Internet quality intelligent judgment and optimization method

Abstract

The invention relates to the field of communication network equipment operation and maintenance optimization, in particular to an internet quality intelligent judgment and optimization method, which comprises the following steps: S1, according to an acquisition machine deployed on OLT (Optical Line Terminal) equipment, communicating with the OLT equipment through an SNMP (Simple Network Management Protocol) to obtain real-time operation state data of the equipment, and preprocessing the acquired data; s2, key features of fault judgment are extracted from the data through the feature extraction technology, and real-time fault judgment is conducted on newly-collected data according to the trained model; s3, analyzing the traffic congestion condition in the network by monitoring the port traffic data in real time; s4, analyzing the attenuation condition of the optical fiber line by monitoring the light attenuation data of the OLT equipment in real time; and S5, providing an optimization scheme through comprehensive evaluation of fault judgment, flow congestion analysis and light attenuation analysis, and the method can improve network performance, troubleshooting and prevention of faults, improve user experience and reasonably utilize resources.

Description

Technical Field

[0001] The present invention relates to the field of communication network equipment operation and maintenance optimization, and in particular to an intelligent judgment and optimization method for Internet access quality. Background Art

[0002] In modern communication networks, optical line terminal equipment plays a key role. As the core equipment of the access network, OLT equipment is responsible for connecting multiple user-end devices. Its operating status directly affects the overall performance of the network and the user's Internet experience. With the continuous expansion of the network scale and the increasing requirements of users for network quality, OLT equipment faces many challenges. On the one hand, network traffic has shown explosive growth, and the demand for bandwidth and stability of various business types has continued to increase, which easily leads to traffic congestion at the ports of OLT equipment. On the other hand, the OLT equipment is connected to the user end through optical fiber, and the optical attenuation problem of the optical fiber line has a vital impact on the quality of signal transmission. At the same time, the failure of the OLT equipment itself will also seriously affect network services.

[0003] The current mainstream network planning and optimization solutions mainly rely on the experience and historical data of network engineers to make manual judgments and decisions. Generally, basic network monitoring and statistical analysis tools are used to conduct preliminary evaluation and planning of network equipment and user data. In the process of network construction and transformation, a phased testing method is usually adopted to gradually adjust and optimize the network configuration according to the test results. These solutions can work to a certain extent, but there are significant shortcomings, including: the experience-based planning method lacks scientific basis, is prone to subjective errors, and is difficult to ensure global optimization; basic monitoring and analysis tools cannot fully mine and utilize network data, and have limited prediction and optimization capabilities; the step-by-step implementation and adjustment process is cumbersome and long, and cannot quickly respond to changes in network requirements. Summary of the invention

[0004] To solve the above problems, the present invention provides an intelligent judgment and optimization method for Internet access quality, which can improve network performance, troubleshoot and prevent faults, improve user experience and rationally utilize resources.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for intelligently judging and optimizing Internet access quality comprises the following steps:

[0007] S1: According to the data collector deployed on the OLT device, it communicates with the OLT device through the SNMP protocol to obtain the real-time operation status data of the device and pre-process the collected data;

[0008] S2: Extract key features for fault diagnosis from data through feature extraction technology, train the fault diagnosis model through historical data, and perform real-time fault diagnosis on newly collected data based on the trained model;

[0009] S3: Analyze the traffic congestion in the network through real-time monitoring of port traffic data and make optimization suggestions;

[0010] S4: Analyze the attenuation of optical fiber lines and provide maintenance suggestions by monitoring the optical attenuation data of OLT equipment in real time;

[0011] S5: Through comprehensive evaluation of fault judgment, traffic congestion analysis and optical attenuation analysis, an optimization solution for OLT equipment is proposed.

[0012] Furthermore, the real-time operating status data of the device is obtained, including: basic information data of the device, port traffic data, optical module information data, device performance data and network connection data; the preprocessing includes data cleaning and data normalization; the real-time collected data is stored by combining local storage and cloud transmission.

[0013] Furthermore, the method of extracting key features for fault diagnosis from data by feature extraction technology, training a fault diagnosis model by historical data, and performing real-time fault diagnosis on newly collected data according to the trained model comprises the following steps:

[0014] S21: Determine key features based on the correlation between the fault type and the data, and extract the determined key features through data mining technology;

[0015] S22: dividing the historical data into a training set and a test set, training the fault judgment model through the training set, and evaluating and optimizing the trained fault judgment model through the test set;

[0016] S23: Continuously collect the real-time operation status data of the OLT device through the data collector, input the pre-processed data into the trained fault judgment model, and output the fault judgment result according to the input feature data.

[0017] Furthermore, determining key features according to the correlation between the fault type and the data, and extracting the determined key features by using data mining technology, comprises the following steps:

[0018] S211: Based on the understanding of the working principle of OLT equipment, network communication knowledge and common failure modes, determine the characteristics related to the failure from the various types of data collected;

[0019] S212: Calculate basic statistics of each feature, and determine the normal fluctuation range of port traffic by analyzing the mean and standard deviation of inbound and outbound traffic;

[0020] S213: Determine the correlation between each feature and the fault label by using the Pearson correlation coefficient, and determine the discrete feature by calculating the information gain of each feature;

[0021] S214: Reduce the number of correlated features after initial screening through principal component analysis, and combine features into new features based on a deeper understanding of the failure mode.

[0022] Furthermore, the historical data is divided into a training set and a test set, and the fault judgment model is trained by the training set; the following steps are included:

[0023] S221: Coordinate the data partitioning operation of each node according to the task scheduling component in the distributed architecture, calculate the hash value of each data record, and distribute the data to different subsets according to the range of the hash value, and divide it into a training set and a test set;

[0024] S222: Each node performs Bootstrap sampling with replacement on the stored data in parallel, and sends the information of the local sub-dataset to the centralized coordination node through the distributed message passing mechanism. The coordination node summarizes the sampling results of each node to construct the sub-dataset of the entire random forest.

[0025] S223: determining the best split feature based on the information gain criterion, and in the process of constructing multiple decision trees, allocating the task of constructing the decision trees according to the computing resources of the cluster;

[0026] S224: The prediction results of the random forest are used as new features and combined with part of the original features to form input features of the logistic regression, and a logistic regression model is constructed.

[0027] Furthermore, the method of determining the best split feature based on the information gain criterion and, in the process of constructing multiple decision trees, allocating the task of constructing the decision trees according to the computing resources of the cluster comprises the following steps:

[0028] S2231: When each node splits, the operation of randomly selecting features from all features is performed locally on each node, and each node calculates information gain for the local data subset and the selected features;

[0029] S2232: After calculating the information gain, each node sends the information gain to the coordination node, which summarizes the calculation results and selects the feature with the largest information gain as the best split feature;

[0030] S2233: In the process of building multiple decision trees, the distributed architecture manages the resource allocation and task scheduling of each node, and allocates the task of building the decision tree according to the computing resources of the cluster.

