Wide fixed network user optical link quality standard analysis and monitoring method

Abstract

The invention belongs to the technical field of data analysis, and particularly relates to a wide fixed network user optical link quality standard analysis and monitoring method. Comprising the following steps: S1.1) collecting optical power data and optical signal spectrum data in an optical link; s2) feature extraction; s3) constructing an One-Class SVM (Support Vector Machine) model, and training the One-Class SVM model by using a large number of feature vector data sets which reach the standard; and S4) the trained One-Class SVM model is deployed in a monitoring system. A One-Class SVM model is adopted, and the advantages of the One-Class SVM model in single-class data anomaly detection are fully utilized; by constructing a boundary hyperplane in a high-dimensional space, high-sensitivity detection of data deviating from normal distribution is realized. The limitation that a traditional classification algorithm needs a large number of positive and negative sample labeling is avoided, and the difficult problem of optical link signal data labeling is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data analysis, and in particular, relates to a method for analyzing and monitoring the quality of broadband fixed network user optical links. Background Art

[0002] With the rapid development of information technology, high-speed fiber-optic communication networks have been widely used around the world. As the core component of modern high-speed fiber-optic communication networks, optical transmission networks (OTNs) play a vital role in ensuring the transmission quality of optical links. However, with the increase in network scale and complexity, the monitoring and quality assessment of optical links face huge challenges.

[0003] Traditional monitoring methods mainly rely on the fiber optic equipment performance monitoring system to provide the operating status and performance data of optical terminals and other fiber optic equipment, and third-party monitoring tools to provide partial network performance data and user feedback through active testing; these methods may lead to high false alarm rates or missed alarm rates in complex environments, and are difficult to adapt to the complexity of changes in optical link quality.

[0004] Another problem is that many optical link quality monitoring technologies rely on a large amount of labeled data to train the model, but in practical applications, the acquisition of labeled data is time-consuming and laborious, especially in the absence of clear negative samples, which poses a great challenge to the construction of anomaly detection models. Conventional classification algorithms usually require both positive and negative samples, but in optical link quality analysis, negative samples are often scarce, further limiting the effectiveness of classification algorithms.

[0005] In order to solve these problems, the present invention proposes an optical link quality compliance analysis and monitoring method based on user experience data, which uses advanced data collection and key feature extraction technology to perform quality compliance analysis and real-time monitoring of optical links to improve network service quality. Summary of the invention

[0006] In order to solve the high false alarm rate or missed alarm rate of traditional methods; and the conventional classification algorithm relies on a large amount of positive and negative sample labeled data to train the model, while in the optical link quality analysis, the scarcity of negative samples limits the technical problem of the classification algorithm, the present invention provides a broadband fixed network user optical link quality compliance analysis and monitoring method.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A method for analyzing and monitoring the quality of broadband fixed network user optical links, comprising the following steps:

[0009] S1) Data collection and preprocessing:

[0010] S1.1) Deploy monitoring equipment to collect optical power data and optical signal spectrum data in the optical link;

[0011] S1.2) Preprocessing the collected data, including data cleaning and denoising;

[0012] Preferably, it also includes: S1.3) detecting and removing abnormal points in the data by statistical methods; the specific process includes:

[0013] The triple standard deviation method is used to detect and eliminate abnormal values ​​in the collected data. The abnormal judgment condition of P(ti) is:

[0014] or

[0015] In the formula, and σ are the mean and standard deviation of the data.

[0016] S2) Feature extraction:

[0017] S2.1) extracting time domain features reflecting the quality of the optical link from the preprocessed optical power data; the time domain features include the mean and variance of the optical power;

[0018] S2.2) extracting frequency domain features reflecting the quality of the optical link from the pre-processed optical signal spectrum data; the frequency domain features include the peak value and fluctuation range of the frequency deviation;

[0019] Preferably, in step S2.1), the process of extracting the time domain features includes:

[0020] Record the optical power data collected in the optical link in a time series format;

[0021] The optical power time series data is divided into time windows of fixed length, expressed as:

[0022] P(t1),P(t2),...,P(t n );

[0023] Where P(t) is the optical power value at timestamp t, t is the corresponding timestamp; n is the time window length;

[0024] For this time window, the following time domain features are calculated:

[0025] The average value of all optical power values ​​in the accumulation window is used to obtain the mean optical power; when the mean optical power is lower than a certain threshold, it indicates signal attenuation or link performance degradation.

