Network fault intelligent prediction system based on dial testing

Abstract

The invention belongs to the field of network fault analysis, and particularly relates to a network fault intelligent prediction system based on dial testing, which comprises a dial testing module, a network testing module, a data acquisition and storage module, a signal prediction module, a fault analysis module and a monitoring and early warning module, according to the method, the principle of a graph neural network is creatively utilized, dynamic spatial-temporal feature extraction is performed on the dial test data, fusion features including overall network features and local features of network nodes are obtained, and rapid fault analysis is realized; according to the method, a bidirectional gating dynamic space-time signal prediction method based on a Pearson's correlation coefficient is creatively adopted, VMD decomposition is utilized, data proximity, periodicity and tendency characteristics are enhanced, Pearson characteristics are fused into a GAT model, a hidden space relation and potential characteristics between network nodes are highlighted, bidirectional gating characteristics are fused into a TCN model, and the prediction precision of the bidirectional gating dynamic space-time signal is improved. The problem of long-time dependence is solved, global features and dynamic changes are effectively extracted, and finally accurate processing of complex fault conditions is improved.

Description

Technical Field

[0001] The present invention belongs to the field of network fault analysis, and specifically is an intelligent prediction system for network faults based on probing measurement. Background Art

[0002] Network fault analysis based on probing measurement refers to the process of deploying probing measurement nodes in the network to perform active performance tests and monitoring on the network regularly or randomly, so as to detect problems such as faults, delays, and packet losses in the network, and analyze and locate network faults. However, existing systems usually require manual intervention to analyze probing measurement data and locate faults, which is time-consuming and laborious, and the fault analysis models and algorithms of existing systems are difficult to handle various complex fault situations. Summary of the Invention

[0003] In view of the above situation, to overcome the defects of the prior art and address the problem that existing systems usually require manual intervention to analyze probing measurement data and locate faults, the present invention creatively designs an intelligent prediction system for network faults based on probing measurement. Using the principle of graph neural networks, by extracting dynamic spatio-temporal features based on graphs from the probing measurement data, a fusion feature containing the overall network features and the local features of each network node is obtained. Based on the fusion feature, rapid fault location and fault analysis of the probing measurement data can be performed; the present invention creatively adopts a bidirectional gated dynamic spatio-temporal signal prediction method based on the Pearson correlation coefficient. Using VMD decomposition, the proximity, periodicity, and trend features of the probing measurement data are strengthened. The Pearson features are incorporated into the GAT model, and the correlation between non-adjacent nodes is incorporated on the basis of the correlation between adjacent nodes, highlighting the hidden spatial relationship and potential features between network nodes. The bidirectional gated features are incorporated into the TCN model. On the basis of solving the long-term dependence of network signals, its global features and dynamic changes are effectively extracted, and finally the accurate handling of complex fault situations is improved.

[0004] The technical solution adopted by the present invention is as follows. An intelligent prediction system for network faults based on probing measurement includes a probing measurement module, a network testing module, a data acquisition and storage module, a signal prediction module, a fault analysis module, and a monitoring and warning module; The probing measurement module regularly sends a detection request to the network testing module, receives the sequential return signal returned by the network testing module, and sends the sequential return signal to the signal prediction module. The network testing module includes multiple probing measurement nodes, receives the detection request sent by the probing measurement module, and returns a sequential return signal; The data acquisition and storage module collects, preprocesses, and stores the sequential return signal of the probing measurement module; The signal prediction module uses a bidirectional gated dynamic spatio-temporal signal prediction method based on the Pearson correlation coefficient to predict the time-series return signal, obtains the final signal prediction result, and sends the final signal prediction result to the fault analysis module; The fault analysis module uses a machine learning model to analyze the network faults of the historical time-series return signals, obtains the correspondence between the time-series return signals and the network faults, and obtains the network fault prediction result according to the correspondence and the final signal prediction result; The monitoring and warning module analyzes the corresponding dial test node positions according to the network fault prediction result and provides fault location information.

