A data-driven communication signal anomaly identification method and system

Abstract

The application provides a data-driven communication signal anomaly recognition method, which comprises the following steps: performing adaptive noise reduction and baseline drift correction on an original communication time domain signal to obtain a corrected time domain signal sequence; calculating a channel environment disturbance coefficient and correcting a baseline parameter set to generate a dynamic reference feature sequence; performing adaptive decomposition and correlation calculation on the corrected signal to construct a signal dynamic topology graph; extracting a decoupling anomaly feature vector through feature aggregation and orthogonal constraint, and calculating a deviation vector of the decoupling anomaly feature vector from the dynamic reference; outputting an anomaly type probability distribution through an anomaly classifier; and completing cause-effect path retrieval and root cause confidence calculation in an anomaly cause-effect knowledge graph, and finally outputting an anomaly recognition and root cause diagnosis result. The application can adapt to a complex channel environment, effectively improves the anomaly recognition accuracy and root cause positioning precision, and is suitable for a communication system real-time monitoring scene.

Description

Technical Field

[0001] This invention relates to the field of communication signal monitoring and fault diagnosis technology, and in particular to a data-driven method and system for identifying communication signal anomalies. Background Technology

[0002] With the continuous increase in the coverage and transmission complexity of modern communication networks, communication signals are susceptible to multiple factors during transmission, such as electromagnetic interference, channel fading, and hardware performance degradation, resulting in various anomalies such as amplitude distortion, phase shift, and rhythm disorder, directly affecting the transmission reliability of communication systems. Traditional communication signal anomaly identification often relies on fixed threshold judgments and single feature matching schemes, which are difficult to adapt to dynamically fluctuating channel environments. False alarms and missed alarms are prominent in low signal-to-noise ratio scenarios, and these methods can only identify anomalies but cannot trace their physical root causes. Existing data-driven signal anomaly identification methods often directly classify time-domain or frequency-domain features without fully exploring the topological correlation information between multiple signal components. They suffer from insufficient feature decoupling capabilities and lack causal reasoning mechanisms, making it difficult to accurately locate the root cause from the anomaly's appearance. This fails to meet the application requirements of high-reliability communication systems for rapid anomaly detection and precise fault location. Summary of the Invention

[0003] The purpose of this invention is to provide a data-driven communication signal anomaly identification method and system to improve the accuracy of anomaly identification and root cause localization.

[0004] To achieve the above objectives, the present invention provides the following solution: A data-driven method for identifying communication signal anomalies includes the following steps: The acquired raw communication time-domain signal is subjected to adaptive noise reduction and baseline drift correction to obtain a corrected time-domain signal sequence. The channel environment disturbance coefficient is calculated based on the characteristics of historical normal signals, a baseline parameter set is constructed, and the baseline parameter set is corrected according to the channel environment disturbance coefficient to obtain a dynamic reference feature sequence. Adaptive decomposition and correlation calculation are performed on the corrected time-domain signal sequence to construct a dynamic topology map of the signal; Feature aggregation and orthogonal constraint processing are performed on the dynamic topology graph of the signal to extract the decoupling anomaly feature vector, and the deviation vector between the decoupling anomaly feature vector and the dynamic reference feature sequence is calculated. The probability distribution of the anomaly type is calculated by performing probability mapping on the deviation vector through an anomaly classifier. Based on the probability distribution of anomaly types, causal path retrieval and confidence calculation are performed in the pre-set anomaly causal knowledge graph to obtain the root cause causal confidence. Based on the confidence level of root cause causality, the results of communication signal anomaly identification and root cause diagnosis are obtained by sorting and filtering.

[0005] Optionally, the acquired raw communication time-domain signal is subjected to adaptive noise reduction and baseline drift correction processing to obtain a corrected time-domain signal sequence, including: The original communication time-domain signal is subjected to sliding window framing to obtain multiple signal segments of equal length. Each frame of signal segment is decomposed into waves, and the decomposed wavelet coefficients are thresholded and shrunk before wavelet reconstruction is performed to obtain the denoised signal frame. Polynomial fitting baseline correction is performed on all denoised signal frames to remove the overall baseline drift of the signal, and the corrected time-domain signal sequence is obtained by splicing them together.

[0006] Optionally, channel environment disturbance coefficients are calculated based on historical normal signal characteristics to construct a baseline parameter set. The baseline parameter set is then corrected according to the channel environment disturbance coefficients to obtain a dynamic baseline feature sequence, including: Time-domain statistical features, frequency-domain distribution features, and rhythmic features are extracted from historical normal communication signals and integrated to construct a baseline parameter set. Calculate the signal-to-noise ratio fluctuation and power spectral flatness of the signal within the current sliding window, and perform a weighted summation to obtain the channel environment disturbance coefficient; The characteristic fluctuation range of the baseline parameter set is proportionally corrected based on the channel environment disturbance coefficient to obtain the dynamic benchmark characteristic sequence.

[0007] Optionally, adaptive decomposition and correlation calculation are performed on the corrected time-domain signal sequence to construct a dynamic topology map of the signal, including: Adaptive empirical mode decomposition is performed on the corrected time-domain signal sequence to obtain several intrinsic mode function components; The intrinsic mode function components are transformed to the time-frequency domain by short-time Fourier transform, and the time-frequency coefficient matrix of each component is extracted as graph node features. Calculate the Pearson correlation coefficient and mutual information value between any two graph node features, and then perform a weighted fusion calculation of the Pearson correlation coefficient and mutual information value to obtain the association weight between nodes; An adjacency matrix is ​​constructed based on the association weights between nodes, and a dynamic signal topology graph is constructed based on the graph node characteristics and the adjacency matrix.

[0008] Optionally, feature aggregation and orthogonal constraint processing are performed on the dynamic topology graph of the signal to extract the decoupling anomaly feature vector, and the deviation vector between the decoupling anomaly feature vector and the dynamic reference feature sequence is calculated, including: The attention weight coefficient is calculated based on the feature vectors of the neighboring nodes of each target node in the dynamic topology graph of the signal. The aggregated feature vector of the target node is obtained by weighted summation of the feature vectors of adjacent nodes based on the attention weight coefficients. The original feature vector of the target node is residually concatenated with the aggregated feature vector and then processed by a nonlinear activation function to obtain the updated feature vector of the target node. The updated feature vectors of all nodes are concatenated into a matrix and a soft orthogonal constraint penalty is applied to obtain the decoupled abnormal feature vector; The deviation vector is obtained by subtracting the decoupled anomaly feature vector from the dynamic benchmark feature sequence element by element.

