A method and system for detecting the quality of transmission of an optical fiber link

Abstract

The application relates to the technical field of intelligent detection, and discloses a kind of optical fiber link transmission quality detection method and system, the method includes collecting optical signal time sequence and obtaining frequency sub-column by adaptive wavelet frequency division reconstruction, after separating high and low frequency components, time domain segmentation and energy distribution analysis are carried out, dynamic change spectrum is generated by adaptive enhancement. From the spectrum, extract high and low frequency energy distribution sequence, generate time-frequency correlation matrix by multi-scale correlation analysis, obtain correlation quantization result by feature extraction and region quantization. If the result is over threshold, analyze the spectral characteristics to generate unstable characteristic components and calculate the root mean square amplitude to obtain the instability measure value, match the fault database to obtain the fault mode, project to the state space to evaluate the link degradation state. If it is a degradation edge state, combine the network topology to locate the fault and generate a link quality detection report. The method can improve the accuracy of fault judgment.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, and in particular to a method and system for detecting the transmission quality of optical fiber links. Background Technology

[0002] Currently, optical fiber communication serves as the core support of modern information networks. The quality of its link transmission directly determines the high-speed, stable, and reliable transmission of data, playing a crucial role in large-scale application scenarios such as optical fiber broadband operation services, 5G, cloud computing, and data center interconnection. However, even minor anomalies in the link can lead to a decline in the performance of the entire network or even network outages, with a wide-ranging impact. Optical signal power data is affected by various factors such as temperature drift and mechanical vibration, exhibiting highly irregular fluctuation characteristics. This poses a challenge to traditional detection methods in capturing early-stage deterioration signals.

[0003] In existing technologies, fiber optic link transmission quality detection methods typically rely on a single power value or simple statistical indicators to determine link status. For example, anomaly alarms are triggered by measuring the average optical power or setting a fixed threshold. This approach is simple to operate and suitable for basic monitoring. However, in complex real-world transmission environments with multiple superimposed interferences, such as the slow drift caused by temperature changes and the rapid fluctuations caused by equipment jitter, existing technologies struggle to capture the subtle fluctuations in signal power over time. Specifically, slow, trend-based drift and rapid, random jitter in the power sequence can mask each other, and existing methods cannot effectively separate these two types of fluctuations. This makes it difficult to accurately distinguish between tolerable disturbances under normal operating conditions and abnormal signals indicating genuine link quality deterioration. Consequently, misjudgments or missed detections are prone to occur in early fault identification. In high-speed application scenarios such as fiber optic broadband operation services, this inaccuracy may delay fault warnings or lead to unnecessary maintenance operations, increasing operational risks.

[0004] In summary, existing technologies suffer from insufficient accuracy in fault diagnosis. Summary of the Invention

[0005] This invention provides a method and system for detecting the transmission quality of optical fiber links, in order to solve the problem of insufficient accuracy in fault diagnosis in the prior art.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for detecting the transmission quality of an optical fiber link, comprising: The optical signal time series is acquired and adaptive wavelet frequency division reconstruction is performed to obtain the frequency-divided subsequence; Low-frequency and high-frequency components are separated from the frequency-divided subsequence, time-domain segmentation is performed on the low-frequency and high-frequency components, energy distribution analysis is performed on the segmented signal data, and dynamic change spectrum is generated by adaptive enhancement processing of the analysis results. Low-frequency and high-frequency energy distribution sequences are extracted from the dynamic spectrum. Multi-scale correlation analysis is performed on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix. By extracting unstable features and quantizing regions from the time-frequency correlation matrix, the correlation quantization result is obtained. If the correlation quantification result exceeds the preset correlation threshold, analyze the spectral distribution characteristics of the dynamic change spectrum, generate unstable characteristic components, calculate the root mean square amplitude of the unstable characteristic components, and obtain the instability metric value. A test sequence is generated based on the instability metric. The test sequence is matched with a preset fault feature database to obtain a fault mode. The fault mode is projected onto the state space, and the link degradation status is evaluated by combining the position change characteristics of the projection point to obtain the link quality evaluation result. If the link quality assessment result is a deterioration edge state, the anomaly type is determined according to the fault mode, the fault location is located by combining network topology analysis, and the anomaly type and the fault location are integrated to generate a link quality detection report.

[0007] Secondly, the present invention provides an optical fiber link transmission quality detection system, comprising: The data acquisition module is used to acquire the time series of optical signals and perform adaptive wavelet frequency division and reconstruction to obtain frequency-divided subsequences; The feature extraction module is used to separate low-frequency and high-frequency components from the frequency subsequence, perform time-domain segmentation on the low-frequency and high-frequency components, perform energy distribution analysis on the segmented signal data, and generate a dynamically changing spectrum by adaptive enhancement processing of the analysis results. The correlation analysis module is used to extract low-frequency and high-frequency energy distribution sequences from the dynamic change spectrum, perform multi-scale correlation analysis on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix, and obtain the correlation quantification result by extracting unstable features and quantizing regions of the time-frequency correlation matrix. The link quantization module is used to analyze the spectral distribution characteristics of the dynamically changing spectrum if the correlation quantization result exceeds a preset correlation threshold, generate unstable characteristic components, calculate the root mean square amplitude of the unstable characteristic components, and obtain an instability metric value. The quality assessment module is used to generate a test sequence based on the instability metric, match the test sequence with a preset fault feature database to obtain a fault mode, project the fault mode onto the state space, and evaluate the link degradation state by combining the position change characteristics of the projection point to obtain the link quality assessment result. The result output module is used to determine the anomaly type based on the fault mode if the link quality assessment result is a deterioration edge state, locate the fault location by combining network topology analysis, and generate a link quality detection report by integrating the anomaly type and the fault location.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains frequency sub-sequences by acquiring optical signal time series and performing adaptive wavelet frequency division reconstruction; this process can dynamically adjust the decomposition scale according to the characteristics of the signal itself, effectively separate the original complex power fluctuations into sub-sequences of different frequency bands, thereby initially separating the long-term trend components and short-term jitter components, laying a clear data foundation for subsequent independent analysis, fundamentally improving the pertinence of fault feature extraction, and solving the problem of initial feature ambiguity caused by the mixing of original signals.

[0009] (2) The present invention separates low-frequency and high-frequency components from the frequency subsequence, constructs feature matrices and performs signal-to-noise ratio analysis based on the low-frequency and high-frequency components, and determines the low-frequency trend trajectory and high-frequency jitter range. By quantifying the trend trajectory and random fluctuation range of the signal respectively and evaluating its signal-to-noise characteristics, the system can more accurately characterize the steady-state and transient characteristics of the signal, significantly enhance the ability to distinguish between normal operating condition disturbances and early abnormal signals, and directly improve the accuracy of preliminary fault identification.

