A network load balancing method based on Xinyuan platform and Xinyuan terminal

Abstract

The application provides a network load balancing algorithm based on a Xinyuan platform and a Xinyuan terminal, and relates to the technical field of Xinyuan, which comprises the following steps: collecting real-time traffic data from multiple ports; performing normalization, outlier elimination and protocol field standardization preprocessing on the collected data; performing multi-dimensional protocol fingerprint feature extraction on the traffic of each port based on a protocol analysis algorithm; modeling the historical and real-time traffic of each port using a time series neural network, and outputting traffic prediction values and abnormal traffic probability evaluation in a future short period; dynamically generating a multi-port distribution weight factor based on the protocol characteristics and traffic change trend of each port; inputting the prediction results and real-time load state into a reinforcement learning scheduling agent; and immediately performing feedforward load distribution adjustment when abnormal traffic trend is detected. The method can improve the data processing efficiency, realize the intelligence of load scheduling, ensure the resource allocation priority of the ports of the Xinyuan terminal, and guarantee service continuity.

Description

Technical Field

[0001] This invention relates to the field of information technology innovation, and in particular to a network load balancing algorithm and an information technology innovation terminal based on an information technology innovation platform. Background Technology

[0002] With the rapid development of information technology, network scale is constantly expanding, and business traffic is experiencing explosive growth, making network load balancing issues in heterogeneous protocol environments increasingly prominent. In the network architecture built on domestic IT innovation platforms, due to the involvement of multiple protocol types, concurrent processing of multiple ports, and complex traffic patterns, traditional load balancing algorithms are gradually revealing many limitations when facing real-time traffic fluctuations, protocol compatibility differences, and abnormal traffic handling. In existing technologies, most load balancing solutions employ static weight allocation or simple round-robin mechanisms, which are difficult to adapt to scenarios with diverse protocol types (such as the coexistence of new protocols like TCP, UDP, and HTTP / 3 with traditional protocols) under domestic IT innovation platforms. These solutions lack deep analysis capabilities of protocol characteristics and cannot establish accurate port-protocol association mappings, resulting in low distribution efficiency and a tendency for port load imbalances in multi-protocol mixed environments. Simultaneously, traditional algorithms do not adequately mine the temporal characteristics of traffic, relying solely on historical load data for static prediction. They cannot capture the dynamic trends of traffic changes in real time, especially when sudden or abnormal traffic occurs. Often, adjustments are only made passively after the actual occurrence of the anomaly, failing to meet the high reliability and low latency requirements of domestic IT innovation platforms.

[0003] Furthermore, existing load balancing technologies have significant shortcomings in handling abnormal traffic. Their anomaly detection mechanisms typically operate on a single dimension, failing to perform multi-level anomaly detection and isolation based on collected traffic data (including protocol type, port number, time-series characteristics, and historical load change information). This can easily lead to the spread of abnormal traffic, impacting the stability of the entire network system. Moreover, traditional solutions lack effective self-learning and parameter optimization mechanisms, making it difficult to continuously adjust strategies based on dynamic changes in protocol type and port distribution. As network environments become more complex, their load balancing accuracy and distribution efficiency gradually decline.

[0004] Therefore, it is necessary to improve the network load balancing algorithm of the existing information technology innovation platform to overcome the shortcomings of the existing technology. Summary of the Invention

[0005] To overcome the problems existing in related technologies, one of the objectives of this invention is to provide a network load balancing algorithm based on a domestic IT innovation platform. This method can improve data processing efficiency, achieve intelligent load scheduling, ensure the priority of resource allocation for ports of domestic IT innovation terminals, and guarantee service continuity.

[0006] A network load balancing algorithm based on a domestic IT innovation platform includes the following steps:

[0007] S1. Collect real-time traffic data from multiple ports, including protocol type, port number, time series characteristics and historical load change information, and perform multi-level anomaly detection and isolation processing on the collected data, and output the processed structured data.

[0008] S2. Perform preprocessing on the collected data, including normalization, outlier removal, and protocol field standardization, to generate structured multi-port traffic data;

[0009] S3. Based on the protocol parsing algorithm, multi-dimensional protocol fingerprint features are extracted from the traffic of each port, and a port-protocol association mapping table is established to realize multi-protocol compatibility analysis;

[0010] S4. Use a time-series neural network to model the historical and real-time traffic of each port, and output the traffic prediction value and abnormal traffic probability assessment in the short future period.

[0011] S5. Based on the protocol characteristics and traffic change trends of each port, dynamically generate multi-port distribution weight factors to achieve adaptive parameter configuration in heterogeneous protocol environments.

[0012] S6. Input the prediction results and real-time load status into the reinforcement learning scheduling agent, which will select a feedforward distribution strategy, including pre-allocating redundant resources, dynamically adjusting the weight of each port, or activating backup links.

[0013] S7. When an abnormal traffic trend is detected, feedforward load distribution adjustment is performed immediately without waiting for the actual abnormality to occur.

[0014] S8. Monitor the actual traffic distribution and system load performance after feedforward adjustment, and feed back the abnormal event handling results to the prediction model and scheduling agent to achieve model self-learning and continuous parameter optimization.

[0015] S9. Based on the protocol type and port distribution, regularly adapt and update the protocol compatibility parameter group to ensure continuous high-precision identification and distribution efficiency in multi-port environments.

[0016] In a preferred embodiment of the present invention, the anomaly detection in step S1 includes packet loss rate detection, port active status heartbeat detection, multi-feature consistency verification, and anomaly clustering analysis.

[0017] In a preferred embodiment of the present invention, in step S2, the data normalization method is Z-score normalization or Min-Max scaling.

[0018] In a preferred embodiment of the present invention, in step S3, the protocol fingerprint features include protocol name, version number, header features, extended flags, session duration distribution, typical message length sequence, and header field variability.

[0019] In a preferred embodiment of the present invention, in step S4, the temporal neural network is a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU).

[0020] In a preferred embodiment of the present invention, in step S5, the distribution weight factors are generated using principal component analysis (PCA) or an adaptive weighted fusion algorithm, and the weight factors are normalized within intervals and anomaly detection is performed to ensure that the total allocation is 1.

[0021] In a preferred embodiment of the present invention, in step S6, the reinforcement learning scheduling agent employs an intelligent decision-making algorithm based on a value function or policy gradient to perform action scoring and optimization for various distribution strategies.

[0022] In a preferred embodiment of the present invention, in step S7, the feedforward adjustment strategy includes temporarily increasing the redundant bandwidth of the affected port, dynamically reducing the weight of high-risk ports, or switching some traffic to a backup link.