[0009] Optionally, the anomaly classifier performs probability mapping on the deviation vector to calculate the anomaly type probability distribution, including: Pre-construct class prototype vectors and corresponding covariance matrices for various exception types; Calculate the Mahalanobis distance between the deviation vector and the class prototype vector, and take the negative value of the Mahalanobis distance as the initial similarity score; Calculate the characteristic distribution dispersion of the deviation vector, and calculate the adaptive temperature adjustment coefficient based on the characteristic distribution dispersion; The initial similarity score is scaled using an adaptive temperature adjustment coefficient, and the scaled similarity score is then input into a normalized exponential function for probability mapping to obtain the probability distribution of the anomaly type.

[0010] Optionally, based on the probability distribution results of the anomaly types, causal path retrieval and confidence calculation are performed in a pre-defined anomaly causal knowledge graph to obtain the root cause causal confidence, including: Construct an anomaly causal knowledge graph that includes anomaly phenomenon nodes, physical root cause nodes, and directed causal edges; Based on the probability distribution of anomaly types, nodes with anomaly values ​​greater than a preset threshold are selected as activation nodes. In the abnormal causal knowledge graph, a reverse traversal is performed along the directed causal edges to retrieve all physical root cause nodes pointing to the active node and form a causal path. Calculate the path transition probability for each causal path, and multiply the path transition probability by the anomaly type probability distribution to obtain the root cause causality confidence score.

[0011] Optionally, the formula for calculating the root cause causality confidence level is: ;in, The root cause causality confidence score for the physical root cause node r. This is the probability distribution result of the anomaly type corresponding to the activation node A. The weights of the directed causal edges on the causal path from the physical root node r to the activation node A. The product of consecutive products, is the index of the directed causal edge on the causal path.

[0012] Optionally, the results of communication signal anomaly identification and root cause diagnosis are obtained by sorting and filtering based on the confidence level of root cause causation, including: Sort the root cause causality confidence scores in descending order of their numerical values; The physical root cause node with the highest root cause confidence is selected as the target root cause node. Extract the identification information of the target root cause node and the corresponding anomaly type identification information, and combine them to generate a structured data text containing anomaly description and root cause diagnosis description.

[0013] A data-driven communication signal anomaly identification system includes: The signal preprocessing module is used to perform adaptive noise reduction and baseline drift correction on the acquired raw communication time-domain signal to obtain a corrected time-domain signal sequence. The baseline calculation module is used to calculate the channel environment disturbance coefficient based on the characteristics of historical normal signals, construct a baseline parameter set, and correct the baseline parameter set according to the channel environment disturbance coefficient to obtain a dynamic baseline feature sequence. The signal topology construction module is used to perform adaptive decomposition and correlation calculation on the corrected time-domain signal sequence to construct a dynamic signal topology graph. The deviation calculation module is used to perform feature aggregation and orthogonal constraint processing on the dynamic topology graph of the signal, extract the decoupling anomaly feature vector, and calculate the deviation vector between the decoupling anomaly feature vector and the dynamic reference feature sequence. The probability calculation module is used to map the deviation vector to a probability using an anomaly classifier and calculate the probability distribution of the anomaly type. The causal path retrieval module is used to perform causal path retrieval and confidence calculation in a preset abnormal causal knowledge graph based on the probability distribution results of abnormal types, and obtain the root cause causal confidence. The anomaly identification module is used to sort and filter based on the confidence level of root cause and causation to obtain communication signal anomaly identification and root cause diagnosis results.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention improves the quality of the original signal through adaptive noise reduction and baseline drift correction, and dynamically corrects the reference features by combining the channel environment disturbance coefficient, which can adapt to complex time-varying communication channel environments; by constructing a dynamic topology graph of the signal and applying orthogonal constraints to extract decoupled abnormal features, the correlation information between multiple signal components is fully explored, feature redundancy and coupling interference are reduced, and the distinguishability of abnormal features is enhanced; combined with probabilistic classification with adaptive temperature adjustment and causal knowledge graph root cause reasoning, physical root cause location can be completed while accurately identifying the anomaly type, effectively reducing the false alarm and false negative rates of anomalies, and significantly improving the accuracy of communication signal anomaly diagnosis and fault diagnosis efficiency. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a data-driven communication signal anomaly identification method according to an embodiment of the present invention; Figure 2 This is a block diagram illustrating the principle of communication signal anomaly identification in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a data-driven communication signal anomaly identification system according to an embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a data-driven method for identifying communication signal anomalies, comprising the following steps: Step 100: Perform adaptive noise reduction and baseline drift correction on the acquired raw communication time-domain signal to obtain the corrected time-domain signal sequence; Step 200: Calculate the channel environment disturbance coefficient based on the characteristics of historical normal signals, construct a baseline parameter set, and correct the baseline parameter set according to the channel environment disturbance coefficient to obtain a dynamic reference feature sequence; Step 300: Perform adaptive decomposition and correlation calculation on the corrected time-domain signal sequence to construct a dynamic topology map of the signal; Step 400: Perform feature aggregation and orthogonal constraint processing on the dynamic topology graph of the signal to extract the decoupling anomaly feature vector, and calculate the deviation vector between the decoupling anomaly feature vector and the dynamic reference feature sequence; Step 500: Perform probability mapping on the deviation vector using an anomaly classifier to calculate the probability distribution of anomaly types; Step 600: Based on the probability distribution results of the anomaly type, perform causal path retrieval and confidence calculation in the preset anomaly causal knowledge graph to obtain the root cause causal confidence. Step 700: Sort and filter according to the confidence of root cause and causation to obtain the results of communication signal anomaly identification and root cause diagnosis.

[0020] In the specific implementation process, step 100 first performs sliding window adaptive framing processing on the original communication time-domain signal. For the original discrete time-domain signal, the initial frame length is preset to N points and the frame shift to M points. The frame length N is adaptively set to 8 to 16 times the carrier period based on the signal carrier frequency and sampling frequency. The frame shift M is set to half of the frame length N to ensure a 50% inter-frame overlap rate, thereby eliminating boundary spectral distortion caused by frame truncation. For incomplete frames at both ends of the signal, a symmetrical extension method is used to supplement sampling points. That is, the mirror sample value of the first segment of the signal is added before the first frame, and the mirror sample value of the last segment of the signal is added after the last frame, ensuring that all output signal frames have the same length and that edge information is not lost. After framing, a total of K equal-length signal segments are obtained. Each frame contains N consecutive sampling points, and adjacent frames share NM overlapping sampling points.

