Intelligent SLA default prediction method and system oriented to government and enterprise special line service

Abstract

The invention relates to the technical field of network service quality monitoring, in particular to an intelligent SLA default prediction method and system oriented to government and enterprise private line services, and the method comprises the following steps: collecting government and enterprise private line network KPI time sequence data and SLA contract text data, carrying out sliding window statistical feature extraction and wavelet transform frequency domain feature extraction on the network KPI data, and carrying out SLA default prediction on the network KPI data; performing semantic coding on the contract text data by using a pre-training language model, constructing a dual-channel deep learning model, and training the model based on a weighted cross entropy loss function; the method has the advantages that the network state and the contract terms are dynamically associated through the attention mechanism, and the problem of data islands in a traditional method is solved; space-time-semantic joint modeling: the CNN-LSTM captures the dynamic change of the KPI, the Transform analyzes the legal semantics of the contract, and fine-grained risk modeling is realized; and an interpretable-driven operation and maintenance decision: quantifying a risk contribution degree by an SHAP value, and directly guiding network optimization and contract management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network service quality monitoring, in particular to an intelligent SLA breach prediction method and system for government-enterprise private line services. BACKGROUND

[0002] The current SLA management of government-enterprise private line services mainly relies on threshold-based alarm systems and manual inspection mechanisms. The core principle is to trigger an alarm through a pre-set network performance indicator (such as latency, bandwidth utilization) threshold, or to regularly check the contract terms and actual performance by operation and maintenance personnel.

[0003] In the prior art, KPI monitoring systems based on SNMP protocol and rule engine-driven alarm configuration methods are widely used. For example, the SLA-Trust model can predict the trend of cloud service indicators through Markov chains, but it only relies on single-dimensional network data and fails to effectively associate the penalty rules in the contract text with dynamic network characteristics. In addition, traditional methods generally use single-modal analysis frameworks, such as LSTM-based traffic prediction models or random forest-based device fault classification algorithms, which only process time series data or structured indicators, resulting in a lack of semantic association between key KPIs such as network jitter rate and packet loss rate and contract constraint clauses, making it difficult to actively identify breach risks.

[0004] The limitations of the prior art mainly manifest in three aspects: first, the lag of threshold alarms makes breach events usually detected after they occur, with an average response time delay of more than 4 hours; second, single-modal models lack the ability to integrate heterogeneous data (such as unstructured contract texts and multi-dimensional time series data), and the semantic gap and dimension difference between multi-modal data significantly reduce the prediction accuracy; third, there is a lack of explainability modules, and existing research such as network performance prediction models based on deep learning are mostly "black box" structures, making it difficult to quantify the specific contribution of each risk factor (such as the weight of the compensation clause, network jitter rate) to the breach result, limiting the pertinence of operation and maintenance decisions.

[0005] The root cause of the above problems lies in the fact that the existing technology does not build a dual-modal deep learning framework, which cannot achieve deep integration of network KPI time series features and contract semantic information, and does not introduce attention mechanisms and SHAP values and other tools to solve the problems of heterogeneous data association and model explainability, resulting in SLA management still being in the passive response stage, making it difficult to meet the stringent requirements of government and enterprise customers for service quality. SUMMARY

[0006] The present application aims to provide an intelligent SLA breach prediction method and system for government-enterprise private line services to solve the problems raised in the background art.

[0007] To achieve the above object, the present application provides the following technical scheme: an intelligent SLA violation prediction method for government-enterprise private line service, comprising the following steps:

[0008] Collecting government-enterprise private line network KPI time series data and SLA contract text data, the network KPI data including bandwidth utilization, delay, packet loss rate, network jitter rate, and the contract text data containing service index threshold and violation clause;

[0009] Performing sliding window statistical feature extraction and wavelet transform frequency domain feature extraction on the network KPI data to generate a spatiotemporal fusion feature vector;

[0010] Using a pre-trained language model to perform semantic encoding on the contract text data, and splicing the mixed semantic feature vector with the threshold value field;

[0011] Building a dual-channel deep learning model, wherein the KPI time series processing channel adopts a CNN-LSTM network to extract dynamic time series features, the contract text processing channel adopts a Transformer network to extract semantic association features, and the dual channels are fused through an attention mechanism to output;

[0012] Training the model based on a weighted cross-entropy loss function, and using SHAP values to analyze feature contribution, to output the SLA violation probability and key risk factors within the next 6 hours.

