Intelligent device connection retention method for automatic switching of multi-modal communication links

Abstract

This invention relates to a smart device connectivity maintenance method for automatic switching of multimodal communication links, aiming to solve the problems of inconsistent state data, context drift, and limited real-time adaptability of end-side models in multi-source heterogeneous links. The technical solution uses a graph neural network model to perform multimodal standardization, dynamic topology mapping, and deep extraction of spatiotemporal features on link state data, including signal quality, service load, device movement trajectory, and environmental interference information, generating a multi-candidate link switching strategy distribution. On the end side, a lightweight temporal graph neural network is used, along with cloud-based knowledge distillation and single-step strategy alignment mechanisms, to calibrate local inference model parameters in real time, improving the accuracy and robustness of switching decisions. This solution significantly improves the real-time performance, stability, and service continuity of communication link switching, and can adaptively adjust the end-side knowledge synchronization frequency to achieve optimal communication experience under end-cloud collaboration.

Description

Technical Field

[0001] This invention relates to the field of multimodal communication and network adaptive decision-making technology for smart terminals, and in particular to a method for maintaining the connection of smart devices by automatically switching multimodal communication links. Background Technology

[0002] Currently, in the field of multimodal communication link switching strategy generation and connection maintenance technology for smart terminals, mainstream systems typically adopt the following technical paths: First, automatic link switching is achieved through matching with a local static rule base on the device, based on preset signal thresholds, switching priority tables, and service type classifications. The response speed of this approach is limited by the frequency of rule base maintenance and is difficult to adapt to complex dynamic environments. Second, full-scale deep learning models, such as LSTM and DQN, are deployed offline for end-to-end link decision training on the device. However, due to limitations in embedded computing power and storage resources, model performance is often constrained by compressed weights and inference latency. Third, centralized decision-making in the cloud involves collecting context data from the device side and generating switching commands in the cloud before sending them to the device. However, this mode is prone to policy command delays or failures in weak network or unstable backhaul links. On the other hand, some industry solutions attempt to optimize strategies through OTA commands for forced synchronization, rapid historical database searches, or fuzzy evaluation mechanisms. However, these lack the ability to continuously adapt to real-time context drift, easily leading to policy lag, frequent link jumps, or connection interruptions in high-availability scenarios.

[0003] In terms of representative technologies, existing publicly available multimodal communication link switching decision systems are commonly used in environments such as high-speed rail, tunnels, and industrial sites, employing a collaborative communication decision architecture between local and cloud environments. Typical applications include intelligent terminal devices generating periodic link switching suggestions based on environmental interference event labels, movement trajectory prediction, and signal quality parameters, or utilizing fixed-point deep neural network models to optimize link priority ranking. However, these technologies mostly face challenges such as limited edge-side model capabilities, long-term drift between local and cloud-based global policy distributions, difficulty in achieving millisecond-level response rates for new events, and the tendency for switching commands to deviate from real-world business loads. Especially in edge-side embedded devices, the deployment of complete models, online training protocols, and secondary confirmation mechanisms for commands can easily lead to response delays and resource conflicts, failing to fully guarantee the requirements for adaptive capabilities and high connection availability.

[0004] Analyzing the shortcomings of existing technologies reveals the following prominent issues: Local decision-making modules lag in perceiving dynamic context changes, and switching strategies often fail to keep pace with sudden changes in the communication environment (such as tunneling, sudden interference, and heavy-load service switching). This leads to delays in link switching timing and the failure to activate high-priority links in a timely manner, thus affecting data transmission continuity and service uninterruption. Simultaneously, cloud-based large-scale model knowledge cannot be efficiently migrated to the edge. Traditional static rule or full-model deployment methods are limited by computing and storage resources, making it difficult to support context adaptation and continuous policy evolution. Furthermore, the lack of an effective feedback loop for link switching effects means that policy performance degradation and distribution drift cannot be dynamically monitored and corrected. This causes edge-side policies to deviate from the optimal cloud solution for extended periods, impacting the overall system reliability and connection stability in high-availability scenarios.

[0005] In summary, existing technologies urgently need to overcome bottlenecks such as delayed response to local link policy generation, insufficient real-time context adaptation capabilities, low efficiency of cloud knowledge transfer, and difficulty in closed-loop feedback of policy execution effects. This paper proposes a novel technical solution that can achieve high-precision dynamic distillation of cloud-based policy knowledge, adaptive evolution of link switching policies, and millisecond-level continuous connection maintenance under resource-constrained conditions. This solution meets the real-time stability and high availability adaptive requirements of multimodal communication link automatic switching systems in complex dynamic environments. Summary of the Invention

[0006] This application provides a smart device connection maintenance method for automatic switching of multimodal communication links, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0007] This application provides a method for maintaining connectivity in intelligent devices with automatic switching of multimodal communication links, specifically including: S1: Obtain historical data on the status of multi-source heterogeneous links, and construct a graph neural network model based on the multi-source heterogeneous link status data; S2: Obtain multi-source heterogeneous link status data, and use the graph neural network model to perform spatiotemporal feature encoding processing on the multi-source heterogeneous link status data to generate a cloud strategy output distribution vector; S3: Deploy the edge graph neural network model, wherein the edge graph neural network model retains the graph structure encoding layer and the temporal attention gating module, and initializes the policy adaptation layer parameters embedded in the edge graph neural network model based on the cloud policy output distribution vector; S4: Monitor local communication context change characteristics. When the device is detected to meet the context drift conditions of entering a tunnel, sudden electromagnetic interference, or service type switching, trigger a single-step policy alignment mechanism based on the cloud policy output distribution vector. S5: Execute the single-step policy alignment mechanism, calculate the KL divergence constraint value between the current output distribution of the edge graph neural network model and the cloud policy output distribution vector, and update only the learnable parameters in the policy adaptation layer to generate the aligned policy adaptation layer. S6: Input the latest local link status data into the end-side graph neural network model loaded with the aligned strategy adaptation layer, and infer to generate a real-time optimal switching strategy instruction that adapts to the current dynamic context.

[0008] S7: Execute the real-time optimal switching strategy instruction to complete the multimodal communication link switching operation, and collect strategy execution effect feedback data including switching time, connection recovery delay and number of service interruption frames.

[0009] S8: Based on the feedback data of the strategy execution effect, compare the difference with the preset baseline. If the difference in the distribution of the end-side strategy exceeds the threshold or the execution effect is lower than the baseline within several consecutive periods, dynamically adjust the distillation frequency and trigger a new round of cloud strategy output distribution vector distribution request.

[0010] The intelligent device connection maintenance method with automatic switching of multimodal communication links provided in this application has the following beneficial effects: (1) By constructing a collaborative architecture of a cloud-based global policy graph neural network and a lightweight temporal graph neural network on the edge, this invention effectively overcomes the lag in dynamic context awareness of traditional switching policy generation methods. Existing technologies generally rely on preset threshold judgment or pattern matching based on historical databases, which is difficult to adapt to complex and ever-changing communication environments. Especially in scenarios such as high-speed device movement, sudden interference, or frequent switching of service types, the policy response delay is high and the decision accuracy is low. This solution uses a cloud model to perform fine-grained modeling of multi-source heterogeneous data (including signal quality, load status, movement trajectory, and environmental interference labels), and outputs a candidate link priority sequence and confidence distribution with spatiotemporal correlation, realizing a deep understanding of the global context state. The edge model, as a knowledge receiver, retains only key structural modules and embeds an adaptation layer that can be updated quickly, completing policy alignment without the need for complete reasoning and backpropagation, significantly improving the real-time performance and accuracy of local decision-making. This novel paradigm of "cloud-based knowledge supply + lightweight edge evolution" overcomes the bottlenecks of traditional DQN online training, such as high computational overhead, LSTM prediction error accumulation, and delays introduced by secondary confirmation of OTA commands. It enables the system to complete policy adjustments within hundreds of milliseconds, effectively ensuring the timeliness and stability of communication link switching in highly dynamic environments.

[0011] (2) By introducing a single-step policy alignment mechanism based on KL divergence constraints and an adaptive distillation triggering strategy driven by context drift, this invention achieves an efficient balance between continuous optimization of switching strategies and resource efficiency. Unlike traditional methods that rely on a large number of local training samples or frequent uploads of raw data to support model iteration, this scheme uses a secure encrypted channel to distribute cloud policy distribution and feature embeddings in a vectorized manner. The edge side achieves policy approximation through single-step parameter updates, avoiding complex gradient calculations and storage consumption, and greatly reducing the computing power requirements and energy consumption of the edge side. At the same time, the policy execution effect feedback mechanism provides the cloud with a closed-loop evaluation basis in real scenarios, which is used to optimize the selection strategy of subsequent distillation samples and enhance the generalization ability of the model in actual deployment. Furthermore, the system dynamically determines whether to trigger a new round of distillation by monitoring the difference between edge and cloud policy distribution and local execution performance, and can automatically adjust the distillation frequency and feature dimension granularity according to network conditions, minimizing bandwidth consumption while ensuring policy timeliness. This mechanism is particularly suitable for edge scenarios with weak networks or even disconnected backhaul, enabling the terminal to maintain a better strategy output even without a stable cloud connection, which significantly improves the system's robustness and autonomous adaptability.

[0012] (3) Overall, this invention constructs an intelligent switching strategy generation system with continuous learning capabilities, high response sensitivity, and strong environmental adaptability. It not only solves the problem of strategy failure caused by static rule setting and model update lag in the prior art, but also realizes a paradigm shift from "passive execution" to "active evolution" at the system level. Compared with traditional methods that rely on fuzzy evaluation or multi-objective weighting, this solution retains the uncertainty expression and multi-path selection potential of strategy decision-making through probability distribution transmission rather than deterministic command issuance, and enhances the flexibility in dealing with unknown disturbances. Its lightweight design is compatible with a wide range of terminal forms and can be deployed without additional hardware upgrades, with good scalability and engineering implementation value. Especially in application scenarios with extremely high requirements for connection continuity, such as vehicle networking, industrial IoT, and emergency communication, it can effectively reduce the number of service interruption frames, shorten connection recovery delay, and comprehensively improve the service quality and user experience of multimodal communication links. Attached Figure Description

[0013] Figure 1 This is the main flowchart of a smart device connection maintenance method for automatic switching of multimodal communication links.

[0014] Figure 2 This is a sub-flowchart of a smart device connection maintenance method for automatic switching of multimodal communication links.

[0015] Figure 3 This is another sub-flowchart of a smart device connection maintenance method for automatic switching of multimodal communication links. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0018] like Figure 1 As shown, this application provides a method for maintaining connectivity of intelligent devices with automatic switching of multimodal communication links, specifically including: S1: Obtain historical data on the status of multi-source heterogeneous links, and construct a graph neural network model based on the multi-source heterogeneous link status data; S2: Obtain multi-source heterogeneous link status data, and use the graph neural network model to perform spatiotemporal feature encoding processing on the multi-source heterogeneous link status data to generate a cloud strategy output distribution vector; S3: Deploy the edge graph neural network model, wherein the edge graph neural network model retains the graph structure encoding layer and the temporal attention gating module, and initializes the policy adaptation layer parameters embedded in the edge graph neural network model based on the cloud policy output distribution vector; S4: Monitor local communication context change characteristics. When the device is detected to meet the context drift conditions of entering a tunnel, sudden electromagnetic interference, or service type switching, trigger a single-step policy alignment mechanism based on the cloud policy output distribution vector. S5: Execute the single-step policy alignment mechanism, calculate the KL divergence constraint value between the current output distribution of the edge graph neural network model and the cloud policy output distribution vector, and update only the learnable parameters in the policy adaptation layer to generate the aligned policy adaptation layer. S6: Input the latest local link status data into the end-side graph neural network model loaded with the aligned strategy adaptation layer, and infer to generate a real-time optimal switching strategy instruction that adapts to the current dynamic context.