[0026] The average of the squares of the differences between all optical power values ​​and the mean within the accumulation window gives the optical power variance. An increase in variance usually indicates that there may be interference or instability in the link, such as insertion loss fluctuations or environmental noise interference.

[0027] Preferably, in step S2.2), the frequency domain feature extraction process includes:

[0028] Perform continuous wavelet transform on spectral data;

[0029] Find the local maximum / minimum of the wavelet coefficients; can locate important peak or valley information in the spectrum.

[0030] According to the scale information of the wavelet coefficients, the position and width of the peak are determined; that is, the wavelet scale corresponding to each local extreme point is determined as a quantitative indicator of the peak width. Combined with time shift, the time position of the peak in the spectrum is obtained.

[0031] Calculate the difference between the frequency corresponding to the peak position and the center frequency to obtain the frequency deviation peak value;

[0032] Taking the peak position as the center, divide the spectrum sequence data into time windows of length n;

[0033] For the spectrum data at each time point in the window, the frequency deviation value is calculated;

[0034] The difference between the maximum and minimum frequency deviations in the window is calculated to obtain the frequency deviation fluctuation range.

[0035] Preferably, the process further includes step S2.3): using the mutual information method to sort the dependencies of various features, and selecting features with strong dependencies to form a feature vector data set; the specific process includes:

[0036] Obtain a data set containing the target variable and all the features that need to be evaluated; the data set can come from experimental acquisition, historical data reading, or other sources. The target variable can be the signal-to-noise ratio (SNR) or the Q factor. Both can reflect the quality of the optical signal in the optical link as a whole.

[0037] Calculate each feature x1,x2,...,x i With the target variable y i The mutual information value of ; mutual information measures the dependency between two variables, and the larger the value, the stronger the correlation between the feature and the target variable.

[0038] The calculation formula of mutual information value is:

[0039]

[0040] In the formula, p(x i ,y i ) is the joint probability distribution; p(xi )p(y i ) is the marginal probability distribution;

[0041] Sort the features from high to low according to the mutual information value. The larger the value, the stronger the dependence of the feature on the target variable.

[0042] According to the sorting results, the first k features are taken as the constituent feature vector data set.

[0043] Preferably, the method further comprises S2.4): normalizing or standardizing the extracted characteristic values;

[0044] S3) Build and train the One-Class SVM model:

[0045] S3.1) construct and initialize a One-Class SVM model, wherein a Gaussian kernel (RBF) or a polynomial kernel is selected as a kernel function; kernel function parameters and an initial parameter ν value are set;

[0046] S3.2) Use a large amount of feature vector data set that has reached the standard to train the One-Class SVM model;

[0047] S3.3) Use known positive and abnormal sample data to evaluate model performance. The evaluation indicators of model performance include accuracy, recall and F1-score;

[0048] Preferably, the method further comprises step S3.4): adjusting the optimal kernel parameters and ν value using a grid search or Bayesian optimization method;

[0049] S4) Model deployment:

[0050] S4.1) deploying the trained One-Class SVM model into the monitoring system;

[0051] S4.2) The monitoring system inputs the new data into the One-Class SVM model for anomaly detection; if the model prediction result is abnormal, an alarm message is generated and sent to the operation and maintenance personnel.

[0052] Preferably, it also includes S5) model updating:

[0053] S5.1) Use the KL divergence method to calculate the distribution change of new and old input data to determine whether to update the model;

[0054] S5.2) Adopt incremental learning method to learn new data distribution rules while retaining the original model data.