[0010] (3) The present invention performs time-domain segmentation on the low-frequency and high-frequency components, performs energy distribution analysis on the segmented signal data, and generates a dynamic spectrum through adaptive enhancement processing. The time-domain segmentation focuses on the local evolution of the signal, and the time-frequency characteristics and energy distribution characterization reveal the time-frequency distribution law of the fault characteristics. The adaptive enhancement highlights the potential abnormal mode, so that the spatiotemporal evolution law of the fault characteristics can be clearly and dynamically presented, providing a high-fidelity and high-resolution basis for subsequent deep correlation analysis.

[0011] (4) The present invention extracts low-frequency and high-frequency energy distribution sequences from the dynamic change spectrum, performs multi-scale correlation analysis to generate a time-frequency correlation matrix, and obtains correlation quantification results by extracting unstable features and quantizing regions of the matrix. Multi-scale correlation analysis can capture the cooperative or disconnected change patterns of different frequency components at different time scales, and the regional quantization of unstable features transforms the complex correlation patterns into stable criteria, thereby realizing an objective and quantitative assessment of abnormal changes in the internal structure of the signal, avoiding the randomness of relying on a single threshold or experience judgment, and greatly improving the reliability of fault criteria.

[0012] (5) When the correlation quantification results are abnormal, the present invention analyzes the spectral features to generate an instability metric value, then matches the fault database to obtain the fault mode, evaluates the degradation state through state space projection, and generates a report by combining topology location when it is confirmed to be at the edge of degradation. This series of operations constitutes a complete decision chain from quantitative evaluation, pattern recognition, state judgment to precise location, realizing a systematic and accurate judgment of fault type, development stage and spatial location, providing a reliable technical basis for proactive link maintenance in fiber broadband operation services, and comprehensively solving the shortcomings of insufficient accuracy in fault judgment in the existing technology. Attached Figure Description

[0013] Figure 1 This is a schematic flowchart of a fiber optic link transmission quality detection method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a fiber optic link transmission quality detection system provided in the second embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Reference Figure 1 The first embodiment of the present invention provides a method for detecting the transmission quality of an optical fiber link, comprising the following steps: S11, acquire the optical signal time series and perform adaptive wavelet frequency division reconstruction to obtain the frequency division subsequence; S12, low-frequency and high-frequency components are separated from the frequency division sub-sequence, time-domain segmentation is performed on the low-frequency and high-frequency components, energy distribution analysis is performed on the segmented signal data, and dynamic change spectrum is generated by adaptive enhancement processing of the analysis results. S13, extract low-frequency and high-frequency energy distribution sequences from the dynamic change spectrum, perform multi-scale correlation analysis on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix, and obtain the correlation quantization result by extracting unstable features and quantizing regions from the time-frequency correlation matrix; S14. If the correlation quantification result exceeds the preset correlation threshold, analyze the spectral distribution characteristics of the dynamic change spectrum, generate unstable characteristic components, calculate the root mean square amplitude of the unstable characteristic components, and obtain the instability metric value. S15, generate a test sequence based on the instability metric, match the test sequence with a preset fault feature database to obtain a fault mode, project the fault mode onto the state space, and evaluate the link degradation state by combining the position change characteristics of the projection point to obtain the link quality evaluation result. S16, if the link quality assessment result is a deterioration edge state, determine the anomaly type according to the fault mode, locate the fault location by combining network topology analysis, and generate a link quality detection report by integrating the anomaly type and the fault location.

[0016] In step S11, the optical signal time series is acquired and adaptive wavelet frequency division reconstruction is performed to obtain the frequency-divided subsequence, including: The instantaneous optical signal is acquired and discretized to obtain the optical signal time series; Calculate the energy distribution characteristics of the optical signal time series, and select target wavelet basis functions based on the energy distribution characteristics; The number of decomposition levels is determined based on the total number of sampling points in the optical signal time series. Based on the target wavelet basis function and the number of decomposition levels, the optical signal time series is subjected to discrete wavelet transform to obtain the transform coefficients. The transform coefficients are reconstructed using a single branch to obtain the frequency-divided subsequence.

[0017] It should be noted that the optical signal time series is obtained by acquiring instantaneous optical signals using a high-precision optical power meter and then discretizing them. First, the optical power meter captures the instantaneous optical power value in the fiber optic link. This process is based on the photoelectric effect principle, where incident photons are converted into photocurrents by photosensitive elements such as avalanche photodiodes, and then converted into voltage signals by transimpedance amplifiers. Subsequently, the analog voltage signal is quantized, typically using a 12-bit successive approximation analog-to-digital converter. Based on the Nyquist sampling theorem, the continuous analog signal is discretized into equally spaced sampling points in the time domain, and 4096 levels of digital encoding are performed in the amplitude domain to generate a high-fidelity digital sequence. Then, the digital sequence is segmented using a sliding window truncation method, for example, using a Hanning window with a length of 1024 points and an overlap rate of 50% to window and truncate the sequence, finally generating a pre-processed, stable optical signal time series.

[0018] Subsequently, the energy distribution characteristics of the optical signal time series are calculated, and target wavelet basis functions are selected based on these characteristics. The specific implementation of this process is as follows: the autocorrelation function or power spectrum of the optical signal time series is calculated, and its time-domain or frequency-domain statistical characteristics such as waveform symmetry, number of oscillations, and zero-crossing interval are extracted; the characteristics are then matched with the time-frequency characteristics (such as symmetry, vanishing moment, and support length) of each wavelet basis in the pre-stored wavelet basis function library (such as db4, sym4, coif3, haar), and the wavelet basis with the highest matching degree is selected as the target wavelet basis function.

[0019] Next, the number of decomposition levels is determined based on the total number of sampling points in the optical signal time series. A logarithmic calculation method is used. First, the base-2 logarithm of the total number of sampling points is calculated as the theoretical maximum level. Because wavelet decomposition is based on the principle of multi-resolution analysis, its decomposition process is essentially binary. Each level approximately halves the signal length, thereby ensuring that the lowest frequency subsequence still contains at least one sampling point at the maximum level, avoiding excessive decomposition that could lead to data loss. For example, 8192 sampling points correspond to a maximum level of 13. The calculation process is achieved by iteratively dividing by a power of 2, i.e., log2(8192)=13, indicating that theoretically 13 levels of decomposition can be performed. However, the actual selection needs to be combined with the actual bandwidth and noise characteristics of the signal.

