A deep learning-based spatial channel dynamic characteristic prediction system

Abstract

The application relates to the technical field of intelligent application of communication networks, in particular to a spatial channel dynamic characteristic prediction system based on deep learning. The system comprises the following modules: a non-stationary topology detection module, which is used for obtaining broken-link communication data; non-stationary topology detection is carried out according to the broken-link communication data, and non-stationary topology data is obtained; a subset cluster feasible communication domain construction module, which is used for carrying out subset cluster feasible communication domain construction according to the non-stationary topology data, and obtaining subset cluster feasible communication domain data; a communication uncertainty prediction module, which is used for carrying out communication uncertainty prediction according to the subset cluster feasible communication domain data, and obtaining communication uncertainty data; a cross-domain opportunity action learning module, which is used for carrying out broken-link topology local learning according to the communication uncertainty data, and obtaining local prediction data; cross-domain opportunity action learning is carried out according to the local prediction data, and broken-link cross-domain reconnection data is obtained. The application can realize dynamic reconnection strategy optimization under the environment of broken-link communication of a UAV cluster.

Description

Technical Field

[0001] This invention relates to the field of intelligent application technology in communication networks, and in particular to a spatial channel dynamic characteristic prediction system based on deep learning. Background Technology

[0002] In highly dynamic network scenarios such as complex terrain, post-disaster emergency response, or drone swarms, the spatial channel connectivity between communication nodes is highly susceptible to multiple influences, including terrain obstruction, electromagnetic interference, energy attenuation, and high-frequency node maneuvering. This often leads to frequent "break-reconstruction-rebreak" processes in the network topology within a short timescale. Such phenomena not only cause instantaneous loss of global connectivity but also leave local sub-clusters in a state of long-term isolation or instability, posing serious challenges to task scheduling, link maintenance, and collaborative control. Traditional communication reconstruction schemes are mostly based on topology assumptions of "eventually connectable" or "central node-dominated," relying on static configuration or global information convergence for link maintenance. These schemes are ill-suited to the aforementioned sudden and localized topology disturbances, thus reducing the network's survivability and autonomous recovery capabilities in complex scenarios. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a deep learning-based spatial channel dynamic characteristic prediction system to solve at least one of the aforementioned technical problems.

[0004] This application provides a deep learning-based spatial channel dynamic characteristic prediction system, including:

[0005] The non-stationary topology detection module is used to acquire broken link communication data; based on the broken link communication data, non-stationary topology detection is performed to obtain non-stationary topology data;

[0006] The sub-cluster feasible communication domain construction module is used to construct the sub-cluster feasible communication domain based on non-stationary topology data, and obtain the sub-cluster feasible communication domain data.

[0007] The communication uncertainty prediction module is used to predict communication uncertainty based on the feasible communication domain data of the sub-cluster, and obtain communication uncertainty data.

[0008] The cross-domain opportunistic action learning module is used to perform local learning of the disconnected topology based on communication uncertainty data to obtain local prediction data; and to perform cross-domain opportunistic action learning based on the local prediction data to obtain cross-domain reconnection data after a disconnection.

[0009] This invention utilizes non-stationary topology detection and feasible communication domain construction for sub-clusters to provide a structured representation of spatial channel states in complex scenarios characterized by frequent communication link breaks and rapid topology evolution. This avoids misjudging instantaneous link breaks as stable failures, thereby improving the accuracy of perception of dynamic channel changes from the source. The system does not make decisions directly based on the current communication state, but rather uses a communication uncertainty prediction module to proactively assess channel stability over future periods. This ensures the learning and decision-making process has temporal continuity and predictive constraints, reducing the risk of strategy oscillations caused by short-term fluctuations. The cross-domain opportunistic action learning module learns and filters cross-domain reconnection behaviors based on local prediction results, avoiding blind or forced reconnection operations that could interfere with existing stable sub-clusters, achieving a balance between success probability and system stability in the communication reconstruction process.

[0010] Preferably, the non-stationary topology detection includes:

[0011] During the operation of the drone swarm, disconnected communication data is periodically collected from each communication node;

[0012] Based on the broken-link communication data, a communication topology snapshot is processed to obtain communication topology snapshot data;

[0013] Topology change features are extracted from communication topology snapshot data to obtain topology change feature data;

[0014] Nonstationary discriminant features are calculated based on topological change feature data to obtain nonstationary discriminant feature data;

[0015] Nonstationary detection is performed on nonstationary discriminant feature data to obtain nonstationary topological data.

[0016] This invention, through the aforementioned non-stationary topology detection process, elevates the frequent link breakage phenomena in UAV swarm communication from "isolated events" to a "temporal topology evolution problem" for processing. By periodically collecting link breakage communication data and constructing communication topology snapshots, the communication state is completely preserved in a structured form, avoiding the one-sidedness of judging based solely on a single link indicator. Through the extraction and evolutionary analysis of topology change characteristics, the system can identify persistent structural changes in the communication topology caused by maneuvering, obstruction, or energy differences, rather than merely responding to instantaneous fluctuations. Combined with a non-stationarity discrimination feature calculation and detection mechanism, the system can effectively distinguish between "short-term disturbance-type changes" and non-stationary states where "evolutionary mechanisms switch," thereby reducing misjudgments and omissions.

[0017] Preferably, the topology change feature extraction includes:

[0018] Structural alignment is performed based on communication topology snapshot data to obtain structural alignment data;

[0019] Connectivity configuration morphology data is obtained by performing connectivity configuration morphology processing on the structure alignment data.

[0020] Cross-temporal offset calculations are performed on the connected configuration morphology to obtain cross-temporal offset data;

[0021] Connected field energy state convergence is performed on cross-time offset data to obtain connected field energy state data;

[0022] Connectivity configuration rearrangement features are extracted from the connected field energy state data to obtain topological structure change feature data.

[0023] This invention, through the aforementioned topology change feature extraction process, transforms the communication topology change process from discrete state comparison to continuous structural evolution modeling, enhancing the ability to characterize topology changes. By performing structural alignment processing on communication topology snapshots, the topology at different times is placed in a unified structural reference system, avoiding error accumulation caused by changes in node identifiers or local missing parts. Through connectivity configuration morphology processing, the originally scattered link changes are mapped into describable structural morphological units, which is beneficial for capturing the internal organization patterns of the topology. Through cross-temporal offset calculation, not only is the topology changed, but the directionality and persistence of structural changes are also quantified, thereby distinguishing between gradual evolution and sudden reconfiguration behavior. The integration of cross-temporal offsets using a connectivity field energy state convergence mechanism can effectively suppress the influence of instantaneous noise or occasional disturbances on the judgment results, highlighting representative structural evolution trends. Through connectivity configuration rearrangement feature extraction, the structural changes brought about by communication path reorganization and connectivity relationship rearrangement can be accurately identified, making the obtained topology change features more stable and discriminative.

[0024] Preferably, the convergence of the connected field energy state includes:

[0025] The offset energy data is obtained by performing connectivity configuration offset energy mapping based on cross-temporal offset data;

[0026] Local energy state coupling of the connected field is performed based on the offset energy data to obtain local energy state coupling data;

[0027] Energy state equilibrium convergence is performed on the local energy state coupling data to obtain energy state equilibrium converged data;

[0028] The connected field steady-state energy levels are extracted based on the energy state equilibrium convergence data to obtain the connected field energy state data.

[0029] This invention, through the aforementioned connected field energy state convergence process, transforms the structural changes reflected by trans-temporal offsets in the communication topology from discrete, noise-sensitive offsets into energy state characteristics with physical meaning and stable expression. By mapping the energy of connected configuration offsets, the trans-temporal offset amplitude and its persistence characteristics are uniformly mapped to an energy scale, making the strength of changes between different configurations comparable, thus avoiding the instability caused by solely relying on offset thresholds. Through connected field local energy state coupling modeling, the mutual influence relationship between adjacent configurations is introduced, allowing local structural changes to be comprehensively considered within a spatial range, avoiding the over-amplification effect of isolated configuration anomalies on the overall judgment. Through energy state equilibrium convergence processing, high-frequency, short-term energy fluctuations are suppressed, causing the energy state distribution to gradually converge towards an equilibrium state reflecting the true structural evolution trend. The extracted connected field steady-state energy levels can stably characterize the overall evolution intensity and structural stability level of the communication topology within the current time window.

[0030] Preferably, the calculation of the nonstationarity discrimination feature includes:

[0031] Based on the topological change feature data, the change feature evolution fragmentation process is performed to obtain evolution fragment data;

[0032] Evolutionary fragment data is processed to ensure content consistency across fragments, resulting in fragmented data.

[0033] The adjacent offset data is obtained by calculating the offset between adjacent segments based on the fragmented data.

[0034] The continuity disruption degree is calculated for adjacent offset data to obtain continuity disruption degree data;

[0035] The fragmented data, adjacent offset data, and continuity disruption data are vectorized to obtain non-stationary discriminant feature data.

