Tobacco informatization scene-oriented local area network double-link quality assurance real-time monitoring method

Abstract

The invention discloses a tobacco informatization scene-oriented local area network double-link quality assurance real-time monitoring method, which comprises the following steps of S1, collecting tobacco core business sensitive data in real time through double-link multi-protocol aiming at a tobacco industry three-level network topology of a prefecture, a county and a purchase point; s2, evaluating the quality of the main link and the standby link; s3, constructing a tobacco business full-scene time sequence sample; s4, constructing an LSTM link quality prediction model adapted to the tobacco double-link time sequence characteristics; s5, performing model training and optimization based on historical operation data; s6, setting a double-link quality assurance real-time monitoring mechanism in combination with the tobacco network topology hierarchy and the service priority; and S7, monitoring health states of the main link and the standby link in real time, calculating a health degree prediction value, and performing early warning, warning or suggestion according to a set double-link quality guarantee real-time monitoring mechanism according to a tobacco core business continuity demand. The comprehensive monitoring of double-link quality, risk preposed prevention and control and rapid fault disposal are realized, and the continuity of core businesses such as tobacco leaf purchasing and logistics scheduling is ensured.

Description

Technical Field

[0001] This invention relates to the field of communication fault monitoring technology, specifically to a real-time monitoring method for dual-link quality assurance in local area networks for tobacco information technology scenarios. Background Technology

[0002] Tobacco informatization is a systematic process by which the tobacco industry utilizes modern information technology to reshape and optimize its entire industry chain operation, management, and service models, ultimately building a data-driven smart tobacco ecosystem. It encompasses all aspects of tobacco planting, production, logistics, sales, and supervision. In the informatization construction of the tobacco industry, from state-level companies to their subordinate county (city) companies and then to lower-level tobacco leaf purchasing points, a dual-link redundancy architecture of "main link + backup link" is generally adopted. The core objective is to ensure the high availability of critical business data transmission. However, when existing network management solutions (such as traditional SNMP polling-based network management systems) are adapted to real-time monitoring scenarios in the tobacco network, most sampling periods are on the order of minutes (30s-60s), failing to capture sudden problems such as fiber optic interruptions and equipment port failures. This results in high fault detection latency and cannot meet the rapid response requirements for link interruptions during peak tobacco business periods.

[0003] Furthermore, traditional methods only focus on basic indicators such as latency and packet loss rate, ignoring the linkage analysis between bandwidth utilization and BFD session status. They have not established a unified link health evaluation system, and the differences in indicator dimensions lead to subjective quality judgments, making it impossible to accurately classify link operation levels. At the same time, they cannot capture the temporal correlation of multiple indicators, cannot achieve risk prevention and control quality assurance, and lack specific solutions for graded alarms and intelligent decision-making in tobacco-related special scenarios. Summary of the Invention

[0004] This invention aims to solve the technical problems existing in the prior art. In particular, it innovatively proposes a real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios, which realizes comprehensive monitoring of dual-link quality, risk prevention and control, and rapid fault handling, thereby ensuring the continuity of core businesses such as tobacco leaf acquisition and logistics scheduling.

[0005] To achieve the above objectives, this invention provides a real-time monitoring method for dual-link quality assurance in local area networks for tobacco information technology scenarios, comprising the following steps:

[0006] S1: For the three-tier network topology of the tobacco industry (state-county-purchasing point), sensitive data of core tobacco business are collected in real time through dual links and multiple protocols;

[0007] S2: Evaluate the quality of the primary and backup links;

[0008] S3: Construct time-series samples for the entire tobacco business scenario;

[0009] S4: Construct an LSTM link quality prediction model adapted to the timing characteristics of tobacco dual links;

[0010] S5: Model training and optimization based on historical operational data;

[0011] S6: Combine the tobacco network topology hierarchy and service priority to set up a dual-link quality assurance real-time monitoring mechanism;

[0012] S7: Monitors the health status of the main link and backup link in real time, calculates the predicted health value, and provides early warnings, alerts or suggestions based on the set dual-link quality assurance real-time monitoring mechanism to meet the continuity requirements of tobacco core business.

[0013] In the above scheme, step S1 further includes:

[0014] The collected data includes metrics across five dimensions: link round-trip latency, latency jitter, packet loss rate, bandwidth utilization, and BFD session status, required for real-time uploading of tobacco leaf acquisition data and low-latency transmission of logistics scheduling instructions; let the first... The link at time The original state vector is: ;

[0015] in, For a moment No. The state vector of the link; For a moment No. The round-trip time of each link, For a moment No. The latency jitter of the link, For a moment No. Packet loss rate of each link, For a moment No. Bandwidth utilization of each link For a moment No. The BFD session state of each link;

[0016] S1-1: By triggering RPING at certain intervals through the active detection layer, the round-trip latency, latency jitter, packet loss rate and bandwidth utilization of the main link and backup link that carry the core business of tobacco leaf acquisition and logistics scheduling are collected in real time.

[0017] S1-2: The passive sensing layer monitors the BFD session status of the main link and the backup link in real time at certain intervals.

[0018] Deploy the BFD session monitoring module and configure the BFD session monitoring cycle. Real-time monitoring of the BFD session status of the primary and backup links; when detected... When the value changes from 1 to 0, a passive alarm is immediately triggered.

[0019] In the above scheme, step S1-1 further includes the following steps:

[0020] S1-1-1: Actively send ICMP probe packets with source address via SNMPRPING protocol to target and collect the round-trip latency of the state-county and county-collection point links. The ICMP probe packet timeout is set to 100ms.