[0019] S7: Execute the real-time optimal switching strategy instruction to complete the multimodal communication link switching operation, and collect strategy execution effect feedback data including switching time, connection recovery delay and number of service interruption frames.

[0020] S8: Based on the feedback data of the strategy execution effect, compare the difference with the preset baseline. If the difference in the distribution of the end-side strategy exceeds the threshold or the execution effect is lower than the baseline within several consecutive periods, dynamically adjust the distillation frequency and trigger a new round of cloud strategy output distribution vector distribution request.

[0021] Step S1: Obtain historical data on the status of multi-source heterogeneous links, and construct a graph neural network model based on the multi-source heterogeneous link status data. The multi-source heterogeneous link status data includes signal quality indicators, service load parameters, device movement trajectory coordinates, and environmental interference event labels, specifically including: S1.1: Collect and standardize multi-source heterogeneous link status data such as signal quality indicators, service load parameters, equipment movement trajectory coordinates, and environmental interference event labels to eliminate differences in sampling frequency and dimensions of different sensors and generate a unified format of multimodal time-series feature vector sequence.

[0022] The system uniformly collects multi-source heterogeneous link status data streams, including signal quality indicators, service load parameters, equipment movement trajectory coordinates, and environmental interference event labels. The specified collection objects include the signal-to-noise ratio and received power values ​​output by the radio frequency status monitoring module, the service queue length and throughput collected by the network interface, the time-series location information output by the positioning unit, and the time and type code of interference events recorded by the environmental monitoring unit.

[0023] For data from different sources, a frequency alignment algorithm is invoked to interpolate and resample the original sampling sequences from different sensors on a unified time scale. A hybrid strategy based on linear interpolation and spline interpolation is adopted to ensure that high-speed changing parameters maintain accuracy while low-frequency changing parameters maintain stability.

[0024] Dimensional normalization is performed on each data channel, with the normalization benchmark specified as the global extreme value range of each parameter over the past M periods, using the following normalization formula: in This is the current sampled value. and These are the minimum and maximum values ​​of the parameter within the reference period, respectively. This formula maps data of different dimensions to the standard range of 0 to 1.

[0025] Noise suppression and outlier removal were performed on the standardized time series. Noise suppression employed a combination of moving mean filtering and bidirectional median filtering. Outlier removal used a three-standard-deviation threshold, determined by the following formula: in The mean of the sequence. The standard deviation of the sequence is used to detect and eliminate sudden abnormal samples.

[0026] Multimodal feature synchronous encoding is performed on the sequence after outlier removal. The sampling vector in each time scale is the concatenation result of four types of features: signal quality, service load, trajectory coordinates, and environmental label. A multimodal time-series feature vector sequence with a unified format is formed by feature concatenation operator.

[0027] Through the above processing method, the original multi-source heterogeneous link state data in the previous step is transformed into a multimodal temporal feature vector sequence with a unified time base, unified dimensions, low noise, and consistent structure, so as to realize the high-quality input required for subsequent topology mapping and spatiotemporal modeling of graph neural networks.

[0028] For example, in an industrial automation wireless monitoring scenario, the radio frequency detection module acquires signals-to-noise ratios ranging from 15 to 35, with a received power of... 80 to 50dBm; network interface monitoring service queue length ranges from 0 to 2000, throughput ranges from 0 to 100Mbps; the trajectory coordinate change radius output by the positioning module at 1Hz frequency is approximately 30 meters; the environmental monitoring module records an average of 2 interference events per hour. When aligning sampling frequencies, RF and network data are interpolated from 5Hz to 1Hz to synchronize with positioning and environmental data; for dimension normalization, signal-to-noise ratio and received power are normalized according to their respective extreme values ​​over the last 600 seconds; service queue length and throughput are also normalized according to their extreme values ​​over the last 600 seconds; for anomaly removal, a threshold of three standard deviations is set to detect and remove trajectory jump values. The resulting multimodal temporal feature vector has 16 dimensions per time scale, covering signal-to-noise ratio, received power, queue length, throughput, trajectory X / Y coordinates, environmental interference type encoding, etc. Quality verification shows that this feature vector sequence significantly improves the stability and accuracy of link state embedding in subsequent graph structure mapping.

[0029] S1.2: Based on the multimodal temporal feature vector sequence, perform dynamic topology mapping processing, use the sliding time window mechanism to convert the feature vectors of continuous time moments into node attributes, and construct edge connection relationships according to the communication association strength between devices to generate dynamic communication link graph structure data containing spatiotemporal attributes.

[0030] Based on the unified format multimodal temporal feature vector sequence generated in the previous sub-step, a sliding time window constructor is called when calculating node attributes to segment the feature vector sequence according to a preset window length and step size, so that each slice corresponds to a single moment of observation in the dynamic communication topology, and this state is encoded as a node attribute matrix. After forming a node set, correlation coefficient calculation and mutual information evaluation are performed on the communication association strength between each node, and an association strength value vector is generated based on the synchronous change ratio of both parties in three dimensions: signal quality, service load, and location trajectory. Based on the association strength value vector, a weighted threshold determination module is applied to establish edge connections for node pairs that exceed a set association threshold. An adjacency matrix is ​​used to store the edge set to ensure the instant retrieval of the spatial topology. To simultaneously preserve the temporal evolution characteristics of the links, a timestamp and a time weight factor are added to each edge of the adjacency matrix. A dynamic matrix update mechanism is used to increment or decrease the edge weights within the sliding window to reflect the changing trend of the communication state over time. Topology formatting is performed under the combination of the above node attribute matrix and the time-weighted adjacency matrix to generate dynamic communication link graph structure data containing node spatial attributes, edge connection relationships, and time series markers. This processing method transforms the temporal feature vector from the previous step into a spatiotemporal attribute graph that can be directly input into the graph neural network, ensuring the accuracy of subsequent high-order neighborhood and temporal evolution feature extraction. For example, in the multimodal communication link monitoring of industrial automation scenarios, standardized feature vector sequences of processing robot A and inspection robot B are collected, and the sliding window length is set to... seconds, step size The sequence is sliced ​​to form a node attribute matrix for each device at consecutive time points. The communication association calculation module is invoked, and the signal quality weighting coefficient is set to... The business load weighting coefficient is The position trajectory weighting coefficient is Calculate the correlation coefficients of the two devices in each dimension, and obtain the overall correlation strength using the following formula: in, The correlation coefficient is the corresponding dimension, and the unit is dimensionless; when Greater than the set threshold At that time, an edge connection is established between node A and node B, and the current time timestamp and initial weight value are added to the edge attributes. During the sliding window process, a time-weight update strategy is invoked, incrementing the weight of edges that maintain high correlation. Perform decay to the lowest possible value on the edges with decreasing correlation. The resulting dynamic communication link graph structure is then input into subsequent graph convolution and temporal attention modules. Verification shows that this dynamic graph significantly improves the ability to capture the decline in link quality before device switching and can support millisecond-level switching strategy prediction.

[0031] S1.3: Perform graph convolution aggregation and temporal attention weighting on the dynamic communication link graph structure data, extract high-order neighborhood dependency features by stacking multiple graph neural network layers, and capture the temporal evolution law by combining long short-term memory gating mechanism to generate a deep link state embedding representation rich in spatiotemporal context information.

[0032] The graph neural network model described is a cloud-based global deep learning model. Its structure consists of an input layer, a multi-layer graph convolutional network, a temporal attention gating module, and an output layer with a fully connected layer and a Softmax function stacked sequentially. The input layer receives multi-source heterogeneous link state data and converts it into dynamic communication link graph structure data. The multi-layer graph convolutional network extracts spatial dependencies and high-order neighborhood features between nodes through layer-by-layer neighborhood aggregation and a learnable spatial weight matrix, and introduces residual connections to retain low-order information. The temporal attention gating module uses causal convolution and long short-term memory gating mechanisms to capture local evolution patterns and long-term dependency rules on the time axis. The output layer maps deep spatiotemporal features to priority scores of candidate links through a fully connected layer, and then normalizes them through a Softmax function to generate a cloud-based policy output distribution vector containing fine-grained switching priority sequences and confidence intervals for multiple candidate links.

[0033] For dynamic communication link graph structure data containing spatiotemporal attributes, a multi-layer graph convolution operation unit is loaded to perform spatial neighborhood aggregation on the attribute vectors of each node. The signal quality and environmental interference labels of adjacent nodes are mapped and fused through a learnable weight matrix to form a preliminary spatial context feature representation.

[0034] Based on the preliminary spatial context feature representation input to the temporal attention weighting module, the correlation weights of feature vectors between consecutive time steps are calculated. Causal convolution kernels are used to extract local evolution patterns on the time axis, and attention enhancement processing is performed on highly correlated sequences to improve the sensitivity to long-term and short-term changes.

[0035] The long short-term memory gating mechanism is invoked to perform gating selection and state update operations on the feature tensor enhanced by temporal attention. The synergistic effect of the input gate, forget gate and output gate is used to retain long-term dependency information, while avoiding the interference of noise on short-term trends, and generating a comprehensive temporal context feature matrix.

[0036] The comprehensive temporal context feature matrix is ​​input into the subsequent stacked multi-layer graph neural network structure to capture high-order neighborhood dependency features layer by layer, forming a high-dimensional spatial semantic representation that can characterize the relationship between distant nodes, and residual connections are introduced between layers to retain low-order feature information.

[0037] Through the above processing methods, the dynamic communication link graph structure data is transformed into a deep link state embedding representation rich in spatiotemporal context information, thereby achieving an accurate characterization of the global spatial correlation and dynamic temporal evolution of the link state.

[0038] For example, in an industrial automation control scenario, multi-source link state data from five types of sensors are standardized to generate node attribute vectors, with each node containing a signal strength value set as follows: 45dBm, service load parameters set to 250kbps, trajectory coordinate sampling frequency set to 10Hz, interference labels set to binary variables. The time window length is set to 8 sampling periods, the convolution kernel size is set to 3, and the initial value of the graph convolution weight matrix is ​​[…]. Random sampling is performed within the interval [0.1, 0.1]. In the temporal attention calculation, the number of attention heads is set to 4, and the correlation weight between time step i and time step j is calculated using the following formula: in The query vector is generated from node feature mapping. The deep link-state embedding generated after LSTM gating significantly improves the response speed to link-state abrupt changes in offline evaluation, maintaining link switching latency below 25ms under interference events. The final output embedding representation can be used for subsequent fully connected mapping to accurately derive priority scores, enabling online generation of high-availability link switching strategies.

[0039] S1.4: Based on the deep link state embedding representation, perform fully connected mapping and probability normalization processing, and use the soft maximum function to map the high-dimensional features into priority scores and confidence intervals of multiple candidate links, generating a graph neural network model to guide the evolution of end-side strategies.

[0040] Step S2: Obtain multi-source heterogeneous link status data, and use the graph neural network model to perform spatiotemporal feature encoding processing on the multi-source heterogeneous link status data to generate a cloud policy output distribution vector. Specifically, this includes: S2.1: Obtain multi-source heterogeneous link status data containing signal quality indicators, service load parameters, device movement trajectory coordinates, and environmental interference event labels. Based on graph structure mapping rules, convert the multi-source heterogeneous link status data into an initial node feature matrix and a dynamic adjacency matrix to construct an original spatiotemporal graph data structure representing the current communication topology.

[0041] For multi-source heterogeneous link status data containing signal quality indicators, service load parameters, equipment movement trajectory coordinates, and environmental interference event labels, a pre-configured data parsing module is invoked to map various inputs to a unified feature dimension space according to source type, so as to ensure consistent feature index positions in the subsequent graph model.