[0036] By employing the aforementioned non-stationarity discrimination feature calculation process, this invention elevates the judgment of communication topology changes from amplitude-level fluctuations to structural analysis at the evolutionary mechanism level, effectively improving the accuracy and reliability of non-stationary state identification. By fragmenting the evolutionary features of topological structural changes, the topology change process is adaptively divided into several time segments based on its inherent evolutionary characteristics, thus avoiding boundary effects caused by fixed time windows. Analysis of the consistency of evolutionary segment content enables the system to identify whether the topology follows a relatively stable evolutionary pattern within a certain time period, providing a structural reference for discrimination. Calculation of the offset between adjacent segments quantifies the differences in changes between different evolutionary stages, highlighting the mode-switching behavior during topology evolution; combined with the measurement of continuity disruption, it effectively distinguishes between gradual structural adjustments and sudden evolutionary interruptions.

[0037] Preferably, the non-stationary detection includes:

[0038] Threshold judgment is performed on the non-stationary discriminant feature data to obtain the first non-stationary topological data;

[0039] Evolutionary migration is performed based on nonstationary discriminant feature data to obtain second nonstationary topological data;

[0040] Based on the first non-stationary topology data and the second non-stationary topology data, a dual-source non-stationary topology fusion determination is performed to obtain non-stationary topology data.

[0041] This invention utilizes threshold judgment on non-stationary discriminant feature data to quickly identify explicit non-stationary states with significant topological changes and concentrated structural disturbances, thus achieving efficient screening under computationally limited conditions. A second non-stationary topology determination based on evolutionary offset enables the system to identify implicit non-stationary states that may not be accompanied by drastic numerical changes but whose topological evolution patterns have shifted, overcoming the insensitivity of single threshold methods to asymptotic or structural changes. Through dual-source non-stationary topology fusion judgment, explicit and implicit discrimination results are jointly analyzed to comprehensively determine non-stationary topological states under constraints of temporal continuity and structural consistency, avoiding misjudgments or omissions caused by the failure of a single criterion.

[0042] Preferably, the construction of the sub-cluster feasible communication domain includes:

[0043] Local connectivity data is obtained by parsing the non-stationary topology data.

[0044] Connectivity skeleton units are extracted based on local connectivity relationship data to obtain connectivity skeleton unit data;

[0045] Sub-cluster stability features are extracted based on the Unicom skeleton unit data to obtain sub-cluster stability feature data;

[0046] Based on the sub-cluster stability characteristic data, the feasible communication domain of the sub-cluster is determined from the data of the connecting skeleton unit, and the feasible communication domain data of the sub-cluster is obtained.

[0047] This invention, through the aforementioned sub-cluster feasible communication domain construction process, avoids relying on global connectivity assumptions to perform coarse-grained partitioning of the communication structure even when the communication topology is in a non-stationary evolution state. By parsing the local connectivity relationships of the non-stationary topology data, the reachability relationships between communication nodes are finely characterized within a local range, thereby reducing the interference of topology breaks on the overall analysis. Through the extraction of connectivity skeleton units, complex and variable local connectivity relationships are compressed into representative structural units, effectively preserving the core connectivity paths and key node relationships in the communication domain. By extracting and analyzing the stability characteristics of the sub-cluster, the ability of each connectivity skeleton unit to maintain communication in a non-stationary environment can be evaluated from the perspectives of time persistence and structural consistency. Based on the stability characteristics, the connectivity skeleton units are screened and confirmed to accurately determine the feasible communication domains of the sub-cluster with actual communication value, avoiding the misjudgment of short-term, accidental connectivity as stable communication areas.

[0048] Preferably, the communication uncertainty prediction includes:

[0049] Link association is performed based on the feasible communication domains of the sub-cluster to obtain link association data;

[0050] Communication feature data is obtained by extracting communication features from the link association data;

[0051] Deep learning is used to predict communication feature data to obtain communication uncertainty data.

[0052] This invention employs correlation modeling of link relationships within the feasible communication domain of sub-clusters. This allows communication prediction to move beyond isolated link metrics, fully considering structural dependencies and collaborative relationships between nodes, thus more accurately reflecting the overall characteristics of communication behavior. Extracting communication features from link correlation data enables a unified expression of link quality changes, stability trends, and correlation structural features, ensuring that input features simultaneously include spatial structural information and temporal evolution information. Deep learning prediction, through nonlinear modeling of these communication features, facilitates the capture of implicit patterns in communication state changes under operating conditions, enabling forward-looking predictions of future communication stability and fluctuation risks.

[0053] Preferably, the broken-chain topology local learning includes:

[0054] Based on communication uncertainty data, a local learning graph of the broken link topology is constructed to obtain local learning graph data;

[0055] Local topological evolution feature encoding is performed on the local learning graph data to obtain local topological evolution feature data;

[0056] Deep learning is applied to local topological evolution feature data to obtain local prediction data.

[0057] This invention constructs a local learning graph of broken-link topology based on communication uncertainty data, focusing the learning process on high-risk links and critical structural regions, thus avoiding redundant computation and noise interference caused by global modeling. By encoding the topology evolution features of the local learning graph, the connection changes between nodes, broken-link trends, and their temporal correlations can be uniformly mapped into learnable feature representations, thereby preserving the structural information of local topology evolution. By learning the above-mentioned local topology evolution features through a pre-set deep learning model, the system can capture the nonlinear relationship between broken-link behavior and topology changes, and achieve prediction of future local topology states.

[0058] Preferably, the cross-domain opportunistic action learning includes:

[0059] Cross-domain reconnection region identification is performed based on local prediction data to obtain cross-domain reconnection region data.

[0060] Cross-domain opportunity action space is constructed from cross-domain reconnection area data to obtain cross-domain opportunity action space data;

[0061] Cross-domain opportunity action state feature encoding is performed on the cross-domain opportunity action space data to obtain action state feature data;

[0062] Deep cross-domain opportunistic action learning is performed based on action state feature data machine to obtain data for broken-chain cross-domain reconnection.

[0063] This invention identifies cross-domain reconnection areas based on local prediction data, ensuring that the system only conducts subsequent analysis within areas with potential communication value and reconnection feasibility, avoiding energy waste and system disturbance caused by indiscriminate attempts. By constructing a cross-domain opportunistic action space, node adjustments, parameter configurations, or relay cooperation behaviors that may trigger cross-domain reconnection are uniformly modeled, giving the reconnection process clear action constraints and selection boundaries. Encoding the state features of the cross-domain opportunistic action space comprehensively reflects the current topology prediction state, regional stability, and action influencing factors, providing a complete and structured decision input for the learning model. Through deep cross-domain opportunistic action learning, the system can automatically select reconnection strategies with a high success probability and minimal impact on the existing communication structure from multiple candidate reconnection schemes, thereby improving the efficiency of cross-domain reconnection after a link break while maintaining the stability and robustness of the overall communication topology.

[0064] The beneficial effects of this invention are as follows: By performing structured analysis of broken-link communication data through a non-stationary topology detection module, the system can identify non-stationary states caused by maneuvering, obstruction, or energy differences from the perspective of communication topology evolution, avoiding misjudgments caused by relying solely on instantaneous link indicators. The sub-cluster feasible communication domain construction module decomposes the complex and fragmented global topology into several local sub-domains with practical communication value, limiting the analysis to a structurally stable and controllable range. Through a communication uncertainty prediction module, the communication state of the sub-cluster is prospectively modeled, enabling the system to perceive potential broken-link risks and fluctuation trends before making decisions, thereby reducing the impact of short-term disturbances on the overall strategy. The cross-domain opportunistic action learning module learns and filters cross-domain reconnection behaviors based on local prediction results, ensuring that reconnection decisions are based on predictive constraints and structural evolution cognition, avoiding blind reconnection that could damage the existing stable communication structure. Attached Figure Description

[0065] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0066] Figure 1 A structural diagram of a deep learning-based spatial channel dynamic characteristic prediction system is shown in one embodiment.

[0067] Figure 2 A flowchart illustrating the steps of a non-stationary topology detection module according to an embodiment is shown.

[0068] Figure 3 A step diagram of a sub-cluster feasible communication domain construction module according to an embodiment is shown;

[0069] Figure 4 A flowchart illustrating the steps of a communication uncertainty prediction module according to one embodiment is shown.

[0070] Figure 5 A step diagram of a broken-chain topology local learning submodule according to an embodiment is shown. Detailed Implementation

[0071] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0072] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0073] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated items listed.

[0074] Please see Figures 1 to 5 This application provides a deep learning-based spatial channel dynamic characteristic prediction system 01, comprising:

[0075] The non-stationary topology detection module 10 is used to acquire broken link communication data; and to perform non-stationary topology detection based on the broken link communication data to obtain non-stationary topology data.

[0076] Specifically, the non-stationary topology detection module is used to identify topological instability in communication networks caused by link breaks. The system collects broken-link communication data based on preset terminal devices / communication equipment / communication centers, with periodic collection intervals of 50 to 200 milliseconds. The data collected for each node includes parameters such as location information, speed information, remaining energy, channel signal strength, interference signal-to-noise ratio, link quality indicators, and distance between nodes. When events such as an abnormal increase in packet loss rate or handshake failure occur, the system automatically triggers event collection and records the relevant broken-link events.