[0021] Simultaneously, preprocessing is required when collecting data to remove outliers; during the removal process, the following methods are used: Criteria for filtering outliers, i.e., when the indicator value exceeds... If the value is outside the range, it is identified as an outlier, and then valid data is supplemented using linear interpolation.

[0022] S1-1-2: Calculate latency jitter;

[0023] Take the link round-trip delay over several adjacent cycles and calculate the average. Substitute this into the following formula:

[0024] ;

[0025] in, The average time delay over several periods prior to time t; For a moment Next The round-trip time of each link; For time delay jitter;

[0026] S1-1-3: Count the number of ICMP probe packets sent and received, and calculate the packet loss rate;

[0027] The formula is as follows:

[0028] ;

[0029] in, For a moment No. Packet loss rate of each link;

[0030] S1-1-4: Read the inbound / outbound traffic data of the network device interface counter via the SNMP protocol, and calculate the bandwidth utilization rate by combining it with the link's rated bandwidth;

[0031] The formula is as follows:

[0032] ;

[0033] in, For a moment No. Bandwidth utilization of each link.

[0034] In the above scheme, step S2 also includes the following:

[0035] The formula is as follows:

[0036] ;

[0037] in, For the first The link at time A comprehensive health assessment score, and ;

[0038] , , , , All are weighting coefficients, and each weighting coefficient is determined using the analytic hierarchy process based on the importance of tobacco business.

[0039] The time-delay normalized mapping function uses an inverse linear mapping, and the specific formula is as follows:

[0040] ;

[0041] in, For optimal latency, Tolerate the maximum latency for business operations For the first The link at time Link round-trip latency;

[0042] The jitter normalization mapping function is defined by the following formula:

[0043] ;

[0044] in, For optimal latency jitter, Tolerate the maximum latency jitter for business operations For the first The link at time The time delay jitter;

[0045] The normalized mapping function for packet loss rate is as follows:

[0046] ;

[0047] in, For the first The link at time The packet loss rate;

[0048] The normalized mapping function for bandwidth utilization is formulated as follows:

[0049] ;

[0050] in, This is the link congestion threshold. For the first The link at time Bandwidth utilization.

[0051] In the above scheme, step S3 also includes the following steps:

[0052] S3-1: Determine the time range covered by the feature window, and match the window length with the business data transmission cycle;

[0053] Take a sliding time window of length T, which contains the original state vector for T consecutive sampling periods;

[0054] S3-2: Collect the original state vector at each time step ;

[0055] S3-3: Concatenate to form a feature matrix;

[0056] ;

[0057] in, For the first The link at time The input feature matrix;

[0058] S3-4: Generate multiple sets of samples, including samples from specific scenarios such as high load during the tobacco purchasing season, low load during the non-purchasing season, and link failure switching;

[0059] S3-5: Predict link health;

[0060] The prediction model constructed in step S4 is used for calculation.

[0061] In the above scheme, step S4 also includes the following steps:

[0062] S4-1: Input the temporal feature matrix into the input layer

[0063] S4-2: Extract preliminary temporal features using the first-layer LSTM unit;

[0064] S4-2-1: Define the first layer of LSTM units, set the number of hidden units in the first layer of LSTM units to 64, and set the activation function to... Activation function;

[0065] S4-2-2: Input the time series feature matrix ;

[0066] S4-2-3: Extract shallow temporal features through the first-layer LSTM unit and increase the dimensionality;

[0067] S4-2-4: Through The activation function constrains the feature values ​​output by the first-layer LSTM unit to avoid excessive numerical fluctuations.

[0068] Output ,in ;

[0069] in, This represents the set of output features of the first LSTM unit. For the first Hidden state features of the first layer LSTM unit corresponding to each time step;

[0070] S4-3: Randomly deactivate some neurons in Dropout layer 1 to prevent model overfitting;

[0071] S4-4: Deep feature fusion is performed through the second-layer LSTM unit;

[0072] S4-4-1: Define the second LSTM layer, set the number of hidden units in the second LSTM layer to 32, and set the activation function to... Activation function;

[0073] S4-4-2: Input Feature Set ;

[0074] S4-4-3: Extract deep temporal features using the second-layer LSTM unit;

[0075] S4-4-4: Through The activation function constrains the feature values ​​output by the second-layer LSTM unit;

[0076] Output ,in ;

[0077] in, This represents the set of output features of the second-layer LSTM unit. For the first Hidden state features of the second-layer LSTM unit corresponding to each time step;

[0078] S4-5: Enhance generalization capability through Dropout layer 2;

[0079] S4-6: Calculate the attention weights of the hidden states at each time step using an attention mechanism;

[0080] The formula is as follows:

[0081] ;

[0082] in, For the first Attention weights corresponding to each time step It is the attention weight matrix, which can be obtained through learning; For the first Hidden state features of the second-layer LSTM unit corresponding to each time step; It is an exponential function. To combine the attention weight matrix with the first Multiply the hidden state features of the second-layer LSTM unit corresponding to each time step;

[0083] Calculate the attention vector of the hidden state at each time step. ;

[0084] The formula is as follows:

[0085] ;

[0086] S4-7: Mapped to 1-dimensional features through a fully connected layer;

[0087] attention vector Perform a linear transformation, mapping it to 1-dimensional features, and then... Activation is performed using an activation function;

[0088] S4-8: Using the Sigmoid activation function, constrain the output to the [0,1] interval to obtain the future... Link health prediction value at time Output;

[0089] The formula is as follows:

[0090] ;

[0091] in, and These are the weights of the fully connected layer, obtained through training. The attention vector for the hidden state at each time step.