[0042] For the parsed single-source dataset, a time-series index alignment algorithm is used to achieve synchronous labeling of multi-source data at the same time step. Signals with different sampling frequencies are resampled to a unified time resolution and interpolated to fill in missing values, so as to eliminate the impact of cross-source device acquisition cycle differences on feature matrix construction.

[0043] In the multi-source dataset after synchronous annotation, according to the preset graph structure mapping rules, the device status features of each time step are merged into the attribute vectors of the corresponding nodes to form an initial node feature matrix, where the matrix row index corresponds to the device node and the column index is arranged in order of feature category.

[0044] After the node feature matrix is ​​constructed, a dynamic adjacency matrix is ​​calculated based on the communication correlation strength between devices, using a weighted similarity formula based on signal quality differences and service load correlation: in, This is the signal quality similarity matrix. This is a business load similarity matrix. The larger the value of each element, the stronger the association between nodes.

[0045] The initial node feature matrix and the dynamic adjacency matrix are combined to form the original spatiotemporal graph data structure. The structure is then encapsulated into a unified data object that meets the input specifications of graph neural networks through a structure serialization mechanism to support subsequent spatial convolution and temporal attention processing.

[0046] By establishing a bidirectional mapping relationship between the initial node feature matrix and the dynamic adjacency matrix, the result of the previous step is transformed into a data structure that can directly represent the current communication topology, thereby realizing a unified spatiotemporal coding input for the state of multi-source heterogeneous links.

[0047] For example, in an industrial automation scenario, the cloud collects signal quality indicators (range -90 to -40 dBm), service load parameters (in Mbps), device movement trajectory coordinates (GPS coordinate resolution 0.5 meters), and environmental interference event labels (integer codes 0-3) from ten devices. The parsing module normalizes the signal quality to the 0-1 range, normalizes the service load to a maximum of 50 Mbps, represents the trajectory coordinates using three-dimensional position vectors, and maps the interference labels to one-hot encoded vectors. The synchronous annotation process resamples all inputs to one time step per second and uses linear interpolation to handle missing values ​​within the sampling interval. The node feature matrix has 10 rows and a feature dimension length of 20. The dynamic adjacency matrix constructs edge weights based on the condition that the absolute value of the signal quality difference is less than 5 dBm and the service load difference is less than 2 Mbps. (The formula is missing from the provided text.) and The edge weights, obtained by back-calculation using normalized differences, range from 0.7 to 0.95. This original spatiotemporal graph data structure, after serialization and encapsulation, is input into a cloud-based global policy graph neural network. In testing, this structure significantly improves the efficiency of neighborhood feature aggregation in the spatial convolution stage and enhances the accuracy of long-term link switching policy generation.

[0048] S2.2: Perform multi-layer graph convolution aggregation operation on the initial node feature matrix in the original spatiotemporal graph data structure, and use the learnable spatial weight matrix to fuse the signal quality indicators and environmental interference event labels of adjacent nodes to generate spatially enhanced node embedding vectors containing local spatial dependencies.

[0049] S2.3: Based on the spatial enhancement node embedding vector input temporal attention gating module, a causal convolution mechanism is used to perform sliding window feature extraction on the time series of device movement trajectory coordinates and business load parameters, calculate the correlation attention weights between time steps, and generate a global spatiotemporal joint feature tensor that integrates long-term and short-term time evolution laws.

[0050] S2.4: Perform multi-head self-attention mechanism transformation on the global spatiotemporal joint feature tensor, calculate the long-distance dependency relationship between different communication link nodes by querying key-value pairs, filter out the high-priority candidate link feature subset, and generate a high-dimensional semantic interaction feature representation that represents the competition relationship of multiple candidate links.

[0051] Based on the global spatiotemporal joint feature tensor, the query-key-value matching module of the multi-head self-attention mechanism is invoked. The feature vector of each communication link node is used as the query vector input, and the feature vectors of other nodes in the global spatiotemporal joint feature tensor are used as key-value pairs for similarity calculation.

[0052] The inner product result of the query and the key is normalized using a scaling dot product operation. Specifically, the attention weight matrix is ​​calculated using the following formula: in, For querying the matrix, The key matrix, The dimension of the key vector.

[0053] Combine the above weight matrix with the value matrix Matrix multiplication is performed to generate weighted feature outputs for each attention head, and feature concatenation is performed between different attention heads to enhance multi-view long-distance dependency information.

[0054] For the spliced ​​multi-head output feature tensor, the feature filtering operator is called to calculate the comprehensive attention score of each candidate link node, and the features of the top M high-priority candidate links are extracted in descending order of scores to form a feature subset.

[0055] The feature subset is subjected to semantic mapping transformation, and the features are nonlinearly reconstructed using a multilayer perceptron, so that the interaction mode of link competition relationship is preserved in the high-dimensional latent space, and a high-dimensional semantic interaction feature representation representing the competition relationship of multiple candidate links is output.

[0056] By using the multi-head self-attention computation and high-dimensional semantic mapping processing methods described above, the global spatiotemporal joint feature tensor of the previous step is transformed into a high-dimensional semantic interaction feature representation with the ability to judge link priority and analyze competitive situation, thereby achieving accurate characterization and priority ranking of long-distance dependencies of multiple candidate links.

[0057] For example, in an industrial automation wireless communication scenario, let the feature vector dimension of each link node be... The value is 64, query matrix It consists of 10 nodes, the key matrix AND-value matrix All are mapping results of the global spatiotemporal joint feature tensor. When performing the scaling dot product calculation, the normalization factor is taken as... After calculating the attention weight matrix, multiplication with the value matrix yields the output of each head. Eight attention heads are used, each calculated individually and then concatenated into a 512-dimensional feature vector. The overall attention scores of each node are ranked, and the three highest-scoring link features are selected into a feature subset. A two-layer perceptron is then used for reconstruction (the first layer outputs 256 dimensions, and the second layer outputs 128 dimensions). In testing, this high-dimensional semantic interaction feature representation successfully reflects the priority decline trend of links under interference within the tunnel and the competitive advantage of backup links, ultimately achieving a significant improvement in the link selection strategy.

[0058] S2.5: Based on the high-dimensional semantic interaction feature representation, the probability distribution projection processing is performed through the fully connected mapping layer. The score values ​​of each link are normalized using the normalized exponential function, and a cloud strategy output distribution vector containing the fine-grained switching priority sequence of multiple candidate links and the corresponding confidence values ​​is output.

[0059] like Figure 2 As shown, step S3 involves deploying an edge-side graph neural network model. This model retains the graph structure encoding layer and the temporal attention gating module, and initializes the policy adaptation layer parameters embedded in the edge-side graph neural network model based on the cloud policy output distribution vector. Specifically, this includes: S3.1: Obtain the network weight parameters and topology definition of the graph neural network model, perform channel sparsification on the complete network weight parameters, and generate a lightweight model architecture description file.

[0060] Obtain the complete network weight parameters and topology definition of the graph neural network model, using the network weight matrix and node connection relationships fixed during the cloud model training phase as the initial input.

[0061] The described edge-side graph neural network model is a lightweight temporal graph neural network, whose specific structure consists of three parts connected in series. The first part is the graph structure encoding layer, which receives standardized multimodal link state data, performs neighborhood aggregation on nodes (i.e., candidate links) through graph convolution operators, extracts spatial topological dependencies between features such as signal quality and service load, and generates a dynamic spatiotemporal graph node embedding matrix. The second part is the temporal attention gating module, which uses a self-attention mechanism and gating units to weight and filter the temporal dimension of the node embedding matrix, captures the evolution of link state over time, and outputs key temporal context feature tensors. The third part is the policy adaptation layer, which is the only part of the model that contains learnable parameters that are updated. Its parameters are not randomly initialized, but are cold-started by inverse mapping the policy distribution vector issued by the cloud. During context drift, only this layer is updated with single-step parameters based on KL divergence constraints, thereby quickly mapping the temporal features extracted by the previous layer into a real-time policy distribution vector containing priority ranking and confidence scores of multiple candidate links, and finally generating executable link switching instructions.

[0062] Channel importance evaluation is performed on the network weight matrix. A two-factor scoring method based on weight magnitude and gradient sensitivity is used to sort the channels according to their contribution to the feature extraction process, generating a channel importance sequence.

[0063] Based on the channel importance sequence, the model pruning algorithm is invoked, and the pruning ratio and retention threshold parameters are set. Channel weights with scores below the threshold are deleted along the channel dimensions of the convolutional layer and attention layer to ensure that the removed feature channels do not significantly reduce the graph structure encoding and temporal feature capture capabilities.

[0064] Non-critical attention heads are removed from the pruned attention layer. The criteria for determining critical and non-critical attention heads are set as a joint index of the mean and variance of attention weights. Non-critical attention heads are deleted when they meet the criteria of low mean and small variance, so as to reduce redundant calculations.

[0065] The structural information of the retained graph structure encoding layer and the optimized temporal attention gating module is extracted to construct a lightweight model architecture description file containing the definitions of the remaining network layers, node index mappings, weight file paths, and input / output tensor specifications.

[0066] Through the above processing method, the result of the previous step is transformed into a lightweight model architecture file that can be loaded on the edge, has a core spatiotemporal feature extraction unit, and has significantly compressed parameters, thereby realizing the simplification of the structure and resource adaptation of the basic inference model deployed on the edge.

[0067] For example, in an industrial automation scenario, a graph neural network model is obtained with 512 channels in its convolutional layers, 8 attention heads, and an average node degree of 3. The pruning ratio is set to 0.95, and the retention threshold is set to a weight magnitude greater than [value missing]. And gradient sensitivity is greater than The channels were analyzed. In the channel importance sequence, 487 low-contribution channels were removed, retaining only 25 high-contribution channels. In the attention layer, channels with an average attention weight lower than [a certain value] were ranked. And the variance is less than The decision criteria eliminated 5 non-critical attention heads and retained 3 critical attention heads. The final lightweight model architecture description file is 2.8MB in size and contains only the node adjacency processing rules of the graph structure encoding layer and the weight mapping path of the temporal attention gating module. When the file is loaded on the edge embedded processor, the CPU usage drops to about one-fifth of that of the original model deployment, and the inference latency is significantly reduced. The verification results show that the output of the lightweight model and the original global policy model achieve a high confidence interval in link priority determination, meeting the requirements for switching strategy generation in industrial real-time communication.

[0068] S3.2: Read the lightweight model architecture description file, instantiate a lightweight temporal graph neural network model in the memory space of the embedded processor on the edge, map the remaining weight parameters after channel sparsification to the graph structure encoding layer and the temporal attention gating module, and generate an uninitialized inference model on the edge with basic spatiotemporal feature extraction capabilities.

[0069] In the embedded processor's memory space, a lightweight model architecture description file generated by the model pruning algorithm is loaded as the input condition for instantiation. By parsing the structural definition fields of this description file, the logical topology relationship between the graph structure encoding layer and the temporal attention gating module is established to ensure that the calling interfaces and data flow between modules match. The remaining weight parameters after channel sparsification are read using the weight mapping controller and filled into the weight matrices of the corresponding graph structure encoding layer node convolution kernels and temporal attention gating modules according to their index order, achieving precise binding of parameters to computational units. Based on the processor's internal memory page allocation mechanism, the mapped modules are loaded into the cache to reduce access latency during the inference stage, thereby improving the computational response capability of the initial model. A weight integrity verification algorithm is used to perform hash verification and comparison on the mapped weight matrix to ensure that the model loaded on the edge is consistent with the parameters defined in the lightweight model architecture description file in the cloud. Through memory pointer redirection technology, the output interfaces of the graph structure encoding layer and the temporal attention gating module are connected to the input buffer of the subsequent policy adaptation layer, providing a data entry point for policy knowledge initialization in the next sub-step. Through the instantiation and parameter mapping processes described above, the lightweight architecture description file from the previous step is transformed into an uninitialized edge inference model with basic spatiotemporal feature extraction capabilities, thereby achieving the structural integrity and initial usability of the edge model.