[0031] Furthermore, the method of combining the prediction results of the random forest as new features with a part of the original features to form input features of the logistic regression and constructing a logistic regression model includes the following steps:

[0032] S2241: The prediction results after random forest training are used as new features, which are combined with the important features determined by feature selection in the original features to form the input features of logistic regression;

[0033] S2242: Determine the probability of failure through the hypothesis function of the logistic regression model, measure the prediction accuracy by minimizing the logarithmic loss function of the loss function, update the model parameters through the stochastic gradient descent algorithm, and calculate the gradient of each sample;

[0034] S2243: Send the locally calculated gradient to the summary node through the distributed message passing mechanism. The summary node accumulates the gradients sent by each node to obtain the global gradient. According to the global gradient and the set learning rate, the model parameters are updated, and the iteration is continued until the stop condition is met.

[0035] In step S2242, the hypothesis function of the logistic regression model is: Where x is the input feature vector, θ is the model parameter vector, and T is the transpose; the logarithmic loss function is: Where m is the number of samples, y i is the true label of the i-th sample, and i is the index variable.

[0036] Furthermore, the method of analyzing the traffic congestion in the network through real-time monitoring of the port traffic data and proposing optimization suggestions comprises the following steps:

[0037] S31: obtaining the pre-processed port flow data, calculating the average flow, peak flow and flow fluctuation range, then calculating the skewness and kurtosis of the flow, and calculating the autocorrelation coefficients under different time lags;

[0038] S32: When the traffic exceeds a preset threshold, a congestion detection algorithm is triggered, and a comprehensive congestion judgment index is constructed by assigning different weights to the traffic surge index and the port utilization excessive index;

[0039] S33: Adjust the traffic allocation strategy based on the service type. For services with high real-time requirements, fixed bandwidth is allocated. For non-real-time but important services, dynamic bandwidth allocation based on priority is adopted. For non-critical services, bandwidth is restricted when congested. According to the real-time traffic monitoring results, dynamic adjustment of traffic shaping is performed.

[0040] S34: Plan backup links for network topology, enable backup links when network is congested, evaluate the degree of relief of network congestion based on the enabling effect, optimize traffic routing through routing algorithms based on traffic engineering, recalculate traffic routing after congestion detection, and implement dynamic optimization of traffic routing based on software-defined network technology;

[0041] S35: During the continuous monitoring process, port traffic data is continuously collected and network performance indicators are monitored. The results of continuous monitoring are fed back to the network management system in a timely manner, and adjustment strategies are formulated based on the feedback results.

[0042] Furthermore, the method of monitoring the optical attenuation data of the OLT device in real time, analyzing the attenuation of the optical fiber line, and providing maintenance suggestions comprises the following steps:

[0043] S41: Collect the pre-processed light attenuation data through the optical module information data, fit the light attenuation data through the least square method, judge the light attenuation trend according to the slope of the fitted straight line, and calculate the standard deviation of the light attenuation data within the time window;

[0044] S42: When the optical attenuation data trend is stable and close to the preset threshold, preventive maintenance is performed; when the optical attenuation fluctuates greatly but has not exceeded the threshold, the environment of the optical fiber line is checked, and at the same time, the optical fiber fixture is checked to see if it is firm; when the optical attenuation is on an upward trend and has not exceeded the threshold, in addition to the above inspections, the age and type of the optical fiber are also evaluated; when the optical attenuation value exceeds the preset threshold, an alarm message is generated;

[0045] S43: Continuously collect and analyze optical attenuation data to monitor changes in the status of the optical fiber line, and provide feedback and adjust maintenance recommendations and alarm thresholds based on the implementation of maintenance measures and subsequent optical attenuation data monitoring results.

[0046] Furthermore, the optimization scheme for OLT equipment is proposed through comprehensive evaluation of fault judgment, traffic congestion analysis and optical attenuation analysis; the steps include:

[0047] S51: Determine the weights of fault judgment, traffic congestion analysis and optical attenuation analysis in comprehensive evaluation through the hierarchical analysis method;

[0048] S52: Quantitatively score each indicator of fault judgment, traffic congestion analysis, and optical attenuation analysis according to the determined weights, and calculate a comprehensive score. According to the comprehensive score and the scores of each indicator, locate the area or component with the problem;

[0049] S53: If the fault judgment result shows that there is a problem with the device hardware, then upgrade the hardware; if the excessive optical attenuation is caused by the degradation of the optical module performance, then replace the optical module; if the traffic congestion analysis shows that there is a traffic congestion problem, then optimize the traffic distribution strategy;

[0050] S54: According to the proposed optimization plan, the implementation steps, responsible persons, and time nodes of each optimization measure are clarified, and an effect evaluation index system is established. The effect evaluation index includes equipment performance index, network performance index, and optical attenuation index.

[0051] The beneficial effects of the present invention are:

[0052] 1. Through the multi-faceted data collection and analysis of OLT equipment and related network components, the source of the fault can be accurately determined, which helps to repair the fault in a timely manner, reduce network interruption time, improve the overall reliability of the network, and ensure users' continuous and stable Internet experience. The intelligent fault judgment model quickly locates the fault point, avoiding the blindness and high cost of traditional manual troubleshooting. Operation and maintenance personnel can accurately maintain the faulty components or areas, reduce unnecessary equipment inspections and repairs, and thus reduce operation and maintenance costs. The fault judgment model trained based on historical data can not only judge the current fault, but also predict potential faults by analyzing data trends.

[0053] 2. Through real-time monitoring and analysis of port traffic data, traffic management can be performed according to various business needs. When congestion occurs, the bandwidth of high-priority services is prioritized to ensure the service quality of these services and improve user satisfaction. Traffic congestion analysis can accurately identify congestion points and traffic bottlenecks in the network. Based on this, measures such as optimizing the network topology, adjusting the traffic path, and reasonably allocating port bandwidth can make more effective use of network resources, avoid some links or ports being idle while other parts are congested, and improve the transmission efficiency of the entire network. Network traffic is dynamic and changes with the number of users and business types. Traffic congestion analysis can adapt to such changes in real time and dynamically adjust traffic allocation strategies.

[0054] 3. Real-time monitoring of optical attenuation data can detect optical attenuation abnormalities in a timely manner. Excessive optical attenuation causes the quality of optical signal transmission to decline, affecting Internet access speed and stability. Optical attenuation analysis can quickly locate problem points in the optical fiber line, and timely measures can be taken to repair them, ensuring high-quality transmission of optical signals and Internet access quality. Regular optical attenuation analysis helps to discover potential problems in the early stages of optical fiber lines. Through timely adjustment of optical fiber laying paths, cleaning or replacement of optical modules and other maintenance measures, problems can be avoided from further deterioration, the service life of optical fiber lines can be extended, and the replacement cost of network infrastructure can be reduced. As an important part of network quality assurance, optical attenuation analysis is combined with other analyses (fault judgment, traffic congestion analysis) to comprehensively improve the stability of the network. Stable optical signal transmission is the basis for the normal operation of the network. Ensuring the quality of optical signals through optical attenuation analysis helps to reduce network fluctuations and failures caused by optical signal problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The present invention is a flow chart of an intelligent judgment and optimization method for Internet access quality.

