Power communication analysis method based on big data

Abstract

The invention discloses an electric power communication analysis method based on big data, and the method comprises the following steps: S1, collecting multi-modal feature data, and constructing a communication feature tensor; s2, constructing a communication atlas structure according to the communication equipment connection relationship and the communication feature tensor; s3, inputting the communication feature tensor and the communication atlas structure into a multi-modal residual image convolutional network, and outputting a link degradation score vector; s4, inputting the link degradation score vector into an improved mole search optimization algorithm to generate an optimal threshold vector; s5, introducing the optimal threshold vector into a multi-modal residual image convolutional network training process, and optimizing the multi-modal residual image convolutional network; and S6, on the basis of the optimized multi-mode residual image convolutional network, generating a degradation judgment result of each link for risk labeling of the power communication network. According to the method, the multi-modal residual image convolutional network and the improved mole optimization algorithm are fused, and intelligent judgment and risk labeling analysis of the power link are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing and intelligent analysis, and in particular to a power communication analysis method based on big data. Background Art

[0002] With the rapid development of new power systems, the ubiquitous power Internet of Things (IoT), and smart grids, power communication networks, as the core foundation for ensuring power system operation, dispatching, and remote control, are increasingly facing challenges. The stability and operational quality of their communication links are becoming key factors influencing grid security, intelligent dispatching, and disaster warning. Power communication networks typically consist of optical cable channels, wireless links, relay protection communication equipment, and network management systems. These networks are complex, operate in a volatile environment, and experience massive data traffic. These links are susceptible to multiple factors, including latency fluctuations, packet loss, bit errors, and interference, leading to performance degradation and even communication interruptions.

[0003] Traditional methods for assessing power communication status and diagnosing link degradation are often based on rule-based threshold setting or empirical feature matching. These methods rely primarily on manually defined criteria, such as identifying a link as abnormal if signal strength falls below a fixed threshold, or identifying anomalies through statistical metrics such as bit error rate and packet loss rate. These methods have poor environmental adaptability and struggle to handle the fuzzy boundaries and nonlinear evolution of link degradation states. Meanwhile, some existing research has introduced classifiers based on neural networks or machine learning models for intelligent link status identification. For example, these methods use decision trees, support vector machines, and LSTM models to model link operational data. However, these methods generally suffer from insufficient feature modeling capabilities, limited input information dimensionality, and a lack of modeling of inter-structural dependencies. This makes it difficult to fully reflect the coupling relationships and structural evolutionary behavior between multiple links in a communication network.

[0004] In recent years, graph neural networks (GNNs) have been gradually introduced into the field of power communication state modeling due to their strong ability to express topological structures. Some solutions attempt to model power communication topologies as graph structures, and achieve modeling and classification of device or link states through information propagation between nodes and edges. However, existing link analysis methods based on GNNs still have the following significant shortcomings: First, they fail to fully integrate the multimodal feature information of the link itself, such as simultaneously considering multiple operating parameters with physical meanings such as signal strength, interference ratio, and hop count; second, the GNN structure design is relatively fixed, lacking dynamic scheduling control of path importance and link degradation history; third, at the scoring output end, fixed thresholds or artificial experience parameters are still commonly used for link state classification, resulting in insufficient generalization ability of the model to the decision boundary, prone to false positives and false negatives, especially in the critical scoring interval. The judgment accuracy drops significantly.

[0005] Furthermore, to address the threshold optimization problem for link degradation determination, some methods have attempted to incorporate evolutionary algorithms, such as genetic algorithms and particle swarm optimization, for parameter search. However, most of these methods focus on model structure tuning or feature selection, lacking a structural design that integrates the scoring mechanism with the determination threshold optimization process. This results in the algorithm being unable to establish a closed-loop feedback loop between link score output and actual risk control, impacting the model's robustness and transferability. Regarding multi-objective assessment, existing methods rarely simultaneously consider multi-dimensional indicators such as recognition accuracy, false positive rate, missed negative rate, and early warning capability, lacking a systematic optimization mechanism for link risk prediction scenarios.

[0006] At the risk labeling level, traditional methods often directly display link determination results as normal or abnormal, lacking a structured depiction of risk propagation relationships between links. For example, when a link degrades, whether it affects adjacent links through the network topology, or whether there is a regional degradation diffusion trend, are not effectively modeled and addressed in existing methods. This makes it difficult for network-level scheduling or early warning responses to reason and intervene based on structured information.

[0007] Therefore, how to provide a power communication analysis method based on big data is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One purpose of the present invention is to propose a power communication analysis method based on big data. The present invention integrates a multimodal residual graph neural network and an improved mole search optimization algorithm to construct a closed-loop analysis mechanism of link status scoring, threshold adaptive optimization and boundary alignment training, thereby realizing intelligent identification and risk labeling of the degradation state of the power communication link. It has the advantages of high scoring accuracy, stable judgment boundary, strong real-time response and strong structural perception ability, and can effectively improve the operational safety and intelligent operation and maintenance capabilities of the power communication network. It is suitable for large-scale communication status monitoring and risk warning tasks in complex power grid scenarios.

[0009] A power communication analysis method based on big data according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect multimodal feature data of each communication link in the power communication network, aggregate them by time window, and construct a communication feature tensor;

[0011] S2. Constructing a communication graph structure based on the communication device connection relationship and the communication feature tensor;

[0012] S3. Input the communication feature tensor and the communication graph structure into a multimodal residual graph convolutional network, extract spatiotemporal dependency features, and output a link degradation score vector;

[0013] S4. Inputting the link degradation score vector into an improved mole search optimization algorithm to perform a multi-target search and generate an optimal threshold vector;

[0014] S5. Introducing the optimal threshold vector into the multimodal residual graph convolutional network training process, constructing a boundary alignment loss function, and optimizing the multimodal residual graph convolutional network;

[0015] S6. Based on the optimized multimodal residual graph convolutional network, the degradation judgment results of each link are generated for risk labeling of the power communication network.