[0021] Furthermore, adaptive wavelet denoising is performed on each frame of signal segment, selecting the db4 wavelet as the mother wavelet function, and adaptively determining the number of wavelet decomposition layers based on the proportion of high-frequency energy in each frame. Specifically, layer-by-layer wavelet decomposition is performed on the signal, calculating the energy proportion of the detail coefficients at each layer. When the ratio of the energy of the j-th layer detail coefficients to the total signal energy is lower than a preset energy threshold (0.5% in this embodiment), deep decomposition is stopped, ultimately obtaining one set of approximate coefficients and j sets of detail coefficients. Subsequently, adaptive threshold shrinkage processing is performed, estimating the noise level of the current frame based on the statistical characteristics of the first layer detail coefficients, calculating the noise standard, and then calculating the adaptive threshold corresponding to each layer detail coefficient based on the noise standard deviation. ,in This refers to the number of sampling points in a single frame of signal. The standard deviation of noise. It is a square root function. A soft thresholding function is used to shrink the coefficients of each level of detail. Coefficients with absolute values ​​greater than the adaptive threshold retain their signs and are reduced by the threshold magnitude, while coefficients with absolute values ​​less than or equal to the adaptive threshold are set to zero. At the same time, a sinusoidal smooth transition function is used in the threshold neighborhood to eliminate the constant deviation of traditional soft thresholding and the discontinuity problem of hard thresholding. After the coefficient shrinking is completed, the processed approximate coefficients and the coefficients of each level of detail are subjected to inverse wavelet transform to reconstruct the denoised signal frame of the k-th frame. After processing frame by frame, the denoised signal of all frames is obtained.

[0022] Further, after completing the denoising process for all frames, baseline drift correction and signal stitching are performed. First, all denoised signal frames are overlapped and stitched together according to the sampling offset corresponding to the frame shift, resulting in a complete denoised time-domain signal. Then, an iterative weighted polynomial fitting method is used to extract the signal baseline. Specifically, the fitting weights are initialized as all-1 vectors, and a p-order polynomial is used to perform weighted least-squares fitting on the denoised time-domain signal to obtain an initial baseline sequence. The fitting residual of this sequence is calculated, and for sampling points whose absolute residual value is greater than 3 times the residual standard deviation, their fitting weight is reduced to 0.1 times the original weight. Then, polynomial fitting is performed again based on the new weights to obtain a new baseline sequence. The iterative process is repeated until the difference between two adjacent fitting baselines is less than the convergence threshold, finally obtaining the optimal baseline sequence. The polynomial order p is determined through adaptive optimization. The sum of squared residuals of polynomials from order 1 to 5 is calculated, and the order with the smallest sum of squared residuals and the lowest order is selected as the final fitting order, thereby avoiding excessive absorption of effective signal components by higher-order fitting. Finally, the denoised time-domain signal is subtracted from the optimal baseline sequence point by point to obtain the corrected time-domain signal sequence.

[0023] It should be noted that step 100 forms a complete signal preprocessing link through a three-stage cascaded process of overlapping frame division, adaptive wavelet threshold denoising, and iterative weighted polynomial baseline correction. This not only dynamically adjusts the denoising intensity according to the real-time noise level of each frame, suppressing channel background noise while preserving the instantaneous distortion characteristics of the signal to the greatest extent, but also eliminates the interference of signal spike anomalies on baseline estimation through the iterative weighted fitting mechanism, accurately removes slowly varying baseline drift, and outputs a corrected signal with higher signal-to-noise ratio and a stable baseline. This effectively reduces the interference of noise and baseline drift on the accuracy of subsequent anomaly identification in complex channel environments and improves the robustness of the algorithm in low signal-to-noise ratio scenarios.

[0024] In the specific implementation process, step 200 first constructs a baseline parameter set. Normal communication signal samples within a preset historical time period, manually labeled and confirmed to be free of anomalies, are selected as the baseline construction data source. The data source covers stable transmission conditions under different time periods and load levels, excluding data from known interference and fault periods. For all historical normal signal samples, three basic features are extracted: time-domain statistical features, frequency-domain distribution features, and rhythmic features. The time-domain statistical features include the mean signal amplitude, root mean square value, peak factor, kurtosis coefficient, and waveform factor, calculated frame-by-frame, and then the statistical mean and standard deviation for the entire time period are taken. The frequency-domain distribution features include the center frequency of the power spectrum, 3dB bandwidth, power spectrum entropy, and high-frequency energy proportion, obtained through fast Fourier transform and then statistically analyzed frame-by-frame. The rhythmic features include the main fluctuation period and envelope volatility of the signal envelope, extracted through Hilbert transform and then calculated using autocorrelation analysis. After performing min-max normalization on all features, the global baseline mean and baseline fluctuation range for each feature are calculated. The baseline fluctuation range is taken as 3 times the standard deviation of the full sample of the feature, corresponding to the 99.7% confidence interval of normal fluctuation. The baseline mean and baseline fluctuation range of all dimensions are combined one by one to form a structured baseline parameter set.

[0025] Further, after completing the baseline parameter set construction, the channel environment perturbation coefficient corresponding to the current window is calculated. The corrected time-domain signal sequence is traversed window by window using a sliding window of the same length as in step 100. For the signal within each current sliding window, two perturbation quantification indicators, signal-to-noise ratio (SNR) fluctuation and power spectral flatness, are calculated. The calculation process for SNR fluctuation is as follows: the signal within the current window is divided into sub-windows, each containing 128 sampling points. The short-time SNR of each sub-window is calculated. The effective signal power is calculated from the energy of the main frequency band within the sub-window, and the noise power is estimated from the residual energy of the high-frequency band within the sub-window. After forming a short-time SNR sequence, the coefficient of variation (i.e., the ratio of standard deviation to mean) of this sequence is calculated as the SNR fluctuation, with a value range mapped to the interval 0 to 1. A larger value indicates a more severe fluctuation in SNR over time and a poorer channel stability. The calculation process for power spectral flatness is as follows: A fast Fourier transform is performed on the current window signal to obtain the power spectral sequence. The ratio of the geometric mean to the arithmetic mean of the power spectral sequence is calculated. The closer this ratio is to 1, the flatter the power spectrum and the weaker the channel frequency-selective fading. The power spectral flatness index is obtained by subtracting this ratio from 1, and then normalized to the interval between 0 and 1. The two normalized indices are then weighted and summed using weighting coefficients of 0.6 and 0.4 to obtain the final channel environment disturbance coefficient. The disturbance coefficient ranges from [0,1]. A larger value indicates a higher degree of disturbance in the current channel environment and a wider acceptable range for normal signal fluctuations.