[0056] Figure 2 It is a flowchart for training the fault judgment model.

[0057] Figure 3 It is a flow chart for analyzing the attenuation of optical fiber lines. DETAILED DESCRIPTION

[0058] See also Figure 1-3 As shown, the present invention relates to an intelligent judgment and optimization method for Internet access quality, comprising the following steps:

[0059] S1: According to the data collector deployed on the OLT device, it communicates with the OLT device through the SNMP protocol to obtain the real-time operation status data of the device and pre-process the collected data;

[0060] S2: Extract key features for fault diagnosis from data through feature extraction technology, train the fault diagnosis model through historical data, and perform real-time fault diagnosis on newly collected data based on the trained model;

[0061] S3: Analyze the traffic congestion in the network through real-time monitoring of port traffic data and make optimization suggestions;

[0062] S4: Analyze the attenuation of optical fiber lines and provide maintenance suggestions by monitoring the optical attenuation data of OLT equipment in real time;

[0063] S5: Through comprehensive evaluation of fault judgment, traffic congestion analysis and optical attenuation analysis, an optimization solution for OLT equipment is proposed.

[0064] Furthermore, the real-time operating status data of the device is obtained, including: basic information data of the device, port traffic data, optical module information data, device performance data and network connection data; the preprocessing includes data cleaning and data normalization; the real-time collected data is stored by combining local storage and cloud transmission.

[0065] Specifically, the basic information data of the device includes: device name, device model and device serial number; the port traffic data includes: port number, port status, port traffic statistics and port error statistics, and the port error statistics reflect the data transmission quality of the port and help to timely discover problem ports; the optical module information data includes: optical module type, optical power, optical attenuation data and optical module temperature; the device performance data includes: CPU utilization, memory utilization and device temperature; the network connection data includes: the number of connected ONUs, the status information of each ONU and the network topology information.

[0066] Furthermore, the method of extracting key features for fault diagnosis from data by feature extraction technology, training a fault diagnosis model by historical data, and performing real-time fault diagnosis on newly collected data according to the trained model comprises the following steps:

[0067] S21: Determine key features based on the correlation between the fault type and the data, and extract the determined key features through data mining technology;

[0068] S22: dividing the historical data into a training set and a test set, training the fault judgment model through the training set, and evaluating and optimizing the trained fault judgment model through the test set;

[0069] S23: Continuously collect the real-time operation status data of the OLT device through the data collector, input the pre-processed data into the trained fault judgment model, and output the fault judgment result according to the input feature data.

[0070] Specifically, in step S23, the fault judgment result is fed back to the network management personnel in a timely manner so that they can take corresponding measures. At the same time, the fault judgment result is continuously monitored, and the model is optimized and updated according to the actual situation. For example, when a new fault type appears and the model cannot accurately judge, it is necessary to collect data again, adjust features or retrain the model.

[0071] Furthermore, determining key features according to the correlation between the fault type and the data, and extracting the determined key features by using data mining technology, comprises the following steps:

[0072] S211: Based on the understanding of the working principle of OLT equipment, network communication knowledge and common failure modes, determine the characteristics related to the failure from the various types of data collected;

[0073] S212: Calculate basic statistics of each feature, and determine the normal fluctuation range of port traffic by analyzing the mean and standard deviation of inbound and outbound traffic;

[0074] S213: Determine the correlation between each feature and the fault label by using the Pearson correlation coefficient, and determine the discrete feature by calculating the information gain of each feature;

[0075] S214: Reduce the number of correlated features after initial screening through principal component analysis, and combine features into new features based on a deeper understanding of the failure mode.

[0076] Specifically, in step S211, for example, from the equipment performance data, based on experience, it is known that a CPU utilization rate that is continuously higher than 80% is related to equipment operation failure, so the CPU utilization rate is a potential key feature; for the optical module information data, if the fluctuation range of the optical power exceeds the normal threshold (such as the fluctuation of the transmitted optical power exceeds ±3dBm), it indicates an optical link failure, so the transmitted optical power is a feature that needs to be focused on.

[0077] In step S212, the basic statistics include mean, median and standard deviation. For port traffic statistics, the normal fluctuation range of port traffic is determined by analyzing the mean and standard deviation of inbound and outbound traffic. If the traffic standard deviation of a port suddenly increases, it indicates that the port has an abnormal traffic pattern, which is valuable for fault diagnosis.

[0078] In step S213, the fault label includes normal or different types of faults. For example, taking the temperature of the optical module and the optical link fault as an example, if the Pearson correlation coefficient shows that the temperature of the optical module is strongly positively correlated with the optical link fault (for example, the correlation coefficient is greater than 0.6), it means that the temperature of the optical module is an important feature in the judgment of the optical link fault. For discrete features, the information gain of each feature is calculated. The information gain measures how much information a feature can provide for the fault judgment classification. For example, assuming that a port fault is to be judged, the port state (discrete features: connected, disconnected) has a higher information gain relative to whether it is faulty or not. Because the port disconnection state is directly related to the port fault, the degree of this correlation is quantified by calculating the information gain.

[0079] In step S214, if the number of features after preliminary screening is still large and there is a certain correlation, principal component analysis is used to perform feature dimensionality reduction. Principal component analysis converts the original features into a set of new unrelated principal components through linear transformation. For example, for multiple indicators in device performance data (CPU usage, memory usage, or device temperature, etc.), principal component analysis converts these related features into a few principal components, which can retain most of the information of the original data, thereby reducing the number of features and improving model training efficiency.

[0080] Based on the in-depth understanding of the failure mode, combining some features will produce more valuable features. For example, the port's inbound and outbound traffic is combined into a new feature "port traffic difference (inbound traffic - outbound traffic)", which can better reflect whether the port is working normally. If the port traffic difference is continuously negative and the absolute value is large, it means that the port has abnormal traffic flow, which is a sign of port failure or malicious attack.

[0081] Furthermore, the historical data is divided into a training set and a test set, and the fault judgment model is trained by the training set; the following steps are included:

[0082] S221: Coordinate the data partitioning operation of each node according to the task scheduling component in the distributed architecture, calculate the hash value of each data record, and distribute the data to different subsets according to the range of the hash value, and divide it into a training set and a test set;

[0083] S222: Each node performs Bootstrap sampling with replacement on the stored data in parallel, and sends the information of the local sub-dataset to the centralized coordination node through the distributed message passing mechanism. The coordination node summarizes the sampling results of each node to construct the sub-dataset of the entire random forest.