[0092] In the above scheme, step S5 also includes the following steps:

[0093] S5-1: Construct a training sample set, which includes historical data of tobacco business dual links during failures and deterioration at the state-county and county-collection point levels.

[0094] S5-2: Define the loss function;

[0095] The loss function is as follows:

[0096] ;

[0097] in, To determine the true health of the sample, For the corresponding predicted health level; The number of samples;

[0098] S5-3: Configure training parameters;

[0099] The initial learning rate η = 0.001 is set. To improve training stability, a learning rate decay strategy is set, where the learning rate decays to 0.9 times the original value every 20 epochs. The batch size is set to 64 to balance training efficiency and gradient stability. The maximum number of training epochs is 200. When the validation set loss does not decrease for 10 consecutive epochs, training is stopped and the current optimal model parameters are saved.

[0100] In the above scheme, step S6 further includes the following steps:

[0101] S6-1: Set early warning thresholds based on the degree of impact of tobacco business interruption;

[0102] Based on the health level classification of each link, set degradation alarm thresholds. When the real-time health status or prediction result is less than the warning threshold When the link degradation alarm is triggered, the system automatically triggers a fault alarm; a fault threshold μ is set, and when the real-time health status or prediction result is less than the fault threshold μ, an emergency fault alarm is triggered.

[0103] S6-2: Configure hierarchical alarm policies;

[0104] Combining the tobacco network topology hierarchy and the importance of link services, the k value of each link is pre-divided into three intervals;

[0105] If the k-value of the link requiring fault or degradation alarms falls within the value range representing the core link of the state-county, a level one alarm will be issued, using a triple notification of "SMS + email + system pop-up", with a response time of ≤5 minutes.

[0106] If the k-value of the link requiring fault or degradation alarm belongs to the value range of the core link representing the county-collection point, a level two alarm will be issued, using "SMS + system pop-up" notification, with a response time of ≤10 minutes.

[0107] If the k-value of the link requiring fault or degradation alarms falls within the range representing non-core links, a system pop-up notification will be used, with a response time of ≤30 minutes.

[0108] S6-3: Set output decision recommendations;

[0109] Based on the prediction results and current status of the primary and backup links, output differentiated handling suggestions or automated emergency handling:

[0110] If the main link prediction is good, that is And the backup link is currently healthy, that is It is recommended to guide subsequent core services to temporarily use the backup link for transmission to reduce the bandwidth pressure on the main link; among which Main link future Health prediction value at any given time For backup links Health value at any given time;

[0111] If the main link is predicted to deteriorate, that is... And the backup link is currently in good or healthy condition, that is If so, it is recommended to immediately redirect core business traffic to the backup link, and at the same time notify the operations and maintenance personnel to check the network environment of the primary and backup links.

[0112] If the main link is predicted to fail, that is... The backup link is currently in good or healthy condition. If this happens, automatic traffic switching will be triggered immediately, and maintenance personnel will be notified to check the network environment of the primary and backup links.

[0113] If both primary and backup links are predicted to degrade, that is... ,and If this happens, an emergency response strategy will be triggered, and operations and maintenance personnel will be notified to investigate problems with link aggregation nodes, core data centers of state / county companies, and network operator lines to avoid affecting business across the entire region.

[0114] In the above scheme, step S7 also includes the following:

[0115] When an alarm is triggered, the alarm trigger time, handling measures, handling results, link recovery, and the actual health status and predicted health status change trends in the 10-second time window before the trigger are recorded to form automatic data collection. At the same time, the fault data is added to the model training set to continuously optimize the prediction accuracy of the LSTM model.

[0116] In summary, the beneficial effects of this invention are as follows: Through a collaborative acquisition mechanism of "active detection + passive perception," with a differentiated detection cycle of 1s for the main link and 5s for the backup link, and 10ms-level BFD session monitoring, the fault capture time is compressed from the traditional minutes to the millisecond level, accurately adapting to the rapid response requirements during peak tobacco business periods and significantly reducing the risk of long-term interruptions in core tobacco information services such as tobacco leaf acquisition and logistics scheduling. Simultaneously, it innovatively constructs a quantitative health evaluation system that integrates round-trip latency, jitter, packet loss rate, bandwidth utilization, and BFD session status. Through scientific weight allocation and normalization, it eliminates dimensional differences, achieving objective and accurate classification of link quality levels. This is further enhanced by a dual-layer LS system incorporating an attention mechanism. The TM model predicts quality trends 10 seconds in advance. Combining the hierarchical alarm and intelligent decision-making logic of tobacco network topology and business importance in tobacco informatization, it provides differentiated handling suggestions for operation and maintenance personnel (such as seamless traffic switching and emergency troubleshooting guidance), which reduces the operation and maintenance response time by 60%. It also continuously optimizes model performance through a closed-loop feedback mechanism, improves the utilization rate of primary and backup link resources, reduces network operating costs, and fully realizes real-time monitoring of link quality, risk prevention and control, and rapid fault handling. This ensures the stable operation of all aspects of tobacco planting, production, logistics, sales, and supervision, and guarantees the availability of tobacco informatization applications, the accuracy of data transmission, and the continuity of business processes, providing a solid guarantee for business continuity. Attached Figure Description

[0117] Figure 1 This is a flowchart of the present invention.

[0118] Figure 2 This invention demonstrates the timing and early prediction effectiveness of link health detection.