[0016] Optionally, the multimodal feature data includes received signal strength, delay, packet loss rate, bit error rate, interference ratio and number of hops.

[0017] Optionally, the aggregation by time window specifically includes dividing the multimodal feature data of each communication link into multiple sliding time windows according to a set time interval, normalizing the feature data in each sliding time window, and calculating the statistics of each feature dimension in the sliding time window, the statistics including the mean, standard deviation, maximum and minimum values, to construct a communication feature tensor with a unified scale and time series structure.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Establishing an initial node set and an initial edge set based on the physical connection topology of the communication equipment in the power communication network;

[0020] S22. Abstract each communication device as a node, abstract the communication links actually connected between the communication devices as edges, and obtain an initial communication graph topology structure;

[0021] S23. Extract link feature information based on the multimodal feature data of each link in the communication feature tensor and generate edge attributes;

[0022] S24. Calculate the correlation between communication links by analyzing the temporal variation trend of link characteristic information and determine the correlation strength between the links;

[0023] S25. Dynamically adjust the edge connection relationship between nodes in the initial communication graph topology structure according to the association strength, including adding, deleting or updating communication link edges;

[0024] S26. The final output is a communication graph structure containing nodes, edges, and edge attributes, which reflects the dynamic dependency between links in the power communication network in real time.

[0025] Optionally, the S3 specifically includes:

[0026] S31, the multimodal residual graph convolutional network includes a cross-modal residual modulation gating unit, a graph structure convolution unit, a spatiotemporal enhancement path scheduling unit and a compressed time perception unit;

[0027] S32. Inputting the communication feature tensor and the communication graph structure into a cross-modal residual modulation gating unit. The cross-modal residual modulation gating unit constructs a gating weight matrix based on inter-modal mutual information estimation. The gating weight matrix performs element-by-element weighting on each modal feature and retains the original modal residual signal to form a parallel path structure.

[0028] S33. Perform a graph structure convolution operation using the graph structure convolution unit to extract spatial dependency features, and combine link degradation trend estimation with multi-scale similarity between nodes to expand the original edge set and generate a spatiotemporal enhanced graph structure.

[0029] S34: Input the spatiotemporal enhancement graph structure into the spatiotemporal enhancement path scheduling unit, and calculate the scheduling weight of each link based on the node modal sensitivity, historical degradation frequency, and graph structure centrality index:

[0030]

[0031] Among them, π (l) represents the scheduling weight of the lth link, represents the historical degradation frequency of the i-th node on the l-th link, γ i Denotes the modal sensitivity factor of the i-th node, Deg i represents the degree centrality of the i-th node, α, β and δ represent adjustment parameters, N represents the total number of links, represents the historical degradation frequency of the i-th node on the k-th link, EC i represents the feature embedding centrality of the i-th node, τ i represents the activation threshold of the i-th node, I[·] represents the indicator function, when The value is 1 when it is, otherwise the value is 0;

[0032] S35. The high-order node feature vector output after path scheduling is input into the compression time perception unit. The compression time perception unit adopts a two-layer nonlinear projection network and introduces the head-tail time difference through the residual connection mechanism to compensate for the long-term dependency attenuation in the compression process, performs information compression across time windows, and finally generates a link degradation score vector.

[0033] Optionally, the S4 specifically includes:

[0034] S41. Construct an initial search space based on the link degradation score vector, where the search dimension is consistent with the number of communication links, and each search dimension corresponds to a threshold value to be optimized for a communication link;

[0035] S42. Construct a multi-objective fitness function based on historical link degradation samples. The multi-objective fitness function includes four objectives: recognition accuracy, false alarm rate, missed alarm rate, and average early warning time. The four objectives are integrated using a weighted summation method to generate a fitness value.

[0036] S43. Initialize the mole search population, design the initial burrowing direction based on the distribution pattern in the link degradation score vector, simulate the mole's jumping exploration behavior in the underground layered path, and form a deep and shallow parallel search strategy;

[0037] S44, setting a geothermal response value corresponding to each search path, which is used as a search guidance signal to guide the mole to control the longitudinal expansion and lateral contraction of the path;

[0038] S45. After the maximum number of iterations is met or the fitness reaches a preset convergence threshold, the path threshold combination corresponding to the mole individual with the highest fitness is selected from the current search population as the optimal threshold vector.

[0039] Optionally, the S5 specifically includes:

[0040] S51, obtaining an optimal threshold vector, where each component in the optimal threshold vector corresponds to an optimal decision threshold for a link;

[0041] S52. During the training of the multimodal residual graph convolutional network, the link degradation score vector is used as the output target, and the optimal threshold vector is used as the boundary label signal, which together with the historical true degradation label constitutes a joint supervision input;

[0042] S53. Define a boundary alignment loss function, comprehensively consider the relative position relationship between the link degradation score vector and the optimal threshold vector, and combine the true degradation label to achieve boundary sensitivity control of the score output;

[0043] S54. The boundary alignment loss function is introduced as an additional term into the training objective of the multimodal residual graph convolutional network, and works together with the original classification loss to jointly drive the update process of the multimodal residual graph convolutional network.