[0070] For example, in the wireless communication equipment of an industrial automated production line, the processor is an ARM Cortex-A53 with 512MB of memory. The lightweight model architecture description file is 3.2MB in size. This file defines a graph-structured coding layer containing 64 convolutional kernels and a temporal attention gating module containing 4 attention heads. The total size of the weight matrix after channel sparsification is 4.8% of the original model, that is, the number of weights is reduced from 2,000,000 to 96,000. The memory mapping relationship between the coding layer and the gating module is established sequentially by parsing the architecture description file. 256KB of space is allocated in the cache for the convolutional kernel data and 128KB of space is allocated for the attention weight matrix. After parameter binding is completed, a hash function is used. The hash values ​​of each weight matrix are calculated and compared with the pre-stored check values ​​in the description file. If the difference is zero, the mapping is confirmed to be correct. In this scenario, the uninitialized inference model on the edge can complete a graph convolution operation with 64 nodes within 8ms. The temporal correlation extraction latency of the gating module is 3ms, and the overall response time is significantly improved, providing an efficient spatiotemporal feature data foundation for the initialization of the subsequent policy adaptation layer.

[0071] S3.3: Receive the cloud policy output distribution vector issued by the cloud. The cloud policy output distribution vector contains a multi-candidate link fine-grained switching priority sequence and confidence distribution. Through inverse transformation operation, the cloud policy output distribution vector is converted into a policy feature embedding vector in a high-dimensional latent space to generate an initial policy knowledge representation for guiding local parameter updates.

[0072] The system receives a cloud-based policy output distribution vector from the cloud. The input is a standardized probability vector containing a fine-grained switching priority sequence and confidence distribution for multiple candidate links. This vector has been generated by a cloud-based global policy graph neural network in a high-precision inference environment. Vector deconstruction is performed on this probability vector to separate the priority sequence and confidence distribution into two sub-domains, and numerical buffers for inverse projection calculations are established in memory for each. The priority sequence is converted into candidate link node encoding indices using a mapping index table. This mapping index table consists of a mapping relationship between link topology nodes and latent space indices pre-stored on the endpoint, ensuring that each link identifier in the sequence corresponds to a unique high-dimensional embedding position. Nonlinear inverse normalization is performed on the confidence distribution using the same soft maximum function inverse operation as the cloud model, according to the formula... in This is the original score. Given the probability value, solve using the inverse operation. To restore the feature scale of the cloud, the edge can update parameters within the same feature space. The inverse-normalized confidence score and the node encoding index are concatenated in multiple dimensions to form a partial feature vector of the candidate link node in the high-dimensional latent space. An inverse mapping projection algorithm is applied to these partial feature vectors. This algorithm utilizes the transpose of the weight matrix of the edge graph structure encoder and the inverse weights of the time-series attention gating module to backproject the optimized output probability distribution from the cloud to the edge latent space, generating a complete policy feature embedding vector matrix. This policy feature embedding vector matrix is ​​stored as a row vector in the input buffer of the policy adaptation layer initialization interface, forming an initial policy knowledge representation to guide the update of locally learnable parameters. Through inverse mapping projection and feature reconstruction, the cloud policy output distribution vector from the previous step is transformed into policy feature embedding data in the high-dimensional latent space, enabling the edge model to obtain the initial parameter mapping of cloud policy knowledge without performing full training.

[0073] For example, in an industrial automation wireless communication system, the cloud-based policy output distribution vector has a length of 50, containing the handover priority scores and corresponding confidence values ​​for 15 candidate links. The edge side sets up a mapping index table to map link numbers to 1024-dimensional embedding index positions in the latent space. The confidence distribution is recovered using a soft maximum inverse operation, where the denominator in the inverse operation uses... The original scale scores with a mean range of [-3.5, 4.2] were calculated. The recovered scores were concatenated with the link embedding index to form a 15×1024 partial feature matrix. Inverse projection was then performed using the inverse weight matrix of the edge graph structure encoding layer (size 1024×1024, sparsity 0.85) and the attention-gated inverse weights (time window of 5 steps) to generate a 15×1024 complete policy feature embedding matrix. This matrix, after caching, was input into the policy adaptation layer initialization interface to drive subsequent cold-start weight allocation. Field tests showed that the edge model could output a link switching priority sequence close to the cloud policy during the first inference, achieving significantly improved context adaptation capability and real-time response.

[0074] S3.4: Call the preset policy adaptation layer interface in the uninitialized inference model on the end side, take the initial policy knowledge representation as the input stimulus signal, perform forward propagation weight allocation operation, calculate the initial bias and scaling factor of the learnable parameters inside the policy adaptation layer, and generate the initialized policy adaptation layer component that has completed parameter cold start.

[0075] The initial policy knowledge representation generated by inverse mapping projection is used as the input signal and imported into the pre-defined policy adaptation layer interface in the uninitialized inference model on the edge. This ensures that the input tensor matches the expected weight matrix inside the adaptation layer in terms of dimension and numerical range. The policy adaptation layer is a learnable parameter module embedded at the output end of the edge graph neural network model. Its function is to adapt the policy knowledge delivered from the cloud to the edge model, enabling the edge to quickly generate link switching decisions consistent with the cloud policy.

[0076] The forward propagation operation is performed on the input tensor. Based on the combined feature tensor of the graph structure encoding layer output and the temporal attention gating module output, the policy knowledge representation excitation signal is distributed to each weight channel of the adaptation layer to form an initial response mapping matrix of learnable parameters.

[0077] The bias of each output node is calculated based on the response mapping matrix. The initial bias is determined by the difference between the accumulated product of the weights and the centered value of the input features, as shown in the following formula: in, For bias, To center the mean of the input features, For the weight vector, This is the input feature vector.

[0078] The scaling factor for each output channel is calculated by combining the confidence interval of the cloud strategy output distribution vector. The scaling factor is obtained through the ratio mapping method, and the formula is as follows: in, This is the scaling factor. This represents the confidence level in the cloud. This is the initial output confidence value at the endpoint.

[0079] The bias and scaling factor are injected into the corresponding learnable parameter area inside the adapter layer to complete the weight initialization and data calibration, and the adapter layer state is transformed from uninitialized to an initialized component with basic mapping capabilities.

[0080] Through the above weight allocation and parameter injection processing, the initial strategy knowledge representation of the previous step is transformed into cold start parameters that can directly drive the output of the edge model, thereby realizing the construction and deployment of the initialization strategy adaptation layer.

[0081] For example, in an industrial automation scenario, the uninitialized inference model on the edge receives the distribution vector output by the cloud policy. The confidence level of a candidate link is 0.82, corresponding to an initial confidence level of 0.64 on the edge. The mean of the input feature center is 0.55, the weight vector length is 128, and the input feature vector length is 128, with element values ​​ranging from -1 to 1. The bias is calculated by multiplying and accumulating the corresponding elements. The scaling factor is 0.07. The result is 1.28125. When multiple links are initialized simultaneously, the bias and scaling coefficients of each link are injected into the corresponding parameter area, ultimately forming an initialization strategy adaptation layer. In online testing, the switching strategy generated by this adaptation layer is completely consistent with the cloud standard strategy in the confidence distribution overlap range, and the switching response time is significantly reduced to the millisecond range.

[0082] S3.5: Integrate the initialized policy adaptation layer component into the output end of the uninitialized inference model on the edge, replace the original randomly initialized output layer, complete the assembly of the overall model structure and parameter solidification, and generate the final edge graph neural network model that can be used for real-time context awareness and policy generation.

[0083] like Figure 3 As shown, step S4 involves monitoring local communication context change characteristics. When a device enters a tunnel, experiences sudden electromagnetic interference, or experiences context drift due to service type switching, a single-step policy alignment mechanism based on the cloud policy output distribution vector is triggered. Specifically, this includes: S4.1: Acquire a multi-source heterogeneous link status data stream containing real-time signal quality indicators, instantaneous service load parameters, current device movement trajectory coordinates, and the latest environmental interference event labels. Perform sliding window sampling and timestamp alignment processing on the multi-source heterogeneous link status data stream to eliminate timing deviations caused by asynchronous sensor acquisition and generate a standardized multimodal timing feature sequence.

[0084] For the multi-source heterogeneous link status data stream output in real time by the link status monitoring unit, a data input set is formed using signal quality sensors, service load acquisition modules, equipment trajectory positioning units, and environmental interference event labeling modules as data sources. Within the time-series acquisition buffer, a segmentation operation is performed according to predetermined sliding window width and step size parameters, dividing the continuously sampled data into window blocks for subsequent synchronous processing. For the data streams from different sensors within each window block, a timestamp remapping is performed according to a unified time base, mapping the original acquisition time to a global reference time axis. An interpolation algorithm is used to fill in missing time points caused by asynchronous sampling, forming a time-axis consistent multimodal data block. For the timestamp-aligned multimodal data blocks, normalization and standardization operations are performed based on the dimensional standards and numerical ranges of each data domain. For example, for signal quality indicators, linear normalization to the [0,1] interval is used; for service load parameters, Z-score standardization is used to eliminate absolute value differences. The standardized value is calculated using the following formula: in These are the original sampled values. This is the mean of the feature within the window block. The standard deviation is given. The trajectory coordinate data is mapped from geographic coordinates to a standardized Cartesian coordinate system, and environmental interference event labels are One-Hot encoded to form discrete feature vectors. The signal quality, service load, trajectory coordinates, and interference labels, after the above multi-domain normalization, are concatenated in time series to form a fixed-length, time-step consistent standardized multimodal temporal feature sequence, providing consistent input for the subsequent local spatiotemporal correlation graph construction module. Through sliding window sampling, timestamp alignment, and cross-domain normalization, the input conditions of the previous edge model are transformed into a standardized multimodal temporal feature sequence that accurately reflects the current communication context state and is unaffected by asynchronous sampling and dimensional differences, thus achieving the construction of a high-precision data base for context change monitoring.

[0085] For example, in an IoT industrial automation scenario, the signal quality sensor sampling frequency is 200Hz, the workload monitoring module sampling frequency is 50Hz, the device trajectory positioning unit outputs position coordinates at a frequency of 20Hz, the environmental interference event tag update frequency is 10Hz, the sliding window width is set to 2 seconds, and the step size is set to 0.5 seconds. Within one processing cycle, a total of 400 signal quality data points, 100 workload data points, 40 trajectory coordinate data points, and 20 interference tags are collected from the four types of sensors. These data are unified to a 400-point time axis through time mapping, and linear interpolation is used to fill in the gaps between the trajectory coordinates and the interference tags. Signal quality data has a mean of -78dBm and a standard deviation of 3dBm within the window block, and is linearly normalized to obtain values ​​in the [0,1] interval. Service load data has a mean of 120kbps and a standard deviation of 15kbps within the window block. Standardized values ​​are calculated using the above formula to obtain a zero-mean, unit-variance distribution. Trajectory coordinates are transformed from geographic coordinates to Cartesian coordinates and normalized to the [-1,1] interval. Interference event labels are converted into sparse vectors of length 5 using a One-Hot encoding strategy. The four types of data are spliced ​​together with time consistency to form a standardized multimodal temporal feature sequence of length 400 and feature dimension (1+1+2+5)=9. This sequence is verified to stably reflect the context changes in the subsequent local spatiotemporal correlation graph construction stage, and the detection accuracy of events such as tunnel entry, electromagnetic interference, and service switching is significantly improved.