[0023] In a preferred embodiment of the present invention, in step S8, the closed-loop feedback data includes the effect of traffic distribution adjustment, anomaly recovery time, and resource utilization, and is used for adaptive parameter optimization of the traffic prediction model and scheduling agent.

[0024] In S9, the protocol compatibility parameter group includes feature extraction rules, protocol category labels, and distribution priority settings, and the port-protocol association mapping table is updated in real time through the interface.

[0025] The second objective of this invention is to provide a domestically developed IT terminal for implementing the network load balancing algorithm based on the domestically developed IT platform as described above.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention provides a network load balancing algorithm and a domestically developed domestically built-in terminal based on a domestically developed domestically built-in platform. The algorithm includes the following steps: collecting real-time traffic data from multiple ports, including protocol type, port number, time series characteristics, and historical load change information; performing multi-level anomaly detection and isolation processing on the collected data; outputting processed structured data; performing preprocessing on the collected data, including normalization, outlier removal, and protocol field standardization, to generate structured multi-port traffic data; extracting multi-dimensional protocol fingerprint features from the traffic of each port based on a protocol parsing algorithm, and establishing a port-protocol association mapping table to achieve multi-protocol compatibility analysis; using a time-series neural network to model the historical and real-time traffic of each port, outputting predicted traffic values ​​and anomaly traffic probability assessments for the short term; and based on each... This algorithm dynamically generates multi-port distribution weight factors based on port protocol characteristics and traffic change trends, enabling adaptive parameter configuration in heterogeneous protocol environments. The prediction results and real-time load status are input into a reinforcement learning scheduling agent, which selects a feedforward distribution strategy, including pre-allocating redundant resources, dynamically adjusting port weights, or activating backup links. Upon detecting abnormal traffic trends, feedforward load distribution adjustments are immediately executed without waiting for actual anomalies to occur. The algorithm monitors the actual traffic distribution and system load performance after feedforward adjustments, feeding back the anomaly handling results to the prediction model and scheduling agent, achieving model self-learning and continuous parameter optimization. Protocol compatibility parameter groups are periodically adapted and updated based on protocol type and port distribution, ensuring continuous high-precision identification and distribution efficiency in multi-port environments. Through a multi-layered anomaly detection and isolation mechanism, this algorithm achieves comprehensive screening of traffic data. Compared to traditional single-dimensional detection, it effectively improves the anomaly identification rate, ensures high reliability of input data, effectively prevents the spread of abnormal traffic, and enhances system stability. Furthermore, the temporal neural network model effectively captures the long-term dependencies and short-term fluctuations in traffic data, providing more accurate decision-making basis for load scheduling. Reinforcement learning scheduling agent continuously optimizes distribution strategies through self-learning mechanisms. In heterogeneous protocol environments, the system load balance (standard deviation / mean) decreases, effectively improving resource utilization, reducing response latency, and dynamically adjusting weight factors to ensure the priority of resource allocation for critical business ports and guarantee service continuity. Attached Figure Description

[0028] Figure 1 The flowchart of the network load balancing algorithm based on the domestic IT innovation platform is provided in the embodiments of this application. Detailed Implementation

[0029] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0030] The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” as used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0031] It should be understood that although the terms "first," "second," "third," etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this invention, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, features defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0032] In existing technologies, most load balancing solutions employ static weight allocation or simple round-robin mechanisms, which are ill-suited to scenarios with diverse protocol types (such as the coexistence of new protocols like TCP, UDP, and HTTP / 3 with traditional protocols) in domestically developed IT platforms. These solutions lack deep analysis capabilities of protocol characteristics and cannot establish accurate port-protocol mappings, leading to low distribution efficiency and port load imbalances in multi-protocol environments. Furthermore, traditional algorithms fail to adequately mine the temporal characteristics of traffic, relying solely on historical load data for static prediction. They cannot capture dynamic traffic trends in real time, especially during sudden or abnormal traffic spikes, often requiring reactive adjustments only after the anomaly actually occurs, failing to meet the high reliability and low latency requirements of domestically developed IT platforms. Moreover, existing load balancing technologies have significant shortcomings in handling abnormal traffic. Their anomaly detection mechanisms typically operate on a single dimension, unable to perform multi-level anomaly detection and isolation of collected traffic data (including protocol type, port number, time-series characteristics, and historical load change information), easily leading to the spread of abnormal traffic and impacting the stability of the entire network system. Furthermore, traditional solutions lack effective self-learning and parameter optimization mechanisms, making it difficult to continuously adjust strategies based on dynamic changes in protocol type and port distribution. As the network environment becomes more complex, their load balancing accuracy and distribution efficiency will gradually decline.

[0033] Based on this, this application provides a network load balancing algorithm based on the domestic IT innovation platform.

[0034] Example

[0035] like Figure 1 As shown, this embodiment provides a network load balancing algorithm based on a domestic IT innovation platform, including the following steps:

[0036] S1. Collect real-time traffic data from multiple ports, including protocol type, port number, time series characteristics and historical load change information, and perform multi-level anomaly detection and isolation processing on the collected data, and output the processed structured data.

[0037] S2. Perform preprocessing on the collected data, including normalization, outlier removal, and protocol field standardization, to generate structured multi-port traffic data;

[0038] S3. Based on the protocol parsing algorithm, multi-dimensional protocol fingerprint features are extracted from the traffic of each port, and a port-protocol association mapping table is established to realize multi-protocol compatibility analysis;

[0039] S4. Use a time-series neural network to model the historical and real-time traffic of each port, and output the traffic prediction value and abnormal traffic probability assessment in the short future period.

[0040] S5. Based on the protocol characteristics and traffic change trends of each port, dynamically generate multi-port distribution weight factors to achieve adaptive parameter configuration in heterogeneous protocol environments.

[0041] S6. Input the prediction results and real-time load status into the reinforcement learning scheduling agent, which will select a feedforward distribution strategy, including pre-allocating redundant resources, dynamically adjusting the weight of each port, or activating backup links.

[0042] S7. When an abnormal traffic trend is detected, feedforward load distribution adjustment is performed immediately without waiting for the actual abnormality to occur.

[0043] S8. Monitor the actual traffic distribution and system load performance after feedforward adjustment, and feed back the abnormal event handling results to the prediction model and scheduling agent to achieve model self-learning and continuous parameter optimization.