[0084] S223: determining the best split feature based on the information gain criterion, and in the process of constructing multiple decision trees, allocating the task of constructing the decision trees according to the computing resources of the cluster;

[0085] S224: The prediction results of the random forest are used as new features and combined with part of the original features to form input features of the logistic regression, and a logistic regression model is constructed.

[0086] Specifically, in step S221, in a distributed architecture, data is stored on multiple nodes (such as different computing nodes in a cluster), and a coordination mechanism is required to determine the data partitioning strategy. For example, assuming the hash value range is 0 - M, it can be stipulated that hash values from 0 - N (N < M) correspond to the training set, and N + 1 - M correspond to the test set. The task scheduling component in the distributed architecture (such as JobTracker in the Map - Reduce framework or a similar component) coordinates the data partitioning operations of each node to ensure that the data partitioning is uniform and random across the entire dataset. Each node is responsible for processing the locally stored data and determining which data belongs to the training set and which belongs to the test set according to the above rules. Then, the partitioning results are aggregated through a distributed message queue or shared storage (such as a distributed file system) to form a complete training set and test set.

[0087] In step S222, during the construction of each decision tree, Bootstrap sampling with replacement needs to be performed in a distributed environment, and each node performs the sampling operation on the locally stored data in parallel. Each node samples according to a given sampling ratio (for example, for local data of size, sample with replacement according to the scale of n) to obtain a local sub - dataset. Then, relevant information about the local sub - dataset (such as data indices or references) is sent to a centralized coordination node (a node dedicated to model construction coordination) through a distributed message passing mechanism (such as a message queue like Apache Kafka). The coordination node aggregates the sampling results of each node to construct the sub - dataset required for the entire random forest. If it is necessary to integrate the data to form a complete sub - dataset sample, the coordination node can obtain the actual data records from each node according to the indices, or instruct each node to send the locally sampled data to a specific storage location for integration.

[0088] Furthermore, for determining the best split feature based on the information gain criterion, during the construction of multiple decision trees, tasks for constructing decision trees are allocated according to the computing resources of the cluster; the steps include:

[0089] S2231: When each node splits, the operation of randomly selecting features from all features is performed locally at each node, and each node calculates the information gain for the local data subset and the selected feature;

[0090] S2232: After each node calculates the information gain, it is sent to the coordination node, and the coordination node aggregates the calculation results and selects the feature with the maximum information gain as the best split feature;

[0091] S2233: During the construction of multiple decision trees, the distributed architecture manages the resource allocation and task scheduling of each node, and allocates tasks for constructing decision trees according to the computing resources of the cluster.

[0092] Specifically, in step S2231, when each node is split, the operation of randomly selecting m features (m < total number of features) from all features can be performed locally on each node. Each node calculates information gain for the local data subset and the selected m features. When calculating information gain, taking node i as an example, for data set D i , its entropy Where pk is the k-th class sample in the dataset D i The ratio in k is the number of categories. Assume that feature A has v different values ​​{a1, a2, ..., a v}, according to feature A, the dataset D i Divide into v subsets D i1 , D i2 , …, D iv , then feature A has a significant impact on data set D i Information gain

[0093] In step S2232, after each node calculates the information gain of each local feature, it sends the result to the coordination node. The coordination node summarizes the calculation results of each node, compares the sum of the information gain of each feature on different nodes (or the average value, depending on the specific aggregation strategy), and selects the feature with the largest information gain as the best split feature. In this process, the message passing mechanism in the distributed architecture ensures the information exchange between each node, and the coordination node is responsible for integrating the results and making decisions.

[0094] In step S2233, for example, under the Map-Reduce framework, the construction of each decision tree is regarded as an independent task (Map task), and each node can execute these tasks in parallel. The coordination node (similar to the function of the Reduce node) is responsible for monitoring the progress of the task, assigning unprocessed tasks to idle nodes, and summarizing the results of the constructed decision trees. According to the computing resources of the cluster (such as the CPU and memory of the node), the task of building the decision tree is reasonably allocated to ensure that the construction process of the entire random forest is carried out efficiently. For example, if the computing power of a node is strong, more decision tree construction tasks can be assigned to the node.

[0095] Furthermore, the method of combining the prediction results of the random forest as new features with a part of the original features to form input features of the logistic regression and constructing a logistic regression model includes the following steps:

[0096] S2241: The prediction results after random forest training are used as new features, which are combined with the important features determined by feature selection in the original features to form the input features of logistic regression;

[0097] S2242: Determine the probability of failure through the hypothesis function of the logistic regression model, measure the prediction accuracy by minimizing the logarithmic loss function of the loss function, update the model parameters through the stochastic gradient descent algorithm, and calculate the gradient of each sample;

[0098] S2243: Send the locally calculated gradient to the summary node through the distributed message passing mechanism. The summary node accumulates the gradients sent by each node to obtain the global gradient. According to the global gradient and the set learning rate, the model parameters are updated, and the iteration is continued until the stop condition is met.

[0099] In step S2242, the hypothesis function of the logistic regression model is: Where x is the input feature vector, θ is the model parameter vector, and T is the transpose; the logarithmic loss function is: Where m is the number of samples, y i is the true label of the i-th sample, and i is the index variable.

[0100] Specifically, each node obtains the prediction result of the random forest on the local data and the original feature data, and each node combines the random forest prediction result and the selected original features into a new feature vector according to the feature combination rule.

[0101] In a distributed environment, the model parameters θ are updated by the stochastic gradient descent algorithm. i ,y i ), calculate the gradient, the formula is:

[0102] In a distributed architecture, the prediction results after random forest training (such as the results of each sample being classified by each tree in the random forest) are combined as new features with part of the original features (important features determined by feature selection) to form the input features of logistic regression. The process is as follows: First, each node obtains the prediction results of the random forest on local data and the original feature data. Then, the feature combination operation is performed locally, and each node combines the random forest prediction results and the selected original features into a new feature vector according to the predetermined feature combination rules. The feature vector data combined by each node is summarized through a distributed message queue or a distributed file system. For example, the data of each node is written to a specific directory of the distributed file system, and then an integration program merges these data into a complete input data set required for the logistic regression model.

[0103] Each node calculates the gradient of local data in parallel. Each node reads the local input feature vector and the corresponding real label data and calculates the gradient. The locally calculated gradient is then sent to a summary node through a distributed message passing mechanism. The summary node accumulates the gradients sent by each node (or processes them according to a specific aggregation algorithm) to obtain a global gradient. According to the global gradient and the set learning rate, the model parameter θ is updated. This process is continuously iterated under the distributed architecture until the stopping condition is met (such as reaching a preset number of iterations or the loss function converges). Through such a distributed architecture, the computing resources of the cluster can be fully utilized to improve the efficiency of model training, especially when processing large-scale OLT equipment operating status data, which can accelerate the construction process of the random forest and logistic regression combination model.