[0077] For any two nodes, if the signal-to-noise ratio is higher than a set threshold, the link loss is lower than a limit, and the distance between the nodes is within the communicable range, then the link is considered valid at the current moment and marked as a valid edge; otherwise, it is marked as invalid. The system generates a sequence of topology snapshots at consecutive moments according to this rule and assigns uniform numbers to the node indices. Missing nodes are marked as placeholders.

[0078] The system extracts the topological change features corresponding to the topological snapshot sequence. It employs a node neighborhood-based structural configuration encoding method to record the structural pattern of each node within a certain adjacency range. This involves constructing a local configuration graph of the node based on a 1-hop or 2-hop neighborhood, and encoding it using elements such as structural labels, degree sequences, and the number of ring structures to form a structural configuration. By calculating the degree of change between structural configurations at adjacent time points, such as Jaccard distance and cosine difference, the system obtains the offset features representing the dynamics of topological evolution, i.e., the topological evolution sequence. The system divides the topological evolution sequence into segments based on the topological similarity between consecutive time points (i.e., using cosine similarity to calculate similarity). If the similarity is below a set threshold, a new evolutionary stage is considered to have begun. The system statistically analyzes the dispersion of structural features within each segment, such as the mean absolute deviation, and analyzes the average feature differences and change directions between adjacent segments to determine if there are abrupt changes in the structural evolution path. If large offsets are continuously observed in multiple segments, accompanied by drastic reversals in the direction of change, the continuity of the topological structure is considered to have been disrupted. The system identifies non-stationary topological states of abrupt changes using dual-source feature determination. If the observed rearrangement intensity or continuity disruption exceeds a set threshold, it is classified as a dominant non-stationary phenomenon. If a decrease in feature consistency or abrupt changes in structural evolution direction occur in multiple segments, it is classified as a latent non-stationary phenomenon. The system combines the two types of results for a fusion output: when both dominant and latent features exist simultaneously, a high-confidence non-stationary determination is output; if only latent features exist and their duration meets the set requirements, a medium-confidence determination is output; if only short-term dominant perturbations are observed, they are not output as non-stationary events to avoid misjudgment.

[0079] The sub-cluster feasible communication domain construction module 20 is used to construct the sub-cluster feasible communication domain based on non-stationary topology data, and obtain the sub-cluster feasible communication domain data.

[0080] Specifically, the sub-cluster feasible communication domain construction module is used to identify and construct sub-cluster regions with stable communication capabilities based on non-stationary topology data. At each non-stationary moment, the system identifies multiple locally connected sets of nodes, i.e., locally reachable regions, based on the node connection relationships in the topology graph. The system constructs an edge-weighted graph to represent link quality for each locally reachable set. This graph integrates multi-dimensional indicators such as link loss, inter-node distance, and channel interference. The link loss value is calculated from the difference between the received signal strength and the transmitted power during inter-node communication; the inter-node distance is obtained by calculating the Euclidean distance through the location information periodically broadcast by the nodes; and the channel interference is mainly characterized by the signal-to-noise ratio (SINR) acquired in real time during communication. In the preferred embodiment, the system sets link loss as the primary factor according to a weighted ratio, and assigns certain weights to the distance between nodes and channel interference. For example, when the link loss value exceeds a set upper limit, the system increases the weight of that edge to prevent it from entering the skeleton path; when the node distance is within the communication critical range, the weight cost of that edge is increased based on a preset distance penalty factor; when the SINR value is lower than the decodeable threshold, the system automatically removes that link as a skeleton candidate, or it obtains the skeleton by normalizing the first three factors and performing a weighted sum based on preset weights. Based on this edge-weighted graph, the system can extract the communication skeleton structure of the sub-cluster in different ways, including constructing a minimum spanning tree to obtain the skeleton with the minimum connection cost, or constructing a multi-edge connected subgraph to retain the backbone link. The system outputs the skeleton edge set and skeleton node set of each sub-cluster as the skeleton structure.

[0081] The system analyzes the stability of each skeleton structure in the time dimension and obtains stability characteristics, including: the retention ratio of the skeleton structure in a continuous time window; the proportion of links in the skeleton that are prone to communication interruption; and the remaining energy level and distribution characteristics of the nodes that make up the skeleton.

[0082] The system sets judgment criteria to filter sub-clusters that can be considered feasible communication domains. These criteria include maintaining the continuity of the skeleton structure within a certain time range, keeping the link ratio within a reasonable range, and ensuring that the overall remaining energy is higher than a set task threshold. Sub-clusters that meet all conditions will be identified as feasible communication domains / sub-cluster feasible communication domain data, and the system will output information including sub-cluster identifier, skeleton node set, skeleton edge set, effective time range, and stability feature vector.

[0083] The communication uncertainty prediction module 30 is used to predict communication uncertainty based on the feasible communication domain data of the sub-cluster to obtain communication uncertainty data;

[0084] Specifically, the communication uncertainty prediction module is used to predict the unstable states that communication links may experience in the future based on feasible communication domain data of the sub-cluster, and output communication uncertainty data. The system constructs a relationship graph between links within each feasible communication domain, forming a link dependency graph. In this graph, if two links share the same node, or are located on the same skeleton path and are spatially adjacent, they are considered to be related, and a correlation weight is assigned accordingly.

[0085] ;

[0086] For link association weights, It is an exponential function. This is the distance function between paths. This represents the subgraph or path number of the skeleton path containing link (i,j). This represents the subgraph or path number of the skeleton path containing link (m,n). , where is the distance attenuation scaling factor, controlling the degree of attenuation of the association weight by the distance between paths, i is the encoding of the i-th node, j is the encoding of the j-th node, m is the encoding of the m-th node, and n is the encoding of the n-th node. The system extracts features for each communication link within a set time window. The features include the changing trend of communication quality indicators (such as signal-to-interference ratio, link loss, rate of change and fluctuation of distance between nodes), link stability performance (such as the frequency of link breakage events and unstable jump behavior of reconnection after breakage), the role of the link in the network structure (whether it is a skeleton edge or a critical connection edge, and its topological centrality in the sub-cluster), and the risk aggregation features obtained by combining the link dependency graph (i.e., the average risk value of other links associated with it).

[0087] The system identifies features based on a pre-trained risk identification model to predict the failure probability and fluctuation risk of each link within a future timeframe, ranging from 0.5 to 3 seconds. The training process uses historical data on actual link failures and communication quality fluctuations as labels. The system outputs two types of communication uncertainty data: a risk score for each backbone link and an overall communication risk assessment result for the entire sub-cluster, such as the average risk value or high-risk boundary value.

[0088] The training phase of the risk identification model uses real-world recorded link failure events and communication quality fluctuations as supervisory labels, achieving link risk prediction capabilities through end-to-end time series modeling. In the training data preparation phase, the system uses each link in the communication topology as the basic analysis unit, constructing a multi-dimensional input sequence with a time window length of 1–3 seconds. Input features include, but are not limited to, the following: link quality trend features (such as signal-to-noise ratio, packet loss rate, slope and fluctuation amplitude of link loss), link stability indicators (such as recent link failure frequency and reconnection bounce rate), structural attributes (whether it is a skeleton edge, a critical cut edge, and the centrality of a node in the subgraph), and aggregated risk values ​​in the link dependency graph (obtained by a weighted average of the risks of its adjacent links). These features are normalized to form the multi-dimensional time series input data of the links. The main structure of the model adopts a temporally enhanced graph neural network architecture. Based on the input link feature sequence, the system uses a set of one-dimensional convolutional or multi-layer GRU / LSTM units to perform feature compression and pattern extraction in the time dimension. The system spatially encodes the risk propagation relationships between adjacent links into the model based on a communication topology graph or link dependency graph using Graph Convolutional Networks (GCNs) or graph attention mechanisms (such as GAT). The intermediate layer nodes of the graph network are represented with dimensions ranging from 64 to 128, and the weight matrix uses learnable parameters, combined with residual connections and batch normalization. The model output layer is a two-branch structure, outputting two key prediction results: first, the probability score of a link failure event occurring within a specified prediction window (e.g., 0.5 seconds, 1 second, or 3 seconds); second, the risk score of drastic fluctuations in link quality over the future period (which can be represented as the variance exceeding the threshold probability). During training, the model uses a cross-entropy loss function and a regression error function (such as mean squared error) to jointly optimize the two types of labels, with weights dynamically adjusted according to the fault tolerance requirements of the scenario task. After training, the system deploys the model to the communication scheduling module, supporting predictive evaluation of the link status at each time step. The output communication uncertainty data includes two levels: the first is the risk score output for each skeleton link; the second is the evaluation index of the overall communication stability of the sub-cluster, including the average risk level, the highest risk link marker, and the proportion of high-risk links.

