Equipment cooperative control method and system based on distributed system, terminal and storage medium

Abstract

The invention discloses a device cooperative control method and system based on a distributed system, a terminal and a storage medium, and the method comprises the steps: obtaining the operation state data and communication link information of a plurality of physical devices, mapping each physical device into a node in a dynamic topological graph, and carrying out the cooperative control of the plurality of physical devices based on the operation state data and the communication link information; constructing feature representation of the distributed system; inputting the feature representation into a collaborative decision-making model for topological adaptive processing and inter-device long-range dependency relationship modeling, and outputting collaborative decision-making instructions corresponding to the physical devices; distributing the collaborative decision instruction to corresponding physical equipment for execution, and collecting system-level efficiency feedback generated when the collaborative decision instruction is executed; and dynamically optimizing the collaborative decision model according to the efficiency feedback and a preset system-level resource constraint so as to continuously adapt to the topology and state change of the distributed system. According to the method, the self-adaption to the dynamic change network topology is realized, and accurate collaborative decision and efficient resource control can be carried out.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-device collaborative control, and particularly to a device collaborative control method, system, terminal and computer-readable storage medium based on a distributed system. Background Art

[0002] The core of device collaborative control is to enable multiple independent devices to work together as a whole to complete complex tasks. The current mainstream technical paths are diverse, mainly including: building decentralized automatic collaboration rules based on blockchain and smart contracts to ensure the transparency and credibility of the process; fusing the data of each device based on multi-modal perception to form a unified environmental state perception for precise linkage; and based on the network topology, abstracting the devices as nodes and performing global joint scheduling through agent learning to optimize system-level objectives (such as energy efficiency, throughput). Although these methods have their own focuses, they are essentially exploring the mechanism of efficient collaboration between devices.

[0003] However, the existing technologies expose common defects when dealing with large-scale and highly dynamic industrial-level collaborative scenarios. First, they are generally insensitive to the dynamic changes of the network topology, and the preset connection relationships are difficult to adjust quickly when devices frequently go online, offline or the links fluctuate violently. Often, system reset or long-term retraining is required, which cannot meet the real-time requirements. Second, their collaborative decisions are mostly limited to the interaction between a device and its direct neighbors, lacking an effective mechanism to quantify and model the long-range implicit influence between non-adjacent nodes, so it is difficult to predict and prevent systemic risks such as cascading failures. Third, the trained models often rely heavily on specific network structures, with poor generalization and migration capabilities, and need to be trained from scratch when in a new environment or scale expansion, with high costs. Finally, when many methods pursue global optimality, they tend to ignore the local resource constraints (such as computing power, power) of individual devices, which may cause some nodes to be overloaded, thus endangering the overall stability.

[0004] Therefore, the existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main purpose of the present invention is to provide a device collaborative control method, system, terminal and computer-readable storage medium based on a distributed system, aiming to solve the problem in the existing technology that due to the frequent change of the device connection relationship in the dynamic topology network, the traditional fixed collaborative strategy fails and high-precision adaptive collaborative decision-making cannot be achieved.

[0006] To achieve the above purpose, the present invention provides a device collaborative control method based on a distributed system, and the device collaborative control method based on a distributed system includes the following steps: The system acquires real-time operational status data and communication link information of multiple physical devices, maps each physical device to a node in a dynamic topology graph, and constructs a feature representation of the distributed system based on the operational status data and the communication link information. The feature representation is input into the trained collaborative decision-making model, and the collaborative decision-making model is used to perform topology adaptation processing and modeling of long-range dependencies between devices, and outputs collaborative decision-making instructions corresponding to each physical device. The collaborative decision-making instructions are distributed to the corresponding physical devices for execution, and system-level performance feedback generated during the execution of the collaborative decision-making instructions is collected. Based on the performance feedback and preset system-level resource constraints, the collaborative decision-making model is dynamically optimized to continuously adapt to the topology and state changes of the distributed system.

[0007] Optionally, the device collaborative control method based on a distributed system, wherein the real-time acquisition of operating status data and communication link information of multiple physical devices specifically includes: By parsing communication messages of industrial fieldbus protocols or industrial Ethernet protocols, the operating status data of multiple physical devices can be read in real time. Each addressable physical device is mapped to a network node in a dynamic topology graph, and each physical device is assigned a unique node identifier. Based on the actual communication connections between devices or the preset network architecture, the communication links between nodes are measured or defined, and each link is assigned a weight value that characterizes the connection strength or communication quality to generate communication link information.