[0104] In summary, the actual application process from step S1 to step S2 is as follows:

[0105] Input the preprocessed new data into the trained random forest model. In a distributed framework, the data can be distributed in parallel to each node storing the random forest model for processing. Each decision tree in the random forest independently classifies the input data. Each decision tree divides the data into different leaf nodes according to its internal node splitting rules, and each leaf node corresponds to a preliminary fault classification or normal state category. In a distributed framework, each node aggregates the classification results of each decision tree to a coordination node. The coordination node determines the preliminary judgment result of the random forest model on the data through a voting mechanism or an averaging mechanism. For example, if the majority of decision trees determine that the device corresponding to the data is in a certain fault state (such as an optical link failure), the preliminary judgment is that fault state; if the majority of decision trees determine that it is normal, the preliminary result is normal.

[0106] Combine the preliminary judgment results of the random forest (expressed in coded form) with some important features in the original data (features that are strongly correlated with fault judgment determined by feature selection). In a distributed framework, each node can perform feature combination operations in parallel, and the combined feature vector will be used as the input of logistic regression. The combined feature vector is input into the trained logistic regression model to calculate the probability that the data belongs to different fault categories or normal states. In a distributed environment, the calculation process can be carried out in parallel on multiple nodes, each node is responsible for processing the probability calculation of a part of the data, and then the results are aggregated to a coordination node. The coordination node makes the final fault judgment based on the probability results calculated by logistic regression. For example, if the calculated probability that the device is in a certain fault state exceeds the preset threshold, the device is judged to be in the fault state; if the probability is lower than the threshold, it is judged to be in a normal state.

[0107] The final fault judgment result is fed back to the network operation and maintenance personnel or related monitoring systems. In the distributed framework, the result can be delivered to the designated receiving end through message queues or distributed file systems. If the judgment result is a fault state, the operation and maintenance personnel can take corresponding measures according to the fault type, such as checking the fiber connection and cleaning the optical module for optical link failures; checking the port configuration and troubleshooting the connected devices for port failures. At the same time, new fault data (if any) can be collected for subsequent updating and optimization of the fault judgment model.

[0108] Furthermore, the method of analyzing the traffic congestion in the network through real-time monitoring of the port traffic data and proposing optimization suggestions comprises the following steps:

[0109] S31: obtaining the pre-processed port flow data, calculating the average flow, peak flow and flow fluctuation range, then calculating the skewness and kurtosis of the flow, and calculating the autocorrelation coefficients under different time lags;

[0110] S32: When the traffic exceeds a preset threshold, a congestion detection algorithm is triggered, and a comprehensive congestion judgment index is constructed by assigning different weights to the traffic surge index and the port utilization excessive index;

[0111] S33: Adjust the traffic allocation strategy based on the service type. For services with high real-time requirements, fixed bandwidth is allocated. For non-real-time but important services, dynamic bandwidth allocation based on priority is adopted. For non-critical services, bandwidth is restricted when congested. According to the real-time traffic monitoring results, dynamic adjustment of traffic shaping is performed.

[0112] S34: Plan backup links for network topology, enable backup links when network is congested, evaluate the degree of relief of network congestion based on the enabling effect, optimize traffic routing through routing algorithms based on traffic engineering, recalculate traffic routing after congestion detection, and implement dynamic optimization of traffic routing based on software-defined network technology;

[0113] S35: During the continuous monitoring process, port traffic data is continuously collected and network performance indicators are monitored. The results of continuous monitoring are fed back to the network management system in a timely manner, and adjustment strategies are formulated based on the feedback results.

[0114] Specifically, in step S31, the skewness reflects the asymmetry of the flow data distribution, and the formula is: Where n is the number of data points, is the average flow rate, and σ is the standard deviation. Skewness can help determine whether the flow distribution is biased to one side. For example, if the skewness is positive, it means that the flow distribution is biased to the right and there is a large flow peak. Kurtosis measures the peak degree of flow data distribution, and the formula is A large kurtosis value indicates that the traffic distribution is relatively concentrated and there are sudden high traffic situations. At the same time, analyze the autocorrelation of the traffic. Calculate the autocorrelation coefficients at different time lags to understand the correlation of traffic in the time series. For example, calculate the autocorrelation coefficient of the current traffic and the traffic 10 seconds ago. If the autocorrelation coefficient is high, it means that the traffic has a certain degree of continuity, which is important for predicting traffic trends and judging congestion.

[0115] In step S32, in addition to simply comparing whether the traffic exceeds the threshold, the traffic growth rate is also analyzed. The traffic growth rate in a short period of time is calculated. If the growth rate exceeds a certain proportion, it is considered a traffic surge. The skewness and kurtosis of the traffic are combined to determine whether the traffic surge is abnormal. If the traffic surge is accompanied by a large change in skewness and kurtosis, it indicates that an abnormal traffic pattern has occurred in the network, such as a sudden large number of small data packets (high skewness) or a small number of large traffic sources (high peak kurtosis), which is a precursor to congestion.

[0116] Port utilization not only considers the simple proportional relationship between the inbound and outbound traffic and the port design capacity, but also considers the port cache usage. Calculate the port cache usage. If the port inbound or outbound traffic is close to the port design capacity and the cache usage is high, the port utilization is considered to be too high. By assigning different weights to different indicators, a comprehensive congestion judgment indicator is constructed. For example, a weight of 0.4 is assigned to the traffic surge indicator, and a weight of 0.6 is assigned to the port utilization indicator. The weight values ​​are determined based on historical data and network characteristics. Calculate the comprehensive indicator. If it exceeds a certain threshold (such as 0.7), it is determined that the network is congested.

[0117] In step S33, for services with extremely high real-time requirements (such as voice calls and video conferences), a fixed bandwidth guarantee is allocated for them. For example, at least 200Kbps of bandwidth is reserved for voice call services to ensure normal communication even when the network is congested. For non-real-time but important services (such as enterprise database access), priority-based dynamic bandwidth allocation is adopted. According to the load of the network, appropriate bandwidth is allocated on the premise of ensuring the bandwidth requirements of other high-priority services. For example, when the network is idle, a larger bandwidth (such as 1Mbps) is given, and when congested, the bandwidth is appropriately reduced according to the degree of congestion, but not less than a minimum guarantee value (such as 200Kbps). The parameters of the dynamic adjustment flow limiter, such as finding that the flow of a certain type of business continues to grow in a short period of time and is close to the congestion threshold, gradually reduce the flow rate limit of the business, rather than suddenly limit it to a very low value, to avoid excessive impact on the business. Traffic shaping is performed through the token bucket algorithm, such as dynamically adjusting the rate of token generation and the capacity of the bucket according to the degree of network congestion. When there is congestion, the token generation rate is reduced to reduce the rate at which traffic passes; when the network is idle, the token generation rate is appropriately increased to allow more traffic to pass.