[0089] The cross-domain opportunistic action learning module 40 is used to perform local learning of the disconnected topology based on communication uncertainty data to obtain local prediction data; and to perform cross-domain opportunistic action learning based on the local prediction data to obtain cross-domain reconnection data after a disconnection.

[0090] Specifically, the cross-domain opportunistic action learning module is used to proactively identify broken link areas based on local prediction data when there is uncertainty in the communication link, and to find feasible cross-domain reconnection solutions through simulation or learning mechanisms. This module consists of two stages: first, local learning and risk prediction for the current topology; and second, modeling and optimizing cross-domain opportunistic actions based on the prediction results.

[0091] In the first phase, the system constructs a local learning graph based on nodes and links within the feasible communication domain. This graph uses drones within the current sub-cluster as nodes, with backbone links and qualified auxiliary links as edges. Communication uncertainty information is embedded into the edge and node attributes. Edge attributes include link disconnection probability, current signal quality, link loss, distance information, and risk aggregation characteristics; node attributes include remaining battery power, movement speed, and whether it is located at the cluster boundary (i.e., close to the outer domain). The system encodes the state evolution characteristics of each link and node within the current time window, including the trend of risk changes, sensitivity markers of link structures (such as whether they are critical cut edges or have high centrality), and the frequency of repeated changes in edge states over time. By processing the graph structure and features using a local communication prediction model, the system can predict the local communication evolution trend in the next time window, outputting the probability of backbone structure preservation, the set of most likely failed critical links, and the cross-domain communication potential score of boundary nodes.

[0092] In the second phase, the system identifies potential cross-domain reconnection areas based on the predicted outreach capabilities of boundary nodes. The system only includes a region as a candidate region if the spatial distance is within the communication reachable range and both boundary nodes spanning the sub-clusters on either side possess outreach potential. The system requires that the current skeleton stability probability of the selected sub-clusters not be lower than a certain threshold. The identified cross-domain regions will include candidate node pairs and the spatial range they cover. For these candidate regions, the system constructs an executable action space. The action types that each candidate node can take include position fine-tuning (such as increasing or decreasing flight altitude, or horizontal translation), communication parameter adjustment (such as adjusting transmit power or modulation / coding level), and introducing a cooperative relay node (provided that the relay node has sufficient remaining energy and is not on a high-risk link). The composition of each action is defined by the executing node, the action type, and its specific parameters. The system encodes the action execution state into a feature vector. This state feature includes: the outreach potential score of the nodes on both sides of the candidate region, the spatial margin between the current and potential connectivity windows, the predicted impact of the action on the stability of the original cluster skeleton, and the estimated energy consumption required for the action. Based on the aforementioned state characteristics, the system can choose to use a deep scoring network or a heuristic optimization strategy to score and rank actions. The deep scoring network is preferably a graph convolutional network model that integrates action features and candidate region context information. The optimization strategy, based on a constructed scoring function, weights and ranks the action execution effects (including expected success rate, communication cost, and stability loss), selecting the action with the highest score. The system sets explicit policy constraints; for example, if the execution of an action causes a decrease in the stability of the domain skeleton exceeding a set threshold, the priority of that action is reduced or it is directly eliminated. The system outputs disconnected cross-domain reconnection data, including the selected cross-domain region, participating key nodes, the type of action executed and its parameter configuration, the expected time window for the action execution, the success prediction probability of the reconnection strategy, and an assessment of its potential impact on the domain's communication structure.

[0093] The local communication prediction model is used to predict the evolution trend of local communication structure within the next time window based on the communication topology and risk characteristics within the current sub-cluster. The model's training process is based on historical topology snapshot data and communication uncertainty data, combined with link state evolution labels, and uses a time-series graph modeling mechanism to predict link failure trends, skeleton preservation capabilities, and cross-domain potential. In the training data preparation phase, the system performs time-series processing on topology snapshots within the feasible communication domain, extracting the evolution features of nodes and edges within multiple consecutive time windows as model input. The input data includes the following: for each link, the input includes its recent link failure probability sequence, signal quality (e.g., SINR), link loss, communication distance, and its risk aggregation value in the link dependency graph; for nodes, the input includes their remaining power change trend, relative speed, whether they are boundary nodes, and their centrality index and communication participation changes. All input features are stacked according to time windows to form multi-dimensional time-series tensor inputs at the link and node levels. The model structure adopts a graph-aware time-series deep network. The system preferably constructs the model backbone using a combination of Graph Convolutional Networks (GCNs) and time series prediction units (such as GRUs or TemporalConvNets). The first layer of the model is a graph convolutional layer, used to extract the spatial association features of nodes and edges in the communication topology at each time step. The parameters of this layer include adjacency matrix weights, node feature mapping dimensions (such as 64 or 128 dimensions), and edge feature fusion mechanisms. Multiple layers of GRU units are used to model the time series to capture the dynamic trend of link risk evolution. Key parameters for this part include the GRU hidden state dimension (such as 128), time window length (such as 5-10 sampling points), and Dropout anti-overfitting rate (such as 0.2). Model training labels are generated based on historical real-world evolution and mainly include: whether each link fails in the next time window (link breakage marker), whether it is removed from the skeleton path (skeleton preservation label), and whether boundary nodes have the possibility of effective cross-domain connections in future time periods (connectivity score). The system employs a joint loss function for multi-task optimization. Link failure prediction uses binary cross-entropy loss, skeleton preservation probability uses regression-based mean squared error loss, and cross-domain potential scoring uses ranking loss or continuous scoring error. After training, the model can be deployed online or offline. Within each prediction period, it takes the evolutionary features of the current topology as input and outputs three types of results: 1) the preservation probability of each skeleton link within a future time window; 2) the identification results of high-failure-risk link sets; and 3) the cross-domain communication potential score of boundary nodes.

[0094] Preferably, the non-stationary topology detection includes:

[0095] 11. During the operation of the drone swarm, periodically collect disconnected communication data for each communication node;

[0096] Specifically, during the periodic sampling phase, each communication node collects communication status information between itself and its neighboring nodes at preset fixed time intervals (preferably 50 to 200 milliseconds). The collected information includes parameters such as whether the link establishment was successful, the current packet loss rate, signal-to-noise ratio, received signal strength, round-trip communication delay, and handshake status. If the system detects an abnormal trend in the communication status, it automatically identifies it as a link failure event and records it immediately. Triggering conditions include: multiple consecutive (preferably 2 to 5) failed communication attempts; the link packet loss rate exceeding a preset threshold (e.g., 30% or higher) within a short time window; or failure of the handshake process and link maintenance mechanism. If any of these conditions are met, the system records the communication interruption of that node pair as a link failure communication event. Based on the above, link failure communication data is obtained.

[0097] 12. Perform communication topology snapshot processing based on the broken link communication data to obtain communication topology snapshot data;

[0098] Specifically, the system defines the basic elements in the topology snapshot. Each UAV communication node is considered a node in the graph; the communication state between any two nodes, provided that certain conditions are met, is defined as a valid communication edge. The system determines the validity of each pair of nodes based on the communication data collected at the current moment. The signal-to-noise ratio between nodes must not be lower than a set minimum communication quality threshold, preferably between 6 and 12 dB; the observed packet loss rate during communication should be within an acceptable range, preferably between 20% and 40%; and the system confirms that the node pair was not recorded as having a broken link in the most recent communication interaction. Only when all three conditions are met simultaneously is the node pair considered to have a valid communication link. After determining the validity of all node pairs, the system integrates all nodes at the current moment with their corresponding valid communication edges to generate communication topology snapshot data.

[0099] 13. Extract topology change features from communication topology snapshot data to obtain topology change feature data;

[0100] Specifically, the system standardizes node identifiers across the topology at each time point, ensuring that identical nodes have consistent numbers across different time periods. For nodes temporarily missing at a given moment, the system fills in the gaps using placeholder markers. For each node, the system analyzes its one-hop or two-hop neighborhood connections and encodes the structural morphology within that neighborhood into a standardized configuration representation, such as using structural fingerprints. The system performs cross-temporal comparisons of the configuration representations at consecutive time points, calculating the degree of change in configuration between adjacent moments to generate offset data representing the evolution trend of the topology. This offset can be measured by the proportion of changes in adjacency relationships or differences in the editing of connection structures, yielding topology change characteristic data.