[0086] S4.2: Construct a local spatiotemporal correlation graph structure based on the standardized multimodal temporal feature sequence, use graph convolution operators to perform aggregation operations on the spatial dependencies between nodes, and combine causal convolution mechanisms to extract the evolutionary rules in the time dimension, generating a dynamic spatiotemporal graph node embedding matrix that represents the current communication topology state.

[0087] S4.3: Apply a multidimensional anomaly detection algorithm to the dynamic spatiotemporal graph node embedding matrix to calculate the signal strength mutation rate, trajectory curvature change value, and service load entropy increase rate, respectively. Compare and analyze the above calculation results with the preset context drift threshold vector to generate a context drift event candidate set including tunnel entry marker, electromagnetic interference marker, and service switching marker.

[0088] Multidimensional anomaly detection is performed on the node embedding matrix of the dynamic spatiotemporal graph, using the signal quality sequence, device movement trajectory sequence, and service load sequence contained in the matrix as input objects for multidimensional detection. A mutation rate detection operator is applied to the signal quality sequence to calculate the ratio of the absolute value of signal strength change between adjacent time steps to the historical mean, obtaining the signal strength mutation rate index, which is achieved through the following formula: in This is the signal strength value. This is a time index. Curvature change value calculation is performed on the device movement trajectory sequence. This is achieved by extracting the coordinates of trajectory points within adjacent windows and calculating the curvature difference, forming a trajectory curvature change value index. The formula is as follows: in This represents the trajectory curvature value. For the service load sequence, entropy increase is calculated. The entropy value is calculated using the probability vector of the service data packet distribution, and the difference between the entropy value of the current window and the previous window is obtained to form the service load entropy increase index. The formula is as follows: in The probability distribution of service load is represented. The signal strength mutation rate, trajectory curvature change value, and service load entropy increase rate are compared with the preset context drift threshold vector at the element-wise difference level. By determining the positions that exceed the threshold, corresponding tunnel entry markers, electromagnetic interference markers, and service switching markers are generated, and these are combined to form a context drift event candidate set. Through multi-dimensional anomaly detection and threshold comparison processing, the dynamic spatiotemporal graph node embedding matrix of the previous step is transformed into structured candidate data containing different types of drift markers, thereby achieving accurate identification of complex dynamic context changes.

[0089] For example, in an industrial automation wireless control scenario, the dynamic spatiotemporal graph node embedding matrix originates from a multimodal link monitoring system. Signal quality is sampled by an RSSI sensor, trajectory coordinates are generated by an inertial navigation unit, and service load is statistically analyzed from the real-time control command stream. The signal strength mutation rate threshold is set to 0.25, the trajectory curvature change threshold to 0.15, and the service load entropy increase threshold to 0.2. Within a certain detection window, the RSSI signal strength changes from... 65dBm jump 80dBm, substitute into the formula to calculate the mutation rate. The value is 0.27, which is higher than the threshold; the trajectory curvature changes from 0.5 to 0.68, and the curvature change value is... The value is 0.18, which is higher than the threshold; the business load entropy increased from 1.2 to 1.45, indicating a significant increase in entropy. It is also higher than the threshold. Through threshold comparison, tunnel entry, electromagnetic interference, and service switching events are marked respectively, and the candidate set contains all three types of markers simultaneously. In the subsequent S4.4 weighted voting decision, this embodiment can combine the confidence levels of the three types of markers to generate trigger signals, thereby significantly improving the accuracy and timeliness of context drift detection.

[0090] S4.4: Perform fusion decision processing based on the aforementioned context drift event candidate set, and use a weighted voting mechanism to comprehensively evaluate the confidence scores of various drift markers. When the comprehensive score exceeds the judgment threshold, the context drift phenomenon is confirmed to have occurred, and a context drift trigger signal is generated to activate the policy update process.

[0091] Using the context drift event candidate set generated by S4.3 as input for the fusion decision processing, the confidence score vectors of tunnel entry markers, electromagnetic interference markers, and service switching markers are read respectively, and a marker category index and score mapping table is established within the decision processor for subsequent combination calculations. Based on this index table, a category weight allocation operation is performed. The weight parameters are synchronized from the cloud to the edge based on historical performance evaluations and preset scene priority values. The confidence scores of various drift markers are multiplied element-wise with their corresponding category weights to form a weighted score matrix. The multi-category voting mechanism module is invoked to perform cumulative summation on the weighted score matrix column-wise, obtaining a comprehensive score value representing the global drift significance assessment. A numerical comparison operation is performed between the comprehensive score and a preset significance judgment threshold. If the comprehensive score is greater than the threshold value, a context drift confirmation flag is generated within the decision processor. The context drift confirmation flag is converted into a context drift trigger signal through the event trigger interface and output to the activation module of the policy update process. By integrating decision processing methods, the multidimensional drift marker data from the previous step is transformed into a confirmatory context drift trigger signal, enabling precise response to dynamic link state changes and timely initiation of strategy adjustments.

[0092] For example, in a wireless communication environment in an industrial automated workshop, the end device receives a candidate set of context drift events including tunnel entry markers, electromagnetic interference markers, and service handover markers, where the confidence score for the tunnel entry marker is 0.82, the confidence score for the electromagnetic interference marker is 0.76, and the confidence score for the service handover marker is 0.64. The category weights are set to 0.5 for tunnel entry, 0.3 for electromagnetic interference, and 0.2 for service handover, forming a weight vector. A weighted calculation is performed internally by the decision processor, and the overall score is calculated using the following formula: in, For the overall score, the multiplication symbol represents element-wise multiplication. The calculation result is: The preset significance threshold is [value]. The comparison results show that the overall score is greater than the threshold, thus generating a context drift confirmation flag. This flag is converted into a context drift trigger signal and output to the policy update module, triggering the alignment process of the cloud policy output distribution vector. In this scenario, the generation of the trigger signal significantly improves the link switching policy's ability to adapt to sudden interference and environmental changes, ensuring the continuity and stability of control command transmission during communication.

[0093] S4.5: In response to the context drift trigger signal, call the pre-stored secure encrypted communication interface, pull the latest cloud policy output distribution vector from the cloud storage area, and encapsulate the cloud policy output distribution vector as the initialization input parameters of the single-step policy alignment mechanism to generate the single-step policy alignment instruction package.

[0094] The single-step policy alignment mechanism is a lightweight parameter update mechanism for edge graph neural network models. Its goal is to quickly approximate the output probability distribution of the edge model to the target policy distribution vector delivered from the cloud through a single local backpropagation without retraining the entire model, thereby achieving real-time adaptation to dynamic context drift.

[0095] Upon receiving a significant drift trigger signal output by the context drift determination module, the handshake initialization function pre-stored in the end-side secure encrypted communication interface module is invoked to perform bidirectional verification of the current device identity and session key and generate a one-time session token to ensure the security of subsequent policy distribution data transmission.

[0096] The TLS layer key negotiation process is driven by a session token, which establishes an encrypted data channel between the client and the cloud policy storage area. After the channel is established, a pull request message for the latest policy distribution vector is sent to the cloud, which includes the current context feature summary and timestamp, for the cloud to retrieve the global optimization policy at the corresponding time.

[0097] When transmitting the policy distribution vector back to the cloud, the end-side communication interface is used to perform segmented decryption and CRC verification on the transmitted data, restoring the encrypted policy output distribution vector to the original data structure containing the fine-grained switching priority sequence of multiple candidate links and the confidence distribution.

[0098] Based on the inverse mapping projection algorithm, the policy output distribution vector is converted into an initial policy feature embedding matrix in a high-dimensional latent space, and then aligned with the input parameter dimension of the edge-initialized inference model to form an initial input parameter set that conforms to the local adaptation layer update input specification.

[0099] The single-step policy alignment mechanism initialization interface is called to encapsulate the above initialization input parameter set into a single-step policy alignment instruction package containing the update starting point identifier, target distribution label reference, and alignment constraint settings, and triggers the edge model parameter fine-tuning process to achieve rapid mapping calibration of the policy adaptation layer.

[0100] This processing method transforms the context drift trigger signal into an executable single-step policy alignment initialization instruction package on the device side, enabling efficient knowledge synchronization between the device side and the cloud, as well as adaptive local policy updates.

[0101] For example, in an industrial automation scenario, the edge device detects a context drift trigger signal. This signal includes characteristics such as a real-time signal quality degradation to -85 dBm, an electromagnetic interference index rising to 2.3, and a switch in workload from control commands to video streaming. The edge-side secure encrypted communication interface performs an ECDH-based key exchange, generates a 256-bit one-time session token, and drives a TLS 1.3 handshake to establish an encrypted channel. A context feature digest vector is attached when sending a pull request message to the cloud. [85, 2.3, 2, 1625123456], where the third item, the business type, is encoded as 2, and the fourth item is a UNIX timestamp. The cloud returns a policy output vector containing priority scores and confidence distributions for six candidate links. After being decrypted in AES-GCM mode and verified by CRC32 on the end side, the original distribution data is recovered. Using the inverse mapping projection algorithm, this distribution data is mapped to a 128-dimensional latent feature space to form an initialization policy feature embedding matrix. It also matches the parameter dimensions of the input layer of the edge adaptation layer. The instruction package encapsulates the update start point identifier code 0x1F, the target label reference, and alignment constraints. This triggers a local parameter fine-tuning process. This process completes the alignment of adaptation layer parameters in a single step, enabling the edge to generate a switching strategy consistent with the cloud within 0.8 seconds after detecting drift, significantly improving communication continuity and strategy response speed in industrial production control processes.

[0102] Step S5: Execute the single-step policy alignment mechanism, calculate the KL divergence constraint value between the current output distribution of the edge graph neural network model and the cloud policy output distribution vector, and update only the learnable parameters in the policy adaptation layer to generate an aligned policy adaptation layer. Specifically, this includes: S5.1: Obtain the cloud-based policy output distribution vector and the embedded current inference state on the edge. Based on the cloud-based policy output distribution vector, construct a target policy distribution label that includes the main link selection probability, the suboptimal link ranking, and the switching timing suggestion interval. Embed the current inference state on the edge and input it into the policy adaptation layer to be updated to generate the initial edge output distribution vector.

[0103] Based on the latest cloud policy output distribution vector delivered from the cloud storage via a secure encrypted communication interface and the embedding of the current inference state of the edge graph neural network model, the initial matching input conditions for the single-step policy alignment mechanism are established. The cloud policy output distribution vector is structured and parsed to extract the priority score, confidence value, and switching timing suggestion interval corresponding to each candidate link. These parsing results are then mapped into a target policy distribution label matrix containing the primary link selection probability, the second-best link ranking, and the time window boundaries. Index mapping and probability normalization operations are used to align the target policy distribution label matrix to an output space dimension acceptable to the edge model, ensuring consistency between the label vector and the local inference vector in feature length and index order. After alignment, the current edge inference state embedding vector is input to the policy adaptation layer interface to be updated, and forward propagation baseline output calculation is performed to generate an initial edge output distribution vector. This vector is used for subsequent calculation of relative entropy constraints and parameter updates. Through the above processing, the results of the previous step are transformed into pairwise probability distribution data that can be used for KL divergence calculation, achieving direct comparability between cloud policy knowledge and the current edge output.