[0044] Optionally, the S53 specifically includes:

[0045] Define the link degradation score vector as s=[s1,s2,…,s N ], the optimal threshold vector is θ=[θ1,θ2,…,θ N ], the true label vector is y=[y1,y2,…,y N ], where N represents the total number of links, s i represents the degradation score of the i-th link, θ irepresents the optimal decision threshold of the i-th link, y i represents the true degradation label of the i-th link. When the i-th link is manually marked as degraded, y i The value is 1; when the i-th link does not trigger any degradation event, y i Taking the value as 0, the boundary alignment loss function is defined as:

[0046]

[0047] Among them, L align represents the boundary alignment loss function, λ1 and λ2 represent the loss weighting factors, σ represents the activation function, ξ represents the boundary sensitive bandwidth, I[·] represents the indicator function, when |s i -θ i The value is 1 when |<ξ, otherwise it is 0.

[0048] Optionally, the S6 specifically includes:

[0049] S61. Input the currently collected multimodal feature data of the communication link into the updated multimodal residual graph convolutional network to obtain an updated link degradation score vector, where each component corresponds to a degradation score value of a communication link;

[0050] S62: Generate an updated optimal threshold vector based on the updated link degradation score vector input, perform link-by-link comparison and judgment, and determine that the link is in a degraded state if the degradation score value of the link is greater than the corresponding optimal threshold;

[0051] S63. Based on the determination result, construct a link degradation state label vector, where each component takes a value of 0 or 1, indicating whether the corresponding link currently has a degradation risk;

[0052] S64. Mark the links in the degraded state as high-risk edges in the communication graph structure, set the corresponding nodes as alarm state nodes, and construct a risk propagation graph structure;

[0053] S65. Evaluate overall communication stability based on the link degradation determination result to determine whether there is a potential regional communication risk;

[0054] S66. Visualize the risk annotation of link degradation judgment results, including link-level risk layer annotation, dynamic alarm update of power communication topology, and link visualization coloring feedback, to assist operation and maintenance personnel in making proactive early warning responses and fault handling decisions.

[0055] Optionally, the overall communication stability includes the proportion of degraded links, the density of degraded areas, and the path connectivity between degraded links.

[0056] The beneficial effects of the present invention are:

[0057] First, the present invention adopts multimodal feature acquisition and a unified time window processing mechanism to realize information fusion of the operating status of the communication link in multiple dimensions. It can construct a feature tensor with time consistency and scale consistency based on physical quantities such as received signal strength, delay, bit error rate, packet loss rate, interference ratio and number of hops, effectively enhancing the model's ability to perceive the overall picture of the link operation.

[0058] Secondly, the present invention introduces a dynamic communication graph construction mechanism in terms of graph structure modeling. It not only converts the physical connection relationship between communication devices into a basic graph structure of nodes and edges, but also extracts the correlation between edge attributes and links by analyzing the trend of the evolution of the multimodal characteristics of the link over time, and dynamically updates the topological structure, thereby realizing the time-series adaptive reconstruction of the communication graph and significantly improving the accuracy and real-time performance of the dependency modeling between links. On this basis, the constructed multimodal residual graph convolutional network further introduces a modal mutual information gating mechanism, a spatiotemporal enhanced path scheduling structure, and a compressed sensing module, which not only improves the model's ability to recognize the importance differences between modalities, but also dynamically adjusts the information propagation path through factors such as node centrality and degradation frequency, realizing the regulation of deep and shallow residual parallel paths, and effectively improving the expression accuracy and stability of the degradation score.

[0059] Furthermore, this paper addresses the shortcomings of traditional methods, such as rigid link status determination boundaries and fixed thresholds, by innovatively introducing an improved mole search optimization algorithm. This algorithm simulates the jumping digging behavior of moles in strata, combines scoring heat to construct a geothermal response mechanism, and introduces collaborative perturbations and jump depth control in group search. This results in a dynamic threshold search model in a high-dimensional multimodal space. This optimal threshold is not only individually adaptive but also forms a closed-loop optimization feedback loop with the link scoring results, thereby improving the model's accuracy and fault tolerance near the determination boundary.

[0060] Finally, the present invention further proposes a boundary alignment loss function design scheme. By jointly introducing the relative offset between the score value and the optimal threshold and the supervisory information of the true label, the boundary region score output behavior is dynamically adjusted during the network training process, so that the model can automatically focus on improving the discrimination ability of critical states and avoid a large number of misjudgments at the junction of degraded and normal states. In addition, the present invention also constructs a graph-level visual annotation scheme for link degradation states. Combined with the communication topology structure, it realizes degradation path coloring, node alarm triggering, and the construction of regional risk propagation maps. This allows the link judgment results to be used for proactive intervention in the scheduling system, enhancing the overall safety and controllability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying 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 of the present invention. In the accompanying drawings:

[0062] Figure 1 This is an overall flow chart of a power communication analysis method based on big data proposed by the present invention;

[0063] Figure 2 This is a schematic diagram of the multimodal residual graph convolutional network structure of a big data-based power communication analysis method proposed in the present invention;

[0064] Figure 3 This is a search behavior flow chart of the improved mole search optimization algorithm for the big data-based power communication analysis method proposed in the present invention. DETAILED DESCRIPTION

[0065] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0066] refer to Figure 1-3 , a power communication analysis method based on big data, comprising the following steps:

[0067] S1. Collect multimodal feature data of each communication link in the power communication network, aggregate them by time window, and construct a communication feature tensor;

[0068] S2. Constructing a communication graph structure based on the communication device connection relationship and the communication feature tensor;

[0069] S3. Input the communication feature tensor and the communication graph structure into a multimodal residual graph convolutional network, extract spatiotemporal dependency features, and output a link degradation score vector;

[0070] S4. Inputting the link degradation score vector into an improved mole search optimization algorithm to perform a multi-target search and generate an optimal threshold vector;

[0071] S5. Introducing the optimal threshold vector into the multimodal residual graph convolutional network training process, constructing a boundary alignment loss function, and optimizing the multimodal residual graph convolutional network;

[0072] S6. Based on the optimized multimodal residual graph convolutional network, the degradation judgment results of each link are generated for risk labeling of the power communication network.