[0101] 14. Calculate nonstationarity discrimination features based on topological structure change characteristic data to obtain nonstationarity discrimination feature data;

[0102] Specifically, the topological change characteristics are organized chronologically, and evolutionary segments are divided based on the feature similarity between adjacent time points. When the similarity of the change characteristics between a certain time point and the previous time point is lower than a preset threshold (e.g., cosine similarity less than 0.8), the system determines that the topological evolution trend has shown a significant turning point, and accordingly divides the evolution into new evolutionary segments, forming stage boundaries in structural continuity. The system statistically analyzes the dispersion of features at each time point within a segment to obtain the evolutionary consistency of each segment. The system calculates the degree of change between adjacent segments, quantifies their overall difference by comparing the feature mean vectors of each segment, and records the difference in the direction of change between two segments, for example, by judging whether the direction of change is consistent or has a reversal trend through the angle between vectors. If the system detects that the feature difference between segments increases (exceeding the threshold) in a short period of time, and that the direction of change frequently reverses, it determines that the continuity of topological evolution in that period has been disrupted, such as... , The weight of the current time segment t, It is an exponential function. This is the index weighting coefficient (controls sensitivity; the larger the value, the faster the weight increases). This represents the degree of feature difference between the current segment and the previous segment. The system combines the evolutionary consistency of each segment, the degree of shift / change between segments, and the degree of disruption to the continuity of the overall evolution to form non-stationary discriminative feature data.

[0103] 15. Perform non-stationary detection on the non-stationary discriminant feature data to obtain non-stationary topological data.

[0104] Specifically, if the degree of topological continuity disruption reaches or exceeds a set judgment threshold within a certain time period (e.g., the degree of continuity disruption is between 0.6 and 0.8), or if the communication structure undergoes a drastic rearrangement within a short period, causing the intensity of structural change to exceed a set upper limit, then the communication topology can be directly determined to have entered a non-stationary state within that time period, and the first type of non-stationary topological data is output. If multiple consecutive topological evolution segments (preferably 2 to 3) all exhibit a structural shift trend, and the directionality of topological change undergoes a structural change (e.g., from local clustering to overall dispersion), the second type of non-stationary topological data is generated. When the first and second types of non-stationary topological data overlap within the same time period, non-stationary topological data is output; when only the second type of non-stationary topology is established and its duration exceeds a preset window, non-stationary topological data is also output.

[0105] Preferably, the topology change feature extraction includes:

[0106] Structural alignment is performed based on communication topology snapshot data to obtain structural alignment data;

[0107] Specifically, the system performs unified node indexing on topology snapshots at different time points, assigning a globally unique identifier to each communication node. For nodes for which communication data has not been collected at certain times, the system handles the missing data by storing them as placeholders in the topology structure and setting their connection information to an "undefined" state. The system uses a topological adjacency matrix template to normalize the existence of edges and supports a sparse storage structure. The output structure-aligned data includes: an aligned time-series topological adjacency matrix sequence, and a mapping table of node numbers and their corresponding actual devices.

[0108] Connectivity configuration morphology data is obtained by performing connectivity configuration morphology processing on the structure alignment data.

[0109] Specifically, for each communication node, the system extracts the connectivity relationships within its one-hop or two-hop neighborhood from its corresponding topological adjacency matrix, constructing the connected configuration graph of that node at the current time. The system performs structural encoding on each local configuration graph. Encoding methods may include, but are not limited to, extracting structural features such as the node degree sequence, edge density between nodes, and the number of existing cycle structures; or graph isomorphism-invariant encoding methods may be used, such as structural labels based on hierarchical label propagation (e.g., the Weisfeiler-Lehman method) or graph structure hash signatures to generate stable configuration feature codes. After the structural encoding of all nodes and time points is completed, the system organizes the configuration features of the nodes in chronological order to form connected configuration morphology data at multiple time points.

[0110] Cross-temporal offset calculations are performed on the connected configuration morphology to obtain cross-temporal offset data;

[0111] Specifically, the system compares the connectivity configuration of each communication node between two consecutive time points and calculates the configuration difference. This difference can be measured using various distance functions, such as the cosine similarity inverse between configuration feature vectors, Hamming distance, structural edit distance, or Jaccard distance. The system sets a change threshold and marks nodes whose difference exceeds the set standard. Based on the node-level offset results, the system constructs a time series, recording the configuration change trajectory of each node throughout the entire time window. The system performs statistical analysis on the structural changes of the overall network, extracting system-level offset trend indicators, including the average configuration change magnitude of the entire network and the proportion of high-variable configuration regions, i.e., the proportion of nodes whose configuration change exceeds the threshold at a specific time. The output cross-time series offset data includes: the structural offset time series of each node at each time point, and the system-level offset trend indicators.

[0112] Connected field energy state convergence is performed on cross-time offset data to obtain connected field energy state data;

[0113] Specifically, the system maps the structural offset value of each node at each time step to the corresponding local energy value. This mapping process is achieved by setting a transformation function, such as a square function or a sigmoid function, to convert the offset into positive energy. The system performs local energy state coupling processing in the spatial dimension. Using the communication topology adjacency graph as a constraint, the system calculates the average neighborhood energy of each node. If, within multiple consecutive time slices, the change in the energy value of a node is lower than a set threshold, or its energy value variance within a sliding time window is lower than a judgment criterion, then the node is considered to have entered a steady state. After the local node stability judgment is completed, the system constructs a connected field energy state graph as a global representation. This graph inherits the original communication topology structure, with communication nodes as nodes in the graph and the original communication links as edges, while assigning the steady-state energy value of each node as an attribute. The system extracts field-level indicators, such as the average energy state level and local high-energy accumulation regions. The connected field energy state data output by the system includes the steady-state energy value sequence of each node (considered as the average neighborhood energy value corresponding to the steady-state state), and the connected field energy state graph organized in a graph structure.

[0114] Connectivity configuration rearrangement features are extracted from the connected field energy state data to obtain topological structure change feature data.

[0115] Specifically, configuration rearrangement refers to the behavior of a communication node undergoing a significant change in its connectivity configuration pattern during the convergence of its local energy state from an unstable state to a steady state. For example, a node may exhibit a star-shaped connection structure before reaching a steady state, but its connection pattern may change to a chain or ring structure after reaching a steady state. The system then determines that the node has experienced a configuration rearrangement event. The system compares the configuration morphology encoding results of each node before and after energy state convergence. If the difference before and after exceeds a set structural change threshold (e.g., based on Jaccard distance, structural label differences, etc.), the system considers that the node's configuration has undergone a structural switch. The system counts the total number of nodes that undergo configuration rearrangement per unit time; in terms of density indicators, it calculates the spatial clustering of these rearranged nodes in the network, such as the average hop distance; in terms of intensity indicators, it evaluates the average and maximum values ​​of the configuration change amplitude of each node. The system structurally organizes the features such as the set of configuration rearranged nodes, change intensity, and spatial density to obtain topology change feature data.

[0116] Preferably, the convergence of the connected field energy state includes:

[0117] The offset energy data is obtained by performing connectivity configuration offset energy mapping based on cross-temporal offset data;

[0118] Specifically, the cross-temporal migration data records the magnitude of the configurational change of each node between adjacent time points in time series form. The system converts the structural migration magnitude of each node at a specific time point into a corresponding energy value to represent its change intensity by setting a mapping function, including squaring the migration magnitude; or using an exponential smoothing amplification mapping, such as... , For offset energy data, This is the magnification factor, with a value ranging from 1 to 3. The input is cross-time series offset data. The system generates a sequence of offset energy values ​​for each node at multiple time points, forming the offset energy data.

[0119] Local energy state coupling of the connected field is performed based on the offset energy data to obtain local energy state coupling data;

[0120] Specifically, the system integrates the energy values ​​of each node's neighboring nodes based on adjacency relationships to form its local coupled energy state. The aggregation method can employ a simple averaging strategy, using the average energy value of all neighboring nodes as the local energy state of the current node; or a weighted strategy, weighting the neighboring node energy values ​​according to the strength or stability of the communication link (such as signal strength indicators or edge existence probability). Each node corresponds to a local coupled energy value at a specific time, and the system aggregates the coupled energy state results of all nodes across the entire graph to form local coupled energy state data.

[0121] Energy state equilibrium convergence is performed on the local energy state coupling data to obtain energy state equilibrium converged data;

[0122] Specifically, the system performs sliding analysis on the local coupling energy state value of each node over a series of consecutive sampling times, typically selecting a fixed-length time window (e.g., 3 to 5 times) as the observation interval. Within this time window, the system evaluates the fluctuation of the node's energy state value. If certain stability criteria are met, the node is considered to have entered an energy state equilibrium state. These criteria include, but are not limited to, small fluctuations in the node's energy state value within the time window (e.g., a low variance value indicating stable state change) and the difference between the maximum and minimum energy state values ​​being within a set range. Nodes meeting any of these conditions are marked as "equilibrium." The system outputs a data structure containing two types of information: a state flag indicating whether each node has reached energy state equilibrium, and the energy state value at the point of equilibrium.

[0123] The connected field steady-state energy levels are extracted based on the energy state equilibrium convergence data to obtain the connected field energy state data.

[0124] Specifically, the system first constructs an energy map of the connected field based on nodes that have reached equilibrium. Each node in this map corresponds to a currently participating unit in communication, and its attribute is its own converged steady-state energy value; edges represent effective communication links between nodes. The system extracts multiple graph-level steady-state features, including the average and fluctuation levels (e.g., standard deviation) of the entire graph's energy levels; spatial clustering indices of high-energy regions (a high-energy region is defined as a region where several high-energy nodes (e.g., nodes whose energy values ​​are higher than the weighted standard deviation of the overall average) are spatially close to each other, forming a clearly dense distribution area), used to determine whether high-energy nodes are concentrated in a specific communication subdomain; density indices of high-energy nodes; and distribution characteristics of nodes that have not yet reached steady state, representing unstable boundaries or transitional regions in the topology. The system then uses these outputs to form connected field energy state data.