[0104] For example, in a certain industrial automation wireless communication scenario, the cloud-based policy output distribution vector contains four candidate links. The priority probabilities of links A, B, C, and D are 0.65, 0.20, 0.10, and 0.05, respectively, with confidence intervals of [0.60, 0.70], [0.18, 0.22], [0.08, 0.12], and [0.04, 0.06], respectively, and suggested switching intervals of [5, 8] seconds, [12, 15] seconds, [20, 23] seconds, and [30, 33] seconds, respectively. The current inference state on the edge is embedded and processed by the graph structure coding layer and the temporal attention gating module to generate a vector of length 4: (0.50, 0.25, 0.15, 0.10). Label mapping is performed on the cloud-based policy output distribution vector to form the target policy distribution label: the selection probability of primary link A is 0.65, the probability of secondary link B is 0.20, and the switching window is [5, 8] seconds. Align the target policy distribution labels with the edge output space and input them into the policy adaptation layer. Forward propagation generates the initial edge output distribution vector (0.50, 0.25, 0.15, 0.10). This step significantly improves the stability of policy matching and its adaptability to dynamic contexts in this scenario.

[0105] S5.2: Receive the target policy distribution label and the initial end-side output distribution vector, and use a divergence calculation algorithm based on relative entropy to quantify the probability density difference between the two to obtain the KL divergence constraint value that characterizes the degree of deviation between the end-side policy and the cloud policy.

[0106] S5.3: Construct a single-step gradient update instruction based on the KL divergence constraint value, and use a local backpropagation mechanism that only targets the learnable weight matrix to fine-tune the parameters inside the policy adaptation layer, so as to generate an aligned policy adaptation layer parameter set with cloud knowledge mapping capability.

[0107] Based on the calculated KL divergence constraint value, this value is used as the core driving parameter of the update instruction and input to the gradient construction module. The gradient construction module generates a single-step gradient update instruction matrix containing the update direction, step size coefficient, and weight selection mask. For the learnable weight matrix within the policy adaptation layer, the weight selection mask in the single-step gradient update instruction matrix is ​​parsed to mask non-learnable parameters in the graph structure encoding layer and the temporal attention gating module, ensuring that the update process only affects a subset of parameters in the policy adaptation layer. Gradient backpropagation based on the local error backpropagation mechanism is performed on the selected weight matrix elements. The local error signal and step size coefficient are multiplied element-wise to generate the weight adjustment amount for this single-step update. The specific calculation method is as follows: in, This is the weight adjustment amount. This is the step size coefficient. This represents the partial derivative of the local loss function with respect to the weights. The aforementioned weight adjustments are accumulated and added to the corresponding positions in the current weight matrix of the policy adaptation layer, forming a new parameter set with cloud-based policy distribution mapping capabilities. The forward inference is then re-executed using this new parameter set, and the output is compared with the target policy distribution label. If the confidence intervals overlap, the updated parameter set is solidified, generating the aligned policy adaptation layer parameter set. Through a local backpropagation mechanism targeting only the learnable weight matrix, the KL divergence constraint values ​​from the previous step are transformed into quantitative weight adjustment data, achieving rapid, single-step alignment between the edge-side policy adaptation layer and the cloud-based policy output distribution vector.

[0108] For example, in an industrial automation wireless control system, the policy adaptation layer of the edge graph neural network model includes two fully connected layers and one normalized layer, with two sets of learnable weight matrices of sizes 64×128 and 128×32. After a context drift is triggered, the KL divergence constraint value between the cloud policy output distribution vector and the edge initial output distribution vector is calculated as follows: The gradient building module sets the step size coefficient based on this value. This generates a weight selection mask to select non-zero elements in the second-layer fully connected matrix. The local backpropagation mechanism calculates the partial derivative of the local loss function with respect to the weights for the selected elements, within a range of... Between these, the weight adjustment amount is generated by combining the step size coefficient within a certain range. Between. The adjustment is accumulated into the original weight matrix to obtain a new parameter set. The absolute value of the difference between the target link selection probability of the real-time optimal switching strategy generated by forward inference and the corresponding item of the cloud strategy distribution label does not exceed... The confidence intervals completely overlap, and the back-end policy adaptation layer of the solidified new parameter set completes global knowledge alignment. In subsequent link switching scenarios, the system can significantly improve policy response speed and maintain high connection stability.

[0109] S5.4: Load the aligned strategy adaptation layer parameter set into the edge graph neural network model, replace the original uncalibrated parameter module, thereby completing the knowledge evolution update of the edge model and forming the aligned strategy adaptation layer.

[0110] After generating the aligned strategy adaptation layer parameter set with cloud-based knowledge mapping capabilities, the structure management module of the graph neural network model inputting this parameter set calls the model's built-in parameter loading interface to perform a weight replacement operation. This clears the randomly initialized weight mappings in the original uncalibrated parameter module and overwrites them with the aligned parameter set. A memory address binding and weight matrix index matching mechanism ensures that the replaced parameters maintain a consistent mapping relationship with the cloud-based strategy at the outputs of the graph structure encoding layer and the temporal attention gating module. A parameter consistency check operator is used to perform dimension-by-dimensional difference detection on each weight element of the replaced adaptation layer. If the detection results are within a preset deviation tolerance range, the integrity of the replacement process is confirmed. The model structure recompilation process is invoked to reconstruct the inference execution graph, ensuring the correctness of the data path and gradient flow between the aligned strategy adaptation layer and its feedforward layer. Precision quantization and latency evaluation calculations verify that the real-time inference capability of the replaced model meets the requirements for link switching strategy generation under dynamic contexts. Through the above-mentioned loading, replacement, verification, recompilation and performance verification processes, the update results of the previous step are transformed into the final usable output structure of the edge-side temporal graph neural network model, realizing the solidified deployment and knowledge evolution update of the aligned strategy adaptation layer.

[0111] For example, in an industrial automation environment, the end-side device is equipped with a four-channel multimodal communication interface. The original uncalibrated strategy adaptation layer uses randomly initialized parameters, while the aligned parameter set calculated by the cloud is a 512-dimensional weight vector with a precision of float32 per dimension. During loading, the device calls the parameter loading interface to map the 512-dimensional weights to the corresponding 512 parameter slots in the output layer according to the index order, replacing the original weights. When performing parameter consistency verification, an error tolerance is adopted, and the difference between the replaced weights and the cloud parameter set is detected dimension by dimension. If all weight differences are less than the tolerance value, the replacement is considered successful. After recompiling the model structure, the inference latency index is compared. The latency before the update is milliseconds, and the latency after the update is milliseconds. The difference is negligible and within the system's real-time requirements. The latest local link status data is input into the updated model. The generated switching strategy confidence distribution completely overlaps with the confidence interval of the target strategy distribution label in the cloud. The verification results show that the model after deploying the aligned strategy adaptation layer can output the optimal link switching strategy command consistent with the cloud within a millisecond response time, significantly improving the connection stability of the device under sudden link degradation.

[0112] S5.5: Verify the output consistency of the aligned policy adaptation layer by using the latest local link state data as test input through the updated model to confirm that the generated real-time optimal switching policy instruction coincides with the confidence interval of the target policy distribution label, so as to finally establish the availability of the aligned policy adaptation layer.

[0113] The edge graph neural network model, updated with parameters from the policy adaptation layer, is used as the validation object. The latest local link state data is loaded as the unified test input, and the input data undergoes multimodal feature standardization to eliminate dimensional differences. During model inference, the test input sequentially flows through the preserved graph structure encoding layer and the temporal attention gating module to extract the spatiotemporal context feature tensor, which is then input to the aligned policy adaptation layer to generate the probability distribution vector of the real-time optimal switching policy command. Using the comparison detection unit, the overlap between this real-time policy distribution vector and the confidence interval of the target policy distribution label generated in S5.1 is calculated.

[0114] The confidence matching degree of the two distributions on each candidate link is calculated using the interval overlap rate formula: in, This represents the confidence score of the edge-side model for the i-th candidate link. This represents the confidence score corresponding to the target strategy label. The overlap rate is set as the confidence interval. This overlap rate is logically compared with a preset verification threshold. When the overlap rate of all candidate links is not lower than the threshold, the output consistency of the aligned strategy adaptation layer is confirmed to meet the technical requirements. Through interval overlap testing, probability matching degree calculation, and logical judgment processing, the parameter update results of the previous step are transformed into usability verification indicators, thereby establishing the usability of the aligned strategy adaptation layer in the local inference scenario.

[0115] For example, an edge gateway device of an industrial automation system is selected as the verification object. This gateway connects to three multimodal communication links simultaneously: a 5G cellular link, an industrial Ethernet link, and a LoRa link. During the verification process, the latest local link status data is acquired, including a real-time signal quality value of -85dBm and an instantaneous service load of 6.2Mbps for the 5G link, a latency of 5ms and a load of 4.8Mbps for the industrial Ethernet link, and a packet loss rate of 0.8% and a latency of 120ms for the LoRa link. After standardization, this data is input into an updated edge-side graph neural network model to obtain a real-time policy distribution vector, where the confidence score is 0.72 for the 5G link, 0.65 for the industrial Ethernet link, and 0.33 for the LoRa link. The corresponding confidence scores in the target policy distribution labels are 0.70, 0.68, and 0.35, respectively. By applying the score values ​​to the interval overlap rate formula, the overlap rates of the three links were calculated separately: the overlap rate for the 5G link was 0.70 / 0.70 = 1.0, for the industrial Ethernet link it was 0.65 / 0.68 ≈ 0.955, and for the LoRa link it was 0.33 / 0.35 ≈ 0.943. After comprehensive calculation, the overall overlap rate was significantly improved, all exceeding the preset threshold of 0.90. Verification results show that the aligned strategy adaptation layer can generate a handover priority sequence and confidence score highly consistent with the target strategy in the real-time inference scenario of this end-side device, achieving high accuracy and low latency response in multi-link handover decisions.

[0116] Step S6: Input the latest local link status data into the end-side graph neural network model loaded with the aligned policy adaptation layer, and infer to generate a real-time optimal switching policy instruction adapted to the current dynamic context. Specifically, this includes: S6.1: Obtain the latest local link status data, which includes real-time signal quality indicators, instantaneous service load parameters, current device movement trajectory coordinates, and the latest environmental interference event labels. Perform multimodal feature standardization processing on the latest local link status data to generate a standardized multimodal link feature vector sequence.

[0117] The system acquires the latest local link status data input from the link status awareness module, including real-time signal quality indicators output by the RF signal strength sensor, instantaneous service load parameters generated by the service monitoring unit, current device movement trajectory coordinates provided by the positioning module, and the latest environmental interference event labels generated by the interference detection unit. For the aforementioned multi-source heterogeneous raw data, the feature normalization operator within the multimodal data preprocessing component is invoked to perform time-axis interpolation resampling processing on the time-scale differences caused by inconsistent sampling frequencies of different sensors, ensuring that all data channels are aligned under a unified time reference. A Z-score-based normalization transformation is applied to scale and center the signal quality, service load, and trajectory coordinate values ​​according to their respective historical means and standard deviations, eliminating the impact of dimensional differences on the model input distribution. For environmental interference event labels, a one-hot encoding method is used in the feature space to transform event categories into fixed-length and mutually exclusive binary feature vectors, achieving the fusion of event category information and numerical feature vectors on the same tensor dimension. The feature fusion operator is invoked to concatenate the normalized and encoded signal quality feature vectors, service load feature vectors, trajectory coordinate feature vectors, and interference event feature vectors according to a predefined sequence, forming a multimodal link feature matrix in a unified format. Multidimensional vector flattening and serialization processing are then performed on this feature matrix, transforming it into a link feature vector sequence that conforms to the input specifications of the edge-side temporal graph neural network. The generation of this sequence data enables the edge-side model to fully perceive the current context in subsequent neighborhood aggregation and temporal attention calculations.

[0118] By using time-aligned interpolation, Z-score normalization, one-hot encoding, and feature fusion, the link context data required by the alignment strategy adaptation layer in the previous step is transformed into a standardized multimodal link feature vector sequence, achieving consistency in input data distribution and dimensionality, and providing an accurate feature foundation for the construction of dynamic graph node embedding in S6.2.