[0005] Furthermore, in the signal prediction module, the bidirectional gated dynamic spatio-temporal signal prediction method based on the Pearson correlation coefficient specifically includes the following steps: Step S1: Graph construction, construct a graph for all dial test nodes to obtain a return signal graph: ; In the formula, represents the graph node, that is, all dial test nodes, represents the edge between nodes, represents the adjacency matrix of the return signal grid graph; Step S2: Variational mode decomposition, use the variational mode decomposition VMD method to decompose the time-series return signals of all nodes in the return signal graph to obtain K return signal subgraphs; Step S3: Subgraph signal prediction, perform time-series return signal prediction on each return signal subgraph to obtain the predicted return signal of each return signal subgraph; Furthermore, in step S3, when performing time-series return signal prediction, it specifically includes the following steps: Step S31: GAT model processing, perform GAT model processing on the return signal subgraph to obtain the multi-head fusion features of each node in the return signal subgraph, specifically including the following steps: Step S311: Attention coefficient calculation, calculate the similarity coefficient between nodes in the return signal subgraph to obtain the spatial attention coefficient between nodes in the return signal subgraph, and normalize all attention coefficients to obtain the normalized coefficient; Step S312: Feature weighted summation, multiply all the normalized coefficients between each node and its adjacent nodes by the weight matrix and then perform weighted summation to obtain the feature vector of each node; Step S313: Multi-head attention processing, perform multi-head attention processing on the feature vector of each node to obtain the multi-head fusion features of each node; Step S32: Spatial interaction feature calculation, specifically including the following steps: Step S321: Pearson correlation coefficient calculation. According to the time-series return signal values collected by each node, calculate the Pearson coefficients between all nodes: ; In the formula, represents the length of the time series, represents two different nodes, represents the return signal value of node at time represents the average value of the return signal values of node in the time series, represents the Pearson coefficient between nodes Step S322: Weighted summation of Pearson coefficients. Perform weighted summation on all Pearson coefficients between each node and its adjacent nodes to obtain the Pearson value of each node; Step S323: Time-series feature extraction. Use wavelet transform to convert the time-series return signal of each node into a frequency-domain graph, and perform feature extraction on the frequency-domain graph to obtain the signal feature of each node; Step S323: Construction of the global correlation feature matrix. Perform non-linear calculation on the Pearson values of all nodes and their signal features to obtain the global correlation feature matrix; Step S324: Construction of the spatial local attention feature matrix. Perform feature fusion on the multi-head fusion features of all nodes and their signal features to obtain the spatial local attention feature matrix; Step S325: Spatial interaction feature calculation. Perform element-wise multiplication of the Hadamard product on the global correlation feature matrix and the spatial local attention feature matrix, and perform non-linear transformation using the rectified linear unit activation function to obtain the spatial interaction feature of the return signal subgraph: ; In the formula, represents the spatial local attention feature matrix, represents the global correlation feature matrix, represents the element-wise multiplication of the Hadamard product, represents the rectified linear unit activation function, represents the spatial interaction feature; Step S33: Temporal interaction feature calculation, which specifically includes the following steps: Step S331: Local time-series feature extraction. Use the TCN model to perform feature extraction on the time-series return signal of each node to obtain local time-series features, where the TCN model includes a convolutional layer and a residual block; Step S332: Global time-domain feature extraction. Use a bidirectional gated recurrent unit to process the signal features of all nodes to obtain the forward hidden state and the backward hidden state. Integrate the forward hidden state and the backward hidden state of all nodes, and use the self-attention mechanism for optimization to obtain the global time-domain feature: ; ; ; In the formula, represents the activation function, , and are preset parameters, corresponds to the query feature Query of the self-attention mechanism, is the integration result of the forward hidden state and the backward hidden state, corresponds to the key feature Key in the self-attention mechanism, represents the corresponding attention weight parameter at different times, represents the global time-domain feature; Step S333: Time interaction feature calculation. Perform a feature concatenation operation on the local time-series feature and the global time-domain feature to obtain the time interaction feature calculation; Step S34: Feature fusion prediction. Concatenate the spatial interaction feature and the time interaction feature, and use two fully connected layers for prediction output to obtain the predicted return signal; Step S4: Predicted weighted summation. Perform a weighted summation on the predicted return signals of all return signal subgraphs to obtain the final signal prediction result.