[0119] Figure 3 This is a comparison chart of the cumulative distribution of fault detection latency between the present invention and traditional network monitoring methods.

[0120] Figure 4 This is a comparison chart showing the differences between the traditional solution and the present invention in terms of average fault detection delay and 95th percentile delay.

[0121] Figure 5 This invention uses the ROC curve for alarm judgment based on link health prediction results.

[0122] Figure 6 This is the curve showing the relationship between precision and recall in the alarm discrimination task of this invention.

[0123] Figure 7 This invention is based on a correspondence diagram between the predicted results and the actual link status level. Detailed Implementation

[0124] The present invention will be further described below with reference to the embodiments and accompanying drawings:

[0125] like Figures 1-7 As shown, a real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios includes the following steps:

[0126] S1: Real-time data collection via dual-link multi-protocol for the three-tiered network topology of the tobacco industry (state-county-collection point);

[0127] The prefecture-level government is responsible for the unified scheduling, regulatory decision-making, and resource coordination of tobacco leaf production and procurement across the entire city / prefecture, and sends corresponding strategies to the county-level government. The county-level government is responsible for distributing prefecture-level tasks to various stations and is the specific organizer and implementer of tobacco leaf procurement. Procurement stations are responsible for tobacco leaf grading, weighing, warehousing, settlement, and data collection and traceability. All levels are connected via a main link plus backup links, ensuring that the most advanced procurement data (such as grade, weight, and tobacco farmer information) is aggregated at the county level and ultimately fed into the prefecture-level management platform in real time. This achieves "data centralization upwards and service extension downwards," realizing tobacco informatization.

[0128] The collected data includes indicators from five dimensions: link round-trip latency, latency jitter, packet loss rate, bandwidth utilization, and BFD session status, which are required to ensure the real-time uploading of tobacco leaf purchase data and the low-latency transmission of logistics scheduling instructions.

[0129] Let the first Link ( =1,2,...,n, where n is the total number of dual links (including primary and backup links) at time... The original state vector is: ;

[0130] in, For a moment No. The state vector of the link; For a moment No. The round-trip time of each link, For a moment No. The latency jitter of the link, For a moment No. Packet loss rate of each link, For a moment No. Bandwidth utilization of each link For a moment No. The BFD session state of each link;

[0131] S1-1: The active detection layer triggers RPING at regular intervals to collect real-time data on the round-trip latency, latency jitter, packet loss rate, and bandwidth utilization of the main and backup links carrying the core business of tobacco leaf acquisition and logistics scheduling. The main link detection period is... The backup link detection cycle is ;

[0132] S1-1-1: Actively send ICMP probe packets with source address via SNMPRPING protocol to target and collect the round-trip latency of the state-county and county-collection point links. The ICMP probe packet timeout is set to 100ms.

[0133] Simultaneously, preprocessing is required when collecting data to remove outliers; during the removal process, the following methods are used: Criteria for filtering outliers, i.e., when the indicator value exceeds... If the value is outside the range, it is identified as an outlier, and then valid data is supplemented using linear interpolation.

[0134] S1-1-2: Calculate latency jitter;

[0135] Take the link round-trip delay over several adjacent cycles and calculate the average. Substitute this into the following formula:

[0136] ;

[0137] In this embodiment The average time delay of the three periods preceding time t; For a moment Next The round-trip time of each link; For time delay jitter;

[0138] S1-1-3: Count the number of ICMP probe packets sent and received, and calculate the packet loss rate;

[0139] The formula is as follows:

[0140] ;

[0141] in, For a moment No. Packet loss rate of each link;

[0142] S1-1-4: Read the inbound / outbound traffic data of the network device interface counter via the SNMP protocol, and calculate the bandwidth utilization rate by combining it with the link's rated bandwidth;

[0143] The formula is as follows:

[0144] ;

[0145] in, For a moment No. Bandwidth utilization of each link;

[0146] S1-2: The passive sensing layer monitors the BFD session status of the main link and the backup link in real time at certain intervals.

[0147] Deploy the BFD session monitoring module and configure the BFD session monitoring cycle. Real-time monitoring of the BFD session status of the primary and backup links; when detected... When the value changes from 1 to 0, a passive alarm is immediately triggered, enabling millisecond-level fault detection.

[0148] S2: Evaluate the quality of the primary and backup links;

[0149] The formula is as follows:

[0150] ;

[0151] in, For the first The link at time A comprehensive health assessment score, and The higher the value, the better the link health, and the higher the stability of data transmission for core tobacco business. The specific levels are: Excellent (…). ),good( ), deterioration ( ),Fault( 4);

[0152] , , , , All are weighting coefficients, and each weighting coefficient is determined using the analytic hierarchy process (AHP) based on the importance of tobacco business, and satisfies the following conditions: The weight configuration in this embodiment is as follows: , , , , ;

[0153] The time-delay normalized mapping function uses an inverse linear mapping, and the specific formula is as follows:

[0154] ;

[0155] in, For optimal latency, this embodiment takes... , To determine the maximum tolerable latency for the service, this embodiment takes... ; For a moment No. The round-trip time of each link;

[0156] The jitter normalization mapping function is defined by the following formula:

[0157] ;

[0158] in, To achieve optimal latency jitter, this embodiment takes... , To accommodate the maximum latency jitter that the service can tolerate, this embodiment takes... ; For the first The link at time The time delay jitter;

[0159] The normalized mapping function for packet loss rate is as follows:

[0160] ;

[0161] in, For the first The link at time The packet loss rate, when hour, That is, it does not contribute to health;

[0162] The normalized mapping function for bandwidth utilization is formulated as follows:

[0163] ;

[0164] in, The link congestion threshold is taken in this embodiment. ; For the first The link at time Bandwidth utilization;

[0165] S3: Construct a full-scenario time-series sample for tobacco; the full-scenario time-series sample for tobacco includes all links and processes in the tobacco industry chain, such as planting, production, logistics, sales, and supervision.