[0125] Preferably, the calculation of the nonstationarity discrimination feature includes:

[0126] Based on the topological change feature data, the change feature evolution fragmentation process is performed to obtain evolution fragment data;

[0127] Specifically, the input topological change feature data is a multidimensional feature sequence arranged in chronological order, representing the evolution of the system's structural state at different times. The system evaluates the similarity of feature states between adjacent time points to determine whether the structural change is stable or abrupt. This similarity evaluation is calculated based on the angle relationship between the overall feature vectors. When the feature similarity between two consecutive time points is lower than a set threshold, or when the feature difference fluctuates strongly and exceeds the threshold more than a set number of times within a certain period, the current structure is considered to have entered a new evolutionary stage, and this moment is taken as the starting point of a new segment. The system divides the entire structural evolution process into several time-continuous evolutionary segments, each segment containing topological change features corresponding to several time points.

[0128] Evolutionary fragment data is processed to ensure content consistency across fragments, resulting in fragmented data.

[0129] Specifically, the system performs internal statistics on the feature data of multiple time points contained in each evolutionary segment, calculates the overall feature mean of the segment, and uses it to represent the central state of the segment. The system evaluates the degree of deviation of the feature data of each time point within the segment from the central state. In the consistency determination stage, the system sets a deviation threshold. When the overall deviation of the features at each time point within a segment is lower than the set threshold, it is considered an internally consistent segment, indicating that the structural evolution trend of the segment is relatively stable within that stage. If the deviation exceeds the threshold, it is marked as an internally unstable segment, indicating that the segment has fluctuations or inconsistent feature states. The system generates corresponding consistency description parameters for each evolutionary segment, marks its internal evolution consistency state, and forms segmented processed data.

[0130] The adjacent offset data is obtained by calculating the offset between adjacent segments based on the fragmented data.

[0131] Specifically, the system selects the mean feature of each evolutionary segment as its representative vector. The system compares the representative vectors of any pair of adjacent segments and calculates the changes between them. At the level of change magnitude, the system evaluates the overall difference between the representative vectors of adjacent segments; while at the level of change direction, the system analyzes the angular relationship between the representative vectors of the two segments. The system output includes offset data containing the change magnitude and direction parameters between each pair of adjacent segments.

[0132] The continuity disruption degree is calculated for adjacent offset data to obtain continuity disruption degree data;

[0133] Specifically, the system treats the temporal evolution of the topology as a process that should change continuously as a whole. If the characteristic offsets between adjacent segments fluctuate drastically within a short period, or if the directions frequently reverse, it indicates that the continuity of evolution has been disturbed. The system evaluates the degree of amplitude variation and directional perturbation based on multiple continuous offset segments. In terms of amplitude, the system calculates the ratio between the maximum amplitude change and the average amplitude change in the current continuous offsets; in terms of direction, the system statistically analyzes the degree of fluctuation in the direction of continuous offsets, indicating whether the local evolutionary trajectory is stable. The system performs a weighted combination / direct normalization summation of these two dimensions to derive a continuity disruption index.

[0134] The fragmented data, adjacent offset data, and continuity disruption data are vectorized to obtain non-stationary discriminant feature data.

[0135] Specifically, the system extracts the internal consistency index (i.e., consistency deviation), the magnitude and direction offset from adjacent segments, and the continuity disruption index of the interval in which the segment is located for each evolutionary segment. These are then integrated with the segments to obtain non-stationary discriminant feature data. Optionally, the parameters / indicators can be normalized.

[0136] Preferably, the non-stationary detection includes:

[0137] Threshold judgment is performed on the non-stationary discriminant feature data to obtain the first non-stationary topological data;

[0138] Specifically, the system uses the consistency deviation, amplitude offset, directional offset angle, and continuity disruption degree corresponding to each time segment or node substructure as input features, and judges them through preset thresholds, such as a consistency deviation threshold of 0.2-0.4, an amplitude offset threshold of 0.3-0.5, a directional offset angle threshold of 30-45 degrees, and a continuity disruption degree threshold of 0.6-0.8. When any indicator exceeds the set abnormal threshold range, the segment is considered to have potential evolutionary instability characteristics, and is thus marked as the first non-stationary candidate segment.

[0139] Evolutionary migration is performed based on nonstationary discriminant feature data to obtain second nonstationary topological data;

[0140] Specifically, based on a sliding time window mechanism, the system extracts trends from the non-stationary features of topological segments or nodes within a set continuous time period (e.g., 3 to 5 evolutionary steps). By incrementally calculating the mean of features within the time window, the system determines the magnitude of change between the current moment and the previous window. When the increase in this trend exceeds a preset threshold, the system considers the region to potentially exhibit non-stationary evolutionary shift behavior. The system reconstructs the trajectory and calculates the slope of directional change indicators (such as inter-segment directional angles) within the features; if the rate of directional change increases, the region is also considered a potential shift indicator. The system aggregates all regions exhibiting continuously increasing or directional reversal trends within the sliding window into a second set of non-stationary topological data.

[0141] Based on the first non-stationary topology data and the second non-stationary topology data, a dual-source non-stationary topology fusion determination is performed to obtain non-stationary topology data.

[0142] Specifically, the system can process data using one of two fusion methods. The first is a set merging method, which merges the first non-stationary region identified based on static threshold determination with the second non-stationary region identified based on trend analysis at the node or fragment level, directly summarizing them into a non-stationary topological region. The second is a weighted decision method, which calculates the score for each candidate region under threshold determination (e.g., continuity disruption index) and the score under trend enhancement mechanism (e.g., evolutionary migration intensity) according to preset weight parameters. The two scores are then weighted and merged (first normalized, then weighted, with weights of 0.5-0.8 and 0.5-0.2 respectively) to obtain a determination score. When this score exceeds a set threshold, the region is confirmed as a non-stationary topological region.

[0143] Preferably, the construction of the sub-cluster feasible communication domain includes:

[0144] 21. Perform local connectivity analysis on non-stationary topology data to obtain local connectivity data;

[0145] Specifically, the system extracts node connectivity information within a specific time period based on identified non-stationary topology data. This data is represented as a temporal topology graph, indicating the actual connection structure of the communication system within a specific time frame. The system filters the topology snapshot graph of each frame, retaining only the set of nodes in a non-stationary state, forming a local non-stationary region. After filtering, the system identifies the connectivity relationships within each subset of nodes in a non-stationary region. The system uses graph search algorithms (such as depth-first search or breadth-first search) to partition the node set based on connectivity, grouping nodes with direct or indirect connections into several disconnected local subgraphs. Each local subgraph represents a subset of nodes that are non-stationary and mutually reachable within that time period, considered as a locally connected subcluster. The system outputs the node list and edge connections of all connected subgraphs as local connectivity data.

[0146] 22. Extract the connection skeleton units based on the local connectivity relationship data to obtain the connection skeleton unit data;

[0147] Specifically, the system can set several characteristic indicators to represent the importance and stability of each node, including the number of connections (i.e., degree), the frequency of acting as a transit point in the network path (e.g., betweenness centrality), and the frequency of continuous appearance in multiple time segments. In the dimension of connection quantity, the system counts the number of direct connections (i.e., node degree) of a node across multiple time segments. If the number of connections of a node consistently exceeds a set threshold for a number of time segments, a score is assigned to this dimension after standardization based on the node degree. In the dimension of transit influence, the system calculates the transit frequency of a node in the shortest path to obtain a betweenness centrality index, and determines whether it has a transit function based on its average value across multiple time segments. If the condition is met, a score is assigned to this dimension after standardization based on the transit frequency. In the dimension of activity persistence, the system counts the proportion of a node that appears continuously throughout the entire observation period. If it exceeds a specified threshold, a score is assigned after standardization based on the proportion of continuous appearance. The system performs weighted fusion based on scores and their weights (e.g., 1 / 3 by default) across various dimensions to form a node ranking index, and then ranks all nodes in the network accordingly. The system can select a certain percentage (e.g., top 10% to 30%) of nodes from each local connectivity subgraph as the skeleton nodes for that local region. The system identifies the shortest connection paths between these skeleton nodes to construct the skeleton subgraph structure. The system outputs connectivity skeleton unit data for each local region, including skeleton nodes and their interconnection paths.

[0148] 23. Extract sub-cluster stability features based on the backbone unit data to obtain sub-cluster stability feature data;

[0149] Specifically, the system calculates node retention rate, representing the degree of overlap between the skeleton node set at the current time and the previous time. The system calculates edge stability, i.e., whether the key connections in the skeleton structure are retained long-term. The system calculates the center drift of the subgraph, i.e., whether the geometric center of gravity of the skeleton structure has shifted significantly. If the system has communication traffic monitoring capabilities, it can calculate the communication flow residual variance index. This involves comparing the historical average traffic at each time point with the actual traffic, calculating the residual value (the error sequence obtained by subtracting the actual value from the predicted value), and calculating the variance of this error sequence. The aforementioned sets yield the sub-cluster stability characteristic data.