[0020] One feasible approach is to determine the upper limit fmax of the frequency range where the main energy of the signal (e.g., above 95%) is concentrated, based on the power spectral density of the signal; and based on the sampling frequency fs and the decomposition level L, the frequency range corresponding to the detail components of the Lth level is approximately [fs / 2]. L+1 ,fs / 2 L By solving for f_max ≥ fs / 2 (L+1) Find the smallest integer L_min, and in conjunction with the computational complexity, select a fixed value (such as 6) in the interval [L_min, 13] as the final level.

[0021] Then, the optical signal time series is subjected to discrete wavelet transform according to the target wavelet basis function and the number of decomposition levels to obtain the transform coefficients. This is achieved through a discrete wavelet transform algorithm, which is based on the principle of multi-resolution analysis. The algorithm constructs a set of low-pass and high-pass filter banks corresponding to the selected wavelet basis to decompose the time series layer by layer. Specifically, the low-pass filter is used to extract the approximate components (low-frequency trends) of the signal, and the high-pass filter is used to extract the detailed components (high-frequency fluctuations) of the signal. The convolution operation captures local features by performing a sliding dot product between the filter kernel and the signal sequence. Then, downsampling is performed, that is, sampling once every other point to halve the amount of data while retaining key frequency information. For example, in a 6-level decomposition, the initial signal first passes through the first level of filtering and downsampling to obtain the cA1 approximate coefficients and cD1 detail coefficients. Then, this process is repeated for the cA1 coefficients to perform the second level of decomposition, and so on until the sixth level, finally obtaining the cA6 approximate coefficients and cD1 to cD6 detail coefficients. These coefficients together constitute the transform coefficients, which fully characterize the multi-scale features of the signal from macroscopic trends to microscopic jitter. Finally, single-branch reconstruction is performed on the transform coefficients to obtain the frequency-divided subsequences, which is accomplished using the single-branch reconstruction technique of inverse discrete wavelet transform. Specifically, for the approximate coefficient (e.g., cA6) and detail coefficient (e.g., cD6 to cD1) sequences obtained from multi-level decomposition, each coefficient sequence is individually upsampled and convolved. Upsampling involves inserting a zero value between every two adjacent samples of the coefficient sequence, thereby doubling the sequence length to restore it to the length of the previous decomposition layer. The convolution operation involves performing a sliding dot product operation between the upsampled sequence and the reconstruction filter corresponding to the selected wavelet basis. The low-pass reconstruction filter is used to process the approximate coefficients to recover the low-frequency signal profile, and the high-pass reconstruction filter is used to process the detail coefficients to recover the high-frequency signal details. For example, in a 6-level decomposition... Within the framework, the cA6 coefficient sequence is first upsampled and convolved with a low-pass reconstruction filter. Simultaneously, the cD6, cD5, and up to cD1 coefficient sequences are upsampled and convolved with a high-pass reconstruction filter. After independent reconstruction, each coefficient sequence generates a set of signal components of the same length as the original signal but carrying information in different frequency bands. These components are arranged from low to high frequency, forming a complete set of frequency-divided subsequences. That is, the cA6 reconstructed subsequence represents the lowest frequency trend component of the signal (such as long-term slow drift), while the cD6 to cD1 reconstructed subsequences correspond to the detailed components from low to high frequency (such as random jitter and transient fluctuations at different time scales). This achieves clear separation and structured representation of the original optical signal time series in the time-frequency domain.

[0022] In step S12, low-frequency and high-frequency components are separated from the frequency-divided subsequence, time-domain segmentation is performed on the low-frequency and high-frequency components, energy distribution analysis is performed on the segmented signal data, and a dynamically changing spectrum is generated by adaptive enhancement processing of the analysis results, including: Based on the frequency division subsequence, calculate the spectral centroid value, and based on the spectral centroid value, separate the low-frequency component and the high-frequency component from the frequency division subsequence; The low-frequency and high-frequency components are processed by frame segmentation, and the instantaneous frequency and instantaneous amplitude of each frame data segment are extracted to construct a time-frequency energy density matrix. The time-frequency energy density matrix is ​​subjected to feature enhancement processing, and the enhanced matrix is ​​mapped to pixel intensity values ​​to generate a dynamic change spectrum; the dynamic change spectrum includes a slow change trend at low frequencies and a rapid jitter fluctuation at high frequencies.

[0023] It should be noted that the calculation of the spectral centroid value based on the frequency-divided subsequence and the separation of low-frequency and high-frequency components based on this value are specifically achieved through power spectral density estimation and spectral centroid calculation. First, the power spectral density of each subsequence is estimated using the periodogram method. The time-domain signal is converted into a frequency-domain representation through a fast Fourier transform, and the squared amplitude of each frequency point is calculated to obtain the power spectral distribution. Subsequently, the spectral centroid is calculated as a weighted average of the frequency distribution, that is, all frequency points are multiplied by their corresponding power values, summed, and then divided by the total power value, thus obtaining the centroid characterizing the frequency. The values ​​are determined based on a preset centroid frequency threshold as the dividing point. Subsequences with centroids below the frequency threshold are classified as low-frequency components, and those above the frequency threshold are classified as high-frequency components. The frequency threshold is determined as follows: a large number of optical signal frequency subsequences under normal operating conditions are collected, and the spectral centroid of each subsequence is calculated to form a centroid distribution. The centroid distribution is divided into two classes using a clustering algorithm (such as K-means). The mean of the cluster centers or the decision boundary value between the two classes can be used as the initial value of the centroid frequency threshold. The threshold is then fine-tuned using a small number of known outlier samples.

[0024] Next, the low-frequency and high-frequency components are processed by framing, and the instantaneous frequency and instantaneous amplitude of each frame data segment are extracted to construct a time-frequency energy density matrix. This step first uses preset feature extraction time window length, such as 512 sampling points and overlap rate, such as 75%, to window and truncate the low-frequency trend drift part and the high-frequency random jitter part respectively. The Hanning window function is used to weight the signal segment to reduce spectral leakage, thereby obtaining a series of continuous and partially overlapping frame data segments. Then, Hilbert transform is performed on each frame data segment. By constructing an analytical signal, the real signal is decomposed into instantaneous amplitude envelope and instantaneous phase components. Then, the instantaneous frequency is calculated as the phase change rate and the instantaneous amplitude is extracted as the amplitude of the signal envelope. Then, combined with the preset time-frequency resolution grid, such as dividing the frequency axis into 128 uniformly spaced bins and the corresponding frame sequence index of the time axis, the square of the instantaneous amplitude of each frame is used as the energy value to fill the corresponding time-frequency grid cell, thereby constructing the initial time-frequency energy density matrix. Subsequently, the time-frequency energy density matrix is ​​subjected to feature enhancement processing, and the enhanced matrix is ​​mapped to pixel intensity values ​​to generate a dynamic variation spectrum. This step uses an adaptive Gaussian kernel smoothing method based on local signal-to-noise ratio to filter the initial matrix. By calculating the signal-to-noise ratio of the area around each grid point, the smoothing intensity of the Gaussian kernel is dynamically adjusted to filter out isolated noise points while retaining the real trend edges and jitter peaks, resulting in an enhanced matrix with more continuous energy distribution and prominent features. Then, the enhanced matrix is ​​logarithmically compressed to expand the dynamic range, and then linearly normalized to map the values ​​to grayscale pixel intensity values ​​from 0 to 255. Finally, a dynamic variation spectrum is generated with the time axis as the horizontal axis and the frequency axis as the vertical axis. The low-frequency trend is represented by a continuous band-shaped area at the bottom, while the high-frequency jitter is presented as an irregular bright spot pattern at the top.