[0055] Preferably, the specific process of step S5.1) includes:

[0056] Select a key feature and sample N data points of the key feature from the old data Do and the new data De respectively;

[0057] Using the histogram method, the data is divided into several intervals and the frequency in each interval is calculated as the probability;

[0058] Calculate the KL divergence between the new and old data distributions; the calculation formula for KL divergence is:

[0059]

[0060] In the formula, x i is a data point; P(x i ) is the probability of the corresponding point of the old data; Q(x i ) is the probability of the corresponding point of the new data;

[0061] Set a threshold; if the KL divergence is higher than the threshold, update the model.

[0062] Preferably, the specific process of step S5.2) includes:

[0063] Setting a data buffer pool, including new data and old data of the optical link;

[0064] A certain proportion of new and old data is sampled from the data buffer pool, where the ratio of new data to old data is 7-6:3-4;

[0065] Retrain the One-Class SVM model using the sampled subset and adjust the support vector and hyperplane boundary of the model.

[0066] Beneficial effects of the present invention:

[0067] 1. Through the feature extraction step, it covers multi-angle data such as time domain and frequency domain; and through mutual information sorting, it selects features with strong dependence, avoiding the interference of redundant features and improving the accuracy and processing efficiency of data analysis.

[0068] 2. The One-Class SVM model is used to fully utilize its advantages in detecting anomalies in single-class data; by constructing a boundary hyperplane in high-dimensional space, highly sensitive detection of data that deviates from normal distribution is achieved. This avoids the limitation of traditional classification algorithms that require a large number of positive and negative sample annotations, and solves the difficulty of data annotation in the field of optical link signal analysis.

[0069] 3. The incremental learning method enables the One-Class SVM model to learn new data patterns while retaining the original model knowledge, and adapt to the optical link quality characteristics that change over time; thereby improving the model's adaptability to equipment aging and environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0071] Figure 1 A flowchart of the steps of a method for analyzing and monitoring the quality of broadband fixed network user optical links according to the present invention;

[0072] Figure 2 This is a flow chart of the steps of updating the model in S5) in a method for analyzing and monitoring the quality of broadband fixed-line user optical links of the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] See also Figure 1-Figure 2 As shown, a method for analyzing and monitoring the quality of broadband fixed network user optical links, comprising the following steps:

[0075] S1) Data collection and preprocessing:

[0076] S1.1) Deploy monitoring equipment to collect optical power data and optical signal spectrum data in the optical link;

[0077] S1.2) Preprocessing the collected data, including data cleaning and denoising;

[0078] Specifically, in the entire optical link quality compliance analysis and monitoring method, step S1 is the basic work, and its core purpose is to provide high-quality and clean input data for subsequent feature extraction and model training. This process mainly includes three key parts: data acquisition, preprocessing, and anomaly detection. The optical power directly reflects the transmission quality of the optical link, and the spectrum data reflects the modulation characteristics, signal integrity, and interference characteristics of the signal.

[0079] First of all, data acquisition is the starting point of the solution, which is responsible for obtaining optical power data and optical signal spectrum data that reflect the link status from the optical link in real time.

[0080] Optical power data: reflects the transmission loss of the optical link and is mainly used to characterize time domain characteristics.

[0081] Optical signal spectrum data: used to analyze the frequency domain characteristics of the optical link, such as noise and signal distortion. Collecting this data through monitoring equipment (such as optical power meters, optical spectrum analyzers, etc.) can fully cover the operating status of the optical link and provide a basis for subsequent analysis.

[0082] In the specific implementation process, monitoring equipment can be deployed at key nodes of the optical link (such as the input and output ends) to collect the operation data of the optical signal in real time. The collected raw data is then marked with timestamps and collection locations and stored in a database for subsequent analysis.

[0083] Secondly, the raw data collected in the optical link may contain noise, missing values ​​or other abnormal interference, which directly affects the subsequent feature extraction and model training. Therefore, data cleaning, denoising and abnormal elimination are required to ensure data quality.