[0044] S9. Based on the protocol type and port distribution, regularly adapt and update the protocol compatibility parameter group to ensure continuous high-precision identification and distribution efficiency in multi-port environments.

[0045] Specifically, one implementation of this method includes the following steps:

[0046] (1) Data Acquisition and Anomaly Handling

[0047] In a domestic IT cloud data center scenario, the system collects real-time traffic data through 16 gigabit network ports. During the collection process, it simultaneously acquires the protocol type (e.g., TCP, UDP), port number (1-65535), timestamp (accurate to milliseconds), and historical port load data for the past 24 hours (sampling interval of 100ms) for each data packet. For the collected data, multi-level anomaly detection is first performed:

[0048] Based on statistical feature detection: calculate the mean and variance of traffic at each port, and mark traffic points exceeding 3 times the standard deviation as anomalies;

[0049] Based on time-series pattern recognition: Flow time-series curves are analyzed using a sliding window (window size of 5 minutes), and abnormal behaviors that deviate from historical patterns are identified through the Dynamic Time Warping (DTW) algorithm;

[0050] Protocol field validation: The validity of TCP protocol flag combinations such as SYN and ACK is checked, filtering packets with abnormal FIN / FIN-ACK combinations. Upon detecting abnormal data, abnormal traffic is immediately redirected to a dedicated cleaning link via hardware isolation, ensuring that the reliability of the output structured data reaches over 99.99%.

[0051] (2) Data preprocessing

[0052] The collected raw traffic data is input into the preprocessing module. First, normalization is performed: the Min-Max normalization method is used for different dimensions of features such as the number of bytes and packets, mapping them to the [0,1] interval; second, outliers outside the range of 1.5 times the upper and lower quartiles are removed using the IQR (interquartile range) method; finally, the protocol fields are standardized, for example, the HTTP protocol request methods (GET, POST, etc.) are converted into enumeration values ​​of a unified format to generate a structured multi-port traffic dataset, providing standardized input for subsequent processing.

[0053] (3) Protocol fingerprint extraction and mapping

[0054] Using domestically developed protocol parsing algorithms, in-depth analysis of traffic on various ports is performed. Taking HTTP / 3 protocol traffic as an example, the following multi-dimensional features are extracted to construct a protocol fingerprint:

[0055] Header field characteristics: QUIC version number, ConnectionID length, Transport Parameters field structure;

[0056] Data segment pattern characteristics: byte distribution entropy value of encrypted payload, time interval pattern of frame type (DATA, ACK, PING, etc.);

[0057] Interaction timing characteristics: the time delay and round-trip time (RTT) variation patterns from the client's initial handshake to data transmission. Based on these characteristics, a protocol fingerprint database is constructed, and a port-protocol association mapping table is established. For example, it is identified that port 8080 primarily carries HTTP / 3 protocol traffic, while port 443 simultaneously carries HTTPS and WebSocket protocol traffic, achieving an accuracy rate of 98.7% in multi-protocol compatibility analysis.

[0058] (4) Traffic forecasting modeling

[0059] An LSTM neural network is used to model the traffic at each port. The network structure consists of two LSTM layers (128 neurons each), one fully connected layer, and a soft max output layer. Input features include traffic data from the past hour (one sampling point per minute), the corresponding protocol type distribution, and port load history. During training, the Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and 50 iterations. The model outputs traffic predictions for the next 5 minutes (including byte and packet count predictions) and an assessment of abnormal traffic probability (e.g., a probability of exceeding a threshold of 0.12). The mean squared error (MSE) of the prediction is controlled within 5%.

[0060] (5) Dynamic generation of weighting factors

[0061] The distribution weight factor for each port is dynamically generated based on the following parameters:

[0062] Resource consumption characteristics of protocol types: HTTP / 3 protocol consumes more CPU resources due to encryption processing, so its weighting coefficient is set to 0.8; TCP protocol's weighting coefficient is set to 1.0.

[0063] Current port load rate: The CPU utilization, memory usage and bandwidth utilization of each port are calculated in real time, and the load rate index is obtained by normalization and weighting.

[0064] Historical traffic fluctuation range: Calculates the standard deviation of traffic over the past 10 minutes, reflecting traffic stability;

[0065] Predicting traffic trends: Adjusting weights based on the expected traffic growth trend (e.g., a projected 20% increase) output by the LSTM model over the next 5 minutes. Using the above parameters, a weighted summation formula is used to generate weighting factors. For example, the weight of port A is calculated as: 0.4 × protocol coefficient + 0.3 × load rate + 0.2 × fluctuation amplitude + 0.1 × predicted trend, achieving adaptive parameter configuration in heterogeneous protocol environments.

[0066] (6) Reinforcement learning scheduling and feedforward strategy

[0067] The scheduling agent module is implemented using the PPO algorithm. Its state space includes 20 dimensions of features such as real-time port load, predicted traffic, and protocol type distribution. The action space includes six actions: adjusting port weights (step size 0.05), allocating redundant CPU cores (1-4 cores), and activating backup links (1-2 links). The reward function is designed as: 0.5 × load balancing degree + 0.3 × response latency optimization + 0.2 × resource utilization, where load balancing degree is defined as 1 - standard deviation / mean of port load. When the LSTM model predicts that the probability of abnormal traffic in the next 5 minutes exceeds the 0.3 threshold, the scheduling agent immediately executes a feedforward strategy: pre-allocating 2 CPU cores to the high-load port, adjusting its weight from 0.7 to 0.5, and activating backup links to share 30% of the traffic. The entire adjustment process is completed within 100ms.

[0068] (7) Model self-learning and parameter optimization

[0069] After feedforward adjustments, the actual traffic distribution of each port is continuously monitored (e.g., the standard deviation of port load decreases from 0.4 to 0.2 after adjustment) and system load performance (response latency decreases from 50ms to 35ms). The results of abnormal event handling (including action type, adjustment parameters, and performance metrics) are used as training samples and fed back into the LSTM model and the PPO scheduling agent. An online learning mechanism is employed; after processing every 100 abnormal events, the weight parameters of the LSTM model are automatically updated (learning rate decays by 0.95 times) and the policy network parameters of the PPO agent, achieving continuous improvement in model accuracy. After 1000 iterations, the accuracy of abnormal traffic prediction increases from 85% to 92%.