[0026] Furthermore, based on the obtained channel environment disturbance coefficients, the baseline parameter set is dynamically corrected proportionally to generate a dynamic reference feature sequence with a corresponding sliding window. The correction logic is as follows: for the reference fluctuation range of each feature in the baseline parameter set, the channel environment disturbance coefficients are used as an adjustment factor for proportional scaling. The corrected dynamic fluctuation range is equal to the original reference fluctuation range multiplied by [the original value]. ,in This represents the channel environment disturbance coefficient for the current window. The disturbance gain coefficient (valued at 1.2 in this embodiment to control the correction magnitude of the disturbance on the fluctuation range, avoiding excessive relaxation that could lead to missed detections) is used. The dynamic baseline mean of each feature dimension remains unchanged from the baseline mean in the original baseline parameter set, only dynamically adjusting the upper and lower boundaries of the fluctuation range. That is, the dynamic upper limit is the baseline mean plus the corrected dynamic fluctuation range, and the dynamic lower limit is the baseline mean minus the corrected dynamic fluctuation range. The baseline mean of all features in all dimensions is combined with the corresponding dynamic fluctuation boundaries in sequence to obtain the dynamic baseline feature sequence corresponding to the current sliding window. As the sliding window moves frame by frame, the channel environment disturbance coefficient is updated in real time, and the dynamic baseline feature sequence is refreshed synchronously, forming a time-varying baseline system that changes with the channel environment in real time.

[0027] It should be noted that step 200 constructs a multi-dimensional baseline parameter set covering multiple time and frequency domains, combines signal-to-noise ratio fluctuation and spectral flatness as dual indicators to quantify the degree of channel environment disturbance, and dynamically corrects the baseline fluctuation range proportionally based on the disturbance coefficient. This achieves adaptive adjustment of the baseline threshold according to the channel environment, which can narrow the judgment interval during stable channel periods to improve anomaly detection sensitivity, and widen the normal fluctuation range during periods of severe channel disturbance to reduce false alarm rate. This effectively solves the defect of traditional fixed benchmarks that are difficult to adapt to time-varying channel environments, and significantly improves the environmental adaptability and judgment accuracy of anomaly identification in complex channel scenarios.

[0028] In the specific implementation process, step 300 first performs adaptive empirical mode decomposition on the corrected time-domain signal sequence. The input is the corrected time-domain signal sequence output in step 100, and the decomposition process is completed window by window in a sliding window unit. The residual signal is initialized as the corrected signal in the current window. During the decomposition iteration process, the current residual signal is subjected to extreme point detection, cubic spline interpolation fitting of upper and lower envelopes, calculation of envelope mean, and extraction of candidate components of intrinsic mode functions in sequence. An adaptive sieving stopping criterion is used to control the number of sievings of a single component: the sieving termination threshold is dynamically adjusted according to the instantaneous frequency fluctuation of the current candidate component. When the instantaneous frequency variation coefficient is less than 0.1, a stricter standard deviation threshold of 0.2 is used as the sieving stopping condition to avoid over-decomposition leading to mode splitting; when the instantaneous frequency variation coefficient is greater than 0.3, a more lenient standard deviation threshold of 0.8 is used as the sieving stopping condition to avoid under-decomposition leading to mode aliasing. When a candidate component meets the screening stopping condition, it is recorded as the intrinsic mode function component with the corresponding index, and the residual signal is updated to the original residual signal minus this component. The above process is repeated until the residual signal becomes a monotonic function or the number of extreme points is less than 2, completing the decomposition. After the decomposition is completed, pseudo-component screening is performed. The Pearson correlation coefficient between each intrinsic mode function component and the original corrected signal is calculated, and noise pseudo-components with a correlation coefficient less than 0.01 are eliminated, thus obtaining the effective intrinsic mode function components.

[0029] Furthermore, after the effective component screening is completed, each intrinsic mode function component is transformed to the time-frequency domain using short-time Fourier transform, and the corresponding graph node features are extracted. Specifically, for each effective intrinsic mode function component, a Hanning window is used as the sliding window function, with a window length of 256 sampling points and a frame shift of 128 sampling points to ensure a 50% inter-frame overlap rate to reduce time-frequency leakage. Short-time Fourier transform is performed frame by frame to obtain the two-dimensional time-frequency coefficient matrix corresponding to the component. The matrix rows correspond to frequency sampling points, and the columns correspond to time frames, thus completely preserving the joint time-frequency distribution information of the component. To adapt to subsequent graph feature aggregation operations and control the computational dimension, this embodiment performs block feature extraction and dimensionality reduction on each time-frequency coefficient matrix: the time-frequency plane is uniformly divided into 8 sub-bands along the frequency axis. Within each sub-band, three statistical features—average energy, peak power coefficient, and spectral entropy—are calculated. Simultaneously, three global statistical features—the time-domain mean, root mean square value, and kurtosis coefficient—are added to supplement the intrinsic mode function components. This results in an initial feature vector for each node, which is then concatenated sequentially to form a node feature matrix. This process preserves the time-frequency distribution details of the signal components while compressing the high-dimensional time-frequency matrix into a fixed-dimensional feature vector.

[0030] Furthermore, the correlation weight between any two graph node features is calculated. This embodiment uses a weighted fusion of linear and nonlinear correlations to improve the accuracy of the correlation measurement. For any two node feature vectors, the Pearson correlation coefficient is first calculated to quantify the degree of linear correlation between them. The absolute value of the calculation result is taken and normalized to the [0,1] interval to obtain the linear correlation strength index. Then, the normalized mutual information value is calculated to quantify the degree of nonlinear correlation between them. The marginal probability density function and joint probability density function of the two feature vectors are obtained based on Gaussian kernel density estimation. The mutual information value is calculated by numerical integration. Then, the joint entropy of the two feature vectors is used as the denominator for normalization to obtain the nonlinear correlation strength index with a value range of [0,1]. The linear correlation strength and nonlinear correlation strength are weighted and summed according to a weight coefficient of 0.4:0.6 to obtain the final correlation weight between the two nodes. The higher proportion of nonlinear weight can better capture the complex nonlinear coupling relationship between communication signal components, making up for the deficiency that the single Pearson correlation coefficient cannot characterize nonlinear correlation.

[0031] Finally, the dynamic signal topology graph is constructed based on node features and inter-node association weights. Specifically, firstly, an M×M undirected weighted adjacency matrix is ​​constructed based on the association weights of all node pairs. The diagonal elements of the matrix are set to zero to eliminate self-loop interference, where the element in the i-th row and j-th column corresponds to the association weight between the i-th node and the j-th node. To improve the numerical stability of subsequent graph feature aggregation, this embodiment also performs symmetric normalization on the adjacency matrix: first, the degree matrix is ​​calculated, which is a diagonal matrix, and the diagonal elements are the sum of the weights of all associated edges of the corresponding node; then, the adjacency matrix is ​​symmetrically normalized using the inverse square root of the degree matrix to obtain the normalized adjacency matrix. Finally, the node feature matrix and the normalized adjacency matrix together constitute the dynamic signal topology graph corresponding to the current sliding window. In the graph, the nodes correspond to the intrinsic mode function components after signal decomposition, and the edge weights correspond to the association coupling strength between the components.