[0006] The beneficial effects achieved by the present invention using the above solution are as follows: (1) Aiming at the problem that the existing system usually requires manual intervention to analyze the probing data and locate faults, the present invention creatively designs an intelligent network fault prediction system based on probing. Using the principle of graph neural network, through the extraction of dynamic spatio-temporal features based on graphs for the probing data, a fusion feature containing the overall network feature and the local features of each network node is obtained. According to the fusion feature, rapid fault location and fault analysis can be performed on the probing data; (2) The present invention creatively adopts a bidirectional gated dynamic spatio-temporal signal prediction method based on the Pearson correlation coefficient. By using VMD decomposition, it strengthens the proximity, periodicity, and trend characteristics of the probing data, integrates the Pearson features into the GAT model, incorporates the correlation between non-adjacent nodes on the basis of the correlation between adjacent nodes, highlights the hidden spatial relationship and potential features between network nodes, integrates the bidirectional gated features into the TCN model, effectively extracts its global features and dynamic changes while solving the long-term dependence of network signals, and finally improves the accurate handling of complex fault situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a schematic diagram of the modules of the network fault intelligent prediction system based on probing provided by the present invention; Figure 2 It is a schematic flowchart of the bidirectional gated dynamic spatio-temporal signal prediction method based on the Pearson correlation coefficient provided by the present invention.

[0008] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0009] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0010] Embodiment 1. Refer to Figure 1 , the network fault intelligent prediction system based on probing provided by the present invention includes a probing module, a network testing module, a data acquisition and storage module, a signal prediction module, a fault analysis module, and a monitoring and warning module; The probing module regularly sends a probing request to the network testing module, receives the time-series return signal returned by the network testing module, and sends the time-series return signal to the signal prediction module The network testing module includes multiple probing nodes, receives the probing request sent by the probing module, and returns the time-series return signal; The data acquisition and storage module collects, preprocesses, and stores the time-series return signal of the probing module; The signal prediction module adopts a bidirectional gated dynamic spatio-temporal signal prediction method based on the Pearson correlation coefficient to predict the time-series return signal, obtains the final signal prediction result, and sends the final signal prediction result to the fault analysis module; The fault analysis module uses a machine learning model to analyze network faults in historical time-series return signals, obtains the correspondence between time-series return signals and network faults, and obtains a network fault prediction result based on the correspondence and the final signal prediction result; The monitoring and warning module analyzes the corresponding probing node positions according to the network fault prediction result and provides fault location information.

[0011] By performing the above operations, in view of the problem that the existing system usually requires manual intervention to analyze probing data and locate faults, the present invention creatively designs an intelligent network fault prediction system based on probing. Using the principle of graph neural network, through the extraction of graph-based dynamic spatio-temporal features from probing data, a fusion feature containing the overall network feature and the local features of each network node is obtained, and the probing data can be quickly fault-located and fault-analyzed according to the fusion feature.

[0012] Example 2, refer to Figure 2 , based on the above example, in the signal prediction module, the two-way gated dynamic spatio-temporal signal prediction method based on Pearson correlation coefficient specifically includes the following steps: Step S1: Graph construction, construct a graph for all probing nodes to obtain a return signal graph: ; In the formula, represents graph nodes, that is, all probing nodes, represents the edges between nodes, represents the adjacency matrix of the return signal grid graph; Step S2: Variational mode decomposition, use the variational mode decomposition VMD method to decompose the time-series return signals of all nodes in the return signal graph to obtain 10 return signal subgraphs; Step S3: Subgraph signal prediction, perform time-series return signal prediction on each return signal subgraph to obtain the predicted return signal of each return signal subgraph; Step S4: Predicted weighted summation, perform weighted summation on the predicted return signals of all return signal subgraphs to obtain the final signal prediction result.

[0013] Example 3, based on the above example, in step S3, perform time-series return signal prediction, which specifically includes the following steps: Step S31: GAT model processing, perform GAT model processing on the return signal subgraph to obtain the multi-head fusion features of each node in the return signal subgraph Step S32: Calculate spatial interaction features; Step S33: Calculate time interaction features; Step S34: Feature fusion prediction. Concatenate the spatial interaction features and the temporal interaction features, and use two fully connected layers for prediction output to obtain the predicted return signal.

