and autonomous system number to generate a sparse binary vector. The sparse vector is then mapped into a 128-dimensional dense feature vector through an embedding layer to eliminate the curse of dimensionality.

[0102] 2.2. Self-Attention Weight Calculation: Construct three sets of parameterized vectors: query matrix, key matrix, and value matrix: Q = Wq·E, K = Wk·E, V = Wv·E;

[0103] Where E is the embedding feature matrix, W q / W k / W v is the trainable weight matrix;

[0104] 2.3. Calculate the correlation weight between features:

[0105] Attention(Q,K,V)=softmax(dk QKT)V

[0106] where d k is the key vector dimension (set to 64), Used to control gradient stability;

[0107] 2.4. Normalization: Use the softmax function to compress the associated weights to the interval [0, 1] to generate a probability distribution;

[0108] 2.5. Weighted feature fusion: Perform Hadamard product on the normalized weights and the original feature vectors:

[0109] Ffused=∑i=1nαi⊙ViFfused=∑i=1n αi⊙Vi

[0110] where α i is the attention weight of the i-th feature, ⊙ represents element-wise multiplication;

[0111] Output the interaction feature matrix with unified dimension (size: m×128, where m is the number of users).

[0112] The construction of the resource optimization model includes:

[0113] The data packet service process is modeled as an M / M / 1 queuing system, and the constraint relationship between channel quality and QoS flow requirements is defined;

[0114] The resource allocation strategy is trained offline through deep reinforcement learning (DQN), and the bandwidth allocation ratio and packet loss priority threshold are generated according to the real-time traffic prediction value during online deployment.

[0115] The execution of the hardware acceleration layer includes:

[0116] Map traffic classification marking results to sub-queues of the multi-queue FIFO chip. Each sub-queue is independently configured with storage depth and almost full / empty offset values.

[0117] The microcode ratio control algorithm is used to schedule packets in descending order of QoS priority, and delay-sensitive data flows are allocated to low-latency queues.

[0118] The updating of the adaptive feature rule base includes:

[0119] Periodically collect malicious traffic feature samples and generate new rules using the random forest model;

[0120] The similarity of redundant feature attributes is calculated. If the weight difference is lower than the threshold, the features are merged to dynamically compress the size of the rule base.

[0121] The online optimization of the resource allocation strategy includes:

[0122] When real-time traffic exceeds the bandwidth limit, the packet loss ratio is calculated in ascending order according to QoS priority, and redundant traffic is released;

[0123] A microcode proportional control algorithm is used for critical state traffic, and the remainder is compensated in sections according to the ratio of (sent traffic s: discarded traffic d).

[0124] In this embodiment, resource pre-allocation and hardware acceleration are as follows:

[0125] 1. Reinforcement learning decision-making:

[0126] Modeled as a Markov Decision Process (MDP):

[0127] Status t : Current bandwidth utilization + QoS prediction parameters;

[0128] Action a t : {Increase queue A bandwidth by 5%, reduce queue B priority,...};

[0129] Reward t :r t = -(average delay + 0.3 × packet loss rate);

[0130] Using the DQN algorithm for offline training (ε-greedy strategy, γ = 0.9), online inference takes < 2ms.

[0131] 2. Microcode ratio control:

[0132] 2.1 Dynamic discard policy when bandwidth exceeds the limit:

[0133] 2.1.1. Priority sorting and weight allocation:

[0134] There are 8 priority queues (0 is the highest and 7 is the lowest) and each queue is configured with a weight coefficient w. i (i=0,7), satisfying ∑w i =1;

[0135] The weight coefficient is strongly related to the business type (as shown in Table 1 below):

[0136] Priority Business Type <![CDATA[Weight w i > 0 Industrial Control Instructions 0.35 1 Live Video 0.25 ... ... ... 7 Software Updates 0.02

[0137] 2.1.2. Discard ratio calculation model:

[0138] When the real-time traffic L current Exceeds bandwidth limit L max hour:

[0139] Calculate the total excess flow: ΔL = L current -L max ;

[0140] Calculate the discarded amount for each queue from low to high priority (7→0):