[0032] It should be noted that step 300 achieves adaptive multi-component decomposition of the signal through adaptive empirical mode decomposition, effectively alleviating the mode mixing and over-decomposition problems of traditional empirical mode decomposition. Combined with the pseudo-component screening mechanism, the effectiveness of the decomposed components is guaranteed. By combining time-frequency domain joint feature extraction and linear-nonlinear dual-index correlation weight fusion, the limitations of traditional single feature and single correlation metric are broken through, which can accurately characterize the complex coupling relationship between signal components. The constructed dynamic topology graph of the signal transforms the one-dimensional signal into graph structure data containing structural information, fully explores the topological correlation information between signal components, provides rich structural support for the subsequent decoupling extraction of abnormal features, significantly enhances the identifiability of weak anomalies and coupling anomalies, and at the same time, sparsification and normalization processing take into account the computational efficiency and numerical stability of topology construction, which can adapt to the time-varying and complex communication signal environment.

[0033] In the specific implementation process, step 400 first calculates the adaptive attention weight coefficient for each target node in the topology graph based on the features of its neighboring nodes and the topology association weights. Specifically, it traverses all nodes in the dynamic topology graph of the signal, taking each node as the target node, and determines the set of all its first-order neighboring nodes according to the normalized adjacency matrix; for the target node and each neighboring node, it calculates the cosine similarity of their feature vectors, and linearly maps the similarity value to the interval [0,1] to obtain the feature association score, while retrieving the association weight between the two nodes calculated in step 300 as the topology structure score; it weights and fuses the feature association score and the topology structure score according to a weight ratio of 0.3:0.7 to obtain the attention score of the corresponding neighboring node; it performs normalization processing on the attention scores of all neighboring nodes of the target node so that the sum of the attention weights of all neighboring nodes is 1, and finally obtains the attention weight coefficient of each neighboring node relative to the target node.

[0034] Furthermore, neighborhood feature aggregation is performed based on attention weight coefficients, and node update feature vectors are generated through residual connections and nonlinear activation. For each target node, the feature vectors of all its neighboring nodes are multiplied by their corresponding attention weight coefficients and then summed element-wise to obtain the neighborhood aggregation feature vector of the target node. During the aggregation process, only neighboring nodes with valid connections are included; unrelated nodes are not included in the feature aggregation. To prevent the loss of original node feature information during the aggregation process, this embodiment uses a residual connection mechanism to add the original feature vector of the target node to the neighborhood aggregation feature vector element-wise to obtain a residual fusion feature vector. Subsequently, the residual fusion feature vector is input into the LeakyReLU nonlinear activation function for transformation, with the negative half-axis slope set to 0.1. This retains weak negative feature information while introducing nonlinear expressive power, outputting the updated feature vector of the target node. After performing the above processing on all M nodes in the topology graph, the updated feature vectors of all nodes are concatenated in component order to form an M-row, D-column updated feature matrix, where D is the feature dimension of a single node.

[0035] Furthermore, a soft orthogonality constraint penalty is applied to the updated feature matrix to extract decoupled anomalous feature vectors. The specific process is as follows: First, the Gram matrix of the updated feature matrix is ​​calculated, which is the matrix product of the updated feature matrix and its transpose. The off-diagonal elements of the Gram matrix correspond to the inner products of the feature vectors of different nodes, reflecting the degree of coupling between features. An orthogonality penalty term is constructed, and the squared F-norm of the difference between the Gram matrix and the identity matrix of the same order is calculated. A larger penalty term value indicates a higher degree of coupling between features. The orthogonality penalty term is incorporated into the feature optimization objective with a weight coefficient of 0.25. The updated feature matrix is ​​iteratively fine-tuned using a small-step gradient descent method, with an iteration step size of 0.01 and a maximum of 50 iterations. The iteration is terminated early when the difference between the penalty terms of two adjacent iterations is less than 1e-6. After orthogonality constraint correction, the feature vectors corresponding to different nodes are approximately orthogonal, significantly reducing feature coupling and redundancy between components. Simultaneously, the soft constraint mechanism avoids the loss of effective features caused by hard orthogonality. Finally, the corrected feature matrix is ​​flattened into a one-dimensional vector by rows to obtain a decoupled anomaly feature vector of dimension M×D, which completely preserves the feature information of all decoupled signal components.

[0036] Further, after extracting the decoupling anomaly feature vector, its deviation vector from the dynamic benchmark feature sequence is calculated. First, an element-wise difference operation is performed; each element of the deviation vector equals the corresponding element of the decoupling anomaly feature vector minus the benchmark mean in the dynamic benchmark feature sequence. In this embodiment, to eliminate dimensional differences between different feature dimensions, the deviation vector is also normalized by dividing each element by the dynamic fluctuation range of the corresponding dimension, resulting in the final normalized deviation vector. The sign of each element in the deviation vector represents the direction of feature deviation, and the absolute value represents the degree of deviation from the normal benchmark. A larger absolute value indicates a more significant anomaly in that dimension, thus fully quantifying the abnormal deviation of the current signal relative to the dynamic normal benchmark.

[0037] It should be noted that step 400 achieves accurate weighted fusion of neighborhood node information through an adaptive attention aggregation mechanism that integrates topological priors, fully exploring the value of topological correlations between multiple signal components; residual connection and nonlinear activation processing balance the preservation of original feature information with the improvement of nonlinear expression capabilities, avoiding feature degradation caused by deep aggregation; the soft orthogonal constraint penalty mechanism effectively removes coupling interference between signal component features, significantly reduces feature redundancy, and enhances the identifiability of weak anomaly features; the finally generated normalized deviation vector achieves standardized quantification of anomaly degree, with clear feature dimensions and clear anomaly directionality, effectively improving the overall accuracy of anomaly identification and the ability to detect weak anomalies in low signal-to-noise ratio scenarios.

[0038] In the specific implementation process, step 500 first completes the pre-construction of basic parameters for the anomaly classifier. Based on the historically labeled communication signal anomaly dataset, corresponding deviation vector samples are collected according to preset anomaly types such as amplitude distortion, phase shift, rhythm disorder, channel fading interference, and hardware performance degradation. At the same time, deviation samples under normal operating conditions are included as normal categories. For each category of samples, its full-dimensional feature mean vector is calculated as the class prototype vector for the corresponding anomaly type. Simultaneously, the feature covariance matrix of each category of samples is calculated, and a regularization coefficient of 1e-4 is added to the diagonal of the covariance matrix to avoid matrix singularity causing calculation failure in small sample scenarios. Then, similarity matching calculation is performed. The current input deviation vector is compared with each type of prototype vector using Mahalanobis distance calculation. The Mahalanobis distance calculation process automatically incorporates the covariance matrix information of the corresponding category, thereby eliminating correlation interference and dimensional differences between feature dimensions. The negative values ​​of each type of Mahalanobis distance are taken to obtain the similarity score for the corresponding anomaly type. The higher the score, the higher the matching degree between the current deviation and the anomaly of that type.