[0013] Preferably, the network KPI data collection is realized through a probe device deployed in the government-enterprise private line gateway, the sampling period is 1 minute, and the data preprocessing includes linear interpolation to fill in missing values and 3σ principle to filter abnormal values.

[0014] Preferably, the attention mechanism fusion layer generates a weighted fusion vector by calculating the cross-attention weight of the KPI features and the contract features, and the formula is:

[0015]

[0016] wherein Q is the KPI feature matrix, K and V are the contract feature matrix, d k is the feature dimension.

[0017] Preferably, in the sliding window statistical feature extraction, the window size is 30 minutes and the sliding step is 5 minutes, and the calculated statistics include mean, variance, and trend slope, and the wavelet transform adopts db4 base function for 3-layer decomposition and extraction of high-frequency component energy.

[0018] Preferably, the contract text semantic encoding specifically inputs the SLA clause into the BERT model to generate a sentence vector, which is reduced in dimension through a fully connected layer and spliced with the threshold value field to form a 32-dimensional semantic feature and a 5-dimensional numerical feature mixed vector.

[0019] An intelligent SLA violation prediction method and system for government-enterprise private line services, comprising:

[0020] A data acquisition module for acquiring government-enterprise private line network KPI time series data and SLA contract text data, wherein the network KPI data includes bandwidth utilization, latency, packet loss rate, and network jitter rate, and the contract text data contains service index thresholds and violation clauses.

[0021] A feature extraction module for performing sliding window statistical feature extraction and wavelet transform frequency domain feature extraction on the acquired network KPI data to generate spatiotemporal fusion feature vectors, and using a pre-trained language model to perform semantic encoding on the contract text data and concatenating with the threshold value field to generate hybrid semantic feature vectors.

[0022] A model construction module for constructing a dual-channel deep learning model, wherein the KPI time series processing channel uses a CNN-LSTM network to extract dynamic time series features, the contract text processing channel uses a Transformer network to extract semantic correlation features, and the dual-channel outputs are fused through an attention mechanism.

[0023] A model training and prediction module for training the model based on a weighted cross-entropy loss function and using SHAP values to analyze feature contribution, and outputting SLA violation probabilities and key risk factors within the next 6 hours.

[0024] Preferably, in the data acquisition module, network KPI data acquisition is achieved through probe devices deployed in government-enterprise private line gateways, with a sampling period of 1 minute, and data preprocessing includes linear interpolation for time series to fill missing values and 3σ principle to filter outliers.

[0025] Preferably, in the model construction module, the attention mechanism fusion layer calculates the cross-attention weights of KPI features and contract features to generate a weighted fusion vector, with the calculation formula being:

[0026]

[0027] where Q is the KPI feature matrix, K and V are the contract feature matrices, d k is the feature dimension.

[0028] Preferably, in the feature extraction module, when performing sliding window statistical feature extraction, the window size is 30 minutes and the sliding step is 5 minutes, and the calculated statistics include mean, variance, and trend slope; when performing wavelet transform frequency domain feature extraction, db4 basis function is used for 3-level decomposition and extraction of high-frequency component energy.

[0029] Preferably, in the feature extraction module, the contract text semantic coding is specifically inputting the SLA clause into the BERT model to generate a sentence vector, splicing the threshold value field after dimension reduction by a full connection layer, and forming a mixed vector of 32-dimensional semantic features and 5-dimensional numerical features.