[0084] The process of data cleaning is to check the integrity of the data, fill in missing values, or delete abnormal records according to business rules. The process of denoising is to remove random noise through filtering techniques (such as mean filtering and Kalman filtering). For example: applying a low-pass filter to a spectrum signal to remove high-frequency noise.

[0085] Furthermore, it also includes: S1.3) detecting and removing abnormal points in the data by statistical methods; the specific process includes:

[0086] The triple standard deviation method is used to detect and eliminate abnormal values ​​in the collected data. The abnormal judgment condition of P(ti) is:

[0087] or

[0088] In the formula, and σ are the mean and standard deviation of the data.

[0089] In actual scenarios, the operation status of optical links is complex, and abnormal points may come from short-term interference or equipment failure. By detecting and eliminating abnormal points through statistical analysis, the accuracy of data can be further improved. Eliminating noise and abnormal points ensures the accuracy of subsequent feature extraction, while reducing the interference of abnormal points on the One-Class SVM model. It can significantly improve the reliability and accuracy of subsequent feature extraction and model training.

[0090] S2) Feature extraction:

[0091] S2.1) extracting time domain features reflecting the quality of the optical link from the preprocessed optical power data; the time domain features include the mean and variance of the optical power;

[0092] S2.2) extracting frequency domain features reflecting the quality of the optical link from the pre-processed optical signal spectrum data; the frequency domain features include the peak value and fluctuation range of the frequency deviation;

[0093] Specifically, the main goal of step S2) is to extract significant features reflecting the operating status of the optical link from the collected and pre-processed raw data, so as to provide a basis for subsequent quality analysis and anomaly detection.

[0094] By extracting features, data representation can be simplified and key variables can be highlighted. The extracted features need to be able to fully characterize the operating status of the optical link, including normal status and potential abnormal status. The present invention selects statistical features related to the characteristics of the optical signal: time domain features and frequency domain features, targeting the characteristics of the optical link (such as loss, noise, distortion, etc.). Among them, the time domain features include the mean and variance of the optical power;

[0095] Mean optical power: reflects the overall optical power level of the optical link.

[0096] Optical power variance: reflects the stability of optical power fluctuations in an optical link.

[0097] Skewness: Determines the symmetry of the optical power distribution.

[0098] Kurtosis: Describes the sharpness of the optical power distribution.

[0099] Standard Deviation: Describes the fluctuation of optical power.

[0100] Frequency domain characteristics include the peak value and fluctuation range of frequency deviation;

[0101] Frequency deviation peak value: refers to the maximum deviation of the signal main frequency position relative to the theoretical center frequency. It reflects whether the optical signal remains near the designed frequency. Excessive frequency deviation may indicate that there is interference in the link or that the laser performance is degraded.

[0102] Frequency deviation fluctuation range: refers to the width of the frequency interval where the signal strength changes significantly. It reflects the signal stability. The larger the fluctuation range, the more likely it is that the signal spectrum is affected by noise or interference.

[0103] Signal-to-noise ratio: refers to the ratio of signal power to noise power. It can fully reflect the transmission quality of the signal in the link; a low SNR usually indicates that the link is severely interfered or the equipment is aging.

[0104] Spectral width: refers to the bandwidth where the signal spectrum power density reaches a certain threshold (such as -3dB). A wider spectrum width may indicate that the signal is distorted or poorly modulated.

[0105] Phase noise: refers to the jitter or random deviation of the signal spectrum near the center frequency; it directly affects the demodulation performance of the phase modulated signal in the optical link.

[0106] By combining the extraction methods of time domain features and frequency domain features, the variables most sensitive to the operating status of the optical link can be mined from the collected raw data. Step S2 not only improves the representativeness of the feature data, but also provides an accurate input basis for subsequent anomaly detection and model analysis.