[0039] Further adaptive temperature adjustment processing is performed. Specifically, the feature distribution dispersion of the current deviation vector is calculated. This involves first calculating the mean absolute value of all elements in the deviation vector and the overall standard deviation. The ratio of the standard deviation to the mean absolute value is used as the dispersion index. This index quantifies the degree of dispersion of the current deviation features. The larger the dispersion, the more ambiguous the abnormal features and the lower the classification confidence. Then, an adaptive temperature adjustment coefficient is calculated based on the dispersion index. The baseline temperature is set to 1.0. The temperature adjustment coefficient and the dispersion are positively correlated and linearly related. The adjustment gain is set to 0.8, meaning the temperature adjustment coefficient is equal to the sum of the baseline temperature multiplied by 1 and the weighted dispersion. The higher the dispersion, the larger the temperature adjustment coefficient, avoiding excessively sharp probability outputs that could lead to misjudgments in low-confidence scenarios. Finally, the similarity scores of all categories are uniformly divided by the adaptive temperature adjustment coefficient to complete the scaling. Then, the scores are input into a normalized exponential function for probability mapping, outputting the probability distribution results of all abnormal types and normal categories. The sum of the probability values ​​of all categories is 1, and the category with the largest value corresponds to the initially determined dominant abnormal type.

[0040] It should be noted that step 500 achieves similarity matching between anomalous features and class prototypes based on Mahalanobis distance, overcoming the impact of feature dimension coupling and dimensional differences on classification accuracy, and adapting to the distribution characteristics of anomalous features after decoupling. The adaptive temperature adjustment mechanism can dynamically adjust the sharpness of the probability mapping according to the dispersion of the current deviation features. When the anomalous features are clear and the dispersion is low, it outputs a high-discrimination probability result, and when the features are ambiguous and the dispersion is high, it outputs a smooth probability distribution to avoid misjudgment due to overconfidence. This significantly improves the calibration of the anomalous classification probability and the classification robustness under different channel scenarios, and can accurately output the probability distribution of various anomalies, further ensuring the accuracy and environmental adaptability of the overall anomaly identification scheme.

[0041] In the specific implementation process, step 600 first relies on the pre-constructed abnormal causal knowledge graph to build a basic carrier for causal reasoning. This graph adopts a hierarchical directed acyclic graph structure. The bottom layer is a set of physical root cause nodes, covering typical physical fault root causes of communication systems such as electromagnetic environmental spurious interference, channel multipath fading, performance degradation of transmitting power devices, noise floor elevation of receiving front-end circuits, and abnormal contact impedance of transmission links. The top layer is a set of abnormal phenomenon nodes. Each node identifier corresponds one-to-one with all the abnormal types in step 500. The directed causal edges between nodes point from physical root causes to abnormal phenomena. The edge weights are obtained based on the causal association frequency statistics of historical fault investigation cases. After normalization, the value range is [0,1]. The higher the weight, the greater the causal transmission strength of the corresponding abnormal phenomenon caused by the root cause. At the same time, hop-by-hop confidence attenuation logic is set for multi-hop transmission paths to ensure the rationality of confidence calculation for long-distance causal paths.

[0042] Then, an activation node filtering operation is performed. The preset anomaly probability threshold is 0.1. All anomalous phenomenon nodes are traversed, and nodes with a probability value greater than the preset threshold are marked as activation nodes. The anomaly type probability value P(A) corresponding to each activation node is recorded simultaneously. Low-probability anomalous nodes below the threshold are not included in the subsequent causal retrieval scope, thereby filtering out invalid inference branches and reducing the retrieval computation space. Next, a reverse causal path retrieval is carried out. Starting from each activation node, a depth-first traversal is performed in the opposite direction of the directed causal edges to trace back all physical root cause nodes that can reach the activation node. During the traversal, the edge sequence and corresponding edge weights of each causal path are fully recorded. Loop paths and duplicate paths are automatically eliminated to ensure that all output paths are unidirectional valid causal paths from physical root causes to anomalous phenomena.

[0043] Finally, the root cause causal confidence score is quantitatively calculated. For each retrieved causal path, the weights of all directed causal edges on the path are first calculated. Performing a multiplication operation yields the path transition probability of the path. For scenarios where multiple causal paths from the same physical root cause node lead to the same activation node, the transition probabilities of multiple paths are combined using a probability summation method. This avoids underestimating the actual causal strength of the root cause by calculating each path independently, thus completing the confidence score calculation. The calculation formula is as follows: ; in, The root cause causality confidence score for the physical root cause node r. This is the probability distribution result of the anomaly type corresponding to the activation node A. The weights of the directed causal edges on the causal path from the physical root node r to the activation node A. The product of consecutive products, is the index of the directed causal edge on the causal path.

[0044] It should be noted that step 600 establishes a quantitative causal relationship between abnormal phenomena and physical root causes by constructing a hierarchical and structured abnormal causal knowledge graph. Combined with a reverse depth-first traversal mechanism, it realizes the full path tracing from abnormal manifestations to physical root causes. It also integrates the probability results of front-end abnormal classification with the weight transmission mechanism of causal paths. Through multi-path probability merging and hop-by-hop confidence transmission logic, it accurately quantifies the confidence level of each physical root cause. This not only breaks through the technical limitations of traditional signal abnormality identification, which can only determine the type of abnormality but cannot trace the underlying physical root cause, but also avoids the root cause positioning deviation caused by a single causal matching mode. It effectively improves the accuracy of root cause positioning and the efficiency of on-site fault diagnosis. At the same time, the low-probability node screening mechanism takes into account both reasoning efficiency and computing resource utilization, and can be well adapted to the needs of rapid root cause diagnosis in real-time monitoring scenarios of communication systems.