[0025] In step S13, low-frequency and high-frequency energy distribution sequences are extracted from the dynamic change spectrum. Multi-scale correlation analysis is performed on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix. By extracting instability features and quantizing regions from the time-frequency correlation matrix, correlation quantization results are obtained, including: Based on the dynamic change spectrum, extract the low-frequency energy distribution sequence and the high-frequency energy distribution sequence; Perform cross-correlation operation on the low-frequency energy distribution sequence and the high-frequency energy distribution sequence to construct a time-frequency correlation matrix; The time-frequency correlation matrix is ​​transformed according to a preset weighted mapping function to obtain a correlation feature map. The energy gradient amplitude in the region exceeding the preset correlation threshold in the correlation feature map is calculated and weighted and accumulated to obtain the correlation quantification result that characterizes the complex correlation between low-frequency trends and high-frequency jitter.

[0026] It should be noted that when extracting the low-frequency energy distribution sequence and the high-frequency energy distribution sequence based on the dynamic change spectrum, the frequency axis of the spectrum is first divided into two parts: a low-frequency region and a high-frequency region. Based on the signal spectrum feature analysis, the low-frequency component usually represents the slowly changing trend information, while the high-frequency component reflects the detailed features of rapid fluctuations. The column vector corresponding to the time index of each frame is integrated piecewise, and the square values ​​of the pixel intensity in the low-frequency region and the high-frequency region are accumulated respectively, thereby forming two energy sequences that evolve over time. The low-frequency energy sequence reflects the slow change of the overall power of the trend drift, while the high-frequency energy sequence captures the instantaneous activity of random jitter.

[0027] Subsequently, cross-correlation operations were performed on the low-frequency energy distribution sequence and the high-frequency energy distribution sequence to construct a time-frequency correlation matrix. A multi-scale sliding window method was employed, which uses observation windows of different time lengths to capture various correlation patterns between signals, ranging from short-term transients to long-term trends. The selected window lengths were 1 / 16, 1 / 8, and 1 / 4 of the total analysis sequence length. These scales cover key time periods. For example, a 1 / 16 window is suitable for analyzing correlation changes within half an hour caused by brief equipment heating or instantaneous business load; a 1 / 8 window is suitable for coupling effects at medium time scales, such as daily temperature fluctuations; and a 1 / 4 window can reveal long-term correlation evolution caused by slow equipment aging or environmental trends. The sliding window at each scale moved in a step size equal to one-quarter of the window length. To ensure that adjacent analysis windows have a 75% overlap rate, computational redundancy is avoided while maintaining temporal continuity. Within each sliding window, the mean of all sample values ​​for both sequences is calculated, then the difference between each sample value and the mean is calculated, and the sum of the dot products of these difference vectors is used as the covariance. The magnitudes of the difference vectors for the two sequences are then calculated separately. Finally, the covariance is divided by the product of the magnitudes to obtain the Pearson correlation coefficient, which is between -1 and 1. This coefficient quantifies the consistency of the direction of change of low-frequency trend energy and high-frequency jitter energy under the current window and scale. By arranging the correlation coefficient values ​​of all windows according to their analysis scale (as row index) and the time position corresponding to the window center (as column index), a two-dimensional numerical array is constructed, which is the final generated time-frequency correlation matrix.

[0028] Next, the time-frequency correlation matrix is ​​transformed according to a preset weighted mapping function to obtain the correlation feature map. The mapping function adopts an exponential decay form, and its specific implementation process is as follows: First, iterate through each element in the time-frequency correlation matrix, i.e., the correlation coefficient r, take its absolute value, and then use it as input to the function w=exp(-k|r|) for calculation. Here, the parameter k is an adjustable sensitivity factor used to control the sensitivity to the strength of the correlation. The larger the value of k, the stronger the suppression of strong correlations. The value of k can be adaptively set according to the monitoring environment noise level: calculate the standard deviation σ of all |r| values ​​in the time-frequency correlation matrix, and let k=1 / (2σ). 2 This relationship, or similar, allows the weighting function to adaptively distinguish between random correlations caused by noise and persistent correlations caused by faults. Therefore, the new matrix generated after point-by-point mapping by this function is the correlation feature map.

[0029] Finally, the energy gradient magnitude in regions exceeding a preset correlation threshold, such as 0.65, in the correlation feature map is calculated and weighted to obtain the correlation quantification result. First, the bright areas are marked as connected components. Then, the gradient difference between the centers of adjacent bright areas is calculated along the time direction. The area of ​​the bright area is then used as the weight for accumulation. Specifically, the total number of pixels in each connected component is divided by the total number of pixels in all connected components as the area weight. This area weight is multiplied by the corresponding gradient difference and the weighted result of all components is summed to obtain a single quantification index. This index is usually low in the normal aging stage, but rises rapidly when low-frequency drift and high-frequency jitter are significantly coupled, thus characterizing the complex correlation between low-frequency trends and high-frequency jitter.

[0030] The correlation threshold is set based on statistical analysis of correlation feature maps of a large number of historical links under normal operation. First, correlation data of normal links in different time periods are collected to form a sample set, and its statistical distribution characteristics are calculated. The 95th percentile value of the high quantile of the normal sample distribution is taken as the initial threshold. Then, feedback optimization is carried out through actual abnormal cases, and finally, the threshold value that can effectively distinguish between normal fluctuations and abnormal correlations is determined.