[0166] S3-1: Determine the time range covered by the feature window. The window length matches the business data transmission cycle and is an integer multiple of the business data transmission cycle.

[0167] In this embodiment, a sliding time window of length T is used, which contains the original state vector for T consecutive sampling periods.

[0168] S3-2: Collect the original state vector at each time step ;

[0169] S3-3: Concatenate to form a feature matrix;

[0170] ;

[0171] in, For the first The link at time The input feature matrix;

[0172] S3-4: Generate multiple sets of samples, covering specific scenarios such as high load during the tobacco purchasing season, low load during the non-purchasing season, and link failure switching;

[0173] S3-5: Predict link health;

[0174] The predicted health of the link is calculated using the prediction model built in step S4. The calculation process is shown in S4-8 below. To predict the step size, this embodiment sets... (The main link corresponds to 10 sampling periods, and the backup link corresponds to 2 sampling periods), enabling the prediction of link quality trends 10 seconds in advance;

[0175] S4: Construct an LSTM link quality prediction model adapted to the timing characteristics of tobacco dual links;

[0176] A multivariate time series prediction model adapted to tobacco dual-link time series data is constructed. A two-layer LSTM network structure is adopted, and Dropout regularization, attention mechanism and fully connected layer are combined to achieve accurate prediction.

[0177] S4-1: Input the temporal feature matrix into the input layer

[0178] S4-2: Extract preliminary temporal features using the first-layer LSTM unit;

[0179] S4-2-1: Define the first layer of LSTM units, set the number of hidden units in the first layer of LSTM units to 64, and set the activation function to... Activation function;

[0180] S4-2-2: Input the time series feature matrix ;

[0181] S4-2-3: Extract shallow temporal features through the first-layer LSTM unit and increase the dimensionality;

[0182] S4-2-4: Through The activation function constrains the feature values ​​output by the first-layer LSTM unit to avoid excessive numerical fluctuations.

[0183] Output ,in ;

[0184] in, This represents the set of output features of the first LSTM unit. For the first Hidden state features of the first layer LSTM unit corresponding to each time step;

[0185] S4-3: Overfitting is prevented by randomly deactivating some neurons in Dropout layer 1; in this embodiment, the dropout rate of Dropout layer 1 is set to... We randomly deactivate 20% of the neurons temporarily to prevent the model from becoming overly reliant on a few features;

[0186] S4-4: Deep feature fusion is performed through the second-layer LSTM unit;

[0187] S4-4-1: Define the second LSTM layer, set the number of hidden units in the second LSTM layer to 32, and set the activation function to... Activation function;

[0188] S4-4-2: Input Feature Set ;

[0189] S4-4-3: Extract deep temporal features using the second-layer LSTM unit;

[0190] S4-4-4: Through The activation function constrains the feature values ​​output by the second-layer LSTM unit;

[0191] Output ,in ;

[0192] in, This represents the set of output features of the second-layer LSTM unit. For the first Hidden state features of the second-layer LSTM unit corresponding to each time step;

[0193] S4-5: Enhance generalization capability through Dropout layer 2; in this embodiment, the dropout rate of Dropout layer 2 is set to... ;

[0194] S4-6: Calculate the attention weights of the hidden states at each time step using an attention mechanism;

[0195] The formula is as follows:

[0196] ;

[0197] in, For the first Attention weights corresponding to each time step It is the attention weight matrix, which can be obtained through learning; For the first Hidden state features of the second-layer LSTM unit corresponding to each time step; It is an exponential function. To combine the attention weight matrix with the first Multiply the hidden state features of the second-layer LSTM unit corresponding to each time step;

[0198] Calculate the attention vector of the hidden state at each time step. ;

[0199] The formula is as follows:

[0200] ;

[0201] S4-7: Mapped to 1-dimensional features through a fully connected layer;

[0202] 32-dimensional attention vector Perform a linear transformation, mapping it to 1-dimensional features, and then... Activation is performed using an activation function;

[0203] S4-8: Using the Sigmoid activation function, constrain the output to the [0,1] interval to obtain the future... Link health prediction value at time Output;

[0204] The formula is as follows:

[0205] ;

[0206] in, and These are the weights of the fully connected layer, obtained through training. This is the attention vector for the hidden state.

[0207] S5: Model training and optimization based on historical operational data;

[0208] S5-1: Construct a training sample set, which includes 12 months of historical data on the dual-link tobacco business, including the acquisition season and non-acquisition season, and covers abnormal scenarios such as state-county and county-acquisition point failures and deterioration.

[0209] We collected 12 months of historical dual-link data, covering all scenarios including normal operation, performance degradation, and failover, and constructed a total sample size. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

[0210] S5-2: Define the loss function;

[0211] The loss function is as follows:

[0212] ;

[0213] in, To calculate the true health level, This corresponds to the predicted value; The number of samples;

[0214] S5-3: Configure training parameters;

[0215] The initial learning rate is set to η = 0.001. To improve training stability, a learning rate decay strategy is set, where the learning rate decreases to 0.9 times its original value every 20 epochs. The batch size is set to 64 to balance training efficiency and gradient stability. The maximum number of training epochs is 200. When the validation set loss does not decrease for 10 consecutive epochs (the decrease is ≤ 0.0001), training is stopped, the current optimal model parameters are saved, and overfitting is avoided.