[0070] (8) Protocol parameter adaptation and update

[0071] The parameter update cycle is dynamically adjusted based on the frequency of network traffic changes (e.g., low traffic periods in the early morning) and protocol type updates (e.g., the addition of new domestic encryption protocols). In this embodiment, a weekly automatic scan of the protocol fingerprint database is set. When a new protocol feature is detected (e.g., an unknown protocol header field appears), an immediate update process is triggered: the port-protocol mapping table is updated through a combination of manual annotation and automatic clustering, and the protocol parsing model is retrained to ensure that the protocol recognition accuracy in a multi-port environment remains above 98%.

[0072] The above describes a network load balancing algorithm based on a domestic IT innovation platform. This algorithm employs a multi-layered anomaly detection and isolation mechanism to achieve comprehensive filtering of traffic data. Compared to traditional single-dimensional detection, it effectively improves the anomaly identification rate, ensures high reliability of input data, effectively prevents the spread of abnormal traffic, and enhances system stability. The temporal neural network model effectively captures the long-term dependencies and short-term fluctuation characteristics of traffic data, providing more accurate decision-making basis for load scheduling. The reinforcement learning scheduling agent continuously optimizes the distribution strategy through a self-learning mechanism. In heterogeneous protocol environments, the system load balancing degree (standard deviation / mean) decreases, effectively improving resource utilization, reducing response latency, and dynamically adjusting weight factors to ensure the priority of resource allocation for critical business ports, guaranteeing service continuity.

[0073] Furthermore, in S1, anomaly detection includes packet loss rate detection, port activity status heartbeat detection, multi-feature consistency verification, and anomaly clustering analysis.

[0074] Specifically, S1: Collects real-time traffic data from multiple ports on the domestic IT innovation platform, including protocol type, port number, time series characteristics, and historical load change information.

[0075] This step is used to synchronously collect real-time network traffic data from multiple physical or virtual ports in a domestic IT platform environment, and record in detail the protocol type, port number, time series characteristics, and historical load change information of each traffic item. As the basic data support for multi-port heterogeneous protocol identification and load balancing optimization, this step ensures that subsequent core algorithm modules such as data preprocessing, protocol feature extraction, and intelligent scheduling have high-quality raw data input, and is the first step in realizing the system's efficient compatibility and resilient distribution capabilities.

[0076] S1 includes:

[0077] S1.1 monitors all target ports deployed on the domestic IT innovation platform in real time, and continuously captures raw network packets that flow in and out through high-performance data acquisition interfaces (such as DPDK, PF_RING, or self-developed high-concurrency acquisition engines) to obtain complete real-time traffic information.

[0078] S1.2 Based on the captured raw data packets, it parses and extracts the port number and session identifier information of each data packet, and uses a protocol parsing library (such as Wireshark distributor or a custom protocol stack analyzer) to automatically identify the specific network protocol type to which the data packet belongs, thereby realizing data classification and attribution in a multi-protocol environment.

[0079] S1.3 slices the data stream collected from each port according to the time window and counts time series characteristic parameters such as traffic rate, peak value, number of packets, and average packet size to obtain a fine-grained time-series load curve at the port level.

[0080] S1.4 associates the data collected at the current moment with the historical load change information of the port, and maintains the port-level historical traffic data based on the ring buffer or persistent database to provide continuous input for subsequent traffic prediction and anomaly detection.

[0081] S1.5 performs real-time monitoring of abnormal situations during the acquisition process (such as packet loss, port failure, and data packet corruption), uses anomaly detection algorithms to determine data integrity, marks and isolates abnormal data, and ensures that subsequent processing modules receive highly reliable input.

[0082] Using the port-level historical traffic data output from step S1.4 and the raw data packets collected in real time as input, real-time detection of abnormal conditions is implemented throughout the entire traffic collection process in a multi-port environment of the domestic IT innovation platform.

[0083] By employing distributed acquisition and monitoring probes, combined with a high-precision clock synchronization mechanism, continuous monitoring of multiple indicators such as data packet collection rate, data packet sequence integrity, and data frame validity is achieved for each target port acquisition link.

[0084] The sliding window packet loss rate detection algorithm (parameters: window length N, threshold P_thresh) compares the data packet sequence of each port within a set time window, calculates the difference between the actual number of collected packets and the theoretical number of packets, and uses the following formula to calculate the packet loss rate:

[0085]

[0086] Among them, LossRate i Let Rece i ved be the packet loss rate of port i. i Expected represents the actual number of data packets received. i This represents the theoretically expected number of data packets to be received.

[0087] Furthermore, the port activity status heartbeat detection mechanism (parameters: heartbeat period T_beat, loss threshold M) is used to continuously detect the online activity of the port. If no heartbeat response is received for M consecutive periods, the port is judged to be faulty and an alarm flag is triggered.

[0088] Using a multi-feature consistency verification algorithm, the integrity of key fields (such as protocol header, checksum, packet length, etc.) of captured data packets is verified. Potential data packet corruption or tampering events are identified through CRC check and hash comparison, and abnormal data is isolated and marked.

[0089] Based on anomaly clustering detection methods (such as DBSCAN or the Isolation Forest algorithm), statistical analysis is performed on the frequency of anomaly events, their spatial distribution, and their similarity to historical anomaly patterns to automatically determine whether they are systemic failures, sudden storms, or occasional noise.

[0090] Through the above multi-level anomaly detection and isolation process, the marked abnormal data is removed from the structured acquisition results, and only the high-reliability traffic data that has been verified for integrity and confirmed for status is retained and output to the subsequent data preprocessing module.

[0091] Through the above-mentioned chain-like anomaly detection and isolation processing method, the raw multi-port traffic data collected in the previous step is transformed into highly complete and highly reliable structured input data, thereby achieving data quality assurance and anomaly resilience enhancement of the multi-port load balancing system of the domestic IT innovation platform under extreme scenarios.

[0092] For example, in a multi-port traffic collection scenario of a bank's IT terminal access platform, the collection window length N is configured to be 10,000 packets, the packet loss rate alarm threshold P_thresh is set to 0.01 (1%), the heartbeat detection period T_beat is 5 seconds, and the packet loss threshold M is 3. In actual operation, within a certain period, port 2 should theoretically receive 10,000 packets, but actually receives 9,800 packets. Therefore, the packet loss rate is:

[0093]

[0094] Because the packet loss rate exceeded the threshold, the system automatically marked the port as having packet loss anomalies. Simultaneously, the port failed to receive a signal for three consecutive heartbeat cycles, indicating a port failure and triggering an alarm. Corruption in some packet headers was detected through CRC and hash verification; this data was isolated and recorded in the anomaly log. During the one-week operation period, the Isolation Forest algorithm detected an anomaly pattern on port 5 similar to previous network storms. After isolation, this effectively prevented large-scale anomaly data from impacting load balancing scheduling. The resulting high-reliability traffic data significantly improved the system's robustness and data credibility in extreme scenarios during subsequent feature modeling and policy generation.