[0031] In step S14, if the correlation quantization result exceeds a preset correlation threshold, the spectral distribution characteristics of the dynamic change spectrum are analyzed, unstable characteristic components are generated, and the root mean square amplitude of the unstable characteristic components is calculated to obtain an instability metric, including: If the correlation quantization result exceeds the preset correlation threshold, the frequency domain amplitude data sequence is extracted from the dynamic change spectrum to construct the spectral probability density function; Calculate the spectral entropy value based on the spectral probability density function, and construct an entropy feature vector based on the spectral entropy value; The entropy feature vector is input into a preset classifier, and the spectral entropy value is modulated according to the output feature pattern category and classification confidence to obtain the unstable characteristic component. The root mean square magnitude of the unstable characteristic component is calculated and used as a measure of the overall link instability.

[0032] It should be noted that if the correlation quantization result exceeds the preset correlation threshold, the frequency domain amplitude data sequence is first extracted to construct the spectral probability density function. By extracting frequency domain amplitude data from multiple consecutive frames from the dynamic spectrum, each frame corresponds to a frequency distribution vector. Then, the frequency of occurrence of amplitude in each preset frequency interval is counted. The probability value of each interval is obtained by dividing the frequency by the total number of samples, thereby forming a probability density function describing the signal energy distribution characteristics. The correlation threshold is set based on the statistical analysis of correlation quantization results under historical normal link operation conditions. The distribution characteristics are calculated by collecting a large number of normal data samples, and a high quantile, such as the 95th percentile, is taken as the initial value. The abnormal case feedback is combined to optimize the detection sensitivity and false alarm rate.

[0033] Subsequently, the spectral entropy value is calculated based on the spectral probability density function, and an entropy feature vector is constructed. Specifically, the probability density function is discretized, and the frequency axis is divided into multiple equal-width frequency bands. The probability value of each frequency band is taken as the base 2 logarithm, multiplied by the probability value, and the sum is negative to obtain the spectral entropy value that characterizes the degree of disorder in the spectral distribution. Then, multiple spectral entropy values ​​calculated from different time periods or different frequency sub-bands are combined into a multi-dimensional vector to form the entropy feature vector, which comprehensively reflects the complex characteristics of the signal at multiple scales.

[0034] Next, the entropy feature vector is input into a preset classifier to calculate the similarity between the feature vector and each category template. It should be noted that the classifier can be built based on common classification models such as support vector machine (SVM) classification model and decision tree classification model. In a preferred embodiment, classification is achieved through a support vector machine (SVM) model. The key hyperparameters of this model include: the width coefficient γ of the radial basis function (RBF) kernel function, which is set as the reciprocal of the feature dimension; the regularization parameter C is optimized between 0.1 and 100 through grid search; the penalty factor is adjusted by weighting according to the sample size of each category; and the distance between the input entropy feature vector and the hyperplane defined by the support vector is calculated using the RDF kernel function, and the distance is mapped to a similarity score, with the closer the value is to 1, the better the match. During training, five-fold cross-validation is used to evaluate the model's generalization ability, and a convex quadratic programming problem is solved using a sequential minimum optimization algorithm. Kernel function techniques are used to avoid explicit high-dimensional mapping, and finally, the support vectors and their corresponding Lagrange multipliers are retained to form the decision function.

[0035] In another feasible embodiment, classification is achieved through a decision tree model: the input vector is compared with the threshold of the tree nodes by recursive segmentation rules, and the path matching degree is calculated as the similarity; the algorithm sorts the similarity between the input vector and all category templates, and selects the category corresponding to the highest score as the output.

[0036] It should be noted that the dataset used to train the model was obtained based on the following rules: a set of entropy feature vectors for normal state and various known fault states were obtained from historical data or simulated data, and each vector was labeled with its corresponding state category, which constituted the training dataset.

[0037] While the model outputs the results, a classification confidence score is generated using the softmax function to reflect the reliability of the judgment. Subsequently, the original spectral entropy value is modulated based on this confidence score. For example, the spectral entropy value is multiplied by a confidence weighting factor (e.g., the factor is 1.2 when the confidence score is higher than 0.8 to enhance the anomalous signal), or a nonlinear function such as the sigmoid transform is used to adjust the entropy value amplitude to highlight the contribution of high-confidence anomalous patterns. Finally, the modulated spectral entropy value sequence is aggregated into a single scalar through this weighted modulation, which yields the unstable characteristic component, used to quantify the instability of link fluctuations.

[0038] Finally, the root mean square magnitude of the unstable characteristic component is calculated as the instability measure of the entire link. Specifically, the squares of all data points in the time series of the unstable characteristic component are taken, the arithmetic mean of these squares is calculated, and the square root of the mean is taken to obtain a scalar value reflecting the average energy level of the fluctuation, which is used to quantify the degree of instability of the link.

[0039] In step S15, a test sequence is generated based on the instability metric value. The test sequence is matched with a preset fault feature database to obtain a fault mode. The fault mode is projected onto the state space, and the link degradation state is evaluated based on the position change characteristics of the projection point to obtain the link quality evaluation result, including: The time distribution data of the instability metric is obtained, a test wave pattern sequence is generated, and the test wave pattern sequence is mapped to a preset fault feature database to extract early fault feature patterns. The early fault feature patterns are projected onto a multi-dimensional state space, and the Euclidean offset distance between the projection point and the preset quality deterioration warning interval boundary is calculated. If the Euclidean offset distance is lower than the preset degradation threshold standard, a state evaluation vector is constructed and a link quality evaluation is performed based on the approximation rate of the projection point trajectory to obtain the link quality evaluation result.

[0040] It should be noted that when acquiring the time distribution data of the instability metric and generating the wave pattern sequence to be tested, the time series data of the instability metric within a specific time period is first extracted from the continuous monitoring system. For example, the metric values ​​per minute within 24 hours are collected to form a time series. Then, the series is smoothed and features are extracted. The moving average method is used to reduce noise in the original time series data. The window width is set to 5 data points and a weighted average is performed using the Hanning window function to reduce spectral leakage. For example, the mean, variance, and trend slope within the moving window are calculated. The trend slope is obtained by fitting the linear regression line of the data points within the window using the least squares method. A standardized wave pattern sequence is generated for subsequent matching.

[0041] During mapping and matching, a dynamic time warping algorithm is used to calculate the similarity distance between the test sequence and each template in the database. The template with the smallest distance is selected as the matching result, thereby extracting the corresponding early fault feature patterns. The preset fault feature database collects fluctuation pattern sequence samples containing typical faults such as fiber bending, connector contamination, and transmitter aging through simulation experiments or field recordings. Features such as instability metrics and trend slopes are extracted from all sequences to form a feature vector set. The elbow rule or silhouette coefficient method is used to determine the optimal number of clusters K, and the K-means++ algorithm is applied to initialize the centers for clustering. The center vector of each cluster is a fault feature template, and the fault type corresponding to the template is recorded.