[0216] S6: Combine the tobacco network topology hierarchy and service priority to set up a dual-link quality assurance real-time monitoring mechanism;

[0217] S6-1: Set early warning thresholds based on the degree of impact of tobacco business interruption;

[0218] Based on the health level classification of each link, set degradation alarm thresholds. In this embodiment The value is 0.6 when the real-time health status or prediction result is less than the warning threshold. When the link degradation alarm is triggered, the system automatically triggers the link degradation alarm; a fault threshold μ is set, which is 0.4 in this embodiment. When the real-time health or prediction result is less than the fault threshold μ, an emergency fault alarm is triggered.

[0219] S6-2: Configure hierarchical alarm policies;

[0220] Based on the tobacco network topology hierarchy and the importance of link services, the k-value of each link is calculated, and the links are pre-divided into three intervals; among them, the k-values ​​of the links between the state, county, and collection point belong to non-overlapping intervals.

[0221] If the k-value of the link requiring fault or degradation alarms falls within the value range representing the core link of the state-county, a level one alarm will be issued, using a triple notification of "SMS + email + system pop-up", with a response time of ≤5 minutes.

[0222] If the k-value of the link requiring fault or degradation alarm belongs to the value range of the core link representing the county-collection point, a level two alarm will be issued, using "SMS + system pop-up" notification, with a response time of ≤10 minutes.

[0223] If the value falls within a range representing non-core links, a system pop-up notification will be used, with a response time of ≤30 minutes.

[0224] S6-3: Set output decision recommendations;

[0225] Based on the prediction results and current status of the primary and backup links, output differentiated handling suggestions or automated emergency handling:

[0226] If the main link prediction is good, that is And the backup link is currently healthy, that is It is recommended to guide subsequent core services to temporarily use the backup link for transmission to reduce the bandwidth pressure on the main link; among which Main link future Health prediction value at any given time For backup links Health value at any given time;

[0227] If the main link is predicted to deteriorate, that is... And the backup link is currently in good or healthy condition, that is If so, it is recommended to immediately redirect core business traffic to the backup link, and at the same time notify the operations and maintenance personnel to check the network environment of the primary and backup links.

[0228] If the main link is predicted to fail, that is... The backup link is currently in good or healthy condition. If this happens, automatic traffic switching will be triggered immediately, and maintenance personnel will be notified to check the network environment of the primary and backup links.

[0229] If both primary and backup links are predicted to degrade, that is... ,and If this happens, an emergency response strategy will be triggered, and operations and maintenance personnel will be notified to investigate common issues such as link aggregation nodes, core data centers of state / county companies, and network operator lines, in order to avoid affecting business across the entire region.

[0230] S7: Monitor the health status of the main link and backup link in real time, calculate the predicted health value, and provide early warnings, alarms or suggestions based on the set dual-link quality assurance real-time monitoring mechanism to meet the continuity requirements of tobacco core business.

[0231] When an alarm is triggered, the alarm trigger time, handling measures, handling results, link recovery, and the actual health status and predicted health status change trends in the 10-second time window before the trigger are recorded to form automatic data collection. At the same time, the fault data is added to the model training set to continuously optimize the prediction accuracy of the LSTM model.

[0232] To better illustrate this technical solution, the following experimental measurements were conducted:

[0233] like Figure 2 As shown, under typical operating scenarios, the curves of link health status over time, constructed based on multi-dimensional link metrics, and the prediction results of link health status for the next 10 seconds based on historical sliding windows are presented. The real-time health curve reflects the link's status changes during normal operation, gradual performance degradation, and sudden failures; the predicted health curve shows a continuous downward trend before the link enters significant degradation or failure, thus triggering early warnings. In this way, this solution can identify potential anomalies before the link is interrupted or experiences severe performance degradation, providing a time window for subsequent link switching or maintenance.

[0234] like Figure 3 As shown in the cumulative distribution comparison results regarding fault detection latency, traditional solutions rely on minute-level polling mechanisms, and the fault detection latency is limited by the polling cycle, resulting in significant random waiting time. In contrast, this invention combines second-level active detection with millisecond-level passive state awareness, enabling the detection of link anomalies in extremely short times in most cases. As can be seen from the distribution curves in the figure, under the same fault scenario, the solution of this invention significantly shortens the overall fault detection latency and improves the real-time performance of network state awareness.

[0235] like Figure 4 As shown, the results demonstrate that, compared to traditional monitoring methods, this solution exhibits lower fault detection latency under both average and extreme conditions, indicating that it not only improves overall performance but also provides more stable response capabilities under complex or sudden anomaly conditions. This characteristic is beneficial for enhancing the reliability of link switching strategies and service assurance mechanisms.

[0236] like Figure 5As shown, this technical solution uses predicted health status as the criterion to classify and judge whether the system will enter an abnormal or faulty state in the future. This achieves a high anomaly identification capability while maintaining a low false alarm rate. The results demonstrate that the health status model and prediction mechanism constructed in this technical solution have good discriminative ability in link status discrimination, which helps reduce the interference of invalid alarms on operation and maintenance personnel.