[0008] Optionally, the device collaborative control method based on a distributed system, wherein mapping each physical device to a node in a dynamic topology graph and constructing a feature representation of the distributed system based on the operating status data and the communication link information specifically includes: The operational status data of each physical device is organized into node status vectors, and the status vectors of all nodes are aggregated into a spatiotemporal feature matrix that represents the overall real-time status of the system. Based on the communication link information, a dynamic weighted adjacency matrix is ​​generated, and the matrix elements of the weighted adjacency matrix represent the connection relationship and communication attributes between the corresponding physical device nodes. The weighted adjacency matrix and the spatiotemporal feature matrix are fused and encoded to form the feature representation of the distributed system.

[0009] Optionally, in the device collaborative control method based on a distributed system, the collaborative decision model is a graph attention-Transformer coupled engine; The process of performing topology adaptation processing and modeling long-range dependencies between devices through the collaborative decision-making model, and outputting collaborative decision-making instructions corresponding to each physical device, specifically includes: The graph attention-transformer coupling engine detects the topological changes reflected in the feature representation, and performs topological adaptive adjustment on the feature representation based on the topological changes to obtain the adaptively processed feature representation. Based on the adaptively processed feature representation, the long-range dependencies between devices are modeled, and based on the long-range dependencies and system-level resource constraints, a set of collaborative decision-making instructions corresponding to each physical device is output.

[0010] Optionally, the device cooperative control method based on a distributed system, wherein detecting the topology change reflected by the feature representation and performing topology adaptive adjustment on the feature representation according to the topology change to obtain an adaptively processed feature representation specifically includes: By verifying topological invariance, the fingerprint features of the current dynamic topology are calculated. When the change of the fingerprint features from the historical state exceeds a preset threshold, it is determined that the topology has changed. When the topology changes, dynamic neighbor extraction and graph structure encoding are performed on the feature representation to obtain an adaptively processed feature representation.

[0011] Optionally, in the device cooperative control method based on a distributed system, the step of performing dynamic neighbor extraction and graph structure encoding on the feature representation specifically includes: Based on a preset or dynamically calculated number of hops, extract the corresponding k-hop neighbor nodes for each node in the dynamic topology graph, and construct a dynamic neighbor set for each node based on the k-hop neighbor nodes. The embedding representation of each node is updated based on its own features and the set features of its corresponding dynamic neighbor set by using a graph neural network.

[0012] Optionally, in the device collaborative control method based on a distributed system, when performing dynamic neighbor extraction, the hop value used to constrain the neighbor range is dynamically determined according to the degree of the node. For nodes whose degree exceeds a first preset threshold, a hop value less than a second preset threshold is automatically used to limit the node's perception and communication range.

[0013] Optionally, the device cooperative control method based on a distributed system, wherein modeling the long-range dependencies between devices based on the adaptively processed feature representation specifically includes: The node embeddings in the adaptively processed feature representation are fused and normalized to generate an enhanced node feature vector. The enhanced node feature vector is input into the multi-head graph attention layer, and an attention mask based on real-time topology connection is introduced to calculate and capture the long-range dependency strength between any two reachable device nodes in the network, and output the context features representing the long-range dependency relationship. The contextual features are mapped to a value estimate representing the global state of the entire distributed system through a feedforward neural network, and the value estimate is used to assist in generating collaborative decision-making instructions.

[0014] Optionally, the device cooperative control method based on a distributed system further includes, before outputting the cooperative decision instructions corresponding to each physical device: The dedicated computing unit of the collaborative decision-making model performs counterfactual advantage calculation based on the global state value estimate generated by the collaborative decision-making model, the probability distribution of the action strategies of each device, and the estimated immediate reward, to obtain the counterfactual advantage value of each device. The change in distributed system performance of each device is calculated based on the counterfactual advantage value. The contribution of each device is calculated based on the change in distributed system performance. Based on all the contribution values, the parameters of the policy network of the collaborative decision-making model are updated to optimize subsequent decisions.

[0015] Optionally, in the device collaborative control method based on a distributed system, the step of distributing the collaborative decision instructions to the corresponding physical devices for execution specifically includes: Obtain the current network conditions. If the current network conditions are a reliable network, then use a broadcast matrix to uniformly distribute the instructions to the corresponding physical devices for execution. If the current network conditions are a communication-restricted or unreliable network, a tree-like relay transmission strategy is adopted to forward the instructions to the corresponding physical devices for execution in a hierarchical manner.