[0042] Next, the early fault feature patterns are projected onto a multi-dimensional state space, and the Euclidean offset distance between the projection point and the preset quality degradation warning interval boundary is calculated. The specific process is as follows: First, a multi-dimensional state space is constructed, consisting of key indicator dimensions of the fault features. Each dimension corresponds to a feature parameter (such as fluctuation amplitude, frequency component energy ratio, etc.). The feature vector representing the early fault feature pattern is located at a specific coordinate point in this space. The quality degradation warning interval is defined as a hyperellipsoidal region in the multi-dimensional state space, centered on the mean vector of historical normal operation data in each dimension, with a semi-axis length equal to a multiple (e.g., 3 times) of the variance in each dimension. The Euclidean offset distance is calculated as the shortest Euclidean distance from the current projection point coordinates to the boundary of this hyperellipsoid.

[0043] The Euclidean offset distance is calculated using the standard vector distance formula, which is to take the square root of the sum of the squares of the differences in each dimension between the coordinates of the projected point and the coordinates of the nearest point on the boundary of the warning interval. This value intuitively reflects the comprehensive degree to which the current state deviates from the normal range.

[0044] If the Euclidean offset distance is lower than the preset deterioration critical point standard, a state evaluation vector is constructed by combining the approximation rate of the projection point trajectory. The specific process is as follows: First, obtain the historical coordinate data of the projection point in the state space in the current and previous monitoring cycles. Fit the trend line of the distance from the projection point to the warning boundary with time using the sliding window method (e.g., taking the 5 most recent time points), and calculate the slope of the trend line. This slope is the approximation rate. When constructing the state evaluation vector, normalize the latest calculated Euclidean offset distance and divide it by the preset critical distance to obtain the relative distance ratio. At the same time, map the approximation rate to the interval of 0 to 1 using the Sigmoid function. These two processed scalars together constitute a two-dimensional state evaluation vector. The deterioration critical point standard is calculated by collecting a large amount of Euclidean offset distance data of projection points under historical normal operating conditions, using statistical distribution analysis to calculate its 95th percentile as the benchmark threshold, and dynamically optimizing and calibrating it in combination with the distribution characteristics of distance mutation points in actual fault cases.

[0045] Finally, the vector is input into a pre-trained classifier (e.g., a support vector machine trained on historical data, which uses labeled normal and abnormal state vector samples during training and optimizes kernel function parameters such as the width coefficient of the radial basis function through grid search and cross-validation to ensure classification accuracy) or compared with a preset multi-dimensional threshold region (this region is automatically divided into risk clusters by collecting historical link state vector samples and applying the K-means clustering algorithm, or by using the 95th percentile of each dimension value as the boundary threshold). If the vector falls into the preset "high-risk region" (e.g., the relative distance ratio is less than 0.3 and the approximation rate after mapping is greater than 0.7), it is determined that the link quality has reached the deterioration edge state; otherwise, it is in a stable or observation state, and the link quality assessment result is output. The multidimensional threshold region is set by collecting state vector samples from historical links under normal and abnormal conditions, based on statistical distribution analysis such as clustering algorithms or percentile methods. Specifically, if a clustering algorithm (such as K-means) is used, the historical state vectors are first standardized, and the optimal number of clusters (usually 3-5) is determined by the elbow rule. The centroid and boundary hypersphere radius of each cluster are calculated, and the boundary of the cluster closest to the centroid and containing abnormal samples is set as the high-risk threshold. If the percentile method is used, the statistical distribution of the feature values ​​of each dimension is calculated, and the 95th percentile of normal samples and the 5th percentile of abnormal samples are taken as the threshold boundaries. The threshold range is dynamically adjusted through confusion matrix testing to optimize the balance between false positive rate and false negative rate.

[0046] In step S16, if the link quality assessment result is a deterioration edge state, the anomaly type is determined according to the fault mode, the fault location is located by combining network topology analysis, and the anomaly type and the fault location are integrated to generate a link quality detection report, including: If the link quality assessment result confirms that the link is in a deterioration edge state, obtain the state assessment vector that confirms the link is in a deterioration edge state, and retrieve the fault feature database to obtain fault matching features. The fault matching features are classified to obtain classification feature codes, which are then mapped to a preset anomaly type mapping table to determine the anomaly type. Based on the anomaly type, the network topology is traversed, the node correlation is calculated, and the fault location is obtained. By aggregating the anomaly type and the fault location, a detailed link quality detection report is generated.

[0047] It should be noted that if the link quality assessment result confirms that the link is on the verge of degradation, the first step is to obtain the state assessment vector confirming the state is on the verge of degradation. This vector typically includes multi-dimensional indicators such as the offset value from the degradation boundary and the approximation rate. In specific implementation, the system directly extracts the calculated state assessment vector from the state assessment result generated in step S16. This vector is based on the coordinate data projected onto the multi-dimensional state space in step S16, and is obtained through processing steps such as Euclidean offset distance calculation and approximation rate analysis, ensuring close connection with the previous judgment step. Subsequently, the fault feature database is retrieved, and a similarity matching algorithm based on dynamic time warping is adopted: first, the state assessment vector is Z-score standardized to eliminate the influence of dimensions, and then the dynamic time warping distance between it and each historical fault template vector in the database is calculated. This algorithm effectively solves the matching problem of inconsistent time series lengths by constructing a cumulative cost matrix and backtracking the optimal path. Finally, the calculated distance value is converted into a similarity score in the range of 0-1, and the template with the highest score is selected as the fault matching feature.

[0048] Next, the fault matching features are classified. This process first inputs the fault matching features, typically a multi-dimensional numerical vector, into a pre-trained classification model. For example, a decision tree classifier is used, employing the C4.5 algorithm to construct a tree structure. Each internal node represents a feature attribute test. The optimal splitting feature is selected based on information gain, and the data is recursively divided into child nodes with the highest purity. The final leaf node stores the fault type probability distribution. Furthermore, the system compares the value of each feature with the threshold of a preset node, recursively traversing downwards from the root node along paths that meet the conditions until a leaf node is reached. During the inference phase, the input vector traverses downwards from the root node along branches that meet the threshold conditions, passing through multiple feature judgment nodes until the final leaf node is reached. The code corresponding to the highest probability in the fault type probability distribution stored at this node is the output result. The category label represented by this leaf node is output as the classification feature code; for example, the code "101" may correspond to a periodic fluctuation fault pattern.

[0049] Subsequently, the system uses the classification feature code as a key to perform a precise query in a preset anomaly type mapping table. This mapping table is essentially a key-value pair database, where the key is the classification feature code and the value is the corresponding anomaly type description string defined by expert knowledge. For example, the code "101" maps to "fiber optic aging" and the code "205" maps to "loose connection point". Through this efficient hash matching, the specific anomaly type can be directly determined from the code.