[0237] like Figure 6 As shown in the figure, under different threshold settings, this technical solution can maintain a reasonable precision level while maintaining a high recall rate, indicating that it strikes a balance between "early detection of problems" and "avoiding excessive alarms" during anomaly identification. This characteristic makes this invention suitable for network operating environments with high requirements for stability and reliability.

[0238] like Figure 7 As shown, this technical solution divides the link status into multiple levels (high quality, good, deteriorated, and faulty) and compares the predicted level with the actual level. It can be seen that the two are consistent in most cases. This result indicates that the health calculation and prediction method of this solution can accurately reflect the actual operating status of the link, possessing good status characterization capability and stability.

Claims

1. A real-time monitoring method for dual-link quality assurance in local area networks for tobacco information technology scenarios, characterized in that, Includes the following steps: S1: For the three-tier network topology of the tobacco industry (state-county-purchasing point), real-time collection of core tobacco business data is achieved through dual-link multi-protocol; S2: Evaluate the quality of the primary and backup links; S3: Construct time-series samples for the entire tobacco business scenario; S4: Construct an LSTM link quality prediction model adapted to the timing characteristics of tobacco dual links; S5: Model training and optimization based on historical operational data; S6: Combine the tobacco network topology hierarchy and service priority to set up a dual-link quality assurance real-time monitoring mechanism; S7: Monitors the health status of the main link and backup link in real time, calculates the predicted health value, and provides early warnings, alerts or suggestions based on the set dual-link quality assurance real-time monitoring mechanism to meet the continuity requirements of tobacco core business.

2. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 1, characterized in that: Step S1 also includes: The collected data includes metrics across five dimensions: link round-trip latency, latency jitter, packet loss rate, bandwidth utilization, and BFD session status, required to ensure real-time uploading of tobacco leaf acquisition data and low-latency transmission of logistics scheduling instructions; let the first... The link at time The original state vector is: ; in, For a moment No. The state vector of the link; For a moment No. The round-trip time of each link. For a moment No. The latency jitter of the link, For a moment No. Packet loss rate of each link, For a moment No. Bandwidth utilization of each link For a moment No. The BFD session state of each link; S1-1: By triggering RPING at certain intervals through the active detection layer, the round-trip latency, latency jitter, packet loss rate and bandwidth utilization of the main link and backup link that carry the core business of tobacco leaf acquisition and logistics scheduling are collected in real time. S1-2: The passive sensing layer monitors the BFD session status of the main link and the backup link in real time at certain intervals. Deploy the BFD session monitoring module and configure the BFD session monitoring cycle. Real-time monitoring of the BFD session status of the primary and backup links; when detected... When the value changes from 1 to 0, a passive alarm is immediately triggered.

3. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 2, characterized in that: Step S1-1 also includes the following steps: S1-1-1: Actively send ICMP probe packets with source address via SNMPRPING protocol to target and collect the round-trip latency of the state-county and county-collection point links. The ICMP probe packet timeout is set to 100ms. Simultaneously, preprocessing is required when collecting data to remove outliers; during the removal process, the following methods are used: The criterion filters outoutliers, i.e., when the indicator value exceeds... If the value is outside the range, it is identified as an outlier, and then valid data is supplemented using linear interpolation. S1-1-2: Calculate latency jitter; Take the link round-trip delay over several adjacent cycles and calculate the average. Substitute this into the following formula: ； in, The average time delay over several periods prior to time t; For a moment Next The round-trip time of each link; For time delay jitter; S1-1-3: Count the number of ICMP probe packets sent and received, and calculate the packet loss rate; The formula is as follows: ； in, For a moment No. Packet loss rate of each link; S1-1-4: Read the inbound / outbound traffic data of the network device interface counter via the SNMP protocol, and calculate the bandwidth utilization rate by combining it with the link's rated bandwidth; The formula is as follows: ； in, For a moment No. The bandwidth utilization of each link.

4. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 1, characterized in that: Step S2 also includes the following: The formula is as follows: ； in, For the first The link at time A comprehensive health assessment score, and ; , , , , All are weighting coefficients, and each weighting coefficient is determined using the analytic hierarchy process based on the importance of tobacco business. The time-delay normalization mapping function uses an inverse linear mapping, and the specific formula is as follows: ； in, For optimal latency, Tolerate the maximum latency for business operations For the first The link at time Link round-trip latency; The jitter normalization mapping function is defined by the following formula: ； in, For optimal latency jitter, Tolerate the maximum latency jitter for business operations For the first The link at time The time delay jitter; The normalized mapping function for packet loss rate is shown in the following formula: ； in, For the first The link at time The packet loss rate; The normalized mapping function for bandwidth utilization is expressed as follows: ； in, This is the link congestion threshold. For the first The link at time Bandwidth utilization.

5. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 1, characterized in that: Step S3 also includes the following steps: S3-1: Determine the time range covered by the feature window, and match the window length with the business data transmission cycle; Take a sliding time window of length T, which contains the original state vector for T consecutive sampling periods; S3-2: Collect the original state vector at each time step ; S3-3: Concatenate to form a feature matrix; ； in, For the first The link at time The input feature matrix; S3-4: Generate multiple sets of samples, including samples from scenarios such as high load during the tobacco purchasing season, low load during the non-purchasing season, and link failure switching. S3-5: Predict link health; The prediction model constructed in step S4 is used for calculation.

6. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 5, characterized in that: Step S4 also includes the following steps: S4-1: Input the temporal feature matrix into the input layer ; S4-2: Extract preliminary temporal features using the first-layer LSTM unit; S4-2-1: Define the first layer of LSTM units, set the number of hidden units in the first layer of LSTM units to 64, and set the activation function to... Activation function; S4-2-2: Input the time series feature matrix ; S4-2-3: Extract shallow temporal features through the first-layer LSTM unit and increase the dimensionality; S4-2-4: Through The activation function constrains the feature values ​​output by the first-layer LSTM unit to avoid excessive numerical fluctuations. Output ,in ; in, This represents the set of output features of the first LSTM unit. For the first Hidden state features of the first layer LSTM unit corresponding to each time step; S4-3: Randomly deactivate some neurons in Dropout layer 1 to prevent model overfitting; S4-4: Deep feature fusion is performed through the second-layer LSTM unit; S4-4-1: Define the second LSTM unit, set the number of hidden units in the second LSTM unit to 32, and set the activation function to... Activation function; S4-4-2: Input Feature Set ; S4-4-3: Extract deep temporal features using the second-layer LSTM unit; S4-4-4: Through The activation function constrains the feature values ​​output by the second-layer LSTM unit; Output ,in ; in, This represents the set of output features of the second-layer LSTM unit. For the first Hidden state features of the second-layer LSTM unit corresponding to each time step; S4-5: Enhance generalization capability through Dropout layer 2; S4-6: Calculate the attention weights of the hidden states at each time step using an attention mechanism; The formula is as follows: ； in, For the first Attention weights corresponding to each time step It is the attention weight matrix, which can be obtained through learning; For the first Hidden state features of the second-layer LSTM unit corresponding to each time step; It is an exponential function. To combine the attention weight matrix with the first Multiply the hidden state features of the second-layer LSTM unit corresponding to each time step; Calculate the attention vector of the hidden state at each time step. ; The formula is as follows: ； S4-7: Mapped to 1-dimensional features through a fully connected layer; attention vector Perform a linear transformation, mapping it to 1-dimensional features, and then... Activation is performed using an activation function; S4-8: Using the Sigmoid activation function, constrain the output to the [0,1] interval to obtain the future... Link health prediction value at time Output; The formula is as follows: ； in, and These are the weights of the fully connected layer, obtained through training. The attention vector for the hidden state at each time step.

7. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 1, characterized in that: Step S5 also includes the following steps: S5-1: Construct a training sample set, which includes historical data of tobacco business dual links during failures and deterioration at the state-county and county-collection point levels. S5-2: Define the loss function; The loss function is as follows: ； in, To determine the true health of the sample, For the corresponding predicted health level; The number of samples; S5-3: Configure training parameters; The initial learning rate η = 0.001 is set. To improve training stability, a learning rate decay strategy is set, where the learning rate decays to 0.9 times the original value every 20 epochs. The batch size is set to 64 to balance training efficiency and gradient stability. The maximum number of training epochs is 200. When the validation set loss does not decrease for 10 consecutive epochs, training is stopped and the current optimal model parameters are saved.

8. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 1, characterized in that: Step S6 also includes the following steps: S6-1: Set early warning thresholds based on the degree of impact of tobacco business interruption; Based on the health level classification of each link, set degradation alarm thresholds. When the real-time health status or prediction result is less than the warning threshold When the link degradation alarm is triggered, the system automatically triggers a fault alarm; a fault threshold μ is set, and when the real-time health status or prediction result is less than the fault threshold μ, an emergency fault alarm is triggered. S6-2: Set a tiered alarm strategy; Combining the tobacco network topology hierarchy and the importance of link services, the k value of each link is pre-divided into three intervals; If the k-value of the link that requires a fault or degradation alarm falls within the value range representing the core link of the state-county, a level one alarm will be issued, using a triple notification of "SMS + email + system pop-up", with a response time of ≤5 minutes. If the k-value of the link requiring fault or degradation alarm belongs to the value range of the core link representing the county-collection point, a level two alarm will be issued, using "SMS + system pop-up" notification, with a response time of ≤10 minutes. If the k-value of the link requiring fault or degradation alarms falls within the range representing non-core links, a system pop-up notification will be used, with a response time of ≤30 minutes. S6-3: Set output decision recommendations; Based on the prediction results and current status of the primary and backup links, output differentiated handling suggestions or automated emergency handling: If the main link prediction is good, that is And the backup link is currently healthy, that is It is recommended to guide subsequent core services to temporarily use the backup link for transmission to reduce the bandwidth pressure on the main link; among which Main link future Health prediction value at any given time For backup links Health value at any given time; If the main link is predicted to deteriorate, that is... And the backup link is currently in good or healthy condition, that is If so, it is recommended to immediately redirect core business traffic to the backup link, and at the same time notify the operations and maintenance personnel to check the network environment of the primary and backup links. If the main link is predicted to fail, that is... The backup link is currently in good or healthy condition. If this happens, automatic traffic switching will be triggered immediately, and maintenance personnel will be notified to check the network environment of the primary and backup links. If both primary and backup links are predicted to degrade, that is... ,and If this occurs, an emergency response strategy will be triggered, notifying operations and maintenance personnel to investigate link aggregation nodes, as well as the state / county company's core data center and network operator lines.

9. The real-time monitoring method for dual-link quality assurance in local area networks for tobacco information scenarios according to claim 1, characterized in that: Step S7 also Includes the following: When an alarm is triggered, the alarm trigger time, handling measures, handling results, link recovery, and the actual health status and predicted health status change trends in the 10-second time window before the trigger are recorded to form automatic data collection. At the same time, the fault data is added to the model training set to continuously optimize the prediction accuracy of the LSTM model.

Images