[0095] Furthermore, in S2, the data normalization method is Z-score normalization or Min-Max scaling.

[0096] S2: Preprocess the collected multi-port traffic data, including data normalization, outlier removal, and protocol field standardization, to improve the accuracy of subsequent identification.

[0097] This step is responsible for the systematic preprocessing of multi-source traffic data collected from various ports of the self-innovation platform, including data normalization, outlier removal, and protocol field standardization. By improving data consistency and validity, it provides a high-quality, structured input foundation for subsequent core steps such as protocol feature extraction, traffic prediction modeling, and distribution strategy optimization. It is a crucial prerequisite for ensuring the accuracy of load balancing algorithm identification and the real-time performance of the system in a heterogeneous protocol environment.

[0098] S2 includes:

[0099] S2.1 performs format standardization processing on the raw traffic data collected from each port, including standard conversion of timestamps, port numbers, and protocol fields, to ensure the consistency of the input structure of multi-port and multi-source data, which facilitates subsequent batch processing and feature alignment.

[0100] S2.2 Normalizes key numerical features in traffic data (such as packet length, traffic rate, connection latency, etc.) based on normalization algorithms (such as Z-score normalization or Min-Max scaling) to eliminate the difference in units between different ports and improve the numerical stability of subsequent modeling processes.

[0101] S2.3 Apply statistical detection methods (such as the 3σ principle, IQR interval detection, etc.) to detect outliers in the historical and real-time traffic data of each port, identify and remove noisy or invalid traffic records, so as to reduce the interference of extreme anomalies on model training and protocol analysis.

[0102] S2.4 performs standardized mapping processing on the protocol fields, mapping various protocol identifiers, version numbers and extended attributes collected from different ports to unified protocol category labels through the protocol dictionary, so as to facilitate subsequent heterogeneous protocol compatibility analysis and automatic identification.

[0103] S2.5 integrates and normalizes the data, which has been freed of anomalies, and standardizes the protocol fields to form a structured multi-port traffic preprocessing result, which is then output to the feature extraction and modeling module to provide high-quality data input for the next step of protocol fingerprint extraction and traffic prediction modeling.

[0104] Furthermore, in S3, the protocol fingerprint features include protocol name, version number, header features, extended flags, session duration distribution, typical message length sequence, and header field variability.

[0105] S3: Based on the heterogeneous protocol characteristics of each port, extract multi-dimensional protocol fingerprint features and establish a port-protocol association mapping table to realize multi-protocol compatibility analysis.

[0106] This step targets multi-port access scenarios on domestic IT innovation platforms. Based on the heterogeneous protocol characteristics of each port, it extracts multi-dimensional protocol fingerprint features and establishes a port-protocol association mapping table to achieve compatibility analysis in a multi-protocol environment. Through the systematic extraction and mapping of protocol features, it effectively supports subsequent automatic protocol identification and dynamic adjustment of load distribution strategies, and is a core component in improving the adaptability of multi-port heterogeneous protocols and the accuracy of intelligent load balancing strategies.

[0107] S3 includes:

[0108] S3.1 analyzes and extracts protocol-related fields, including but not limited to protocol name, version number, header features, and extended flags, for preprocessed multi-port traffic data based on protocol parsing algorithms (such as Deep Packet Inspection (DPI) and rule-based protocol identification engines) to obtain the basic protocol attribute set for each data stream.

[0109] S3.2 For different types of protocol traffic, feature engineering methods (such as statistical distribution analysis, temporal behavior modeling, and content entropy calculation) are used to extract multi-dimensional protocol fingerprint features, including session duration distribution, typical message length sequence, header field variability, encryption characteristic indicators, etc., to form high-dimensional feature vectors that can distinguish different protocols and their variants.

[0110] S3.3 uses clustering algorithms (such as DBSCAN, hierarchical clustering) or classifiers (such as decision trees, SVM) to classify and label multiple protocol instances existing in the same port based on the extracted protocol fingerprint features, generating protocol category labels and their feature expressions, thereby improving the automatic identification capability in heterogeneous protocol environments.

[0111] S3.4 establishes a port-protocol association mapping table, which structurally archives the main protocol category, fingerprint feature vector and compatibility level corresponding to each port. It enables efficient querying and dynamic maintenance of port and protocol attributes through relational database or high-performance key-value storage, providing underlying support for subsequent distribution weight generation and adaptive strategy configuration.

[0112] S3.5 performs periodic consistency checks and real-time updates to the port-protocol mapping table. Based on newly added traffic samples or detected new protocol fingerprints, it automatically expands the fingerprint database and updates port attributes to ensure the system's continuous compatibility with unknown or mutated protocol types, thereby improving long-term identification accuracy and scalability in multi-port environments.

[0113] Furthermore, in S4, the temporal neural network is a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU).

[0114] S4: Use a time-series neural network to model the historical and real-time traffic of different ports, and output the traffic prediction value and abnormal traffic probability assessment for each port in the short future period.

[0115] This step utilizes a temporal neural network to model the historical and real-time traffic of different ports, aiming to predict the traffic trends and the probability of abnormal traffic occurrences in the short term for each port. Within the overall technical solution, it plays a core role in providing feedforward decision support for the multi-port dynamic distribution and load balancing scheduling modules. Through traffic trend prediction and risk assessment, it enables proactive detection of sudden abnormal events, providing a scientific basis for the scheduling agent to adjust distribution strategies in advance, and improving the system's adaptability and resilience in extreme scenarios.

[0116] S4 includes:

[0117] S4.1 performs time series feature alignment processing on the structured multi-port historical and real-time traffic data, and uses a sliding window algorithm to synchronously organize the data from different ports according to a unified time segmentation, so as to obtain multi-dimensional time series samples that meet the modeling input requirements.

[0118] Based on the aforementioned time-series samples, S4.2 employs normalization preprocessing methods (such as Z-score or Min-Max scaling) to standardize numerical features such as traffic rate, packet count, and average message length, eliminating scale differences between different ports and providing numerically stable input for neural network model training.

[0119] S4.3 uses a time-series neural network (such as LSTM or GRU) to model and train the historical and real-time traffic sequences of each port. It fits the port-level traffic change pattern through supervised learning and periodically optimizes the model parameters to improve the accuracy of traffic trend prediction.