[0118] In step S34, a comprehensive analysis of the network topology is performed to plan backup links. Determine between which nodes backup links need to be added to improve the redundancy of the network. When the network is congested, the backup links are enabled. For example, when a main link (a link connecting two important areas) is detected to be congested, if a backup link exists, part of the traffic is transferred to the backup link. Calculate the expected improvement in network performance after enabling the backup link. By simulating the effect of enabling the backup link under different traffic distribution conditions, the degree of relief for network congestion is evaluated.

[0119] Traffic routing is optimized using a routing algorithm based on traffic engineering. For example, multi-protocol label switching (MPLS) traffic engineering calculates the optimal traffic path based on factors such as network traffic demand, link bandwidth, and latency. After congestion detection, the traffic route is recalculated. Taking into account factors such as the load of different links in the network, the remaining bandwidth, and the latency of the link, the traffic is redistributed to the optimal path. For example, if a path from area A to area B is congested, while another path has sufficient bandwidth and low latency, the routing strategy is adjusted so that the traffic from area A to area B is transmitted through this better path. Combined with software-defined network technology, dynamic optimization of traffic routing is achieved. The software-defined network controller obtains the network's traffic information and link status in real time, and dynamically issues new routing instructions to network devices (such as switches and routers) based on the congestion detection results, to achieve rapid traffic adjustment and improve the overall performance of the network.

[0120] In step S35, in addition to continuing to collect port traffic data during continuous monitoring, other performance indicators of the network are also monitored, such as network delay, packet loss rate, etc. These indicators can reflect the effectiveness of traffic adjustment and network path optimization measures. Monitor the network from different levels (port level, link level, network level). At the port level, pay attention to the traffic changes and number of error packets of each port; at the link level, monitor the bandwidth utilization and delay of the link; at the network level, analyze the traffic distribution of the overall network, the effectiveness of the network topology, etc. According to the feedback results, formulate adjustment strategies. If it is found that the traffic in a certain area is still congested, even though traffic distribution and network path adjustment measures have been taken, it is necessary to further analyze the reasons, such as whether there are undiscovered traffic sources or unreasonable network topology structures.

[0121] Furthermore, the method of monitoring the optical attenuation data of the OLT device in real time, analyzing the attenuation of the optical fiber line, and providing maintenance suggestions comprises the following steps:

[0122] S41: Collect the pre-processed light attenuation data through the optical module information data, fit the light attenuation data through the least square method, judge the light attenuation trend according to the slope of the fitted straight line, and calculate the standard deviation of the light attenuation data within the time window;

[0123] S42: When the optical attenuation data trend is stable and close to the preset threshold, preventive maintenance is performed; when the optical attenuation fluctuates greatly but has not exceeded the threshold, the environment of the optical fiber line is checked, and at the same time, the optical fiber fixture is checked to see if it is firm; when the optical attenuation is on an upward trend and has not exceeded the threshold, in addition to the above inspections, the age and type of the optical fiber are also evaluated; when the optical attenuation value exceeds the preset threshold, an alarm message is generated;

[0124] S43: Continuously collect and analyze optical attenuation data to monitor changes in the status of the optical fiber line, and provide feedback and adjust maintenance recommendations and alarm thresholds based on the implementation of maintenance measures and subsequent optical attenuation data monitoring results.

[0125] Specifically, in step S41, in order to analyze the long-term trend of light attenuation, the light attenuation data is fitted by the least square method. Assume that the light attenuation data is (x i ,y i ), where x i is the acquisition time (which can be converted into a sequence number in the time series), y i is the corresponding light attenuation value. For the univariate linear regression model y=ax+b, the least squares method is used to solve the parameters a and b. in and They are x i and i The average value of n is the number of data points. The trend of optical attenuation is determined by the slope a of the fitted line. If a>0, it means that the optical attenuation is on the rise, and there is a situation of fiber aging or gradual deterioration of the connection point; if a=0, the optical attenuation is basically stable; if a<0, there is a problem with data collection or processing (under normal circumstances, optical attenuation rarely decreases naturally).

[0126] Calculate the standard deviation of light attenuation data within a certain time window where y i is the light attenuation value, is the average value of the optical attenuation in the time window, and n is the number of data points in the time window. A larger standard deviation indicates that the optical attenuation fluctuates greatly, which is caused by external interference (such as vibration, temperature change, etc.) on the optical fiber or loose connection points.

[0127] In step S42, the preventive maintenance includes checking whether the optical fiber connection points are clean and tight, and performing an appearance inspection on the optical fiber line to check whether there are any bends or squeezes. The inspection of the environment of the optical fiber line includes the control of temperature and humidity, and whether there are strong electromagnetic interference sources. At the same time, check whether the optical fiber fixture is firm to avoid micro-bending or vibration of the optical fiber due to external factors. The alarm information should include detailed information such as the optical fiber port number, the specific value of the optical attenuation, and the timestamp of exceeding the threshold. The alarm information is sent in a variety of ways, such as to the console of the network management system, to the maintenance personnel in the form of text messages or emails, etc.

[0128] In step S43, the optical attenuation data is continuously collected and analyzed to monitor the state changes of the optical fiber line. The data is continuously updated according to the previously determined collection frequency, and the analysis step is repeated. According to the implementation of the maintenance measures and the subsequent optical attenuation data monitoring results, the maintenance suggestions and alarm thresholds are fed back and adjusted. For example, if the optical attenuation data is still unstable after maintenance, it is necessary to further reduce the alarm threshold or take more in-depth maintenance measures, such as replacing the optical fiber connector or conducting a comprehensive inspection of the optical fiber.

[0129] Furthermore, the optimization scheme for OLT equipment is proposed through comprehensive evaluation of fault judgment, traffic congestion analysis and optical attenuation analysis; the steps include:

[0130] S51: Determine the weights of fault judgment, traffic congestion analysis and optical attenuation analysis in comprehensive evaluation through the hierarchical analysis method;

[0131] S52: Quantitatively score each indicator of fault judgment, traffic congestion analysis, and optical attenuation analysis according to the determined weights, and calculate a comprehensive score. According to the comprehensive score and the scores of each indicator, locate the area or component with the problem;

[0132] S53: If the fault judgment result shows that there is a problem with the device hardware, then upgrade the hardware; if the excessive optical attenuation is caused by the degradation of the optical module performance, then replace the optical module; if the traffic congestion analysis shows that there is a traffic congestion problem, then optimize the traffic distribution strategy;

[0133] S54: According to the proposed optimization plan, the implementation steps, responsible persons, and time nodes of each optimization measure are clarified, and an effect evaluation index system is established. The effect evaluation index includes equipment performance index, network performance index, and optical attenuation index.