[0073] The present invention constructs a set of intelligent analysis processes for power communication links driven by big data, forming a complete closed loop from feature acquisition, graph structure modeling, deep scoring, threshold optimization to risk judgment and visualization. It breaks through the limitations of existing technologies that rely on static thresholds, single-dimensional features and rule-driven methods, and significantly improves the perception of communication link status and the accuracy of anomaly recognition. In particular, the deep integration of multimodal residual graph convolutional networks and improved mole search algorithms realizes the linkage of boundary sensitivity learning and scoring optimization, effectively solving the problems of fuzzy degradation state judgment and high false alarm rate. This method is applicable to various high-reliability communication scenarios, providing more efficient and intelligent technical support for power grid communication monitoring and risk warning.

[0074] In this embodiment, the multimodal feature data includes received signal strength, delay, packet loss rate, bit error rate, interference ratio and hop count.

[0075] In this embodiment, the aggregation by time window specifically includes dividing the multimodal feature data of each communication link into multiple sliding time windows according to a set time interval, normalizing the feature data in each sliding time window, and calculating the statistics of each feature dimension in the sliding time window, wherein the statistics include the mean, standard deviation, maximum and minimum values, and constructing a communication feature tensor with a unified scale and time series structure.

[0076] By introducing a sliding time window and normalization processing mechanism, this paper takes into account both the temporal continuity and scale consistency of data when constructing the communication feature tensor, effectively avoiding the distortion and drift problems of traditional static data aggregation when dealing with the dynamic evolution of links. By statistically calculating the mean, standard deviation, maximum, and minimum values of the features within each time window, the constructed tensor has good time sensitivity and feature differentiation, providing a more representative input expression for graph neural networks to learn the evolution of link states, improving the overall model's time adaptability and recognition accuracy.

[0077] In this embodiment, S2 specifically includes:

[0078] S21. Establishing an initial node set and an initial edge set based on the physical connection topology of the communication equipment in the power communication network;

[0079] S22. Abstract each communication device as a node, abstract the communication links actually connected between the communication devices as edges, and obtain an initial communication graph topology structure;

[0080] S23. Extract link feature information based on the multimodal feature data of each link in the communication feature tensor and generate edge attributes;

[0081] S24. Calculate the correlation between communication links by analyzing the temporal variation trend of link characteristic information and determine the correlation strength between the links;

[0082] S25. Dynamically adjust the edge connection relationship between nodes in the initial communication graph topology structure according to the association strength, including adding, deleting or updating communication link edges;

[0083] S26. The final output is a communication graph structure containing nodes, edges, and edge attributes, which reflects the dynamic dependency between links in the power communication network in real time.

[0084] In constructing the communication graph structure, this paper proposes a mechanism for evolving from a static physical topology to a dynamic, associative topology. This not only leverages the physical connections between devices as the basis for graph modeling, but also dynamically updates the edge structure by leveraging the correlations between multimodal features and temporal trends, thereby improving the real-time and robustness of the communication graph. This mechanism effectively reflects potential degraded propagation paths between links, providing graph neural network modeling with topological input that better reflects the actual network operating state, significantly enhancing the model's ability to model and respond efficiently to abnormal propagation behaviors in complex network structures.

[0085] In this embodiment, S3 specifically includes:

[0086] S31, the multimodal residual graph convolutional network includes a cross-modal residual modulation gating unit, a graph structure convolution unit, a spatiotemporal enhancement path scheduling unit and a compressed time perception unit;

[0087] S32. Inputting the communication feature tensor and the communication graph structure into a cross-modal residual modulation gating unit. The cross-modal residual modulation gating unit constructs a gating weight matrix based on inter-modal mutual information estimation. The gating weight matrix performs element-by-element weighting on each modal feature and retains the original modal residual signal to form a parallel path structure.

[0088] S33. Perform a graph structure convolution operation using the graph structure convolution unit to extract spatial dependency features, and combine link degradation trend estimation with multi-scale similarity between nodes to expand the original edge set and generate a spatiotemporal enhanced graph structure.

[0089] S34: Input the spatiotemporal enhancement graph structure into the spatiotemporal enhancement path scheduling unit, and calculate the scheduling weight of each link based on the node modal sensitivity, historical degradation frequency, and graph structure centrality index:

[0090]

[0091] Among them, π (l) represents the scheduling weight of the lth link, represents the historical degradation frequency of the i-th node on the l-th link, γi Denotes the modal sensitivity factor of the i-th node, Deg i represents the degree centrality of the i-th node, α, β and δ represent adjustment parameters, N represents the total number of links, represents the historical degradation frequency of the i-th node on the k-th link, EC i represents the feature embedding centrality of the i-th node, τ i represents the activation threshold of the i-th node, I[·] represents the indicator function, when The value is 1 when it is, otherwise the value is 0;

[0092] S35. The high-order node feature vector output after path scheduling is input into the compression time perception unit. The compression time perception unit adopts a two-layer nonlinear projection network and introduces the head-tail time difference through the residual connection mechanism to compensate for the long-term dependency attenuation in the compression process, performs information compression across time windows, and finally generates a link degradation score vector.