[0050] Then, based on the anomaly type, the network topology is traversed, and a breadth-first search algorithm is used to visit all nodes: starting from a critical node in the network (such as a gateway or a node with a high incidence of historical failures), a queue is initialized and the starting node is added to the queue and marked as visited; then the queue is processed in a loop, each time a node is taken from the head of the queue, and after visiting the node, all its directly connected and unvisited neighbor nodes are added to the tail of the queue in order and marked as visited. This process is repeated until the queue is empty, ensuring that all nodes are systematically visited, thereby covering the entire topology. During the traversal, the topology node correlation degree is calculated for each visited node. Specifically, the connection weight parameter is first normalized by dividing the current link bandwidth by the total network bandwidth; the historical fault frequency parameter is normalized by dividing the number of faults at that node in the past thirty days by the maximum number of faults within the monitoring period; and the physical distance parameter is normalized by dividing the current fiber length by the longest fiber distance in the network. Then, the three normalized parameters are weighted and synthesized with weights of 0.4, 0.4, and 0.2 respectively. The physical distance parameter is rounded (i.e., 1 minus the normalized value) to reflect the principle that shorter distances indicate higher correlation. The weighting method is as follows: weights are determined using statistical analysis based on historical fault data. Real-world fault cases and their corresponding node feature vectors are collected, and logistic regression or the analytic hierarchy process (AHP) is used to calculate the weight coefficients of each feature's influence on fault occurrence. Finally, all node correlation values ​​are sorted in descending order, and the node with the highest correlation is selected as the potential fault location. If multiple nodes have similar correlations, further filtering is performed based on geographical location priority, thus outputting the specific fault location variable.

[0051] Finally, the system aggregates anomaly types and fault locations to generate a link quality inspection report. The specific process is as follows: First, the system creates a structured report data object, storing the anomaly type string (e.g., "fiber optic aging") and fault location identifier (e.g., "node A - port 3") as core fields. Simultaneously, based on preset risk assessment rules, the system automatically calculates the risk level by combining the severity weight of the anomaly type and the historical correlation score of the fault node. For example, the severity weight is multiplied by the correlation score and mapped to "high risk," "medium risk," or "low risk" levels using a piecewise function. Then, the system queries the knowledge base bound to the anomaly type and risk level to obtain preset targeted maintenance recommendations, such as "It is recommended to clean the connector and test return loss within 48 hours." Next, the report generation engine calls a preset XML or Markdown template and, through variable substitution, automatically fills the corresponding placeholders in the template with information such as the anomaly type description, fault location coordinates, risk level, timestamp, calculated key indicators, and recommended measures from the data object, generating a complete, directly readable, or distributable document—the final link quality inspection report.

[0052] In summary, this invention provides a method for detecting the transmission quality of optical fiber links. It involves acquiring time-series optical signals and reconstructing them using adaptive wavelet frequency division. After separating high- and low-frequency components, time-domain segmentation and energy distribution analysis are performed, followed by adaptive enhancement to generate a dynamically changing spectrum. High- and low-frequency energy distribution sequences are extracted from the spectrum, and multi-scale correlation analysis generates a time-frequency correlation matrix. Feature extraction and region quantization yield correlation quantization results. If the result exceeds a threshold, spectral features are analyzed to generate unstable characteristic components, and the root mean square amplitude is calculated to obtain an instability metric. This metric is then matched against a fault database to obtain fault modes, which are projected onto the state space to assess the link degradation state. If the state is at the edge of degradation, the fault is located by combining network topology information, and a link quality detection report is generated, thus improving the accuracy of fault diagnosis.

[0053] Reference Figure 2 The second embodiment of the present invention provides an optical fiber link transmission quality detection system, comprising: The data acquisition module is used to acquire the time series of optical signals and perform adaptive wavelet frequency division and reconstruction to obtain frequency-divided subsequences; The feature extraction module is used to separate low-frequency and high-frequency components from the frequency subsequence, perform time-domain segmentation on the low-frequency and high-frequency components, perform energy distribution analysis on the segmented signal data, and generate a dynamically changing spectrum by adaptive enhancement processing of the analysis results. The correlation analysis module is used to extract low-frequency and high-frequency energy distribution sequences from the dynamic change spectrum, perform multi-scale correlation analysis on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix, and obtain the correlation quantification result by extracting unstable features and quantizing regions of the time-frequency correlation matrix. The link quantization module is used to analyze the spectral distribution characteristics of the dynamically changing spectrum if the correlation quantization result exceeds a preset correlation threshold, generate unstable characteristic components, calculate the root mean square amplitude of the unstable characteristic components, and obtain an instability metric value. The quality assessment module is used to generate a test sequence based on the instability metric, match the test sequence with a preset fault feature database to obtain a fault mode, project the fault mode onto the state space, and evaluate the link degradation state by combining the position change characteristics of the projection point to obtain the link quality assessment result. The result output module is used to determine the anomaly type based on the fault mode if the link quality assessment result is a deterioration edge state, locate the fault location by combining network topology analysis, and generate a link quality detection report by integrating the anomaly type and the fault location.

[0054] It should be noted that the optical fiber link transmission quality detection system provided in this embodiment of the invention is used to execute all the process steps of the optical fiber link transmission quality detection method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0055] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0056] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for detecting the transmission quality of an optical fiber link, characterized in that, include: The optical signal time series is acquired and adaptive wavelet frequency division reconstruction is performed to obtain the frequency-divided subsequence; Low-frequency and high-frequency components are separated from the frequency-divided subsequence, time-domain segmentation is performed on the low-frequency and high-frequency components, energy distribution analysis is performed on the segmented signal data, and dynamic change spectrum is generated by adaptive enhancement processing of the analysis results. Low-frequency and high-frequency energy distribution sequences are extracted from the dynamic spectrum. Multi-scale correlation analysis is performed on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix. By extracting unstable features and quantizing regions from the time-frequency correlation matrix, the correlation quantization result is obtained. If the correlation quantification result exceeds the preset correlation threshold, analyze the spectral distribution characteristics of the dynamic change spectrum, generate unstable characteristic components, calculate the root mean square amplitude of the unstable characteristic components, and obtain the instability metric value. A test sequence is generated based on the instability metric. The test sequence is matched with a preset fault feature database to obtain a fault mode. The fault mode is projected onto the state space, and the link degradation status is evaluated by combining the position change characteristics of the projection point to obtain the link quality evaluation result. If the link quality assessment result is a deterioration edge state, the anomaly type is determined according to the fault mode, the fault location is located by combining network topology analysis, and the anomaly type and the fault location are integrated to generate a link quality detection report.