[0120] S4.4 inputs the latest real-time traffic data into the trained temporal neural network, performs rolling predictions on the traffic values ​​of each port in the short future period, and outputs the traffic trend in the future time window as the input for subsequent decisions.

[0121] S4.5 combines the prediction residuals, historical anomaly labels, and contextual statistical distributions from the model output, and uses probabilistic inference algorithms (such as Bayesian anomaly detection or statistical thresholding) to perform probability assessments on possible abnormal traffic during the prediction period, outputting the probability of future abnormal events occurring at each port.

[0122] S4.6 structures and archives the future traffic prediction values ​​and corresponding anomaly probability results for each port, and outputs them synchronously to the multi-port distribution weight generation and reinforcement learning scheduling agent module, providing high-quality prediction basis for dynamic weight adjustment and feedforward distribution decisions.

[0123] Furthermore, in step S5, the distribution weight factors are generated using principal component analysis (PCA) or an adaptive weighted fusion algorithm, and the weight factors are normalized within intervals and anomaly detection is performed to ensure that the total allocation is 1.

[0124] S5: Dynamically generates multi-port distribution weight factors based on the protocol type and traffic change trends of different ports, enabling adaptive parameter configuration in heterogeneous protocol environments.

[0125] This step dynamically generates multi-port distribution weight factors based on the protocol type and traffic change trends of different ports, enabling adaptive parameter configuration in heterogeneous protocol environments. This step plays a core bridging role in the overall technical solution, connecting the analysis results of front-end protocol characteristics and traffic prediction modeling, and providing a real-time adjustable distribution weight basis for subsequent intelligent scheduling and dynamic distribution, thereby supporting the implementation of efficient and refined load balancing strategies in multi-port environments.

[0126] S5 includes:

[0127] S5.1, based on the protocol type and compatibility level of each port output by the protocol fingerprint extraction module, structurally collects the protocol feature data of each port, and constructs a port-protocol attribute vector through the protocol label mapping algorithm, providing the input basis for the generation of distribution weight factors.

[0128] S5.2 combines the future traffic trends and anomaly probability assessment results of each port output by the traffic prediction model with the real-time traffic change curve, and uses a weighted time-series sliding window algorithm to dynamically calculate the traffic pressure index of each port to reflect the current and short-term future port load risk.

[0129] S5.3 utilizes multi-feature fusion algorithms (such as principal component analysis (PCA) or adaptive weighted fusion) to comprehensively model the protocol attribute vector and traffic pressure index, forming a port-level comprehensive load feature set, providing a high-dimensional feature foundation for weight factor parameterization.

[0130] Based on the constructed comprehensive load feature set, S5.4 uses an adaptive weight adjustment algorithm (such as dynamic allocation based on interval normalization, Bayesian optimization, etc.) to generate distribution weight factors for each port in real time, so as to realize the adaptive adjustment of parameters under different protocol types and different traffic trends.

[0131] S5.5 performs boundary constraints and anomaly detection on the generated distribution weight factors. It uses threshold judgment and robustness analysis methods to remove mutated or abnormal parameters, ensuring the legality of all weight factors and that the total allocation is 1, thus providing a guarantee for the safe invocation of subsequent scheduling modules.

[0132] S5.6 outputs the final calculated multi-port distribution weight factor to the load balancing scheduling and policy proxy module, and synchronizes the current weight configuration and related original feature data through the interface to realize end-to-end adaptive parameter transmission and subsequent feedback closed-loop optimization.

[0133] Furthermore, in step S6, the reinforcement learning scheduling agent employs an intelligent decision-making algorithm based on a value function or policy gradient to score and optimize actions across multiple distribution strategies. Step S6 includes:

[0134] S6.1 collects the future traffic trends and anomaly probability assessment results of each port output by the traffic prediction model, and simultaneously obtains the current real-time load status parameters of all ports (such as bandwidth utilization, number of connections, response latency, etc.) to form a complete input feature set required for scheduling decisions.

[0135] Based on the collected prediction and real-time load data, S6.2 uses a feature fusion algorithm to perform multi-dimensional quantitative assessment of the risk level of each port, extracts key indicators that affect the selection of distribution strategies (such as high-risk port identification and resource bottleneck location), and provides interpretable input for the scheduling agent.

[0136] S6.3 inputs the fused features into the reinforcement learning scheduling agent and uses an intelligent decision-making algorithm based on the value function or policy gradient to score and select candidate actions for different distribution strategies (such as pre-allocation of redundant resources, dynamic adjustment of weights, activation of backup links, etc.) to obtain the optimal feedforward distribution strategy.

[0137] S6.4, based on the optimal strategy output by the scheduling agent, combines the current port compatibility parameters and system resource constraints to perform feasibility verification and parameter refinement on the specific distribution actions involved in the strategy, generating a feedforward resource allocation and traffic adjustment instruction set that can be directly issued and executed.

[0138] S6.5 synchronously sends the refined feedforward distribution instructions to the multi-port load balancing scheduling module and records the key decision parameters, strategy selection basis, and expected performance indicators in this round of decision-making process, providing data support for subsequent execution monitoring and closed-loop feedback optimization.

[0139] S7: Upon detecting a sudden surge in abnormal traffic trends, immediately implement feedforward load balancing adjustments without waiting for the actual anomaly to occur, thereby improving system response speed and load resilience in extreme scenarios. This step aims to enable the system to immediately implement feedforward load balancing adjustments when it detects a sudden surge in abnormal traffic trends in the network, without waiting for the anomaly to actually occur. This function is crucial in the overall technical solution, as it proactively mitigates the impact of extreme events such as network storms or traffic surges on the system's load balancing performance, significantly improving the system's response speed and resilience to extreme scenarios, thus ensuring high availability and service continuity of the domestic IT innovation platform in a multi-port environment.

[0140] Furthermore, in S7, the feedforward adjustment strategy includes temporarily increasing the redundant bandwidth of affected ports, dynamically reducing the weight of high-risk ports, or switching some traffic to backup links.

[0141] S7 includes:

[0142] S7.1 uses the port-level future abnormal traffic probability assessment results output by the traffic prediction model to determine the abnormal trend of all monitored ports. When the abnormal probability of a port or port combination exceeds the preset threshold, a feedforward load adjustment decision process is triggered to achieve early detection of potential emergencies.

[0143] For ports identified as exhibiting abnormal trends and their associated traffic characteristics, S7.2 utilizes real-time load status and scheduling agent feedback to quickly generate targeted distribution adjustment strategies, including but not limited to temporarily increasing the redundant bandwidth of affected ports, dynamically reducing the weight of high-risk ports, or switching some traffic to backup links, thereby achieving feedforward resource reallocation.