[0141] Di=min(ΔL×wi,Qilen)Di=min(ΔL×wi,Qilen)

[0142] in is the number of packets currently backlogged in the i-th level queue;

[0143] Dynamic update of remaining excess limit: ΔL=ΔL-D i ;

[0144] 2.1.3. Critical state compensation mechanism:

[0145] If there is still a margin ΔL after weight distribution residual >0:

[0146] Principle of segmented compensation:

[0147] Margin ΔL residual ≤5%L max When: All packets are discarded by the lowest priority queue (level 7);

[0148] Margin ΔL residual >5%L max When: Press (S i :D i )=(3:1) ratio to perform send-drop mixed scheduling on the second lowest priority queue (level 6);

[0149] Among them, S i : Number of normal sent messages, D i : Number of extra discarded packets;

[0150] 2.2 Hardware execution process:

[0151] 2.2.1.NPU microcode logic:

[0152] Deployment of microinstruction sets in programmable network chips (such as Intel Tofino):

[0153] Periodic monitoring hardware queue count register (sampled every 10μs);

[0154] Trigger condition: When queue t otal c ount>threshold m When ax is used, the discard calculation is started and the discard operations are performed in parallel, as follows:

[0155] Directly manipulate the message descriptor ring through the queue management unit (QMU);

[0156] Inject the drop_flag mark into the target queue (without affecting the forwarding of other queues);

[0157] Delay control guarantee:

[0158] The calculation-execution full-link delay is ≤ 0.8μs (measured data);

[0159] Ensure that high-priority queues are not disturbed by low-priority drop operations;

[0160] 3. Hardware accelerated execution:

[0161] Implementation on Intel Tofino chip:

[0162] Traffic classification results are mapped to 8 hardware queues (FIFO depth is configurable);

[0163] The high priority queue sets almost_empty_offset=10% to trigger early scheduling.

[0164] The method also includes an adaptive feedback loop:

[0165] Monitor network congestion indicators and user equipment delay data after QoS execution;

[0166] When the prediction error exceeds the threshold, the feature rule base is retrained and the long-term and short-term prediction model parameters are updated.

[0167] In this embodiment, the adaptive feedback closed loop is as follows:

[0168] 1. Monitoring indicators:

[0169] Core indicator: Forecast error rate

[0170] Auxiliary indicators: user device RTT fluctuation variance;

[0171] 2. Triggering retraining conditions:

[0172] When E>10% for 3 consecutive cycles:

[0173] Update the feature rule base: inject the latest DDoS attack samples (such as SYN Flood features);

[0174] Adjust LSTM hyperparameters: reduce the learning rate from 0.001 to 0.0005 and expand the window to 100 steps.

[0175] In this embodiment, random tests were conducted more than three times in an intelligent network traffic control network or a high-performance network (compared with traditional SDN QoS solutions). Based on the OMNeT++6.0 simulation platform, the load traffic used the 3GPP TR38.901 standard model. The average value of multiple test results was taken and the decimal point was removed. The specific data is shown in the comparison in Table 2 below.

[0176] index Traditional solutions The present invention Improvement Average latency 28ms 9ms 67.8％ Burst traffic packet loss rate 15.2％ 2.1％ 86.2％ Policy response delay 120ms 0.2ms 99.8％ Bandwidth utilization 72％ 89％ 23.6％