[0045] In the specific implementation process, step 700 first performs confidence ranking and initial screening. All physical root cause nodes are sorted in descending order according to the value of the root cause causal confidence score to form a root cause candidate ranking list. At the same time, the effective confidence threshold is set to 0.05. Root cause nodes with confidence scores lower than the effective threshold after ranking are removed, and invalid candidates with low confidence scores are screened out to reduce output redundancy. Then, a hierarchical screening of main root cause determination and candidate root cause supplementation is performed. First, the physical root cause node with the highest confidence score in the ranking list is selected as the main target root cause node. At the same time, the confidence score difference between the first and second ranked root causes is calculated. If the difference is less than the preset difference threshold of 0.1, the second ranked root cause node is listed as a candidate root cause node and included in the output range to adapt to complex fault scenarios with multiple root causes. If the difference is greater than or equal to the threshold, only the main target root cause node is output, which takes into account both diagnostic accuracy and adaptability to concurrent scenarios. Further, after node screening, multi-dimensional information extraction is performed. This includes extracting the unique identifiers and standardized physical root cause descriptions of the primary and candidate root cause nodes. Simultaneously, it extracts the anomaly type identifiers and descriptions of the anomaly phenomena corresponding to each root cause, as well as the complete causal transmission path with the highest confidence level and the weights of each edge along the path. It also associates contextual information such as the signal timestamp, channel number, and channel environmental disturbance coefficient corresponding to the current sliding window. Finally, the final diagnostic result text is generated according to a pre-defined structured specification. A standardized field structure is used to divide the system into anomaly identification, root cause diagnosis, confidence quantification, and causal path modules. The anomaly identification module records the probability distribution and dominant anomaly type of each anomaly type within the current window. The root cause diagnosis module distinguishes between the primary and candidate root causes and outputs descriptive information. The confidence quantification module outputs the precise confidence level value of the corresponding root cause. The causal path module fully presents the transmission link from the physical root cause to the anomaly phenomenon. The final output is a structured communication signal anomaly identification and root cause diagnosis result that can be directly integrated into the operation and maintenance management system.

[0046] like Figure 3 As shown, this embodiment of the invention also provides a data-driven communication signal anomaly identification system, including: The signal preprocessing module is used to perform adaptive noise reduction and baseline drift correction on the acquired raw communication time-domain signal to obtain a corrected time-domain signal sequence. The baseline calculation module is used to calculate the channel environment disturbance coefficient based on the characteristics of historical normal signals, construct a baseline parameter set, and correct the baseline parameter set according to the channel environment disturbance coefficient to obtain a dynamic baseline feature sequence. The signal topology construction module is used to perform adaptive decomposition and correlation calculation on the corrected time-domain signal sequence to construct a dynamic signal topology graph. The deviation calculation module is used to perform feature aggregation and orthogonal constraint processing on the dynamic topology graph of the signal, extract the decoupling anomaly feature vector, and calculate the deviation vector between the decoupling anomaly feature vector and the dynamic reference feature sequence. The probability calculation module is used to map the deviation vector to a probability using an anomaly classifier and calculate the probability distribution of the anomaly type. The causal path retrieval module is used to perform causal path retrieval and confidence calculation in a preset abnormal causal knowledge graph based on the probability distribution results of abnormal types, and obtain the root cause causal confidence. The anomaly identification module is used to sort and filter based on the confidence level of root cause and causation to obtain communication signal anomaly identification and root cause diagnosis results.

[0047] The beneficial effects of this invention are as follows: 1) Through a three-stage cascaded preprocessing process of sliding window framing, adaptive wavelet threshold denoising, and iterative weighted polynomial baseline correction, the denoising intensity can be dynamically adjusted according to the real-time noise level of each frame of signal. This suppresses channel background noise while preserving the instantaneous distortion characteristics of the signal to the greatest extent. The iterative weighted fitting mechanism can eliminate the interference of signal spikes on baseline estimation, accurately remove slowly varying baseline drift, and output a high signal-to-noise ratio and stable baseline correction signal. This effectively reduces the interference of noise and baseline drift on subsequent anomaly identification and significantly improves the robustness of the algorithm in low signal-to-noise ratio scenarios. 2) By narrowing the judgment interval during stable channel periods, the sensitivity of anomaly detection is improved, and the normal fluctuation range is widened during periods of severe channel disturbance, reducing the false alarm rate. This solves the problem that traditional fixed threshold schemes are difficult to adapt to time-varying channel environments, and improves the accuracy and environmental adaptability of anomaly judgment in complex channel scenarios. 3) By transforming one-dimensional signals into graph-structured data containing structural correlation information, the linear and nonlinear coupling relationships between signal components are fully explored, breaking through the limitations of traditional single feature recognition in the time or frequency domain, and significantly enhancing the identifiability of weak anomalies and multi-component coupling anomalies. 4) An adaptive attention aggregation mechanism that integrates topological priors is adopted to balance feature similarity and physical coupling strength to complete the weighted fusion of neighborhood features. It also combines residual connections and nonlinear activation to preserve the original feature information and improve expressive power. Furthermore, a soft orthogonal constraint penalty mechanism is used to remove the coupling interference between component features, which significantly reduces feature redundancy while avoiding the loss of effective features, thereby improving the recognition accuracy of different anomaly types and the ability to detect weak anomalies. 5) The adaptive temperature adjustment coefficient is calculated by the feature distribution dispersion to dynamically control the sharpness of the probability mapping. When the abnormal features are clear, a high-discrimination probability result is output. When the features are ambiguous, a smooth distribution is output to avoid overconfidence and misjudgment. This significantly improves the calibration and robustness of anomaly classification under different channel environments. 6) By constructing a hierarchical and structured abnormal causal knowledge graph and combining it with reverse depth-first traversal, the entire path from abnormal phenomena to physical root causes can be traced. The front-end abnormal classification probability and the causal path weight transfer mechanism are integrated, and the confidence of each physical root cause is quantified. This breaks through the technical limitation of traditional signal abnormality identification, which can only determine abnormal phenomena and cannot trace the underlying physical root causes, thus improving the efficiency of fault diagnosis and the accuracy of fault location.

[0048] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0049] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A data-driven method for identifying communication signal anomalies, characterized in that, Includes the following steps: The acquired raw communication time-domain signal is subjected to adaptive noise reduction and baseline drift correction to obtain a corrected time-domain signal sequence. The channel environment disturbance coefficient is calculated based on the characteristics of historical normal signals, a baseline parameter set is constructed, and the baseline parameter set is corrected according to the channel environment disturbance coefficient to obtain a dynamic reference feature sequence. Adaptive decomposition and correlation calculation are performed on the corrected time-domain signal sequence to construct a dynamic topology map of the signal; The signal dynamic topology graph is subjected to feature aggregation and orthogonal constraint processing to extract the decoupling anomaly feature vector, and the deviation vector between the decoupling anomaly feature vector and the dynamic reference feature sequence is calculated. The probability distribution of the anomaly type is calculated by performing probability mapping on the deviation vector using an anomaly classifier. Based on the probability distribution of the anomaly types, causal path retrieval and confidence calculation are performed in the preset anomaly causal knowledge graph to obtain the root cause causal confidence. Based on the confidence level of the root cause causality, the results of communication signal anomaly identification and root cause diagnosis are obtained by sorting and filtering.