[0014] Example 4. Based on the above example, step S31 specifically includes the following steps: Step S311: Calculate the attention coefficient. Calculate the inter-node similarity coefficient of the return signal subgraph to obtain the spatial attention coefficient between the nodes of the return signal subgraph, and normalize all the attention coefficients to obtain the normalized coefficient. Step S312: Feature weighted summation. Multiply all the normalized coefficients between each node and its adjacent nodes by the weight matrix and then perform weighted summation to obtain the feature vector of each node. Step S313: Multi-head attention processing. Perform multi-head attention processing on the feature vector of each node to obtain the multi-head fusion feature of each node.

[0015] Example 5. Based on the above example, step S32: Calculate the spatial interaction features, specifically including the following steps: Step S321: Calculate the Pearson correlation coefficient. According to the time-series return signal values collected by each node, calculate the Pearson coefficients between all nodes: ; In the formula, represents the length of the time series, represents two different nodes, represents the return signal value of node at time represents the average value of the return signal values of node represents node the Pearson coefficient between Step S322: Pearson coefficient weighted summation. Perform weighted summation on all the Pearson coefficients between each node and its adjacent nodes to obtain the Pearson value of each node. Step S323: Temporal feature extraction. Use wavelet transform to convert the time-series return signal of each node into a frequency-domain graph, and perform feature extraction on the frequency-domain graph to obtain the signal feature of each node. Step S323: Construct the global correlation feature matrix. Perform non-linear calculation on the Pearson values of all nodes and their signal features to obtain the global correlation feature matrix. Step S324: Construct the spatial local attention feature matrix. Perform feature fusion on the multi-head fusion features of all nodes and their signal features to obtain the spatial local attention feature matrix. Step S325: Spatial interaction feature calculation. Perform element-wise multiplication of the global correlation feature matrix and the spatial local attention feature matrix using the Hadamard product, and perform non-linear transformation using the rectified linear unit activation function to obtain the spatial interaction features of the returned signal subgraph: ; In the formula, represents the spatial local attention feature matrix, represents the global correlation feature matrix, represents the element-wise multiplication of the Hadamard product, represents the rectified linear unit activation function, represents the spatial interaction feature.

[0016] Example 6. This example is based on the above example. Step S33: Temporal interaction feature calculation, which specifically includes the following steps: Step S331: Local temporal feature extraction. Use the TCN model to extract features from the temporal return signal of each node to obtain local temporal features. Among them, the TCN model includes a convolutional layer and a residual block; Step S332: Global temporal feature extraction. Use a bidirectional gated recurrent unit to process the signal features of all nodes to obtain a forward hidden state and a backward hidden state. Fuse the forward hidden state and the backward hidden state of all nodes, and use the self-attention mechanism for optimization to obtain the global temporal feature: ; ; ; In the formula, represents the activation function, , and are preset parameters, corresponds to the query feature Query of the self-attention mechanism, is the fusion result of the forward hidden state and the backward hidden state, corresponds to the key feature Key in the self-attention mechanism, represents the attention weight parameter corresponding to different moments, represents the global temporal feature; Step S333: Temporal interaction feature calculation. Perform feature concatenation operation on the local temporal feature and the global temporal feature to obtain the temporal interaction feature calculation.

[0017] By performing the above operations, the present invention creatively adopts a two-way gated dynamic spatio-temporal signal prediction method based on the Pearson correlation coefficient. Using VMD decomposition, it strengthens the proximity, periodicity, and trend characteristics of the probing data, integrates the Pearson features into the GAT model, incorporates the correlation between non-adjacent nodes on the basis of the correlation between adjacent nodes, highlights the hidden spatial relationships and potential features between network nodes, and integrates the two-way gated features into the TCN model. On the basis of solving the long-term dependence of network signals, it effectively extracts their global features and dynamic changes, and finally improves the accurate handling of complex fault situations.

[0018] Example Seven. Based on the above example, the detailed rules of the returned timing signals participating in the probing are as follows: RTT round-trip delay: The round-trip time for sending a probe packet to the target node; Used to judge packet loss, congestion, or device anomalies in the network; Packet loss rate: The proportion of probe packets sent that did not receive a response; Used to judge link failures or device problems; Throughput: The amount of data successfully transmitted per unit time; Used to evaluate whether the link bandwidth reaches the expected performance; Jitter: The change in network latency; Excessive jitter will affect real-time applications, such as video conferencing; Service response time: The response time of an HTTP request; Used to evaluate service performance.