[0107] Further, in step S2.1), the extraction process of the time domain features includes:

[0108] Record the optical power data collected in the optical link in a time series format;

[0109] The optical power time series data is divided into time windows of fixed length, expressed as:

[0110] P(t1),P(t2),...,P(t n );

[0111] Where P(t) is the optical power value at timestamp t, t is the corresponding timestamp; n is the time window length; to facilitate batch processing and feature calculation, the time series data is divided into windows of fixed length, and the length of each window is n. It can be determined according to the specific application scenario: a shorter window is suitable for capturing fast-changing signal dynamics (such as high-frequency noise). A longer window is helpful for analyzing global trends or slowly changing characteristics.

[0112] For this time window, the following time domain features are calculated:

[0113] The average value of all optical power values ​​in the accumulation window is used to obtain the mean optical power; when the mean optical power is lower than a certain threshold, it indicates signal attenuation or link performance degradation.

[0114] The average of the squares of the differences between all optical power values ​​and the mean within the accumulation window gives the optical power variance. An increase in variance usually indicates that there may be interference or instability in the link, such as insertion loss fluctuations or environmental noise interference.

[0115] Further, in step S2.2), the frequency domain feature extraction process includes:

[0116] Perform continuous wavelet transform on the spectrum data; through continuous wavelet transform (CWT), the temporal distribution of different frequency components in the spectrum can be extracted, and the local characteristics of the signal, such as mutation points, peaks, etc., can be revealed. Select appropriate wavelet basis functions (such as Morlet wavelet or Mexican Hat wavelet) to capture the local characteristics of the spectrum in the time-frequency domain.

[0117] Find the local maximum / minimum of the wavelet coefficients; can locate important peaks or valleys in the spectrum. These points usually correspond to mutations, noise or other features in the signal.

[0118] According to the scale information of the wavelet coefficients, the position and width of the peak are determined; that is, the wavelet scale corresponding to each local extreme point is determined as a quantitative indicator of the peak width. Combined with time shift, the time position of the peak in the spectrum is obtained.

[0119] Calculate the difference between the frequency corresponding to the peak position and the center frequency to obtain the frequency deviation peak value;

[0120] The spectrum sequence data is divided into time windows of length n with the peak position as the center; the purpose is to extract data fragments near the peak for in-depth analysis of the change characteristics of the frequency deviation value over time.

[0121] For the spectrum data at each time point in the window, the frequency deviation value is calculated;

[0122] The difference between the maximum and minimum frequency deviations in the window is calculated to obtain the frequency deviation fluctuation range.

[0123] It is worth noting that step S2.2) captures the local characteristics of the spectrum data through continuous wavelet transform, and extracts the peak position, frequency deviation value and fluctuation range of the spectrum by combining the extreme value detection and scale information analysis of the wavelet coefficients. This process realizes the accurate analysis and calculation of the operation status of the optical link, and provides key feature data for subsequent quality monitoring and anomaly detection.

[0124] Furthermore, the method further includes step S2.3): using the mutual information method to sort the dependencies of various features, and selecting features with strong dependencies to form a feature vector data set; the specific process includes:

[0125] Obtain a data set containing the target variable and all the features that need to be evaluated; the data set can come from experimental acquisition, historical data reading, or other sources. The target variable can be the signal-to-noise ratio (SNR) or the Q factor. Both can reflect the quality of the optical signal in the optical link as a whole.

[0126] Calculate each feature x1,x2,...,x i With the target variable y i The mutual information value of ; mutual information measures the dependency between two variables, and the larger the value, the stronger the correlation between the feature and the target variable.

[0127] The calculation formula of mutual information value is:

[0128]

[0129] In the formula, p(x i ,y) is the joint probability distribution; p(x i )p(y) is the marginal probability distribution;

[0130] Sort the features from high to low according to the mutual information value. The larger the value, the stronger the dependence of the feature on the target variable.

[0131] According to the sorting results, the first k features are taken as the constituent feature vector data set.