[0093] The multimodal residual graph convolutional network structure proposed in this invention integrates multiple innovative modules such as cross-modal gating, graph convolution, and path scheduling to construct a graph neural structure with strong dynamic adjustment capabilities for inter-link dependencies and inter-modal weights. Among them, the cross-modal residual modulation gating unit can achieve differential capture and residual preservation of different feature modes, and the path scheduling module uses centrality and degradation history to dynamically select propagation paths, thereby improving the pertinence and efficiency of information transmission. This network structure significantly enhances the robustness and link scoring accuracy of the model in multimodal and multi-link interference environments, and improves the expressiveness and interpretability of degradation states.

[0094] In this embodiment, the S4 specifically includes:

[0095] S41. Construct an initial search space based on the link degradation score vector, where the search dimension is consistent with the number of communication links, and each search dimension corresponds to a threshold value to be optimized for a communication link;

[0096] S42. Construct a multi-objective fitness function based on historical link degradation samples. The multi-objective fitness function includes four objectives: recognition accuracy, false alarm rate, missed alarm rate, and average early warning time. The four objectives are integrated using a weighted summation method to generate a fitness value.

[0097] S43. Initialize the mole search population, design the initial burrowing direction based on the distribution pattern in the link degradation score vector, simulate the mole's jumping exploration behavior in the underground layered path, and form a deep and shallow parallel search strategy;

[0098] S44, setting a geothermal response value corresponding to each search path, which is used as a search guidance signal to guide the mole to control the longitudinal expansion and lateral contraction of the path;

[0099] S45. After the maximum number of iterations is met or the fitness reaches a preset convergence threshold, the path threshold combination corresponding to the mole individual with the highest fitness is selected from the current search population as the optimal threshold vector.

[0100] This paper applies an improved mole search algorithm to the optimization process of link scoring thresholds. It proposes a geothermal feedback mechanism based on scoring heat and simulates jumping search behavior, constructing a multi-target search algorithm structure that better suits the characteristics of power communication environments. By introducing a fitness function to evaluate recognition accuracy, false alarm rate, missed alarm rate, and early warning performance, a more discriminative search target is constructed. At the same time, heat response guides path adjustment, improving the search's global exploration and local convergence capabilities. This optimization mechanism significantly enhances the adaptability of threshold setting and the accuracy of classification boundaries, reducing the errors caused by manually setting thresholds.

[0101] In this paper, the geothermal response value serves as an important feedback signal guiding the path jump direction and search depth in the improved mole search algorithm. It is used to simulate the response behavior of moles to geothermal gradients in reality, enhancing the adaptability and target sensitivity of the search path. The geothermal response value is jointly set based on three factors: the link score distribution characteristics, the score stability, and the historical activity of the path. Specifically, the following steps are included:

[0102] First, the link degradation scores of the relevant nodes in each search path are obtained. The average score of the path in the current search round is calculated to measure the basic level of popularity of the path. A higher score indicates a more obvious degradation trend of the links contained in the path and a higher popularity.

[0103] Secondly, the variance and extreme value ratio of the score values for each link on the path are calculated to assess the volatility of the score. If the variance is large and the score distribution is unstable, the corresponding path may have drastic changes or inconsistent degradation signals. The system will adjust the geothermal value accordingly to reduce the risk of search resource bias.

[0104] Next, we count the jump activity of the path in the historical search rounds, that is, record the frequency of the path being visited or jumped by the search individuals. If a path is frequently jumped in history, the system will reduce its geothermal response value by introducing an activity suppression factor to avoid over-exploitation of the path and thus reduce the global search efficiency.

[0105] Ultimately, these three factors are weighted together to determine the geothermal response value for each path. This value serves as a reference signal for the current searcher's digging direction, depth expansion strategy, and jump range decisions. A higher geothermal response value indicates a searcher's tendency to make deeper jumps along the path, while a lower value indicates a tendency to dig laterally or compress the path.

[0106] Based on the traditional mole search optimization algorithm, this paper proposes a number of structural improvements targeting the high-dimensional non-convex search characteristics of the power communication link degradation threshold determination problem. The specific improvements include the following aspects:

[0107] Introducing a score-driven mechanism for setting the initial burrowing direction: In traditional mole-and-rat search algorithms, individual initial search directions are typically randomly generated or set using simple heuristics, lacking integration with target data characteristics. This invention designs a risk-gradient-based burrowing initiation strategy based on the distribution of link degradation score vectors. This prioritizes the search process in high-risk link regions, improving the effectiveness and focus of the initial global search.

[0108] Design of a dynamic path control mechanism driven by geothermal response value: To simulate the response behavior of moles to temperature gradients in multi-layer geological structures, the present invention constructs a path-level geothermal response value as a search guidance signal. This response value is dynamically calculated based on the link score mean, score stability (variance), and path historical activity, and is used to control the vertical jump depth and lateral expansion range of the mole's search path, thereby realizing thermal perception regulation and structural adaptive adjustment of the search path.

[0109] Constructing a multi-objective fitness function: Traditional mole search algorithms often focus on minimizing error as a single optimization objective, failing to account for the multiple judgment criteria used in real-world scenarios. This new approach incorporates four performance metrics—recognition accuracy, false positive rate, false negative rate, and average lead time—into the fitness function. These metrics are then converted into a unified optimization objective using a weighted summation, ensuring optimal search results across multiple metrics.