[0134] Specifically, in step S51, first, a hierarchical model is constructed. The target layer is a comprehensive optimization evaluation of OLT equipment, and the criterion layer includes fault judgment indicators, traffic congestion indicators, and optical attenuation indicators. Fault judgment indicators can be further subdivided into sub-indicators such as basic equipment information, port status, and optical module information; traffic congestion indicators include sub-indicators such as port traffic characteristics (average traffic, peak traffic, etc.) and network path status; optical attenuation indicators include sub-indicators such as optical attenuation values ​​and optical attenuation trends. Then, a judgment matrix is ​​constructed through expert scoring or statistical analysis of historical data. Finally, a consistency check is performed. Calculate consistency indicators Where n is the order of the judgment matrix, find the corresponding average random consistency index RI, and calculate the consistency ratio When CR<0.1, the judgment matrix is ​​considered to have satisfactory consistency and the weight determination is valid.

[0135] In step S52, various indicators of fault judgment, traffic congestion analysis and optical attenuation analysis are quantitatively scored according to the weights determined above. For example, for fault judgment, if the device has a critical fault (such as excessive CPU usage causing device performance to degrade), a lower score is given according to the severity of the fault (such as 0-3 points, 0 points for the most serious fault); for traffic congestion analysis, if the port traffic is close to the congestion threshold, a certain score is given according to the degree of proximity (such as 3-6 points, 6 points for no congestion); for optical attenuation analysis, if the optical attenuation value is within the normal range and fluctuates little, a higher score is given (such as 6-9 points, 9 points for the best optical attenuation status). Calculate the comprehensive score where w i is the weight of the ith indicator, s i is the score of the ith indicator, and n is the total number of indicators.

[0136] According to the comprehensive score and the scores of each indicator, locate the area or component with problems. If the comprehensive score is low and the fault judgment indicator score is low, there is a problem with the internal hardware or software of the device, and it is necessary to further check the configuration and hardware status of the device; if the traffic congestion indicator score is low, focus on the network topology, port traffic distribution and other aspects; if the optical attenuation indicator score is low, focus on checking the optical fiber line.

[0137] In step S53, if the fault judgment result shows that there is a problem with the device hardware, consider upgrading the hardware. For the problems found in the optical attenuation analysis, if the excessive optical attenuation is caused by the degradation of the optical module performance, replace the optical module. According to the type and specification of the optical module, select a suitable optical module for replacement to ensure that the optical power is within the normal range. According to the service life of the equipment, operating environment and other factors, formulate a preventive hardware maintenance plan.

[0138] If the traffic congestion analysis shows that there is a traffic congestion problem, optimize the traffic distribution strategy. Use a class-based weighted fair queue algorithm to classify traffic according to service type (such as voice, video, data, etc.), and assign different bandwidth weights to different categories of services. Adjust the network topology to optimize the traffic path. Use the link state routing protocol to collect the link state information of the network, and calculate the optimal traffic path based on factors such as link bandwidth and delay. If it is found that some links are often congested, adjust the routing strategy so that the traffic bypasses these congested links. According to the port traffic data and device performance, reasonably configure the port parameters. For example, if the inbound traffic of a port is large for a long time, appropriately increase the cache size of the port to reduce the packet loss rate. At the same time, for underutilized ports, you can adjust their working mode (such as from half-duplex mode to full-duplex mode) to improve the transmission efficiency of the port.

[0139] When the optical attenuation analysis shows that the optical attenuation is abnormal, in addition to replacing the optical module, conduct a comprehensive inspection of the optical fiber line. Use an optical time domain reflectometer to detect the location of the optical fiber's loss point, breakpoint and other fault points. If the optical fiber is found to be bent or squeezed, adjust the laying path of the optical fiber to prevent the optical fiber from being subjected to unnecessary external forces. At the same time, ensure that the optical fiber connection points are clean and tightened, and clean and reconnect the optical fiber connection points regularly to reduce optical attenuation.

[0140] In step S54, a detailed implementation plan is formulated according to the proposed optimization plan. The implementation steps, responsible persons, time nodes, etc. of each optimization measure are clarified. During the implementation process, relevant operating specifications and safety standards are followed to ensure that the optimization process does not have a greater impact on the normal operation of the network.

[0141] Establish an effect evaluation index system, including equipment performance indicators (such as CPU usage, memory usage, etc.), network performance indicators (such as port traffic, network delay, packet loss rate, etc.) and optical attenuation indicators. Collect data on these indicators before and after the implementation of the optimization plan. Use comparative analysis methods to calculate the improvement of indicators. Adjust and improve the optimization plan based on the effect evaluation results. If some indicators have not been significantly improved, analyze the reasons and take further optimization measures, such as readjusting the traffic distribution strategy or replacing other hardware components.

[0142] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for intelligently judging and optimizing Internet access quality, characterized in that: The following steps are involved: S1: According to the data collector deployed on the OLT device, it communicates with the OLT device through the SNMP protocol to obtain the real-time operation status data of the device and pre-process the collected data; S2: Extract key features for fault diagnosis from data through feature extraction technology, train the fault diagnosis model through historical data, and perform real-time fault diagnosis on newly collected data based on the trained model; S3: Analyze the traffic congestion in the network through real-time monitoring of port traffic data and make optimization suggestions; S4: Analyze the attenuation of optical fiber lines and provide maintenance suggestions by monitoring the optical attenuation data of OLT equipment in real time; S5: Through comprehensive evaluation of fault judgment, traffic congestion analysis and optical attenuation analysis, an optimization solution for OLT equipment is proposed.

2. The method for intelligently judging and optimizing Internet access quality according to claim 1 is characterized in that: The real-time operation status data of the device is obtained, including: basic information data of the device, port traffic data, optical module information data, device performance data and network connection data; the preprocessing includes data cleaning and data normalization; the real-time collected data is stored by combining local storage and cloud transmission.

3. The method for intelligently judging and optimizing Internet access quality according to claim 1, characterized in that: The method extracts key features for fault judgment from data by using feature extraction technology, trains a fault judgment model by using historical data, and performs real-time fault judgment on newly collected data according to the trained model; the method comprises the following steps: S21: Determine key features based on the correlation between the fault type and the data, and extract the determined key features through data mining technology; S22: dividing the historical data into a training set and a test set, training the fault judgment model through the training set, and evaluating and optimizing the trained fault judgment model through the test set; S23: Continuously collect the real-time operation status data of the OLT device through the data collector, input the pre-processed data into the trained fault judgment model, and output the fault judgment result according to the input feature data.

4. The method for intelligently judging and optimizing Internet access quality according to claim 3 is characterized in that: Determining key features according to the correlation between the fault type and the data, and extracting the determined key features by using data mining technology; comprises the following steps: S211: Based on the understanding of the working principle of OLT equipment, network communication knowledge and common failure modes, determine the characteristics related to the failure from the various types of data collected; S212: Calculate basic statistics of each feature, and determine the normal fluctuation range of port traffic by analyzing the mean and standard deviation of inbound and outbound traffic; S213: Determine the correlation between each feature and the fault label by using the Pearson correlation coefficient, and determine the discrete feature by calculating the information gain of each feature; S214: Reduce the number of correlated features after initial screening through principal component analysis, and combine features into new features based on a deeper understanding of the failure mode.