[0132] Furthermore, it also includes S2.4) normalizing or standardizing the extracted various eigenvalues ​​to form a characteristic vector data set;

[0133] Specifically, the mutual information value between each feature and the target variable can be automatically calculated using statistical tools or Python libraries such as sklearn.feature_selection or mutual_info_regression.

[0134] The mutual information method provides a strong basis for feature selection by measuring the nonlinear dependency between features and target variables. Its advantage is that it does not make strong assumptions about data distribution and is particularly suitable for measuring the nonlinear dependency between features and target variables. Using the selected highly correlated features to train the model can improve the generalization of the model.

[0135] S3) Build and train the One-Class SVM model:

[0136] S3.1) construct and initialize a One-Class SVM model, wherein a Gaussian kernel (RBF) or a polynomial kernel is selected as a kernel function; kernel function parameters and an initial parameter ν value are set;

[0137] S3.2) Use a large amount of feature vector data set that has reached the standard to train the One-Class SVM model;

[0138] S3.3) Use known positive and abnormal sample data to evaluate model performance. The evaluation indicators of model performance include accuracy, recall and F1-score;

[0139] Specifically, the present invention adopts a One-Class SVM (single-class support vector machine) model, which is designed to process a single-class data distribution and identify samples that are abnormal or deviate from the normal distribution. This method is particularly suitable for unsupervised learning scenarios or situations where there is only normal data (positive class) but no abnormal data (negative class).

[0140] In the specific implementation process, the One-Class SVM model can be constructed using existing machine learning frameworks (such as Scikit-learn, libsvm or other open source libraries). One-Class SVM maps data to a high-dimensional feature space by introducing a Gaussian kernel function K to find a hyperplane that maximizes the distance between the boundary and the origin. The One-Class SVM model is trained with normal data to learn its feature distribution. The training data is mapped to a high-dimensional feature space using a kernel function. A boundary hyperplane that maximizes the distance from the origin is found to distinguish normal data from potential abnormal data.

[0141] Use the validation set (including known normal data and abnormal data) to make model predictions. Accuracy: the overall prediction accuracy. Recall: the proportion of detected abnormal samples to all true abnormalities. F1-score: evaluates the overall performance of the model.

[0142] Further, the method further comprises step S3.4): using a grid search or Bayesian optimization method to adjust the optimal kernel parameters and ν value;

[0143] Specifically, it is crucial to choose a suitable parameter ν: The choice of the ν value controls the sensitivity of the model to abnormal data. It is recommended to start with a smaller value (for example, 0.01) and then gradually adjust it according to the results of cross-validation. The role of the construction and training process of the One-ClassSVM model in the present invention is to utilize the single-class data distribution, capture the characteristics of normal samples, and identify abnormal samples that deviate from the distribution. Its implementation method involves key steps such as data preparation, model initialization, parameter setting, training and optimization, and each step needs to be adjusted according to the specific scenario to achieve the best effect.

[0144] S4) Model deployment:

[0145] S4.1) deploying the trained One-Class SVM model into the monitoring system;

[0146] S4.2) The monitoring system inputs the new data into the One-Class SVM model for anomaly detection; if the model prediction result is abnormal, an alarm message is generated and sent to the operation and maintenance personnel.

[0147] Specifically, the present invention is used for optical link signal anomaly monitoring, and the vibration signal data generated by the normally operating equipment is used as training data. By training the One-Class SVM model, the abnormal signals in the data can be monitored, and the warning of whether the optical link signal quality meets the standard and the alarm of sudden optical link failure can be given.

[0148] Furthermore, it also includes S5) model update:

[0149] S5.1) Use the KL divergence method to calculate the distribution change of new and old input data to determine whether to update the model;

[0150] S5.2) Adopt incremental learning method to learn new data distribution rules while retaining the original model data.