[0144] S7.3 sends the generated feedforward adjustment strategy to the multi-port load balancing scheduling module. Based on the current number of connections, protocol compatibility and hardware resource status of each port, it automatically executes specific traffic redistribution actions to ensure that the adjustment strategy can be implemented and take effect in the shortest possible time.

[0145] S7.4 monitors key parameters (such as policy delivery delay, traffic switching time, and actual traffic distribution changes) in the feedforward adjustment process in real time. It combines reinforcement learning scheduling agent to evaluate the policy implementation effect in real time and dynamically fine-tune the adjustment range or response range accordingly to maximize distribution efficiency and system resilience in abnormal scenarios.

[0146] S8: Monitor the actual traffic distribution and system load performance after feedforward adjustment, and feed back the results of abnormal event handling to the prediction model and scheduling agent to achieve model self-learning and continuous parameter optimization.

[0147] This step is responsible for continuously monitoring the actual traffic distribution and overall system load performance of each port after performing feedforward traffic distribution adjustments, and feeding back the handling results of sudden abnormal events as feedback data to the traffic prediction model and reinforcement learning scheduling agent. Through a closed-loop feedback mechanism, the prediction and scheduling models achieve self-learning and continuous parameter optimization, improving the system's adaptability to extreme scenarios and the accuracy of subsequent load balancing decisions. It is the core component for achieving highly resilient, multi-port adaptive load balancing.

[0148] Furthermore, in step S8, the closed-loop feedback data includes the traffic distribution adjustment effect, anomaly recovery time, and resource utilization, and is used for adaptive parameter optimization of the traffic prediction model and scheduling agent; step S8 includes:

[0149] S8.1 collects and records the actual traffic data of each port and the overall system load performance in real time after the feedforward load adjustment is executed. It uses a high-resolution monitoring module to obtain port-level traffic changes, system bandwidth utilization, response latency and abnormal alarm information to comprehensively reflect the real-time operating status after the distribution strategy adjustment.

[0150] Based on real-time collected traffic and load performance data, S8.2 applies statistical analysis algorithms (such as moving average and anomaly detection) to compare the differences in traffic distribution before and after feedforward adjustment, determine the effectiveness of abnormal event handling, identify potential distribution bottlenecks, and provide quantitative evaluation indicators for subsequent feedback.

[0151] S8.3 takes the actual results of each abnormal event handling, including the effect of traffic distribution adjustment, abnormal recovery time, resource utilization and other key performance parameters, and organizes them into training samples. These samples are then used as feedback data to input into the traffic prediction model, and the model parameters are adaptively fine-tuned to improve the accuracy of future abnormal trend prediction.

[0152] S8.4 For the abnormal event handling results of the feedback and sorting, the relevant state characteristics and policy execution results are simultaneously input into the reinforcement learning scheduling agent. By updating the reward function or policy value evaluation, the scheduling policy parameters are optimized online, and the adaptive response capability of the scheduling agent to extreme scenarios is improved.

[0153] S8.5 performs backtesting and cross-validation periodically on the optimized prediction model and scheduling agent model, ensuring that the model parameters are always in optimal condition and can dynamically adapt to the changing trends of protocol and traffic distribution in a multi-port environment through retraining on historical data and incremental learning on new samples. In S9, the protocol compatibility parameter group includes feature extraction rules, protocol category labels, and distribution priority settings, and updates the port-protocol association mapping table in real time via an interface.

[0154] S9: Based on different protocol types and port distribution, the protocol compatibility parameter group is regularly adapted and updated to ensure continuous high-precision identification and distribution efficiency in multi-port environments.

[0155] This step aims to periodically adapt and update protocol compatibility parameter groups based on different protocol types and port distribution, thereby ensuring continuous high-precision protocol identification and traffic distribution efficiency in multi-port environments. It plays a crucial role in the overall technical solution by dynamically maintaining heterogeneous protocol compatibility, supporting adaptive adjustment of port-protocol mapping, and improving the long-term scalability of the system. It provides the underlying data foundation and compatibility guarantee for continuously outputting optimal configuration parameters for load balancing scheduling strategies and distribution algorithms. S9 includes:

[0156] S9.1 performs periodic statistical analysis on the protocol fingerprint feature data and port distribution information collected in the current multi-port environment, and uses clustering and distribution detection algorithms (such as K-means and distribution entropy analysis) to identify the trend of protocol type changes and newly emerging protocol instances in order to obtain a real-time snapshot of the protocol distribution status.

[0157] S9.2 performs compatibility assessments on newly detected or changed protocol types based on the protocol fingerprint database and port-protocol mapping table. It uses the protocol semantic parsing engine to automatically determine the matching degree between the feature structure and interaction process of the new protocol and the existing distribution mechanism, and outputs a compatibility assessment report.

[0158] For new protocol types that are assessed as needing adaptation, S9.3 calls the parameter template generation module to automatically generate or optimize compatibility parameter groups based on the protocol field structure, communication characteristics, and historical recognition accuracy. These parameters include feature extraction rules, protocol category labels, and distribution priority settings, in order to achieve rapid adaptation to new protocols.

[0159] S9.4 synchronizes newly generated or updated protocol compatibility parameter groups to the multi-port protocol parsing and identification module through the interface, and updates the port-protocol association mapping table in real time to ensure that all subsequent identification and distribution links can perform high-precision protocol differentiation and traffic classification based on the latest parameters.

[0160] S9.5 conducts periodic retrospective evaluations of the actual application effects of the adapted parameter groups, using metrics such as accuracy and recall to analyze the performance of each protocol type in the multi-port identification process. Combined with anomaly identification logs, it identifies declines in identification rate or misallocated ports, triggering parameter fine-tuning or manual review mechanisms. S9.6 archives key data generated during parameter group adaptation and updates (such as records of changes between old and new parameters, identification accuracy analysis results, and dynamic port distribution) into the system management database, providing data support and auditing basis for subsequent model retraining, anomaly tracking, and platform operation and maintenance.

[0161] Example 2

[0162] This embodiment provides a domestically developed terminal, which is used to implement the network load balancing algorithm based on the domestically developed platform as described above.

[0163] Specifically, the hardware of this domestically developed terminal includes a processor, storage devices, network interface cards, and other peripherals such as power modules and heat dissipation devices. The processor's architecture meets the requirements of domestically developed technologies and possesses powerful computing capabilities, enabling it to quickly handle large amounts of network traffic data collection, analysis, and load balancing algorithm calculations, providing a solid computing foundation for the terminal's efficient operation.