[0016] Optionally, in the device collaborative control method based on a distributed system, the step of collecting system-level performance feedback generated during the execution of the collaborative decision-making instruction specifically includes: Real-time operational status data is collected from each physical device after the execution of the collaborative decision-making instructions. The data includes the device's own performance indicators and resource consumption. Collect actual interaction data of communication links between devices in the distributed system, including communication latency and link reliability indicators; Based on the collected real-time operating status data and the actual interaction data, calculations are performed and fused according to the preset global optimization goals to generate system-level performance feedback as the basis for subsequent optimization.

[0017] Optionally, the device collaborative control method based on a distributed system, wherein the step of dynamically optimizing the collaborative decision-making model based on the performance feedback and preset system-level resource constraints specifically includes: Based on the instruction execution data in the performance feedback, the power consumption constraint parameters in the system-level resource constraints are dynamically updated; Based on the updated resource constraints, the policy gradient is weighted according to the degree of network nodes to generate a policy update direction that adapts to different network densities. Based on the system state prediction data in the performance feedback, a long-term reward is calculated, which is used to enhance the optimization objective's consideration of long-term risks. By combining the strategy update direction adapted to different network densities with the enhancement optimization objective incorporating long-term risk considerations, the parameter update of the collaborative decision-making model is driven, thereby achieving dynamic optimization of the collaborative decision-making model.

[0018] Optionally, the device collaborative control method based on a distributed system is deployed on an industrial Internet of Things edge computing node to perform real-time collaborative control of physical devices in a smart energy network, a smart manufacturing production line, or a distributed communication network.

[0019] Furthermore, to achieve the above objectives, the present invention also provides a device collaborative control system based on a distributed system, wherein the device collaborative control system based on a distributed system includes: The perception construction module is used to acquire the operating status data and communication link information of multiple physical devices in real time, map each physical device as a node in a dynamic topology graph, and construct a feature representation of the distributed system based on the operating status data and the communication link information. An adaptive decision module is used to input the feature representation into a trained collaborative decision model, perform topology adaptive processing and model long-range dependencies between devices through the collaborative decision model, and output collaborative decision instructions corresponding to each physical device. The execution and feedback module is used to distribute the collaborative decision-making instructions to the corresponding physical devices for execution, and to collect system-level performance feedback generated when the collaborative decision-making instructions are executed. The dynamic optimization module is used to dynamically optimize the collaborative decision-making model based on the performance feedback and preset system-level resource constraints, so as to continuously adapt to the topology and state changes of the distributed system.

[0020] Optionally, in the device collaborative control system based on a distributed system, the perception construction module includes a data perception unit and a feature construction unit; The data sensing unit is used to read the operating status data of multiple physical devices in real time by parsing the communication messages of the industrial fieldbus protocol or industrial Ethernet protocol, map each addressable physical device as a network node in the dynamic topology, assign a unique node identifier to each physical device, measure or define the communication links between nodes based on the actual communication connection between devices or the preset network architecture, and assign a weight value representing the connection strength or communication quality to each link to generate communication link information. The feature construction unit is used to organize the operating status data of each physical device into a node state vector, aggregate all node state vectors into a spatiotemporal feature matrix that represents the overall real-time status of the system, generate a dynamic weighted adjacency matrix based on the communication link information, wherein the matrix elements of the weighted adjacency matrix represent the connection relationship and communication attributes between the corresponding physical device nodes, and fuse and encode the weighted adjacency matrix with the spatiotemporal feature matrix to form the feature representation of the distributed system.

[0021] Optionally, in the device collaborative control system based on a distributed system, the adaptive decision-making module includes a topology adaptive adjustment unit and a long-range dependency modeling unit. The topology adaptive adjustment unit is used to detect the topological changes reflected by the feature representation through the graph attention-transformer coupling engine, and to perform topology adaptive adjustment on the feature representation according to the topological changes to obtain the adaptively processed feature representation. The long-range dependency modeling unit is used to model the long-range dependencies between devices based on the adaptively processed feature representation, and output a set of collaborative decision instructions corresponding to each physical device based on the long-range dependencies and system-level resource constraints.

[0022] Optionally, in the device collaborative control system based on a distributed system, the execution and feedback module includes an instruction execution unit and a performance feedback unit; The instruction execution unit is used to obtain the current network conditions. If the current network conditions are a reliable network, the instruction is uniformly distributed to the corresponding physical device for execution using a broadcast matrix. If the current network conditions are a communication-restricted or unreliable network, the instruction is forwarded to the corresponding physical device for execution using a tree-like relay transmission strategy. The performance feedback unit is used to collect real-time operating status data from each physical device after executing the collaborative decision-making command. The data includes the device's own performance indicators and resource consumption. It also collects actual interaction data of the communication links between devices in the distributed system. The actual interaction data includes communication latency and link reliability indicators. Based on the collected real-time operating status data and the actual interaction data, it calculates and integrates the data according to the preset global optimization target to generate system-level performance feedback as the basis for subsequent optimization.