[0030] Compared with the prior art, the present application has the following beneficial effects:

[0031] The intelligent SLA breach prediction method and system for government-enterprise private line services provided by the present application dynamically associate network status and contract clauses through an attention mechanism, solve the data island problem in traditional methods, perform spatio-temporal-semantic joint modeling, capture KPI dynamic changes through CNN-LSTM, analyze contract legal semantics through a Transformer, and realize fine-grained risk modeling, and the operation and maintenance decision is driven by explainability, SHAP values quantify risk contribution, and directly guide network optimization and contract management. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme of the present application clear, complete and the advantages more clear and obvious, the embodiments of the present application are further described in detail below in combination with the drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present application, not all embodiments, and are only used to explain the embodiments of the present application, and do not limit the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0034] Embodiment one, please refer to Figure 1 The present application provides a technical scheme: an intelligent SLA breach prediction method for government-enterprise private line services, comprising the following steps:

[0035] Collecting government-enterprise private line network KPI time series data and SLA contract text data, the network KPI data including bandwidth utilization, delay, packet loss rate, network jitter rate, and the contract text data containing service index threshold and breach clause; network KPI data collection is realized by deploying a probe device on a government-enterprise private line gateway, the sampling period is 1 minute, and data preprocessing includes linear interpolation filling of missing values and 3σ principle filtering of abnormal values.

[0036] The network KPI data is subjected to sliding window statistical feature extraction and wavelet transform frequency domain feature extraction to generate a time-space fusion feature vector; in the sliding window statistical feature extraction, the window size is 30 minutes, the sliding step is 5 minutes, and the statistical quantities calculated include mean, variance, and trend slope; the wavelet transform uses db4 base function for 3-layer decomposition and extracts high-frequency component energy.

[0037] The contract text data is subjected to semantic encoding using a pre-trained language model, and is spliced with a threshold value field to generate a hybrid semantic feature vector; the contract text semantic encoding specifically involves inputting the SLA clause into a BERT model to generate a sentence vector, which is reduced in dimension by a fully connected layer and spliced with the threshold value field to form a 32-dimensional semantic feature and a 5-dimensional numerical feature hybrid vector.

[0038] A dual-channel deep learning model is constructed, wherein a CNN-LSTM network is used in the KPI time series processing channel to extract dynamic time series features, a Transformer network is used in the contract text processing channel to extract semantic correlation features, and the dual channels are fused through an attention mechanism to output;

[0039] The model is trained based on a weighted cross-entropy loss function, and SHAP values are used to analyze feature contribution, to output the SLA violation probability within the next 6 hours and key risk factors.

[0040] The attention mechanism fusion layer generates a weighted fusion vector by calculating the cross-attention weight of the KPI features and the contract features, and the formula is:

[0041]

[0042] wherein Q is the KPI feature matrix, K and V are the contract feature matrix, d k is the feature dimension.

[0043] In Embodiment Two, based on Embodiment One, an intelligent SLA violation prediction method and system for government-enterprise special line services includes:

[0044] (1) Data collection and preprocessing module: a) Network KPI data collection. Through the deployment of probe equipment in the government enterprise private line gateway, 12 KPI indicators such as bandwidth utilization, end-to-end delay, packet loss rate, and network jitter rate are collected with a sampling period of 1 minute to form a time series data stream, which needs to cover the key performance indicators defined in the SLA protocol. b) Contract text data analysis. i. Extract the SLA protocol text in PDF / Word format from the operator contract management system; ii. Use regular expression matching to match threshold clauses such as "service availability rate ≥ 99.9%" and "delay ≤ 50ms", and extract the corresponding penalty amount and responsibility attribution fields; iii. Convert unstructured text into structured data table, fields include index name, threshold, monitoring period, penalty coefficient, used to quantify contract constraints, provide semantic features input for model. c) Data preprocessing. i. Missing value processing: use time series linear interpolation method to fill in missing points for network KPI data; ii. Abnormal value filtering: based on the 3σ principle, eliminate data points that exceed the mean value ± 3 times the standard deviation; iii. Text normalization: convert time units such as "working day" and "natural day" in contract clauses to hours to improve data quality and reduce noise interference on model training.