2. The fiber optic link transmission quality detection method according to claim 1, characterized in that, The process of acquiring the optical signal time series and performing adaptive wavelet frequency division and reconstruction to obtain the frequency-divided subsequence includes: The instantaneous optical signal is acquired and discretized to obtain the optical signal time series; Calculate the energy distribution characteristics of the optical signal time series, and select target wavelet basis functions based on the energy distribution characteristics; The number of decomposition levels is determined based on the total number of sampling points in the optical signal time series. Based on the target wavelet basis function and the number of decomposition levels, the optical signal time series is subjected to discrete wavelet transform to obtain the transform coefficients. The transform coefficients are reconstructed using a single branch to obtain the frequency-divided subsequence.

3. The fiber optic link transmission quality detection method according to claim 1, characterized in that, The process of separating low-frequency and high-frequency components from the frequency-divided subsequence, performing time-domain segmentation on the low-frequency and high-frequency components, analyzing the energy distribution of the segmented signal data, and generating a dynamically changing spectrum through adaptive enhancement processing of the analysis results includes: Based on the frequency division subsequence, calculate the spectral centroid value, and based on the spectral centroid value, separate the low-frequency component and the high-frequency component from the frequency division subsequence; The low-frequency and high-frequency components are processed by frame segmentation, and the instantaneous frequency and instantaneous amplitude of each frame data segment are extracted to construct a time-frequency energy density matrix. The time-frequency energy density matrix is ​​subjected to feature enhancement processing, and the enhanced matrix is ​​mapped to pixel intensity values ​​to generate a dynamic change spectrum; the dynamic change spectrum includes a slow change trend at low frequencies and a rapid jitter fluctuation at high frequencies.

4. The fiber optic link transmission quality detection method according to claim 1, characterized in that, The process involves extracting low-frequency and high-frequency energy distribution sequences from the dynamically changing spectrum, performing multi-scale correlation analysis on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix, and obtaining correlation quantization results by extracting unstable features and quantizing regions from the time-frequency correlation matrix. This includes: Based on the dynamic change spectrum, extract the low-frequency energy distribution sequence and the high-frequency energy distribution sequence; Perform cross-correlation operation on the low-frequency energy distribution sequence and the high-frequency energy distribution sequence to construct a time-frequency correlation matrix; The time-frequency correlation matrix is ​​transformed according to a preset weighted mapping function to obtain a correlation feature map. The energy gradient amplitude in the region exceeding the preset correlation threshold in the correlation feature map is calculated and weighted and accumulated to obtain the correlation quantification result that characterizes the complex correlation between low-frequency trends and high-frequency jitter.

5. The fiber optic link transmission quality detection method according to claim 1, characterized in that, If the correlation quantization result exceeds a preset correlation threshold, the spectral distribution characteristics of the dynamically changing spectrum are analyzed to generate unstable characteristic components. The root mean square amplitude of the unstable characteristic components is calculated to obtain an instability metric, including: If the correlation quantization result exceeds the preset correlation threshold, the frequency domain amplitude data sequence is extracted from the dynamic change spectrum to construct the spectral probability density function; Calculate the spectral entropy value based on the spectral probability density function, and construct an entropy feature vector based on the spectral entropy value; The entropy feature vector is input into a preset classifier, and the spectral entropy value is modulated according to the output feature pattern category and classification confidence to obtain the unstable characteristic component. The root mean square magnitude of the unstable characteristic component is calculated and used as a measure of the overall link instability.

6. The fiber optic link transmission quality detection method according to claim 1, characterized in that, The process of generating a test sequence based on the instability metric, matching the test sequence with a preset fault feature database to obtain a fault mode, projecting the fault mode onto a state space, and evaluating the link degradation state based on the position change characteristics of the projection points to obtain a link quality assessment result includes: The time distribution data of the instability metric is obtained, a test wave pattern sequence is generated, and the test wave pattern sequence is mapped to a preset fault feature database to extract early fault feature patterns. The early fault feature patterns are projected onto a multi-dimensional state space, and the Euclidean offset distance between the projection point and the preset quality deterioration warning interval boundary is calculated. If the Euclidean offset distance is lower than the preset degradation threshold standard, a state evaluation vector is constructed and a link quality evaluation is performed based on the approximation rate of the projection point trajectory to obtain the link quality evaluation result.

7. The fiber optic link transmission quality detection method according to claim 6, characterized in that, If the link quality assessment result is a deterioration edge state, the anomaly type is determined based on the fault mode, the fault location is located by combining network topology analysis, and the anomaly type and fault location are integrated to generate a link quality detection report, including: If the link quality assessment result confirms that the link is in a deterioration edge state, obtain the state assessment vector that confirms the link is in a deterioration edge state, and retrieve the fault feature database to obtain fault matching features. The fault matching features are classified to obtain classification feature codes, which are then mapped to a preset anomaly type mapping table to determine the anomaly type. Based on the anomaly type, the network topology is traversed, the node correlation is calculated, and the fault location is obtained. By aggregating the anomaly type and the fault location, a detailed link quality detection report is generated.

8. A fiber optic link transmission quality detection system, characterized in that, include: The data acquisition module is used to acquire the time series of optical signals and perform adaptive wavelet frequency division and reconstruction to obtain frequency-divided subsequences; The feature extraction module is used to separate low-frequency and high-frequency components from the frequency subsequence, perform time-domain segmentation on the low-frequency and high-frequency components, perform energy distribution analysis on the segmented signal data, and generate a dynamically changing spectrum by adaptive enhancement processing of the analysis results. The correlation analysis module is used to extract low-frequency and high-frequency energy distribution sequences from the dynamic change spectrum, perform multi-scale correlation analysis on the low-frequency and high-frequency energy distribution sequences to generate a time-frequency correlation matrix, and obtain the correlation quantification result by extracting unstable features and quantizing regions of the time-frequency correlation matrix. The link quantization module is used to analyze the spectral distribution characteristics of the dynamically changing spectrum if the correlation quantization result exceeds a preset correlation threshold, generate unstable characteristic components, calculate the root mean square amplitude of the unstable characteristic components, and obtain an instability metric value. The quality assessment module is used to generate a test sequence based on the instability metric, match the test sequence with a preset fault feature database to obtain a fault mode, project the fault mode onto the state space, and evaluate the link degradation state by combining the position change characteristics of the projection point to obtain the link quality assessment result. The result output module is used to determine the anomaly type based on the fault mode if the link quality assessment result is a deterioration edge state, locate the fault location by combining network topology analysis, and generate a link quality detection report by integrating the anomaly type and the fault location.

Images