[0164] The application layer of this domestically developed terminal deploys a data acquisition module, a data processing module, a protocol parsing and mapping module, a traffic prediction module, a load balancing and scheduling module, and a monitoring and feedback module. The data acquisition module is responsible for collecting real-time traffic data from multiple ports on the network interface card, including protocol type, port number, time series characteristics, and historical load change information. It also performs preliminary processing and format conversion on the collected data according to predefined rules, providing raw data for subsequent data processing.

[0165] The data processing module performs multi-level anomaly detection and isolation, normalization, outlier removal, and protocol field standardization on the collected data, converting the raw data into structured multi-port traffic data to ensure that the data input into the algorithm model is accurate and effective.

[0166] Protocol parsing and mapping module: Based on the protocol parsing algorithm, multi-dimensional protocol fingerprint features are extracted from the traffic of each port, and a port-protocol association mapping table is established to realize the identification and management of traffic of different protocols, providing protocol-level information support for the formulation of load balancing strategies.

[0167] The traffic prediction module uses a time-series neural network to model the historical and real-time traffic of each port. By training the model, it learns the time series patterns of traffic data and outputs the traffic prediction value and abnormal traffic probability assessment for the short term, providing forward-looking data for load balancing decisions.

[0168] The load balancing scheduling module dynamically generates multi-port distribution weight factors based on the protocol characteristics and traffic change trends of each port, and inputs the prediction results and real-time load status into the reinforcement learning scheduling agent. The agent then selects a feedforward distribution strategy to achieve adaptive load balancing of network traffic.

[0169] The monitoring and feedback module continuously monitors the actual traffic distribution and system load performance after feedforward adjustment, and feeds back the results of abnormal event handling to the prediction model and scheduling agent to optimize model parameters and strategy selection, so that the terminal can continuously adapt to changes in the network environment and improve the load balancing effect.

[0170] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A network load balancing method based on a domestically developed information technology platform, characterized in that, Includes the following steps: S1. Collect real-time traffic data from multiple ports, including protocol type, port number, time series characteristics and historical load change information, and perform multi-level anomaly detection and isolation processing on the collected data, and output the processed structured data. S2. Perform preprocessing on the collected data, including normalization, outlier removal, and protocol field standardization, to generate structured multi-port traffic data; S3. Based on the protocol parsing algorithm, multi-dimensional protocol fingerprint features are extracted from the traffic of each port, and a port-protocol association mapping table is established. The protocol fingerprint features include protocol name, version number, header features, extended flags, session duration distribution, typical packet length sequence, and header field variability. S4. Use a time-series neural network to model the historical and real-time traffic of each port, and output the traffic prediction value and abnormal traffic probability assessment in the short future period. S5. Based on the protocol characteristics and traffic change trends of each port, dynamically generate multi-port distribution weight factors to achieve adaptive parameter configuration in heterogeneous protocol environments. S3-S5 includes: extracting multi-dimensional protocol fingerprint features based on the heterogeneous protocol characteristics of each port, and establishing a port-protocol association mapping table; The extracted protocol fingerprint features are used to perform feature classification and labeling on multiple protocol instances existing in the same port, generating protocol category labels and their feature expressions; The main protocol category, fingerprint feature vector and compatibility level corresponding to each port are structured and archived, and efficient query and dynamic maintenance between port and protocol attributes are realized through relational database or high-performance key-value storage. By using a multi-feature fusion algorithm, protocol attribute vectors and traffic pressure index are comprehensively modeled to form a port-level comprehensive load feature set; Based on the established comprehensive load feature set, an adaptive weight adjustment algorithm is adopted to generate a distribution weight factor for each port in real time, so as to realize the adaptive adjustment of parameters under different protocol types and different traffic trends. S6. Input the prediction results and real-time load status into the reinforcement learning scheduling agent, which will then select a feedforward distribution strategy. The reinforcement learning scheduling agent uses an intelligent decision-making algorithm based on the value function or policy gradient to score and optimize the actions of various distribution strategies. The feedforward distribution strategies include pre-allocating redundant resources, dynamically adjusting the weight of each port, or activating backup links. Based on the optimal strategy output by the scheduling agent, and combined with the current port compatibility parameters and system resource constraints, the feasibility of the specific distribution actions involved in the strategy is verified and the parameters are refined to generate a feedforward resource allocation and traffic adjustment instruction set that can be directly issued and executed. S7. When an abnormal traffic trend is detected, immediately implement feedforward load distribution adjustment; the feedforward adjustment strategy includes temporarily increasing the redundant bandwidth of the affected ports, dynamically reducing the weight of high-risk ports, or switching some traffic to backup links. S8. Monitor the actual traffic distribution and system load performance after feedforward adjustment, and feed back the abnormal event handling results to the prediction model and scheduling agent. S9. Based on the protocol type and port distribution, regularly adapt and update the protocol compatibility parameter group.

2. The network load balancing method based on the domestic IT innovation platform as described in claim 1, characterized in that: In S1, anomaly detection includes packet loss rate detection, port active status heartbeat detection, multi-feature consistency verification, and anomaly clustering analysis.

3. The network load balancing method based on the domestic IT innovation platform as described in claim 1, characterized in that: In S2, the data normalization method is Z-score normalization or Min-Max scaling.

4. The network load balancing method based on the domestic IT innovation platform as described in claim 1, characterized in that: In S4, the temporal neural network is a long short-term memory network or a gated recurrent unit.

5. The network load balancing method based on the domestic IT innovation platform as described in claim 1, characterized in that: In step S5, the distribution weight factors are generated using principal component analysis or an adaptive weighted fusion algorithm, and the weight factors are normalized within intervals and anomaly detection is performed to ensure that the total distribution sum is 1.

6. The network load balancing method based on the domestic IT innovation platform as described in claim 1, characterized in that: In S8, the closed-loop feedback data includes the effect of traffic distribution adjustment, anomaly recovery time, and resource utilization, and is used for adaptive parameter optimization of the traffic prediction model and scheduling agent. In S9, the protocol compatibility parameter group includes feature extraction rules, protocol category labels, and distribution priority settings, and the port-protocol association mapping table is updated in real time through the interface.

7. A domestically developed information technology application (ITAI) terminal, characterized in that: Used to implement the network load balancing method based on the domestic IT innovation platform as described in any one of claims 1-6.

Images