[0045] (2) Feature engineering module: a) Network KPI spatio-temporal feature extraction. i. Sliding window statistics: calculate dynamic features such as delay mean square error and packet loss rate change slope with a window size of 30 minutes and a sliding step of 5 minutes; ii. Wavelet transform: 3-layer db4 wavelet decomposition is performed on the bandwidth utilization sequence, and the high-frequency component energy is extracted as a burst traffic feature. b) Contract semantic feature encoding: i. Input SLA clause text into pre-trained BERT model (bert-base-chinese) to obtain 768-dimensional sentence vector; ii. Map the sentence vector to a 32-dimensional low-dimensional space through a fully connected layer, and concatenate it with the threshold value field (such as "99.9%") to form a mixed feature vector.

[0046] (3) Dual-modal fusion prediction model: a) Model architecture. i. KPI time series processing channel: the input layer receives pre-processed KPI time series data (dimension: 12x30, i.e., 12 indicators within a 30-minute window), the convolution layer uses 3 ID convolution kernels (size 3x1) to extract local time-dependent features; the LSTM layer uses a 64-unit bidirectional LSTM to capture long and short-term time series correlations; ii. Contract text processing channel: the input layer receives a contract semantic mixed feature vector (dimension: 32+5, including 32-dimensional semantic features and 5 numerical threshold values), the Transformer encoding layer uses 4 attention mechanisms, and outputs context-enhanced contract features; iii. Fusion layer: calculate the cross-attention weight of KPI features and contract features, generate a weighted fusion vector, and output the default probability (0-1 interval value). b) Model training: use weighted cross-entropy loss, set a 4 times weight coefficient for default samples (accounting for <5%), use TimeGAN to generate synthetic default samples, balance the positive and negative class data distribution, use AdamW optimizer, set the initial learning rate to 0.001, and decay by 50% every 10 epochs.

[0047] (4) Explainable output module: a) SHAP value analysis. i. Calculate SHAP values for test set samples to quantify the contribution of each feature (such as network jitter rate, contract compensation coefficient) to the prediction result; ii. Generate a feature importance heat map to mark the Top3 risk factors. b) Dynamic threshold adjustment. i. When the model predicts the default probability exceeds the preset threshold (such as 0.85), automatically trigger the resource scheduling strategy: ii. Bandwidth reservation: temporarily allocate additional bandwidth from the shared resource pool to high-risk private lines; iii. Route switching: switch traffic to a backup path through the SDN controller.

[0048] (5) System interaction process: a) Network probe continuously uploads KPI data to the data lake. b) Contract parsing engine incrementally updates clause features daily. c) Model performs batch prediction every hour, outputting default probability for the next 6 hours. d) High-risk private lines trigger alarms and push to the operation and maintenance ticket system. e) Dynamic resource scheduling module completes emergency response within 15 minutes.

[0049] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An intelligent SLA breach prediction method for government and enterprise dedicated line services, characterized by: The following steps are involved: Collect KPI time series data and SLA contract data for government and enterprise dedicated line networks. The network KPI data includes bandwidth utilization, latency, packet loss rate, and network jitter rate. The contract data includes service indicator thresholds and breach of contract clauses. Perform sliding window statistical feature extraction and wavelet transform frequency domain feature extraction on network KPI data to generate spatiotemporal fusion feature vectors; The contract text data is semantically encoded using a pre-trained language model and concatenated with the threshold value field to generate a mixed semantic feature vector; A dual-channel deep learning model was constructed. The KPI time series processing channel used a CNN-LSTM network to extract dynamic time series features, while the contract text processing channel used a Transformer network to extract semantic association features. The dual-channel outputs were then fused through an attention mechanism. The model is trained based on the weighted cross-entropy loss function, and the SHAP value is used to analyze the feature contribution, outputting the SLA default probability and key risk factors within the next 6 hours.