[0110] Introducing a collaborative perturbation mechanism to enhance global escape capabilities: During the iterative search process, to prevent individuals from being trapped in local optimal solutions, this paper proposes a collaborative perturbation mechanism based on degenerate score correlation. By comparing the similarity of score heat between paths, individual moles are allowed to share high-quality path information. Furthermore, perturbation jumps are introduced within the group, enhancing global exploration capabilities and the diversity of the solution space.

[0111] Supports closed-loop interaction between the optimal threshold vector and the neural network: Traditional optimization algorithms often rely on independent solution structures, making it difficult to couple optimization processes with deep models. This method incorporates the optimal threshold vector output by the mole search algorithm as a supervisory signal into the graph neural network training process, constructs a boundary alignment loss function, and forms a closed-loop learning mechanism between the score, threshold, and model, achieving joint optimization of the soft decision boundary.

[0112] In this embodiment, the S5 specifically includes:

[0113] S51, obtaining an optimal threshold vector, where each component in the optimal threshold vector corresponds to an optimal decision threshold for a link;

[0114] S52. During the training of the multimodal residual graph convolutional network, the link degradation score vector is used as the output target, and the optimal threshold vector is used as the boundary label signal, which together with the historical true degradation label constitutes a joint supervision input;

[0115] S53. Define a boundary alignment loss function, comprehensively consider the relative position relationship between the link degradation score vector and the optimal threshold vector, and combine the true degradation label to achieve boundary sensitivity control of the score output;

[0116] S54. The boundary alignment loss function is introduced as an additional term into the training objective of the multimodal residual graph convolutional network, and works together with the original classification loss to jointly drive the update process of the multimodal residual graph convolutional network.

[0117] The present invention reversely introduces the optimal threshold vector into the graph neural network training process, and by constructing a boundary alignment loss function, it achieves adaptive correction and dynamic fine-tuning of the classification boundary during model training. This design breaks through the bottleneck of the traditional training process that ignores the sensitivity of the classification critical interval, enabling the model to enhance its judgment ability near the boundary, especially improving the recognition accuracy of "critical state" links. The introduction of the boundary alignment mechanism also helps to stabilize the score output range, avoid large fluctuations around the critical value, and improve the overall training quality and judgment stability of the model.

[0118] In this embodiment, the S53 specifically includes:

[0119] Define the link degradation score vector as s=[s1,s2,…,s N ], the optimal threshold vector is θ=[θ1,θ2,…,θ N ], the true label vector is y=[y1,y2,…,y N ], where N represents the total number of links, s i represents the degradation score of the i-th link, θ i represents the optimal decision threshold of the i-th link, y irepresents the true degradation label of the i-th link. When the i-th link is manually marked as degraded, y i The value is 1; when the i-th link does not trigger any degradation event, y i Taking the value as 0, the boundary alignment loss function is defined as:

[0120]

[0121] Among them, L align represents the boundary alignment loss function, λ1 and λ2 represent the loss weighting factors, σ represents the activation function, ξ represents the boundary sensitive bandwidth, I[·] represents the indicator function, when |s i -θ i The value is 1 when |<ξ, otherwise it is 0.

[0122] By explicitly constructing a ternary relationship between the link score, optimal threshold, and true label, this paper introduces a Sigmoid score fitting term and a critical interval boundary penalty term into the boundary alignment loss function, achieving dual error control based on boundary-sensitive intervals. This structure not only enables the score to be more accurately aligned with the true state label, but also introduces additional loss weights near the decision boundary, guiding the model to focus more on the critical point learning effect. In practical applications, this mechanism can effectively reduce false positives and false negatives, improving the model's engineering usability and prediction credibility.

[0123] In this embodiment, S6 specifically includes:

[0124] S61. Input the currently collected multimodal feature data of the communication link into the updated multimodal residual graph convolutional network to obtain an updated link degradation score vector, where each component corresponds to a degradation score value of a communication link;

[0125] S62: Generate an updated optimal threshold vector based on the updated link degradation score vector input, perform link-by-link comparison and judgment, and determine that the link is in a degraded state if the degradation score value of the link is greater than the corresponding optimal threshold;

[0126] S63. Based on the determination result, construct a link degradation state label vector, where each component takes a value of 0 or 1, indicating whether the corresponding link currently has a degradation risk;

[0127] S64. Mark the links in the degraded state as high-risk edges in the communication graph structure, set the corresponding nodes as alarm state nodes, and construct a risk propagation graph structure;

[0128] S65. Evaluate overall communication stability based on the link degradation determination result to determine whether there is a potential regional communication risk;

[0129] S66. Visualize the risk annotation of link degradation judgment results, including link-level risk layer annotation, dynamic alarm update of power communication topology, and link visualization coloring feedback, to assist operation and maintenance personnel in making proactive early warning responses and fault handling decisions.

[0130] After the model outputs the scoring results, the present invention further constructs a link degradation state map. Link-level risk labels are generated based on the score and threshold comparison results. High-risk links and alarm nodes are explicitly labeled based on the map structure, creating a structured risk propagation diagram. This fusion of state determination and structural mapping allows link risk to be reflected not only in the results of single-link determinations, but also in the local or regional degradation trends of the entire communication network, helping operations and maintenance personnel achieve visual decision-making and proactive intervention within the scheduling system.

[0131] In this embodiment, the overall communication stability includes the proportion of degraded links, the density of degraded areas, and the path connectivity between degraded links. By introducing the proportion of degraded links, the density of degraded areas, and the link path connectivity indicators, the present invention constructs a set of communication stability quantitative indicator systems suitable for power communication systems, which can provide network-level stability assessment methods based on link risk assessment. This assessment structure can support regional-level communication health measurement, risk distribution thermal analysis, and macro-warning mechanisms, providing quantifiable, calculable, and early-warning technical support for system-level operation and maintenance and resource scheduling, significantly improving the overall network management intelligence level.