[0119] For example, in an industrial automation workshop, the edge device collects the Signal Quality Index (RSSI) once per second, ranging from -80dBm to -40dBm. The workload parameter is the current PLC task queue length, with a value between 0 and 50. The device movement trajectory coordinates are provided by a UWB positioning system with an accuracy of 0.1 meters. Environmental interference event tags are derived from electromagnetic radiation sensors, categorized into "no interference," "low-frequency interference," and "high-frequency interference." During time interpolation, the trajectory sampling frequency is reduced from 10Hz to 1Hz to match the RSSI update frequency. Z-score normalization is used to process the RSSI, ensuring the normalization formula... In this model, μ represents the mean RSSI over the past minute, and σ represents the corresponding standard deviation. Service load is normalized using the same method, and trajectory coordinates are normalized along the X, Y, and Z axes respectively. Interference event categories are one-hot encoded, with the three categories mapped to (1,0,0), (0,1,0), and (0,0,1) respectively. The four normalized / encoded features are concatenated into a multimodal feature vector of length N, where N is the sum of the dimensions of the features in each domain. This generates a standardized link feature sequence of length 20, which is then input to the end-side model. In validation, this standardization process significantly improves feature consistency during the model's neighborhood aggregation stage, avoids attention weight shifts caused by differences in scale, and makes the link priority score generated by the real-time switching strategy more stable.

[0120] S6.2: Based on the standardized multimodal link feature vector sequence, the neighborhood aggregation operation is performed using the graph structure encoding layer retained in the end-side graph neural network model to construct a dynamic spatiotemporal graph node embedding matrix that represents the topological association relationship of multiple candidate links.

[0121] The standardized multimodal link feature vector sequence is input as the initial node feature parameters of the end-side graph neural network model, and the graph structure encoding layer retained in the model is called to parse the topological association index between candidate links.

[0122] A dynamic adjacency matrix is ​​constructed based on the topological association index obtained from the parsing. The neighborhood aggregation operation is performed on the adjacency matrix by the graph convolution operator, and the real-time signal quality indicators, service load parameters and environmental interference labels of adjacent nodes are weighted and fused according to spatial correlation.

[0123] The aggregation result is linearly transformed using a learnable spatial weight matrix to generate a set of node embedding vectors containing local spatial dependency information, and the original device movement trajectory coordinates are retained as time-dimensional features within the embedding vector set.

[0124] Batch normalization and nonlinear activation are applied to the node embedding vector set to enhance feature stability and improve nonlinear expressive power. The processed features are then mapped to a unified dimensional space for subsequent temporal extraction.

[0125] The mapped node embedding vectors are arranged in chronological order to form a dynamic spatiotemporal graph node embedding matrix, which serves as a complete representation of the topological associations of multiple candidate links.

[0126] By using neighborhood aggregation and spatial weight fusion processing, the standardized multimodal link feature vectors from step S6.1 are converted into dynamic spatiotemporal graph node embedding matrices, thereby achieving explicit encoding of the current link network spatial structure.

[0127] For example, in an industrial automation scenario, the input standardized multimodal link feature vector sequence contains 8 candidate link nodes, each with a feature dimension of 16 and an adjacency matrix size of 8×8. Non-zero elements indicate communication associations between links. The graph structure encoding layer uses a two-layer graph convolutional network, with each layer outputting a dimension of 32. The spatial weight matrix is ​​initialized with a random distribution and updated during training to adapt to the edge network structure. During neighborhood aggregation, the embedding vector of node i... The update formula for the l-th layer is ,in Let i be the set of neighbors of node i. This is the spatial weight matrix. The ReLU activation function is used. After aggregation, batch normalization is performed to ensure the consistency of the numerical stability of the embedding vectors at different time steps. The final dynamic spatiotemporal graph node embedding matrix has a size of 8×32. In subsequent S6.3, this matrix will be used as the input of the temporal attention module to extract evolution trend features using trajectory coordinates and service load time series, and significantly improve the accuracy of link switching strategies in the inference results.

[0128] S6.3: Apply the temporal attention gating module to the dynamic spatiotemporal graph node embedding matrix to perform time-dimensional weighted filtering in order to extract key temporal context feature tensors that reflect the evolution trend of link states.

[0129] Based on the temporal attention gating module inside the calling terminal side graph neural network model of the dynamic spatiotemporal graph node embedding matrix, multi-scale sliding window convolution operation is performed on the node feature sequence in the time dimension to separate the link state change patterns at different time granularities.

[0130] The correlation weights between time steps are calculated for the time series features processed by convolution, and a self-attention mechanism is used to generate an attention coefficient matrix for time weighting, thereby highlighting the time segments that contribute significantly to the evolution trend of the link state in the time dimension.

[0131] By performing a weighted product operation between the attention coefficient matrix and the time series features, a time-weighted feature matrix that suppresses short-term fluctuation noise and retains key evolutionary information is obtained.

[0132] Normalization is performed on the time-weighted feature matrix to eliminate the inconsistency of feature amplitudes caused by differences in device sampling frequencies and to maintain the comparability of weights between features at different times.

[0133] A gating mechanism is used to filter the normalized time-weighted feature matrix. By dynamically selecting or masking some time-dimensional features through learnable switching parameters, the feature enhancement output of the link state evolution trend is realized.

[0134] Through the above processing method, the dynamic spatiotemporal graph node embedding matrix of the previous step is transformed into a key temporal context feature tensor that reflects the evolution trend of the link state, so as to realize the accurate capture of the time evolution pattern on the end side and the input optimization of subsequent policy reasoning.

[0135] For example, in an industrial automation wireless communication scenario, the dynamic spatiotemporal graph node embedding matrix of a certain end-side graph neural network model has a dimension of 64×20, where 64 represents the node feature dimension and 20 represents the time step. A sliding window convolution kernel with a length of 5 and a stride of 1 is used to perform one-dimensional convolution operations on the time dimension to extract multi-scale link state change patterns. The calculated attention coefficient matrix has a dimension of 20×20, and a self-attention mechanism is used to capture state dependencies over long time intervals. In the attention weighting stage, the coefficient matrix is ​​multiplied element-wise with the convolution output matrix to obtain a time-weighted feature matrix. This matrix is ​​then Z-score normalized along the time dimension to eliminate the influence of inconsistent feature values ​​across different time periods. The gating mechanism sets the learnable switch parameter to an initial value of 0.5, which is dynamically adjusted during training to enhance the link state evolution trend features. Under this configuration, subsequent inference demonstrates that the model can stably output key temporal feature tensors with high availability, enabling the downstream policy adaptation layer to generate high-confidence switching policy commands even under frequent equipment movement and complex electromagnetic environment changes.

[0136] S6.4: Input the key temporal context feature tensor into the aligned policy adaptation layer that has completed parameter updates, and perform forward inference calculation based on probability distribution mapping to generate a real-time policy distribution vector containing the priority ranking of multiple candidate link switching and confidence score.

[0137] The key temporal context feature tensor is fed as an input signal into the aligned strategy adaptation layer whose parameters have been updated. Vector dimension matching and normalization transformation are performed at the input interface to ensure that the feature components of the tensor are consistent with the dimensions of the weight matrix inside the adaptation layer and are in a unified dimension.

[0138] The normalized feature tensor is mapped to the multi-head probability distribution mapping layer inside the policy adaptation layer. The weighted feature aggregation value of the candidate link node is calculated based on the spatial dependency weight matrix of each attention head, and nonlinear activation is applied to each aggregation value to enhance the discriminability between different link features.

[0139] The high-dimensional link aggregation feature vector output from the probability distribution mapping layer is used as input to the fully connected projection matrix. Matrix multiplication is performed to obtain the original priority scores of each candidate link, while batch normalization is used to suppress numerical drift.

[0140] Applying the soft maximum normalization function to the original priority scores above, the confidence score for each candidate link is calculated using the following formula: in, For the first The confidence score of each candidate link, where the numerator e is the exponent of the link's priority score, and the denominator... This is the summation of the priority score index values ​​for all candidate links.

[0141] The confidence scores and corresponding priority ranking results of all candidate links are encapsulated into a real-time policy distribution vector and transmitted to the main policy generation module through the adaptation layer output port for subsequent decision threshold determination and instruction encapsulation processing.

[0142] By using probability distribution mapping, normalization, and priority sorting, the key temporal context feature tensor from the previous step is transformed into a real-time policy distribution vector containing priority sorting and confidence scores for multiple candidate link switching, thereby enabling high-speed generation and accurate matching of link switching policies in the current dynamic context.

[0143] For example, in an industrial automated workshop environment, the sampling frequency of real-time signal quality indicators is 100Hz, the instantaneous service load parameters include the number of task instructions per second and instruction delay time, the equipment movement trajectory coordinates are GNSS data in three-dimensional space, and the environmental interference event labels include two categories: mechanical vibration and radio interference. When performing this step, the extracted key temporal context feature tensor (dimension 128×256) is compressed to 128×128 through a normalization interface and mapped to the multi-head probability distribution mapping layer inside the adaptation layer. The dimension of each attention head weight matrix is ​​set to 128×64, and the aggregated value of the candidate link features is calculated to be in the range [0, 2.5]. The projection matrix dimension is set to 64×10, corresponding to 10 candidate links, and the calculated original priority scores are respectively in the range [...]. Between [3.2, 4.5]. After applying the soft maximum function, the confidence scores of each link are mapped to the interval [0.05, 0.32]. The link with the highest confidence score is numbered L4, and the priority ranking results are L4, L2, L9, etc. The real-time policy distribution vector output in this way can drive the link execution module to select the most stable communication link in the dynamic context of strong wireless interference and high-speed equipment movement in this workshop, which significantly reduces the delay of control commands and greatly improves the stability of task execution.

[0144] S6.5: Based on the highest confidence score item and the handover priority ranking result in the real-time policy distribution vector, perform decision threshold determination and instruction encapsulation operations to generate a real-time optimal handover policy instruction that can directly drive the action of the link execution module.

[0145] The real-time policy distribution vector, with updated parameters, is input into the parsing module to extract the candidate link identifier and its switching priority order corresponding to the highest confidence score. Based on the extraction results, a decision threshold determination algorithm is invoked. The system's preset decision threshold matrix is ​​read, and the highest confidence score value and its corresponding priority order are matched within the threshold space to generate a decision result label. Priority correction rules are applied to the decision result label, and combined with the interface status and available resource markers of the current link execution module, availability filtering is performed on the candidate links, eliminating link options that do not meet the physical link switching conditions. The filtered link identifier, switching priority, and confidence score are combined into a decision output vector. The instruction encapsulation component is invoked to map the decision output vector into control semantics that the link execution module can parse, forming a real-time optimal switching policy instruction containing the target link identifier, switching trigger time window, and execution mode marker. Through the above multi-stage processing, the real-time policy distribution vector from the previous step is transformed into a structured control instruction that can drive the actions of the link execution module, realizing the generation of link switching instructions under dynamic context conditions.

[0146] For example, in an industrial automation wireless communication environment, the highest confidence score in the local real-time policy distribution vector is... The corresponding candidate link is identified as Link A, and the handover priority order is... The system's preset judgment threshold matrix has the following priority levels. The confidence threshold is After interval matching calculation, the high confidence score meets the threshold condition. When invoking the priority correction rule, it was found that the current availability status of link A interface is marked as "online" and the resource is marked as "idle". After availability filtering, link A is retained. The decision output vector consists of ( , , The instruction encapsulation component maps it to control semantics, including the target link identifier "LINK_A" and the switching trigger time window. The milliseconds and execution mode are marked as "seamless switching". The generated real-time optimal switching strategy command, after being parsed by the link execution module, completes the migration from the source link to link A within the switching trigger window. Verification results show that the arrival delay of the first packet of the business data stream is reduced to [missing information]. Within milliseconds, the number of service interruption frames dropped to The frame rate significantly improves connection stability.