[0177] Table 2

[0178] Although an embodiment of the present invention has been shown and described, this specific embodiment is merely an explanation of the present invention and is not a limitation of the invention. The specific features, structures, materials or characteristics described may be combined in an appropriate manner in any one or more embodiments or examples. After reading this specification, those skilled in the art may make modifications, substitutions and variations to the embodiment without creative contribution as needed without departing from the principles and purpose of the present invention. However, as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A QoS intelligent prediction and adjustment method based on traffic characteristics adaptation, characterized in that: The following steps are involved: (1) Collect network data flows in real time, dynamically classify and mark traffic based on protocols, port numbers, application characteristics, and user behavior data, and generate multi-dimensional feature vectors; (2) Processing the multidimensional feature vector based on the long-term and short-term prediction fusion model: stratifying historical QoS parameters, network status information, and user device behavior data by time series, separating high-frequency fluctuation components and low-frequency trend components, and inputting them into the LSTM short-term prediction module and the periodic trend analysis module respectively, and outputting QoS prediction parameters for a preset time period in the future; (3) Based on the QoS prediction parameters, a resource optimization model is constructed with the goal of minimizing the average packet delay. Through reinforcement learning, a resource pre-allocation scheme in discrete space is solved, and bandwidth allocation and priority strategies are dynamically adjusted. (4) Execute resource pre-allocation scheme through hardware acceleration layer: Use NPU to perform hardware offload of traffic identification, rate limiting and priority queue, and output QoS adjustment instructions to network devices in real time.

2. The method for QoS intelligent prediction and adjustment based on traffic characteristics adaptation according to claim 1 is characterized in that: The dynamic classification and labeling in step (1) includes: Extract the protocol type, source / destination port number, and packet size distribution characteristics of traffic data; Associate user behavior characteristics, including access frequency, geographic location activity, and historical abnormal behavior tags; Based on the symmetric uncertainty algorithm, feature attributes that are strongly related to QoS requirements are screened and an adaptive feature rule base is constructed.

3. The method for QoS intelligent prediction and adjustment based on traffic characteristics adaptation according to claim 1, characterized in that: The processing of the long-term and short-term prediction fusion model includes: The high-frequency fluctuation component is input into the LSTM network to learn the short-term burst traffic characteristics; The low-frequency trend component is analyzed through the kernel function ridge regression model to analyze the periodicity, and the user context features (autonomous system number, service provider, regional topology) are combined to weighted fusion prediction results.

4. The method for QoS intelligent prediction and adjustment based on traffic characteristics adaptation according to claim 1, characterized in that: The construction of the resource optimization model includes: The data packet service process is modeled as an M / M / 1 queuing system, and the constraint relationship between channel quality and QoS flow requirements is defined; The resource allocation strategy is trained offline through deep reinforcement learning (DQN), and the bandwidth allocation ratio and packet loss priority threshold are generated according to the real-time traffic prediction value during online deployment.

5. The method for QoS intelligent prediction and adjustment based on traffic characteristics adaptation according to claim 1, characterized in that: The execution of the hardware acceleration layer includes: Map traffic classification marking results to sub-queues of the multi-queue FIFO chip. Each sub-queue is independently configured with storage depth and almost full / empty offset values. The microcode ratio control algorithm is used to schedule packets in descending order of QoS priority, and delay-sensitive data flows are allocated to low-latency queues.

6. The method for QoS intelligent prediction and adjustment based on traffic characteristics adaptation according to claim 2, characterized in that: The updating of the adaptive feature rule base includes: Periodically collect malicious traffic feature samples and generate new rules using the random forest model; The similarity of redundant feature attributes is calculated. If the weight difference is lower than the threshold, the features are merged to dynamically compress the size of the rule base.

7. The method for QoS intelligent prediction and adjustment based on traffic characteristics adaptation according to claim 3 is characterized in that: The encoding of the user context feature includes: One-hot encode the user ID, service ID, and region ID and map them into dense vectors; The relationship weights between context features are calculated through the self-attention mechanism to generate a normalized interaction feature matrix.

8. The method for QoS intelligent prediction and adjustment based on traffic characteristics adaptation according to claim 4 is characterized in that: The online optimization of the resource allocation strategy includes: When real-time traffic exceeds the bandwidth limit, the packet loss ratio is calculated in ascending order according to QoS priority, and redundant traffic is released; A microcode proportional control algorithm is used for critical state traffic, and the remainder is compensated in sections according to the ratio of (sent traffic s: discarded traffic d).

9. The method for intelligent QoS prediction and adjustment based on traffic characteristics adaptation according to claim 1, characterized in that: Also includes an adaptive feedback loop: Monitor network congestion indicators and user equipment delay data after QoS execution; When the prediction error exceeds the threshold, the feature rule base is retrained and the long-term and short-term prediction model parameters are updated.

Images