5. The method for intelligently judging and optimizing Internet access quality according to claim 3 is characterized in that: The method of dividing the historical data into a training set and a test set, and training the fault judgment model through the training set, comprises the following steps: S221: Coordinate the data partitioning operation of each node according to the task scheduling component in the distributed architecture, calculate the hash value of each data record, and distribute the data to different subsets according to the range of the hash value, and divide it into a training set and a test set; S222: Each node performs Bootstrap sampling with replacement on the stored data in parallel, and sends the information of the local sub-dataset to the centralized coordination node through the distributed message passing mechanism. The coordination node summarizes the sampling results of each node to construct the sub-dataset of the entire random forest. S223: determining the best split feature based on the information gain criterion, and in the process of constructing multiple decision trees, allocating the task of constructing the decision trees according to the computing resources of the cluster; S224: The prediction results of the random forest are used as new features and combined with part of the original features to form input features of the logistic regression, and a logistic regression model is constructed.

6. The method for intelligently judging and optimizing Internet access quality according to claim 5 is characterized in that: The optimal splitting feature is determined based on the information gain criterion, and in the process of constructing multiple decision trees, the task of constructing the decision trees is allocated according to the computing resources of the cluster; The following steps are involved: S2231: When each node splits, the operation of randomly selecting features from all features is performed locally on each node, and each node calculates information gain for the local data subset and the selected features; S2232: After calculating the information gain, each node sends the information gain to the coordination node, which summarizes the calculation results and selects the feature with the largest information gain as the best split feature; S2233: In the process of building multiple decision trees, the distributed architecture manages the resource allocation and task scheduling of each node, and allocates the task of building the decision tree according to the computing resources of the cluster.

7. The method for intelligently judging and optimizing Internet access quality according to claim 5, characterized in that: The method of combining the prediction results of the random forest as new features with a part of the original features to form input features of the logistic regression and constructing a logistic regression model comprises the following steps: S2241: The prediction results after random forest training are used as new features, which are combined with the important features determined by feature selection in the original features to form the input features of logistic regression; S2242: Determine the probability of failure through the hypothesis function of the logistic regression model, measure the prediction accuracy by minimizing the logarithmic loss function of the loss function, update the model parameters through the stochastic gradient descent algorithm, and calculate the gradient of each sample; S2243: Send the locally calculated gradient to the summary node through the distributed message passing mechanism. The summary node accumulates the gradients sent by each node to obtain the global gradient. According to the global gradient and the set learning rate, the model parameters are updated, and the iteration is continued until the stop condition is met. In step S2242, the hypothesis function of the logistic regression model is: Where x is the input feature vector, θ is the model parameter vector, and T is the transpose; the logarithmic loss function is: Where m is the number of samples, y i is the true label of the i-th sample, and i is the index variable.

8. The method for intelligently judging and optimizing Internet access quality according to claim 2 is characterized in that: The method analyzes the traffic congestion in the network by real-time monitoring of port traffic data and proposes optimization suggestions; the method includes the following steps: S31: obtaining the pre-processed port flow data, calculating the average flow, peak flow and flow fluctuation range, then calculating the skewness and kurtosis of the flow, and calculating the autocorrelation coefficients under different time lags; S32: When the traffic exceeds a preset threshold, a congestion detection algorithm is triggered, and a comprehensive congestion judgment index is constructed by assigning different weights to the traffic surge index and the port utilization excessive index; S33: Adjust the traffic allocation strategy based on the service type. For services with high real-time requirements, fixed bandwidth is allocated. For non-real-time but important services, dynamic bandwidth allocation based on priority is adopted. For non-critical services, bandwidth is restricted when congested. According to the real-time traffic monitoring results, dynamic adjustment of traffic shaping is performed. S34: Plan backup links for network topology, enable backup links when network is congested, evaluate the degree of relief of network congestion based on the enabling effect, optimize traffic routing through routing algorithms based on traffic engineering, recalculate traffic routing after congestion detection, and implement dynamic optimization of traffic routing based on software-defined network technology; S35: During the continuous monitoring process, port traffic data is continuously collected and network performance indicators are monitored. The results of continuous monitoring are fed back to the network management system in a timely manner, and adjustment strategies are formulated based on the feedback results.

9. The method for intelligently judging and optimizing Internet access quality according to claim 2, characterized in that: The method of monitoring the optical attenuation data of the OLT device in real time, analyzing the attenuation of the optical fiber line, and providing maintenance suggestions comprises the following steps: S41: Collect the pre-processed light attenuation data through the optical module information data, fit the light attenuation data through the least square method, judge the light attenuation trend according to the slope of the fitted straight line, and calculate the standard deviation of the light attenuation data within the time window; S42: When the optical attenuation data trend is stable and close to the preset threshold, preventive maintenance is performed; when the optical attenuation fluctuates greatly but has not exceeded the threshold, the environment of the optical fiber line is checked, and at the same time, the optical fiber fixture is checked to see if it is firm; when the optical attenuation is on an upward trend and has not exceeded the threshold, in addition to the above inspections, the age and type of the optical fiber are also evaluated; when the optical attenuation value exceeds the preset threshold, an alarm message is generated; S43: Continuously collect and analyze optical attenuation data to monitor changes in the status of the optical fiber line, and provide feedback and adjust maintenance recommendations and alarm thresholds based on the implementation of maintenance measures and subsequent optical attenuation data monitoring results.

10. The method for intelligently judging and optimizing Internet access quality according to claim 2, characterized in that: The method proposes an optimization scheme for OLT equipment through comprehensive evaluation of fault judgment, traffic congestion analysis and optical attenuation analysis; the method comprises the following steps: S51: Determine the weights of fault judgment, traffic congestion analysis and optical attenuation analysis in comprehensive evaluation through the hierarchical analysis method; S52: Quantitatively score each indicator of fault judgment, traffic congestion analysis, and optical attenuation analysis according to the determined weights, and calculate a comprehensive score. According to the comprehensive score and the scores of each indicator, locate the area or component with the problem; S53: If the fault judgment result shows that there is a problem with the device hardware, then upgrade the hardware; if the excessive optical attenuation is caused by the degradation of the optical module performance, then replace the optical module; if the traffic congestion analysis shows that there is a traffic congestion problem, then optimize the traffic distribution strategy; S54: According to the proposed optimization plan, the implementation steps, responsible persons, and time nodes of each optimization measure are clarified, and an effect evaluation index system is established. The effect evaluation index includes equipment performance index, network performance index, and optical attenuation index.

Images