[0147] Step S7: Execute the real-time optimal handover strategy instruction to complete the multimodal communication link handover operation, and collect strategy execution effect feedback data including handover time, connection recovery delay, and number of service interruption frames. Specifically, this includes: S7.1: Perform protocol parsing on the real-time optimal handover strategy command to extract the target communication link identifier, handover trigger time window, and service priority flag, and generate a standardized set of link handover control parameters.

[0148] S7.2: Based on the link switching control parameter set, perform atomic disconnect and reconnect sequence control on the currently active multimodal communication interface to complete the seamless migration operation from the source link to the target link and generate a link switching completion status signal.

[0149] S7.3: Using the link switching completion status signal as the starting point of the time reference, perform high-precision timestamp differential calculation on the arrival delay of the first packet in the data plane transmission process to obtain the connection recovery delay value characterizing the connection recovery speed.

[0150] S7.4: Based on the service data stream interval corresponding to the connection recovery delay value, perform sequence integrity verification statistics on the number of lost application layer video frames or control command packets to obtain a service interruption frame count index that characterizes the degree of service interruption.

[0151] S7.5: Integrate the connection recovery delay value, the service interruption frame count indicator, and the original value of the actual switching time read from the system's underlying registers, and perform multi-source heterogeneous data encapsulation processing to generate a structured feedback message containing complete policy execution effect feedback data.

[0152] Step S8: Based on the feedback data of the strategy execution effect, a difference comparison is performed with a preset baseline. If the difference in the end-side strategy distribution exceeds a threshold or the execution effect is lower than the baseline within multiple consecutive periods, the distillation frequency is dynamically adjusted and a new round of cloud-based strategy output distribution vector distribution requests is triggered. Specifically, this includes: S8.1: Obtain policy execution effect feedback data including handover time, connection recovery delay, and number of service interruption frames; perform normalization preprocessing and time-series alignment operations on the policy execution effect feedback data to generate a standardized multidimensional performance index sequence, so as to eliminate the influence of different units on subsequent evaluation calculations and establish a unified time benchmark.

[0153] S8.2: Based on the standardized multidimensional performance index sequence and the pre-stored system stability preset baseline, perform dimension-by-dimensional deviation calculation processing, use the weighted Euclidean distance algorithm to quantify the performance deviation degree in the current period, and generate a local strategy execution quality deviation value to clearly characterize the degradation level of the current operating state of the end side relative to the ideal baseline.

[0154] S8.3: Perform a statistical consistency test on the local policy execution quality deviation value and the end-side policy distribution data within the historical sliding window. Use the chi-square goodness-of-fit test method to analyze the matching degree between the current output probability distribution and the cloud reference distribution, and generate a significant difference marker for the end-side policy distribution to determine whether there is a risk of policy failure due to context drift.

[0155] S8.4: Based on the significance marker of the distribution difference of the end-side policy and the deviation value of the execution quality of the local policy, execute the adaptive threshold decision logic, apply the dynamic step size adjustment algorithm to update the trigger interval parameter of cloud knowledge distillation, and generate a dynamically adjusted distillation frequency configuration instruction to establish an optimal balance between resource consumption and response real-time performance.

[0156] S8.5: Monitor the significant difference status of the end-side policy distribution and the trend of the local policy execution quality deviation value within multiple consecutive periods. When it is determined that the number of consecutive periods exceeds the preset fault tolerance threshold and the performance index is continuously lower than the baseline, construct an encrypted request message based on the dynamically adjusted distillation frequency configuration instruction, generate a new round of cloud policy output distribution vector distribution request, so as to force the synchronization of the latest global optimal policy knowledge in the cloud to the end side.

[0157] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0158] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0159] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for maintaining connection in intelligent devices with automatic switching of multimodal communication links, specifically comprising: S1: Obtain historical data on the status of multi-source heterogeneous links, and construct a graph neural network model based on the multi-source heterogeneous link status data; S2: Obtain multi-source heterogeneous link status data, and use the graph neural network model to perform spatiotemporal feature encoding processing on the multi-source heterogeneous link status data to generate a cloud strategy output distribution vector; S3: Deploy the edge graph neural network model, wherein the edge graph neural network model retains the graph structure encoding layer and the temporal attention gating module, and initializes the policy adaptation layer parameters embedded in the edge graph neural network model based on the cloud policy output distribution vector; S4: Monitor local communication context change characteristics. When the device is detected to meet the context drift conditions of entering a tunnel, sudden electromagnetic interference, or service type switching, trigger a single-step policy alignment mechanism based on the cloud policy output distribution vector. S5: Execute the single-step policy alignment mechanism, calculate the KL divergence constraint value between the current output distribution of the edge graph neural network model and the cloud policy output distribution vector, and update only the learnable parameters in the policy adaptation layer to generate an aligned policy adaptation layer. S6: Input the latest local link status data into the end-side graph neural network model loaded with the aligned strategy adaptation layer, and infer to generate a real-time optimal switching strategy instruction that adapts to the current dynamic context.

2. The intelligent device coupling retention method for automatic switching of a multi-modal communication link of claim 1, wherein, The process of adding steps S7-S8 after step S6 includes: S7: Execute the real-time optimal switching strategy instruction to complete the multimodal communication link switching operation, and collect strategy execution effect feedback data including switching time, connection recovery delay and number of service interruption frames; S8: Based on the feedback data of the strategy execution effect, compare the difference with the preset baseline. If the difference in the distribution of the end-side strategy exceeds the threshold or the execution effect is lower than the baseline within several consecutive periods, dynamically adjust the distillation frequency and trigger a new round of cloud strategy output distribution vector distribution request.

3. The intelligent device connection preservation method for automatic switching of a multi-modal communication link according to claim 1, wherein, Step S3 specifically includes: Obtain the network weight parameters and topology definition of the graph neural network model, perform channel sparsification on the complete network weight parameters, and generate a lightweight model architecture description file; Read the lightweight model architecture description file, instantiate a lightweight temporal graph neural network model in the memory space of the embedded processor on the edge, and map the remaining weight parameters after channel sparsification to the graph structure encoding layer and the temporal attention gating module to generate an uninitialized inference model on the edge with basic spatiotemporal feature extraction capabilities. The system receives a cloud policy output distribution vector sent from the cloud. The cloud policy output distribution vector contains a multi-candidate link fine-grained switching priority sequence and confidence distribution. The system then converts the cloud policy output distribution vector into a policy feature embedding vector in a high-dimensional latent space through an inverse transformation operation, thereby generating an initial policy knowledge representation to guide local parameter updates. Call the preset policy adaptation layer interface in the uninitialized inference model on the end side, take the initial policy knowledge representation as the input stimulus signal, perform forward propagation weight allocation operation, calculate the initial bias and scaling factor of the learnable parameters inside the policy adaptation layer, and generate an initialized policy adaptation layer component that has completed parameter cold start. The initialized policy adaptation layer component is integrated into the output end of the uninitialized inference model on the edge, replacing the original randomly initialized output layer, completing the assembly of the overall model structure and parameter solidification, and generating the final edge graph neural network model that can be used for real-time context awareness and policy generation.

4. The intelligent device connection preservation method for automatic switching of a multi-modal communication link according to claim 1, wherein, Step S4 specifically includes: Acquire a multi-source heterogeneous link status data stream containing real-time signal quality indicators, instantaneous service load parameters, current device movement trajectory coordinates, and the latest environmental interference event labels. Perform sliding window sampling and timestamp alignment processing on the multi-source heterogeneous link status data stream to eliminate timing deviations caused by asynchronous sensor acquisition and generate a standardized multimodal timing feature sequence. Based on the standardized multimodal temporal feature sequence, a local spatiotemporal correlation graph structure is constructed. The graph convolution operator is used to perform aggregation operations on the spatial dependencies between nodes, and the evolution law in the time dimension is extracted by combining the causal convolution mechanism to generate a dynamic spatiotemporal graph node embedding matrix that represents the current communication topology state. A multidimensional anomaly detection algorithm is applied to the dynamic spatiotemporal graph node embedding matrix to calculate the signal strength mutation rate, trajectory curvature change value, and service load entropy increase rate, respectively. The above calculation results are compared and analyzed with the preset context drift threshold vector to generate a context drift event candidate set including tunnel entry marker, electromagnetic interference marker, and service switching marker. Based on the aforementioned context drift event candidate set, a fusion decision processing is performed. A weighted voting mechanism is used to comprehensively evaluate the confidence scores of various drift markers. When the comprehensive score exceeds the decision threshold, the occurrence of context drift is confirmed, and a context drift trigger signal is generated to activate the policy update process. In response to the context drift trigger signal, the pre-stored secure encrypted communication interface is invoked to retrieve the latest cloud policy output distribution vector from the cloud storage area, and the cloud policy output distribution vector is encapsulated as the initialization input parameters of the single-step policy alignment mechanism to generate the single-step policy alignment instruction package.

5. A method for intelligent device attachment retention for automatic switching of multi-modal communication links as described in claim 1, wherein: Standardization of multi-source heterogeneous link status data includes frequency alignment, dimensional normalization, noise suppression, and outlier removal. Frequency alignment uses linear interpolation and spline interpolation. Dimensional normalization is achieved based on the historical extreme value range of the sampling parameters. Noise suppression uses moving mean filtering and bidirectional median filtering. Outlier removal uses a threshold of three times the standard deviation.

6. The intelligent device connection maintenance method for automatic switching of multimodal communication links as described in claim 1, characterized in that: Dynamic topology mapping uses a sliding time window to construct node attributes and adjacency relationships. The communication association strength is calculated based on the correlation weight of feature dimensions such as signal quality, service load, and location trajectory. Edge connections are established based on thresholds, and edge attributes include timestamps and time weights. The dynamic update mechanism reflects the evolution trend of the link over time.

7. The intelligent device connection maintenance method for automatic switching of multimodal communication links as described in claim 1, characterized in that: The graph neural network model consists of stacked multi-layer graph convolutional networks and temporal attention gating modules. Spatial features are obtained through neighborhood aggregation and weight fusion, while temporal features are extracted through causal convolution and multi-head attention mechanisms. Standardized candidate link priority scores and confidence intervals are output through fully connected layers and softmax functions.

8. The intelligent device connection maintenance method for automatic switching of multimodal communication links as described in claim 1, characterized in that: The structure of the edge graph neural network model is generated by pruning the global model in the cloud based on the pruning algorithm, retaining the graph structure encoding layer and the temporal attention gating module. The policy adaptation layer parameters are cold-started by inverse mapping of the policy distribution vector distributed in the cloud. During the instantiation of the edge model, weight mapping, parameter consistency verification and interface binding are performed.

9. The intelligent device connection maintenance method for automatic switching of multimodal communication links as described in claim 1, characterized in that: The policy execution effect feedback data includes handover time, connection recovery latency, and number of service interruption frames. Through protocol parsing, status signal acquisition, latency monitoring, frame loss statistics, and data encapsulation, structured feedback messages are generated to provide performance basis for system adaptive optimization and a new round of knowledge distillation.

10. The intelligent device connection maintenance method for automatic switching of multimodal communication links as described in claim 1, characterized in that: The real-time optimal switching strategy command output is generated by the aligned strategy distillation head combined with the current standardized multimodal link status input. The strategy distribution vector is decided by the parsing module, and the target link is selected based on the confidence score and priority ranking. After availability filtering and command encapsulation, control commands that can be directly executed by the link switching module are generated.

Images