[0151] Further, the specific process of step S5.1) includes:

[0152] Select a key feature and sample N data points of the key feature from the old data Do and the new data De respectively;

[0153] Using the histogram method, the data is divided into several intervals and the frequency in each interval is calculated as the probability;

[0154] Calculate the KL divergence between the new and old data distributions; the calculation formula for KL divergence is:

[0155]

[0156] In the formula, x i is a data point; P(x i ) is the probability of the corresponding point of the old data; Q(x i ) is the probability of the corresponding point of the new data;

[0157] Set a threshold; if the KL divergence is higher than the threshold, update the model.

[0158] Specifically, distribution deviation indicators (such as KL divergence) are used to monitor the distribution changes of new and old input data during the monitoring process to determine whether further updates are needed. KL divergence is a method to measure the difference between the two probability distributions of the new and old data sets. A low KL value indicates that the new and old distributions are similar and there is no need to update the model. A significant increase in the KL value indicates that new optical link modes or anomalies may appear, triggering a model update. This method can efficiently monitor data distribution changes and ensure that the One-Class SVM model is highly adaptable and updated in a timely manner.

[0159] Further, the specific process of step S5.2) includes:

[0160] Setting a data buffer pool, including new data and old data of the optical link;

[0161] A certain proportion of new and old data is sampled from the data buffer pool, where the ratio of new data to old data is 7-6:3-4;

[0162] Retrain the One-Class SVM model using the sampled subset and adjust the support vector and hyperplane boundary of the model.

[0163] Specifically, the incremental learning method enables the One-Class SVM model to learn new data patterns while retaining the original model knowledge, adapting to the optical link quality characteristics that change over time, and avoiding overfitting and forgetting old patterns.

[0164] The present invention proposes a method for analyzing and monitoring the quality of broadband fixed-line users' optical links; a method for analyzing and monitoring the quality of optical links based on feature extraction, mutual information sorting, One-Class SVM model and incremental learning. This method comprehensively extracts data features in multiple dimensions such as time domain and frequency domain, and uses the mutual information method to sort the features by their dependency strength, thereby eliminating redundant features and improving the accuracy and processing efficiency of data analysis. In addition, as an anomaly detection method for a single type of data, the One-Class SVM model can construct a boundary hyperplane in a high-dimensional space, perform sensitive detection on data that deviates from the normal distribution, and avoid the reliance of traditional classification algorithms on a large amount of labeled data. It is particularly suitable for the unlabeled problem in optical link signal analysis.

[0165] Furthermore, the present invention also adopts an incremental learning method, so that the One-Class SVM model can continuously learn new data patterns on the basis of retaining the original knowledge, and adapt to the dynamic characteristics of the optical link quality changing over time. This method can not only effectively cope with the impact of equipment aging and environmental changes, but also continuously improve the adaptability of the model in a complex network environment, thereby achieving efficient monitoring and intelligent analysis of the optical link quality.

[0166] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A method for analyzing and monitoring the quality of broadband fixed-line user optical links, characterized in that: The following steps are involved: S1.1) Deploy monitoring equipment to collect optical power data and optical signal spectrum data in the optical link; S1.2) Preprocessing the collected data, including data cleaning and denoising; S2.1) extracting time domain features reflecting the quality of the optical link from the preprocessed optical power data; The time domain features include the mean and variance of the optical power; S2.2) extracting frequency domain features reflecting the quality of the optical link from the preprocessed optical signal spectrum data; Frequency domain characteristics include the peak value and fluctuation range of frequency deviation; S2.4) normalizing or standardizing the extracted eigenvalues ​​to form a characteristic vector data set; S3) building a One-Class SVM model and using a large number of qualified feature vector data sets to train the One-Class SVM model; S4) Deploy the trained One-Class SVM model to the monitoring system; the monitoring system inputs new data into the One-Class SVM model for anomaly detection.

2. According to claim 1, a broadband fixed network user optical link quality compliance analysis and monitoring method is characterized by: Also includes: S1.3) Detect and remove outliers in the data using statistical methods; The specific process includes: The triple standard deviation method is used to detect and eliminate abnormal values ​​in the collected data. The abnormal judgment condition of P(ti) is: P(ti)>P+3σ or P(ti)<P-3σ; Where P and σ are the mean and standard deviation of the data.