[0132] Example 1:

[0133] To verify the feasibility of this invention, we applied it to the power communication subnet of a 220kV transmission and transformation regional dispatch center of the State Grid Corporation of China. This subnet covers an area of approximately 120 square kilometers. The communication links are primarily fiber-optic ring networks, supplemented by wireless frequency-hopping links and power carrier channels, totaling 172 communication links and 36 communication exchange nodes. This network has long faced problems such as intermittent link degradation, difficulty in timely warnings, and low manual inspection efficiency. Traditional rule-based triggering mechanisms often suffer from missed and false alarms, especially at night or during high-load periods.

[0134] In this example, we implemented the big data-based power communication analysis method proposed in this invention. The system first collects multimodal feature data for each communication link through optical communication monitors, network exchange logs, and link quality sensing modules. These data primarily include received signal strength (RSSI), link latency, packet loss rate, bit error rate, interference ratio, and hop count. To ensure data timeliness, we aggregate the data using a 5-minute sliding window. We then normalize and statistically process the features within each window, ultimately forming a unified-scale communication feature tensor.

[0135] The system then constructs an initial communication graph structure based on the power communication topology and dynamically adjusts the connection relationships between nodes based on the correlation of inter-link features, forming a time-varying graph. The communication feature tensor and the graph structure are fed into a multimodal residual graph convolutional network, which extracts the spatial correlation and temporal evolution patterns of nodes and outputs a degradation score for each link.

[0136] Based on the scoring results, we constructed a multi-objective search model using an improved mole search optimization algorithm. This model considers the false positive rate, false negative rate, accuracy, and lead time to search for the optimal scoring threshold combination. Notably, this algorithm incorporates a geothermal response mechanism, calculating the adaptability of the search direction based on the popularity of the path score, score fluctuations, and historical jump frequency, effectively preventing the search from becoming trapped in a local optimum.

[0137] After comparing the scoring output with the threshold, the system determines whether a link is degraded and annotates the risk propagation diagram on the communication map. Finally, the degraded links are colored and displayed with dynamic alarms on the visualization platform, providing dispatchers with actionable linkage suggestions and accurate fault warnings.

[0138] The experimental results show that within 72 hours of continuous operation, the system detected a total of 28 degraded links, of which 21 were real degraded links and 7 were weakly degraded links that did not develop into a fault state after early warning. Compared with the results of manual maintenance, the comprehensive recognition accuracy reached 92.4%, the false alarm rate was less than 3.7%, and an average early warning of about 13 minutes was achieved, which is significantly better than the traditional static threshold method (with an average accuracy of less than 80% and no early warning capability). Therefore, the present invention can realize adaptive identification of link degradation trends, dynamic adjustment of scoring boundaries, and auxiliary judgment of risk propagation in high-complexity, large-scale link status monitoring scenarios. It has the advantages of strong practicality, high stability, and low deployment cost, and has good promotion value.

[0139] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A power communication analysis method based on big data, characterized in that: The steps include: S1. Collect multimodal feature data of each communication link in the power communication network, aggregate them by time window, and construct a communication feature tensor; S2. Constructing a communication graph structure based on the communication device connection relationship and the communication feature tensor; S3. Input the communication feature tensor and the communication graph structure into a multimodal residual graph convolutional network, extract spatiotemporal dependency features, and output a link degradation score vector; S4. Inputting the link degradation score vector into an improved mole search optimization algorithm to perform a multi-target search and generate an optimal threshold vector; S5. Introducing the optimal threshold vector into the multimodal residual graph convolutional network training process, constructing a boundary alignment loss function, and optimizing the multimodal residual graph convolutional network; S6. Based on the optimized multimodal residual graph convolutional network, the degradation judgment results of each link are generated for risk labeling of the power communication network.

2. The power communication analysis method based on big data according to claim 1, characterized in that: The multimodal feature data includes received signal strength, delay, packet loss rate, bit error rate, interference ratio and hop count.

3. The power communication analysis method based on big data according to claim 1, characterized in that: The aggregation by time window specifically includes dividing the multimodal feature data of each communication link into multiple sliding time windows according to a set time interval, normalizing the feature data in each sliding time window, and calculating the statistics of each feature dimension in the sliding time window. The statistics include the mean, standard deviation, maximum and minimum values, and constructing a communication feature tensor with a unified scale and time series structure.

4. The power communication analysis method based on big data according to claim 1, characterized in that: The S2 specifically includes: S21. Establishing an initial node set and an initial edge set based on the physical connection topology of the communication equipment in the power communication network; S22. Abstract each communication device as a node, abstract the communication links actually connected between the communication devices as edges, and obtain an initial communication graph topology structure; S23. Extract link feature information based on the multimodal feature data of each link in the communication feature tensor and generate edge attributes; S24. Calculate the correlation between communication links by analyzing the temporal variation trend of link characteristic information and determine the correlation strength between the links; S25. Dynamically adjust the edge connection relationship between nodes in the initial communication graph topology structure according to the association strength, including adding, deleting or updating communication link edges; S26. The final output is a communication graph structure containing nodes, edges, and edge attributes, which reflects the dynamic dependency between links in the power communication network in real time.