2. The intelligent SLA default prediction method for government and enterprise dedicated line services according to claim 1 is characterized by: Network KPI data collection is achieved through probe devices deployed on government and enterprise dedicated line gateways. The sampling period is 1 minute. Data preprocessing includes time series linear interpolation to fill missing values ​​and 3σ principle to filter outliers.

3. The intelligent SLA default prediction method for government and enterprise dedicated line services according to claim 2, characterized in that: The attention mechanism fusion layer generates a weighted fusion vector by calculating the cross-attention weights of the KPI features and the contract features. The formula is: Where Q is the KPI feature matrix, K and V are the contract feature matrices, d k is the feature dimension.

4. The intelligent SLA breach prediction method for government and enterprise dedicated line services according to claim 3 is characterized by: In the sliding window statistical feature extraction, the window size is 30 minutes, the sliding step is 5 minutes, and the calculated statistics include mean, variance, and trend slope. The wavelet transform uses the db4 basis function to perform a three-layer decomposition and extract the high-frequency component energy.

5. The intelligent SLA breach prediction method for government and enterprise dedicated line services according to claim 4 is characterized by: The semantic encoding of the contract text is specifically to input the SLA terms into the BERT model to generate a sentence vector, which is then reduced in dimension by the fully connected layer and concatenated with the threshold numerical field to form a mixed vector of 32-dimensional semantic features and 5-dimensional numerical features.

6. An intelligent SLA breach prediction method and system for government and enterprise dedicated line services according to claim 5, characterized in that: include: A data collection module is used to collect KPI time series data and SLA contract data for government and enterprise dedicated line networks. The network KPI data includes bandwidth utilization, latency, packet loss rate, and network jitter rate. The contract data includes service indicator thresholds and breach of contract clauses. The feature extraction module is used to perform sliding window statistical feature extraction and wavelet transform frequency domain feature extraction on the collected network KPI data to generate a spatiotemporal fusion feature vector; and to perform semantic encoding on the contract text data using a pre-trained language model and concatenate it with the threshold value field to generate a hybrid semantic feature vector; The model building module is used to build a dual-channel deep learning model. The KPI time series processing channel uses a CNN-LSTM network to extract dynamic time series features, and the contract text processing channel uses a Transformer network to extract semantic association features. The dual-channel outputs are then fused through an attention mechanism. The model training and prediction module is used to train the model based on the weighted cross-entropy loss function, and uses SHAP values ​​to analyze feature contributions, outputting the SLA default probability and key risk factors within the next 6 hours.

7. A system according to claim 6, characterized in that: In the data collection module, network KPI data collection is achieved through probe devices deployed on the government and enterprise dedicated line gateway. The sampling period is 1 minute, and data preprocessing includes time series linear interpolation to fill missing values ​​and 3σ principle to filter outliers.

8. A system according to claim 7, characterized in that: The attention mechanism fusion layer in the model building module generates a weighted fusion vector by calculating the cross-attention weights of the KPI features and the contract features. The calculation formula is: Where Q is the KPI feature matrix, K and V are the contract feature matrices, d k is the feature dimension.

9. A system according to claim 8, characterized in that: In the feature extraction module, when extracting statistical features using a sliding window, the window size is 30 minutes, the sliding step is 5 minutes, and the calculated statistics include mean, variance, and trend slope; when extracting features in the wavelet transform frequency domain, the db4 basis function is used to perform a three-layer decomposition and extract the high-frequency component energy.

10. A system according to claim 9, characterized in that: In the feature extraction module, the semantic encoding of the contract text is specifically to input the SLA terms into the BERT model to generate a sentence vector, which is then reduced in dimension by the fully connected layer and concatenated with the threshold numerical field to form a mixed vector of 32-dimensional semantic features and 5-dimensional numerical features.

Images