[0150] 24. Based on the sub-cluster stability characteristic data, determine the feasible communication domain of the interconnection skeleton unit data to obtain the feasible communication domain data of the sub-cluster.

[0151] Specifically, the system filters each sub-cluster based on a set of preset feasibility judgment rules. Sub-clusters that meet the following conditions will be marked as feasible communication domains: First, all indicators of the sub-cluster stability characteristic data meet preset thresholds; second, the skeleton structure appears at least twice in three consecutive time segments; third, the positional change of the geometric center is within a set range, not exceeding three meters or a set node spacing unit. After the feasible communication region is determined, the system spatially encapsulates the selected core skeleton nodes and their adjacent boundary regions to form feasible communication domain units, which will include information such as the spatial boundary range and communication structure characteristics of the domain. The system outputs feasible communication domain data for the sub-cluster, whose data structure includes communication domain number, subgraph number, core skeleton node list, sub-cluster stability characteristic data, and spatial boundary description of the communication domain.

[0152] Preferably, the communication uncertainty prediction includes:

[0153] 31. Perform link association based on the feasible communication domains of the sub-cluster to obtain link association data;

[0154] Specifically, the system takes nodes in each sub-cluster communication domain as input and extracts information such as their spatial location data, historical communication records, and current topology status. Communication records include, but are not limited to, the number of successful link establishments, received signal strength indicators, the frequency of communication interruptions, and the average communication delay. The system performs link candidate construction judgment on any pair of nodes. If the spatial distance between two nodes is less than the system-defined effective communication radius, and there is a precedent of successful communication establishment in the historical record, then the node pair is considered a link candidate. Even if the two nodes are not currently directly connected, if they are indirectly reachable in the network topology through one or two intermediate nodes, and the current spectrum conditions are free of interference or conflict, they can also be marked as potential reconnectable links. The system organizes node pairs that meet the above conditions into link association data.

[0155] 32. Extract communication features from the link-related data to obtain communication feature data;

[0156] Specifically, the system designs various feature metrics from the physical layer, link layer, and structural correlation dimensions. Physical layer features include average received signal strength (RSSI) between links, channel quality metrics (such as signal-to-noise ratio), communication link length, and the noise power level of the link's environment. Link layer features cover the average number of outages per unit time, the average throughput of historical observations, and the link's packet loss rate. Structural correlation features include the difference in the number of neighbors of connected nodes (reflecting the structural distribution balance) and the difference in sub-cluster stability feature data of connected nodes (used to determine whether a link connects to a weakly stable node). The system standardizes various features, such as interval normalization or standard deviation normalization. The system organizes the above features into structured data output, represented in the form of a matrix or tensor, where each row corresponds to a link and its extracted multidimensional feature vector.

[0157] 33. Perform deep learning prediction on communication feature data to obtain communication uncertainty data.

[0158] Specifically, the system can flexibly select an appropriate model structure based on the characteristics of the scenario. For example, for prediction tasks involving static link features, a multilayer perceptron model can be used to directly map features to risk scores. If the links have complex topological relationships, a graph neural network structure (such as a convolutional network) can be introduced to model the spatial relationships between nodes while maintaining feature representation capabilities. For link behavior prediction involving time-series evolution trends, a recurrent neural network (such as LSTM) or a one-dimensional convolutional network (CNN-1D) can be used to model the historical feature sequences of the links. During the training phase, the system constructs prediction labels based on the historical link states. For example, a fixed time window (such as the next 3 seconds) is set. If the link experiences interruption or strong signal fluctuations within this time period, it is labeled as high uncertainty; otherwise, it is labeled as low uncertainty. The label format can be a binary label (such as 1 for high risk and 0 for low risk) or a probability distribution value representing the prediction confidence. The prediction results are output on a per-link basis. The system generates an uncertainty score for each link, with the score ranging from 0 to 1. A higher score indicates a more unstable future communication state and higher uncertainty for the link.

[0159] Preferably, the broken-chain topology local learning includes:

[0160] 41. Construct a local learning graph of the broken link topology based on communication uncertainty data to obtain local learning graph data;

[0161] Specifically, the system inputs two types of data: first, communication uncertainty score data, representing the estimated communication stability between node pairs in the future; and second, non-stationary topology data, providing the connection relationships between nodes in the current network and the structural state before a link break. The system sets an uncertainty screening threshold; when the communication uncertainty score of a link exceeds this threshold, it indicates that the link has potential instability or a risk of link breakage. The system includes the two endpoint nodes corresponding to links that meet the screening criteria into the node set of the local learning graph; and includes the edge relationships formed by these links themselves into the edge set. The system sets a one-hop adjacency rule to supplement any missing adjacent node connections in the edge set. The system encapsulates the constructed node set and edge set into a local learning graph, with each edge accompanied by key attribute information such as communication uncertainty value and historical topology label.

[0162] 42. Encode the local topological evolution features of the local learning graph data to obtain local topological evolution feature data;

[0163] Specifically, the system constructs a continuous temporal snapshot sequence of the local learning graph within a set time window (e.g., the most recent few frames). Each snapshot records the activity state of nodes in the graph, the existence state of edges (e.g., addition or disappearance), and corresponding communication quality data, including signal strength, number of link interruptions, etc. During feature encoding, the system extracts features from both the node and edge levels. The node level includes the number of node connections evolving over time (i.e., the node degree sequence), the degree of change in betweenness centrality (representing the change in the bridging role of nodes in the local graph), and the frequency of node sending or receiving behaviors. The edge level includes the duration of continuous connections, the magnitude of signal strength changes (e.g., moving average difference), and the time series of packet loss rate or communication interruption frequency on the link. These features are uniformly encoded into a tensor format suitable for input into the deep learning model. The local topology evolution feature data output by the system is presented in a structured tensor form, recording the multidimensional feature evolution trajectory of each node or link within the time interval.

[0164] 43. Deep learning is performed on local topological evolution feature data to obtain local prediction data.

[0165] Specifically, the system selects appropriate deep learning models for training and inference based on the required modeling granularity and task complexity. If the focus is on static topological features, fully connected neural networks such as multilayer perceptrons can be used for modeling; if the structural relationships between nodes need to be considered, graph convolutional networks (such as GCN) or graph attention networks (such as GAT) are used; if dynamic evolution over time is involved, temporal graph convolutional networks (such as T-GCN), graph convolutional recurrent networks (GCRN), or graph-based Transformer models are introduced. The system can set different prediction objectives according to specific application scenarios. For example, at the node level, it can predict the probability of connection loss at each node within a future period; at the edge level, it can predict whether a link will be interrupted or re-established; at the local structure level, it can determine whether the subgraph is in a stable state or has a high risk of structural change. If the model is used for supervised learning, the training phase needs to extract real broken or reconnected events as label samples based on historical topological evolution data. For example, the disappearance of an edge within several subsequent frames can be considered a broken link. The local prediction data output by the system is presented in a structured form, including but not limited to the connection stability score of each node, the prediction results of the link breakage trend, and the stability evaluation index of the entire local structure.

[0166] Preferably, the cross-domain opportunistic action learning includes:

[0167] Cross-domain reconnection region identification is performed based on local prediction data to obtain cross-domain reconnection region data.

[0168] Specifically, the system uses the current topology snapshot as a reference and combines it with node and link breakage prediction information to screen unstable parts of the communication structure. For links with a predicted breakage probability exceeding a set threshold, if the two nodes they connect are located at the edges of different sub-clusters, they are considered potential cross-domain boundary candidate links. If a node has a high disconnection score (1 - connection stability score) and its neighboring area lacks stable link support, the node will be identified as a potential isolated node. These boundary links and isolated nodes together constitute a candidate region set. The system performs connectivity expansion and fusion processing on the candidate regions. Using a graph traversal algorithm, starting from the aforementioned nodes, the system gradually expands outwards along links that are structurally weakly stable, constructing a range of regions with potential reconnection needs. For regions that are spatially adjacent and topologically related, the system merges them into a unified cross-domain reconnection region. In the cross-domain reconnection region data output by the system, each region includes a list of boundary nodes within the region, a description of the current internal structural state (such as the number of nodes, link density, etc.), and an overall stability score for the region.