[0019] Example Eight. Based on the above example, the variational mode decomposition (VMD) method is used to decompose the timing return signals, which specifically includes the following steps: Signal preprocessing: Preprocess the timing signal: remove noise and normalize; Construct the Hilbert transform: Perform the Hilbert transform on the timing signal to obtain the analytic signal.

[0020] Initialization: Initialize the intrinsic mode function, frequency bandwidth, and analytic signal frequency; Iterative solution: Solve each intrinsic mode function (IMF) through iterative optimization; Update the frequency bandwidth: In each iteration, update the frequency bandwidth parameter according to the current signal's frequency bandwidth; Regularization processing: Introduce a regularization term to help with convergence and ensure the stability of the decomposition result; Convergence judgment: Judge whether the iterative process converges by setting a threshold; Extract IMFs: When the iteration ends, 10 intrinsic mode functions (IMFs) are obtained. These IMFs contain different frequency components of the original signal; Reconstruct the signal: Add up each IMF to obtain the reconstructed time series signal, and selectively retain some IMFs for the purpose of signal noise reduction or analysis of specific frequency components.

[0021] Example 9. This example is based on the above example, and the steps of the GAT model are described as follows: Input data: The input data is data represented as a graph structure, including node features and the connection relationships between nodes. The input data of the present invention is the returned signal subgraph; Feature propagation: For each node, calculate the new feature representation of the node itself according to the features of its neighbor nodes, and introduce an attention mechanism to enable the node to focus on important neighbor nodes; Attention mechanism: When calculating the information propagation between nodes, introduce an attention mechanism to assign different importance weights to different neighbor nodes; the weights are obtained through learning and can be dynamically adjusted according to node features; Calculate the attention weight: For each pair of connected nodes, calculate the attention weight between them: Calculate the attention coefficient: Calculate the relative importance between node pairs through the weight parameters obtained by learning; Convert the attention coefficient: Normalize the calculated attention coefficient to ensure that its range is between 0 and 1; Weighted summation: Perform weighted summation on the features of neighbor nodes according to the attention coefficient to obtain the new feature representation of the node itself; Multi-head attention: To increase the expressive power of the model, introduce a multi-head attention mechanism, that is, learn multiple groups of attention weights simultaneously and fuse their results.

[0022] It should be noted that, in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0023] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

[0024] The above describes the present invention and its embodiments, and such description is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of the present invention, design similar structural modes and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. The network fault intelligent prediction system based on dial-up testing is characterized by: It includes dial test module, network test module, data acquisition and storage module, signal prediction module, fault analysis module and monitoring and early warning module; The dial test module periodically sends a detection request to the network test module, receives a timing return signal returned by the network test module, and sends the timing return signal to the signal prediction module. The network test module includes a plurality of dial test nodes, receives a detection request sent by the dial test module, and returns a timing return signal; The data acquisition and storage module acquires, pre-processes and stores the timing return signal of the dial test module; The signal prediction module uses a bidirectional gated dynamic spatiotemporal signal prediction method based on the Pearson correlation coefficient to predict the time series return signal, obtain the final signal prediction result, and send the final signal prediction result to the fault analysis module; The fault analysis module uses a machine learning model to perform network fault analysis on historical timing return signals, obtain a corresponding relationship between timing return signals and network faults, and obtain a network fault prediction result based on the corresponding relationship and the final signal prediction result; The monitoring and early warning module analyzes the corresponding dial-up node location according to the network fault prediction result and provides fault location information.

2. The network fault intelligent prediction system based on dial-up testing according to claim 1 is characterized in that: In the signal prediction module, the bidirectional gated dynamic spatiotemporal signal prediction method based on the Pearson correlation coefficient specifically includes the following steps: Step S1: Graph construction, construct a graph for all dial-test nodes to obtain a return signal graph: ； In the formula, Represents graph nodes, that is, all dial-up test nodes, represents the edges between nodes, Represents the adjacency matrix of the returned signal grid graph; Step S2: variational mode decomposition, using the variational mode decomposition (VMD) method to decompose the time series return signals of all nodes in the return signal graph to obtain K return signal subgraphs; Step S3: Sub-graph signal prediction, performing time series return signal prediction on each return signal sub-graph to obtain a predicted return signal of each return signal sub-graph; Step S4: Prediction weighted summation, weighted summation of the predicted return signals of all return signal subgraphs to obtain the final signal prediction result.