5. The power communication analysis method based on big data according to claim 1, characterized in that: The S3 specifically includes: S31, the multimodal residual graph convolutional network includes a cross-modal residual modulation gating unit, a graph structure convolution unit, a spatiotemporal enhancement path scheduling unit and a compressed time perception unit; S32, inputting the communication feature tensor and the communication graph structure into a cross-modal residual modulation gating unit, wherein the cross-modal residual modulation gating unit constructs a gating weight matrix based on inter-modal mutual information estimation; S33. Perform a graph structure convolution operation using the graph structure convolution unit to extract spatial dependency features, and combine link degradation trend estimation with multi-scale similarity between nodes to expand the original edge set and generate a spatiotemporal enhanced graph structure. S34: Input the spatiotemporal enhancement graph structure into the spatiotemporal enhancement path scheduling unit, and calculate the scheduling weight of each link based on the node modal sensitivity, historical degradation frequency, and graph structure centrality index: Among them, π (l) represents the scheduling weight of the lth link, represents the historical degradation frequency of the i-th node on the l-th link, γ i Denotes the modal sensitivity factor of the i-th node, Deg i represents the degree centrality of the i-th node, α, β and δ represent adjustment parameters, N represents the total number of links, represents the historical degradation frequency of the i-th node on the k-th link, EC i represents the feature embedding centrality of the i-th node, τ i represents the activation threshold of the i-th node, I[·] represents the indicator function, when The value is 1 when it is, otherwise the value is 0; S35 , inputting the high-order node feature vector representation output after path scheduling into the compressed time perception unit, performing information compression across time windows, and finally generating a link degradation score vector.

6. The power communication analysis method based on big data according to claim 1, characterized in that: The S4 specifically includes: S41. Construct an initial search space based on the link degradation score vector, where the search dimension is consistent with the number of communication links, and each search dimension corresponds to a threshold value to be optimized for a communication link; S42. Construct a multi-objective fitness function based on historical link degradation samples. The multi-objective fitness function includes four objectives: recognition accuracy, false alarm rate, missed alarm rate, and average early warning time. The four objectives are integrated using a weighted summation method to generate a fitness value. S43. Initialize the mole search population, design the initial burrowing direction based on the distribution pattern in the link degradation score vector, simulate the mole's jumping exploration behavior in the underground layered path, and form a deep and shallow parallel search strategy; S44, setting a geothermal response value corresponding to each search path, which is used as a search guidance signal to guide the mole to control the longitudinal expansion and lateral contraction of the path; S45. After the maximum number of iterations is met or the fitness reaches a preset convergence threshold, the path threshold combination corresponding to the mole individual with the highest fitness is selected from the current search population as the optimal threshold vector.

7. The power communication analysis method based on big data according to claim 1, characterized in that: The S5 specifically includes: S51, obtaining an optimal threshold vector, where each component in the optimal threshold vector corresponds to an optimal decision threshold for a link; S52. During the training of the multimodal residual graph convolutional network, the link degradation score vector is used as the output target, and the optimal threshold vector is used as the boundary label signal, which together with the historical true degradation label constitutes a joint supervision input; S53. Define a boundary alignment loss function, comprehensively consider the relative position relationship between the link degradation score vector and the optimal threshold vector, and combine the true degradation label to achieve boundary sensitivity control of the score output; S54. The boundary alignment loss function is introduced as an additional term into the training objective of the multimodal residual graph convolutional network, and works together with the original classification loss to jointly drive the update process of the multimodal residual graph convolutional network.

8. The power communication analysis method based on big data according to claim 7, characterized in that: The S53 specifically includes: Define the link degradation score vector as s=[s1,s2,…,s N ], the optimal threshold vector is θ=[θ1,θ2,…,θ N ], the true label vector is y=[y1,y2,…,y N ], where N represents the total number of links, s i represents the degradation score of the i-th link, θ i represents the optimal decision threshold of the i-th link, y i represents the true degradation label of the i-th link. When the i-th link is manually marked as degraded, y i The value is 1; when the i-th link does not trigger any degradation event, y i Taking the value as 0, the boundary alignment loss function is defined as: Among them, L align represents the boundary alignment loss function, λ1 and λ2 represent the loss weighting factors, σ represents the activation function, ξ represents the boundary sensitive bandwidth, I[·] represents the indicator function, when |s i -θ i The value is 1 when |<ξ, otherwise it is 0.

9. The method for analyzing power communication based on big data according to claim 1, characterized in that: The S6 specifically includes: S61. Input the currently collected multimodal feature data of the communication link into the updated multimodal residual graph convolutional network to obtain an updated link degradation score vector, where each component corresponds to a degradation score value of a communication link; S62: Generate an updated optimal threshold vector based on the updated link degradation score vector input, perform link-by-link comparison and judgment, and determine that the link is in a degraded state if the degradation score value of the link is greater than the corresponding optimal threshold; S63. Based on the determination result, construct a link degradation state label vector, where each component takes a value of 0 or 1, indicating whether the corresponding link currently has a degradation risk; S64. Mark the links in the degraded state as high-risk edges in the communication graph structure, set the corresponding nodes as alarm state nodes, and construct a risk propagation graph structure; S65. Evaluate overall communication stability based on the link degradation determination result to determine whether there is a potential regional communication risk; S66. Visualize the risk annotation of link degradation judgment results, including link-level risk layer annotation, dynamic alarm update of power communication topology, and link visualization coloring feedback, to assist operation and maintenance personnel in making proactive early warning responses and fault handling decisions.

10. The power communication analysis method based on big data according to claim 9, characterized in that: The overall communication stability includes the proportion of degraded links, the density of degraded areas, and the path connectivity between degraded links.

Images