[0169] Cross-domain opportunity action space is constructed from cross-domain reconnection area data to obtain cross-domain opportunity action space data;

[0170] Specifically, the system defines several executable communication optimization action types, constituting a basic action space. These action types include, but are not limited to, deploying relay nodes in the broken area to repair the communication link; adjusting the routing path to bypass the failed node; increasing node transmission power to enhance signal coverage; moving mobile nodes closer to the target area to restore connectivity; and switching to a less congested backup channel in a multi-channel environment. The system determines the communication status within each cross-domain reconnection area to filter feasible action types under the current conditions. For example, when an isolated node exists in the area, the system will activate actions such as "relay deployment" or "node proximity"; when the communication stability of the link within the area is low, actions such as "power enhancement" or "channel switching" will be activated; and when there is a structurally non-directly connected but line-of-sight-reachable potential path, "path switching" actions are permitted. The system generates all possible action combinations that satisfy the current environmental constraints for each cross-domain region, and configures execution costs (such as energy consumption and communication latency) and feasibility scores for each action group (e.g., based on node resource availability, regional terrain limitations, etc., using one index or a normalized weighted sum of multiple indices). The system encapsulates and generates a complete cross-domain opportunity action space data structure containing action type, triggering basis, cost evaluation, and implementation conditions. It then encodes the cross-domain opportunity action space data using cross-domain opportunity action state features to obtain action state feature data.

[0171] Specifically, the system treats each group of "cross-domain regions and their corresponding actions" as an encoding unit, extracting both the region-level and action-level features of that group. Action types are represented using one-hot encoding. Numerical features (such as probability, cost, and density) are uniformly normalized to a unified range. All feature fields are concatenated into a structured vector in a preset order.

[0172] Deep cross-domain opportunistic action learning is performed based on action state feature data machine to obtain data for broken-chain cross-domain reconnection.

[0173] Specifically, the system selects a policy model structure suitable for action learning tasks, preferably including deep neural networks, attention mechanism models (such as Transformer-based structures), or reinforcement learning policy networks. The model input is pre-encoded action state feature data, and the output is the probability distribution of the corresponding action selection or the expected utility score. Regarding the learning mechanism, the system can employ reinforcement learning methods, such as Deep Q-Networks (DQN) or Proximal Policy Optimization (PPO), as the core policy learning framework. Here, the current action state features serve as the state input, and candidate action combinations serve as the set of available actions. The system optimizes the policy according to a predefined reward function, which can be comprehensively evaluated based on multiple dimensions such as the probability of successful communication recovery, expected energy consumption, and the degree of topological perturbation introduced by the action. If a supervised learning approach is used, the system can utilize historical successful reconnection data or simulated optimal action samples as labels to train the model to fit the optimal action selection rules. For each cross-domain reconnection region, the system outputs a set of optimal opportunity action combinations, i.e., disconnected cross-domain reconnection data, to guide communication recovery or reconstruction operations in real-world network environments.

[0174] Taking a Deep Q-Network (DQN) as an example, the input data is the action-state feature vector constructed by the system. This feature vector integrates information from each region-action pair, such as one-hot encoding of the action type, region-level link failure probability score, number of affected nodes involved in the action, execution cost (such as energy consumption or latency), and expected communication recovery probability. This input vector is uniformly normalized and formed into a fixed-length vector, which serves as the state input to the Q-network. The intermediate structure is a deep neural network used for Q-value estimation, and its basic structure includes several fully connected layers (FC). For example, the network structure is set such that the input layer receives a state feature vector of dimension d, the first layer is a ReLU activation layer with 128 neurons, the second layer is a ReLU activation layer with 64 neurons, and the third layer outputs a linear layer with k neurons, where k is the number of candidate actions, corresponding to the Q-value score of each action. Specifically, the network structure is set as follows: Input layer: dimension d (e.g., d=10, representing the length of the action state features); First hidden layer: 128 neurons, activation function ReLU; Second hidden layer: 64 neurons, activation function ReLU; Output layer: k neurons, outputting the Q-value of each action, no activation function. The output data is a vector of Q-values ​​for all available actions in the current state. The system selects actions according to an ε-greedy policy: it randomly explores with a probability of ε (initial preset value) and selects the action with the largest Q-value as the current policy action with a probability of 1-ε. During training, the system generates state-action-reward-next state (s,a,r,s′) quadruplets through interaction with the simulation environment and stores them in the experience replay pool. Each training iteration samples a batch of data from the experience pool, inputs the current state s into the Q-network, outputs the Q-value of the corresponding action set, and extracts the predicted Q-value for the current action. Simultaneously, the next state s′ is input into the target Q-network, the maximum Q-value is obtained as the expected value, the TD error is calculated using the Bellman equation, and then gradient descent optimization is performed using the mean squared error loss function. The system periodically synchronizes the parameters of the main Q-network and the target Q-network. The entire training process iterates until the Q-value converges or the policy performance meets the task requirements. The system can output the optimal action selection policy for different reconnection regions, realizing a cross-domain reconnection scheduling mechanism with adaptive learning capabilities.

Claims

1. A spatial channel dynamic characteristic prediction system based on deep learning, characterized in that, include: The non-stationary topology detection module is used to acquire broken link communication data; Non-stationary topology detection is performed based on broken link communication data to obtain non-stationary topology data; The sub-cluster feasible communication domain construction module is used to construct the sub-cluster feasible communication domain based on non-stationary topology data, and obtain the sub-cluster feasible communication domain data. The sub-cluster feasible communication domain construction is that at each non-stationary time, the system identifies multiple locally connected node sets, i.e. locally reachable regions, based on the node connection relationship in the topology graph. The system constructs an edge weight graph to represent the link quality for each locally reachable set. This edge weight graph integrates multi-dimensional indicators such as link loss, distance between nodes, and channel interference. Based on this edge weight graph, the system uses different methods to extract the communication skeleton structure of the sub-cluster, including constructing a minimum spanning tree to obtain the skeleton with minimum connection cost, or constructing a multi-edge connected subgraph to retain the backbone link. The sub-cluster feasible communication domain data is the skeleton edge set and skeleton node set of each sub-cluster. The communication uncertainty prediction module is used to predict communication uncertainty based on the feasible communication domain data of the sub-cluster, and obtain communication uncertainty data. The cross-domain opportunistic action learning module is used to perform local topology learning for broken links based on communication uncertainty data to obtain local prediction data. Based on the local prediction data, cross-domain opportunistic action learning is performed to obtain cross-domain reconnection data for broken links. The cross-domain opportunistic action learning is based on the predicted outward connection capability of boundary nodes. Only when the spatial distance is within the communication reachable range and the boundary nodes crossing both sub-clusters have outward connection potential are included in the candidate region. An executable opportunistic action space is constructed, which is jointly defined by the execution node, action type and its specific parameters, and encoded as a feature vector. A deep scoring network or heuristic optimization strategy is used to score and rank the actions, and the highest-scoring action is selected. The cross-domain reconnection data for broken links includes the selected cross-domain region, the key nodes involved, the type of action to be executed and its parameter configuration, the time window for the expected execution action, the success prediction probability of the reconnection strategy and the potential impact assessment on the communication structure of the local domain.

2. The system according to claim 1, characterized in that, The non-stationary topology detection includes: During the operation of the drone swarm, disconnected communication data is periodically collected from each communication node; Based on the broken-link communication data, a communication topology snapshot is processed to obtain communication topology snapshot data; Topology change features are extracted from communication topology snapshot data to obtain topology change feature data; Nonstationary discriminant features are calculated based on topological change feature data to obtain nonstationary discriminant feature data; Nonstationary detection is performed on nonstationary discriminant feature data to obtain nonstationary topological data.

3. The system according to claim 2, characterized in that, The topology change feature extraction includes: Structural alignment is performed based on communication topology snapshot data to obtain structural alignment data; Based on the structure alignment data, connectivity configuration morphology processing is performed to obtain connectivity configuration morphology data. The connectivity configuration morphology processing involves extracting the connection relationships within one-hop or two-hop neighborhood of each communication node from its corresponding topological adjacency matrix and constructing the connectivity configuration graph of that node at the current moment. The connectivity configuration morphology data is the data formed by extracting the connection relationships within one-hop or two-hop neighborhood of its corresponding topological adjacency matrix, performing structural encoding, and organizing the data in chronological order. Cross-time offset calculation is performed on the connectivity configuration morphology to obtain cross-time offset data. The cross-time offset calculation is to compare the connectivity configuration of each communication node between two consecutive time points and calculate the configuration difference. The cross-time offset data is the structural offset time series of each node at each time point, as well as the system-level offset trend index. Connected field energy state convergence is performed on cross-time series offset data to obtain connected field energy state data. Connected field energy state convergence is to map the structural offset value of each node at each time step to the corresponding local energy value. With the communication topology adjacency graph as a constraint, the system calculates the average neighborhood energy of each node, judges the stability of local nodes, and constructs a connected field energy state graph as a global representation. The connected field energy state data is the steady-state energy value sequence of each node and the connected field energy state graph organized in the form of a graph structure. Connectivity configuration rearrangement features are extracted from the connected field energy state data to obtain topological structure change feature data.

4. The system according to claim 1, characterized in that, The communication uncertainty prediction includes: Link association is performed based on the feasible communication domains of the sub-cluster to obtain link association data; Communication feature data is obtained by extracting communication features from the link association data; Deep learning is used to predict communication feature data to obtain communication uncertainty data.

Images