[0023] Optionally, in the device collaborative control system based on a distributed system, the dynamic optimization module includes a resource constraint update unit, a strategy update unit, a long-range reward calculation unit, and a dynamic optimization unit. The resource constraint update unit is used to dynamically update the power consumption constraint parameters in the system-level resource constraints based on the instruction execution data in the performance feedback; The policy update unit is used to generate policy update directions that adapt to different network densities by weighting the policy gradient according to the network node degree based on the updated resource constraints. The long-term reward calculation unit is used to calculate long-term rewards based on the system state prediction data in the performance feedback. The long-term rewards are used to enhance the optimization target's consideration of long-term risks. The dynamic optimization unit is used to combine the strategy update direction adapted to different network densities with the enhancement optimization objective that incorporates long-term risk considerations, thereby driving the parameter update of the collaborative decision-making model and realizing the dynamic optimization of the collaborative decision-making model.

[0024] In addition, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a distributed system-based device collaborative control program stored in the memory and executable on the processor, wherein when the distributed system-based device collaborative control program is executed by the processor, it implements the steps of the distributed system-based device collaborative control method as described above.

[0025] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a device cooperative control program based on a distributed system, and when the device cooperative control program based on a distributed system is executed by a processor, it implements the steps of the device cooperative control method based on a distributed system as described above.

[0026] In this invention, the method includes: acquiring real-time operational status data and communication link information of multiple physical devices; mapping each physical device to a node in a dynamic topology graph; and constructing a feature representation of the distributed system based on the operational status data and the communication link information; inputting the feature representation into a trained collaborative decision-making model; performing topology adaptation processing and modeling long-range dependencies between devices through the collaborative decision-making model; and outputting collaborative decision-making instructions corresponding to each physical device; distributing the collaborative decision-making instructions to the corresponding physical devices for execution; and collecting system-level performance feedback generated during the execution of the collaborative decision-making instructions; and dynamically optimizing the collaborative decision-making model based on the performance feedback and preset system-level resource constraints to continuously adapt to changes in the topology and state of the distributed system. This invention achieves adaptation to dynamically changing network topologies, enabling precise collaborative decision-making and efficient resource control. Attached Figure Description

[0027] Figure 1 This is a flowchart of a preferred embodiment of the device collaborative control method based on a distributed system according to the present invention; Figure 2 This is a schematic diagram of the principle architecture of the device collaborative control method based on a distributed system according to the present invention; Figure 3 This is a flowchart illustrating the real-time acquisition of operating status data and communication link information of multiple physical devices in the device collaborative control method based on a distributed system according to the present invention. Figure 4 This is a flowchart illustrating the characteristic representation of a distributed system constructed in the device collaborative control method for distributed systems of this invention; Figure 5 This is a flowchart of the device collaborative control method based on a distributed system of the present invention, which outputs collaborative decision instructions corresponding to each physical device. Figure 6 This is a flowchart of the topology adaptive adjustment of the feature representation in the device cooperative control method based on a distributed system according to the present invention; Figure 7 This is a flowchart of the process of performing dynamic neighbor extraction and graph structure encoding on the feature representation in the device cooperative control method based on a distributed system according to the present invention; Figure 8 This is a flowchart illustrating the long-range dependency relationship between devices based on the adaptively processed feature representation in the device cooperative control method of the distributed system of the present invention. Figure 9 This is a flowchart illustrating the calculation of the counterfactual advantage value and the contribution of the system efficiency in the device cooperative control method based on a distributed system according to the present invention. Figure 10This is a flowchart illustrating how the collaborative decision-making instructions are distributed to the corresponding physical devices for execution in the device collaborative control method based on a distributed system according to the present invention. Figure 11 This is a flowchart illustrating the process of collecting system-level performance feedback generated during the execution of the collaborative decision-making instructions in the device collaborative control method based on a distributed system according to the present invention. Figure 12 This is a flowchart of the dynamic optimization of the collaborative decision-making model in the device collaborative control method based on a distributed system according to the present invention; Figure 13 This is a schematic diagram of a device collaborative control system based on a distributed system according to the present invention; Figure 14 This is a schematic diagram of the perception construction module in the device collaborative control system based on a distributed system of the present invention; Figure 15 This is a schematic diagram of the adaptive decision-making module in the device collaborative control system based on a distributed system of the present invention; Figure 16 This is a schematic diagram of the execution and feedback module in the device collaborative control system based on a distributed system of the present invention; Figure 17 This is a schematic diagram of the dynamic optimization module in the device collaborative control system based on a distributed system of the present invention; Figure 18 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0028] This application provides a device collaborative control method and related equipment based on a distributed system. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0029] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0030] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0031] The preferred embodiment of the device cooperative control method based on a distributed system described in this invention, such as... Figure 1 and Figure 2 As shown, the device cooperative control method based on a distributed system includes: Step S10: Real-time acquisition of operating status data and communication link information of multiple physical devices, mapping each physical device to a node in a dynamic topology graph, and constructing a feature representation of the distributed system based on the operating status data and the communication link information.