3. The network fault intelligent prediction system based on dial-up testing according to claim 2 is characterized in that: In step S3, a timing return signal prediction is performed, which specifically includes the following steps: Step S31: GAT model processing, GAT model processing is performed on the return signal subgraph to obtain the multi-head fusion feature of each node in the return signal subgraph Step S32: Calculating spatial interaction features, specifically including the following steps: Step S321: Pearson correlation coefficient calculation: according to the time series return signal value collected by each node, the Pearson coefficient between all nodes is calculated: ； In the formula, Represents the length of the time series, Represents two different nodes. represent Node in The return signal value at the moment, represent The average value of the return signal value of the node in the time series, Representative Node Pearson coefficient between ; Step S322: weighted summation of Pearson coefficients, weighted summation of all Pearson coefficients between each node and its adjacent nodes to obtain the Pearson value of each node; Step S323: extracting time series features, using wavelet transform to convert the time series return signal of each node into a frequency domain graph, and extracting features from the frequency domain graph to obtain the signal features of each node; Step S323: constructing a global correlation feature matrix, performing nonlinear calculation on the Pearson values ​​of all nodes and their signal features to obtain a global correlation feature matrix; Step S324: constructing a spatial local attention feature matrix, fusing the multi-head fusion features of all nodes with their signal features to obtain a spatial local attention feature matrix; Step S325: Calculate the spatial interaction features, perform element-wise multiplication of the global correlation feature matrix and the spatial local attention feature matrix using the Hadamard product, and use the rectified linear unit activation function for nonlinear transformation to obtain the spatial interaction features of the returned signal subgraph: ； In the formula, represents the spatial local attention feature matrix, represents the global correlation feature matrix, represents the element-wise multiplication of the Hadamard product, represents the rectified linear unit activation function, Represents spatial interaction characteristics; Step S33: time interaction feature calculation; Step S34: feature fusion prediction, concatenating the spatial interaction features with the temporal interaction features, and using two fully connected layers for prediction output to obtain a prediction return signal.

4. The network fault intelligent prediction system based on dial-up testing according to claim 3 is characterized in that: Step S31 specifically includes the following steps: Step S311: Calculate the attention coefficient, calculate the similarity coefficient between nodes of the returned signal subgraph, obtain the spatial attention coefficient between nodes of the returned signal subgraph, and normalize all the attention coefficients to obtain the normalized coefficient; Step S312: weighted summation of features, multiplying all normalized coefficients between each node and its adjacent nodes by the weight matrix and performing weighted summation to obtain a feature vector for each node; Step S313: multi-head attention processing, perform multi-head attention processing on the feature vector of each node to obtain the multi-head fusion feature of each node.

5. The network fault intelligent prediction system based on dial-up testing according to claim 4 is characterized in that: Step S33: Time interaction feature calculation, specifically including the following steps: Step S331: extract local time series features, using the TCN model to extract features from the time series return signal of each node to obtain local time series features, wherein the TCN model includes a convolutional layer and a residual block; Step S332: Extract global time domain features. Use a bidirectional gated recurrent unit to process the signal features of all nodes to obtain a forward hidden state and a backward hidden state. Fuse the forward hidden state and the backward hidden state of all nodes, and use a self-attention mechanism to optimize them to obtain global time domain features: ； ； ； In the formula, represents the activation function, , and is the preset parameter, Query corresponding to the query feature of the self-attention mechanism, is the fusion result of the forward hidden state and the backward hidden state, Corresponding to the key feature Key in the self-attention mechanism, Represents the corresponding The attention weight parameter, Represents the global time domain characteristics; Step S333: Time interaction feature calculation, performing feature concatenation operation on the local time series feature and the global time domain feature to obtain the time interaction feature calculation.

Images