3. According to claim 1, a broadband fixed network user optical link quality compliance analysis and monitoring method is characterized by: In step S2.1), the extraction process of the time domain features includes: Record the optical power data collected in the optical link in a time series format; The optical power time series data is divided into time windows of fixed length, expressed as: P(t1),P(t2),...,P(t n ); Where P(t) is the optical power value at timestamp t, t is the corresponding timestamp; n is the time window length; For this time window, the following time domain features are calculated: The average value of all optical power values ​​in the accumulation window is obtained to obtain the mean optical power; The average of the squares of the differences between all optical power values ​​in the accumulation window and the mean is used to obtain the optical power variance.

4. According to claim 1, a broadband fixed network user optical link quality compliance analysis and monitoring method is characterized by: In step S2.2), the frequency domain feature extraction process includes: Perform continuous wavelet transform on spectral data; Find the local maximum of the wavelet coefficients; According to the scale information of the wavelet coefficients, the position and width of the peak are determined; Calculate the difference between the frequency corresponding to the peak position and the center frequency to obtain the frequency deviation peak value; Taking the peak position as the center, divide the spectrum sequence data into time windows of length n; For the spectrum data at each time point in the window, the frequency deviation value is calculated; The difference between the maximum and minimum frequency deviations in the window is calculated to obtain the frequency deviation fluctuation range.

5. The method for analyzing and monitoring the quality of broadband fixed network user optical links according to claim 1, characterized in that: The step S2.3) is also included: using the mutual information method to sort the dependencies of various features, and selecting features with strong dependencies to form a feature vector data set; the specific process includes: Get a dataset containing the target variable and all the features that need to be evaluated; Calculate each feature x1,x2,...,x i With the target variable y i The mutual information value of Sort the features from high to low according to the mutual information value. The larger the value, the stronger the dependence of the feature on the target variable. According to the sorting results, the first k features are taken as the constituent feature vector data set.

6. The method for analyzing and monitoring the quality of broadband fixed network user optical links according to claim 1, characterized in that: The specific process of step S3) includes: S3.1) construct and initialize a One-Class SVM model, wherein a Gaussian kernel or a polynomial kernel is selected as a kernel function; kernel function parameters and an initial parameter ν value are set; S3.2) Use a large amount of feature vector data set that has reached the standard to train the One-Class SVM model; S3.3) Use known positive samples and abnormal sample data to evaluate model performance. The evaluation indicators of model performance include accuracy, recall rate and F1-score.

7. The method for analyzing and monitoring the quality of broadband fixed network user optical links according to claim 1, characterized in that: Also includes: S5.1) Use the KL divergence method to calculate the distribution change of new and old input data to determine whether to update the model; S5.2) Adopt incremental learning method to learn new data distribution rules while retaining the original model data.

8. The method for analyzing and monitoring the quality of broadband fixed network user optical links according to claim 7, characterized in that: The specific process of step S5.1) includes: Select a key feature and sample N data points of the key feature from the old data and the new data respectively; Using the histogram method, the data is divided into several intervals and the frequency in each interval is calculated as the probability; Calculate the KL divergence between the new and old data distributions; the calculation formula for KL divergence is: In the formula, x i is a data point; P(x i ) is the probability of the corresponding point of the old data; Q(x i ) is the probability of the corresponding point of the new data; Set a threshold; if the KL divergence is higher than the threshold, update the model.

9. The method for analyzing and monitoring the quality of broadband fixed network user optical links according to claim 7, characterized in that: The specific process of step S5.2) includes: Setting a data buffer pool, including new data and old data of the optical link; Sampling a certain proportion of new and old data from the data buffer pool; Retrain the One-Class SVM model using the sampled subset and adjust the support vector and hyperplane boundary of the model.

10. The method for analyzing and monitoring the quality of broadband fixed network user optical links according to claim 9, characterized in that: The sampling ratio of the new data: old data is 7-6:3-4.

Images