2. The data-driven communication signal anomaly identification method according to claim 1, characterized in that, The acquired raw communication time-domain signal is subjected to adaptive noise reduction and baseline drift correction to obtain a corrected time-domain signal sequence, including: The original communication time-domain signal is subjected to sliding window framing to obtain multiple signal segments of equal length. Each signal segment in the frame is subjected to wavelet decomposition, and the decomposed wavelet coefficients are subjected to threshold shrinkage, and then wavelet reconstruction is performed to obtain the denoised signal frame. Polynomial fitting baseline correction processing is performed on all the denoised signal frames to remove the overall baseline drift of the signal, and the corrected time-domain signal sequence is obtained by splicing them together.

3. The data-driven communication signal anomaly identification method according to claim 1, characterized in that, Channel environment disturbance coefficients are calculated based on historical normal signal characteristics, a baseline parameter set is constructed, and the baseline parameter set is corrected according to the channel environment disturbance coefficients to obtain a dynamic baseline feature sequence, including: The baseline parameter set is constructed by extracting time-domain statistical features, frequency-domain distribution features, and rhythmic features from historical normal communication signals and integrating them. Calculate the signal-to-noise ratio fluctuation and power spectral flatness of the signal within the current sliding window, and perform a weighted summation to obtain the channel environment disturbance coefficient; The characteristic fluctuation range of the baseline parameter set is proportionally corrected based on the channel environment disturbance coefficient to obtain the dynamic benchmark feature sequence.

4. The data-driven communication signal anomaly identification method according to claim 1, characterized in that, Adaptive decomposition and correlation calculation are performed on the corrected time-domain signal sequence to construct a dynamic signal topology graph, including: Adaptive empirical mode decomposition is performed on the corrected time-domain signal sequence to obtain several intrinsic mode function components; The intrinsic mode function components are transformed to the time-frequency domain by short-time Fourier transform, and the time-frequency coefficient matrix of each component is extracted as graph node features. Calculate the Pearson correlation coefficient and mutual information value between any two graph node features, and then perform a weighted fusion calculation on the Pearson correlation coefficient and mutual information value to obtain the node association weight; An adjacency matrix is ​​constructed based on the association weights between the nodes, and the dynamic topology graph of the signal is constructed based on the graph node features and the adjacency matrix.

5. The data-driven communication signal anomaly identification method according to claim 1, characterized in that, The signal dynamic topology graph is subjected to feature aggregation and orthogonal constraint processing to extract decoupling anomaly feature vectors, and the deviation vector between the decoupling anomaly feature vectors and the dynamic reference feature sequence is calculated, including: The attention weight coefficient is calculated based on the feature vectors of the neighboring nodes of each target node in the dynamic topology graph of the signal. The feature vectors of the adjacent nodes are weighted and summed based on the attention weight coefficients to obtain the aggregated feature vector of the target node. The original feature vector of the target node is residually concatenated with the aggregated feature vector and then processed by a nonlinear activation function to obtain the updated feature vector of the target node. The updated feature vectors of all nodes are concatenated into a matrix and a soft orthogonal constraint penalty is applied to obtain the decoupling anomaly feature vector; The deviation vector is obtained by subtracting the decoupled anomaly feature vector from the dynamic benchmark feature sequence element by element.

6. The data-driven communication signal anomaly identification method according to claim 1, characterized in that, The anomaly classifier performs probability mapping on the deviation vector to calculate the anomaly type probability distribution, including: Pre-construct class prototype vectors and corresponding covariance matrices for various exception types; Calculate the Mahalanobis distance between the deviation vector and the class prototype vector, and take the negative value of the Mahalanobis distance as the initial similarity score; Calculate the characteristic distribution dispersion of the deviation vector, and calculate the adaptive temperature adjustment coefficient based on the characteristic distribution dispersion; The initial similarity score is scaled according to the adaptive temperature adjustment coefficient, and the scaled similarity score is input into the normalized exponential function for probability mapping to obtain the probability distribution result of the anomaly type.

7. The data-driven communication signal anomaly identification method according to claim 1, characterized in that, Based on the probability distribution of the aforementioned anomaly types, causal path retrieval and confidence calculation are performed in a pre-defined anomaly causal knowledge graph to obtain the root cause causal confidence, including: Construct the abnormal causal knowledge graph that includes abnormal phenomenon nodes, physical root cause nodes, and directed causal edges; Based on the probability distribution of the anomaly type, nodes of the anomaly phenomenon with a probability value greater than a preset threshold are selected as activation nodes. In the abnormal causal knowledge graph, a reverse traversal is performed along the directed causal edges to retrieve all the physical root cause nodes pointing to the activated node, and a causal path is formed. Calculate the path transition probability for each causal path, and multiply the path transition probability by the anomaly type probability distribution result to obtain the root cause causal confidence score.

8. The data-driven communication signal anomaly identification method according to claim 7, characterized in that, The formula for calculating the root cause causality confidence level is as follows: ;in, The root cause causality confidence score for the physical root cause node r. This is the probability distribution result of the anomaly type corresponding to the activation node A. The weights of the directed causal edges on the causal path from the physical root node r to the activation node A. The product of consecutive products, is the index of the directed causal edge on the causal path.

9. The data-driven communication signal anomaly identification method according to claim 7, characterized in that, Based on the root cause causal confidence scores, the results of communication signal anomaly identification and root cause diagnosis are obtained, including: The root cause causal confidence scores are arranged in descending order of their numerical values; The physical root cause node with the highest root cause confidence is selected as the target root cause node. Extract the identification information of the target root cause node and the corresponding abnormality type identification information, and combine them to generate a structured data text containing an abnormal phenomenon description and a root cause diagnosis description.

10. A data-driven communication signal anomaly identification system, characterized in that, include: The signal preprocessing module is used to perform adaptive noise reduction and baseline drift correction on the acquired raw communication time-domain signal to obtain a corrected time-domain signal sequence. The baseline calculation module is used to calculate the channel environment disturbance coefficient based on the characteristics of historical normal signals, construct a baseline parameter set, and correct the baseline parameter set according to the channel environment disturbance coefficient to obtain a dynamic baseline feature sequence. The signal topology construction module is used to perform adaptive decomposition and correlation calculation on the corrected time-domain signal sequence to construct a dynamic signal topology map. The deviation calculation module is used to perform feature aggregation and orthogonal constraint processing on the dynamic topology graph of the signal, extract the decoupling anomaly feature vector, and calculate the deviation vector between the decoupling anomaly feature vector and the dynamic reference feature sequence. The probability calculation module is used to perform probability mapping on the deviation vector through the anomaly classifier and calculate the probability distribution of the anomaly type. The causal path retrieval module is used to perform causal path retrieval and confidence calculation in a preset abnormal causal knowledge graph based on the probability distribution results of the anomaly type, and obtain the root cause causal confidence. The anomaly identification module is used to sort and filter according to the root cause causal confidence level to obtain communication signal anomaly identification and root cause diagnosis results.

Images