[0032] Understandably, this step forms the data foundation and logical starting point for the entire collaborative control method. It achieves a precise transformation from the physical world to a digital model, providing structured input for subsequent intelligent decision-making. The purpose of this step is not simply to collect data, but to construct a real-time, computable, dynamic digital mapping for the entire distributed system. It unifies and abstracts the scattered, heterogeneous physical device states and complex network connections into a standard graph data structure.

[0033] like Figure 3 As shown, the real-time acquisition of operating status data and communication link information of multiple physical devices specifically includes: Step S21: Read the operating status data of multiple physical devices in real time by parsing the communication messages of the industrial fieldbus protocol or industrial Ethernet protocol.

[0034] In this embodiment, the operational status data refers to the inherent attributes and performance indicators of each physical device, such as CPU / memory utilization, sensor readings (temperature, pressure), task queue length, remaining battery power, and local controller output values. This data is typically obtained by reading device registers, parsing device protocol messages (such as OPC UA, MQTT topics), or calling the device API.

[0035] Furthermore, this embodiment clarifies that the source of the operational status data is real industrial communication protocol messages. This directly anchors the method to specific industrial scenarios such as manufacturing and energy, eliminating the possibility of obtaining data from simulations or abstract interfaces.

[0036] Step S22: Map each addressable physical device to a network node in the dynamic topology graph, and assign a unique node identifier to each physical device.

[0037] In this embodiment, mapping rules from physical devices to graph nodes are defined. Addressability means that each device must have a unique identifier (such as IP address or device ID). Unique node identifiers ensure that each device entity can be unambiguously located and tracked in the graph model, which is the foundation for building an accurate topology model.

[0038] Step S23: Based on the actual communication connection between devices or the preset network architecture, measure or define the communication links between nodes, and assign a weight value representing the connection strength or communication quality to each link to generate communication link information.

[0039] In this embodiment, how to transform physical or logical connections into mathematical weights is defined. "Measurement" applies to dynamic networks (e.g., calculating weights based on real-time latency), while "definition" applies to static architectures (e.g., pre-setting weights based on physical cabling). Assigning weights to links is crucial, upgrading the topology graph from a binary "connected or not" relationship to a quantitative model describing "connection quality," providing input for subsequent weight-based algorithms (e.g., weighted graph convolution).

[0040] Furthermore, such as Figure 4 As shown, the process of mapping each physical device to a node in a dynamic topology graph and constructing a feature representation of the distributed system based on the operational status data and the communication link information specifically includes: Step S31: Organize the operating status data of each physical device into a node state vector, and aggregate the state vectors of all nodes into a spatiotemporal feature matrix that represents the overall real-time status of the system.

[0041] It is understandable that organizing a node into a state vector refers to normalizing and concatenating multiple state parameters of a single device to form the feature vector of that node. Aggregating into a spatiotemporal feature matrix represents stacking the feature vectors of all nodes row by row. The spatiotemporal feature matrix encapsulates the overall state of the system at a certain moment.

[0042] Furthermore, the spatiotemporal feature matrix is ​​represented as: ; in, , This represents the characteristic representation of the spatiotemporal feature matrix at time t. This represents the state vector of device i at time t. Indicates the total number of devices. Indicates the operating status of the equipment This indicates the device's power consumption. This indicates the device's latency.

[0043] Step S32: Generate a dynamic weighted adjacency matrix based on the communication link information. The matrix elements of the weighted adjacency matrix represent the connection relationship and communication attributes between the corresponding physical device nodes.

[0044] In this embodiment, a dynamic weighted adjacency matrix A ∈ R^(N×N) is generated based on the communication link information, where the value of element A[i][j] is the weight of the link from node i to j. This matrix is ​​the mathematical equivalent of the graph structure and is the core input for all graph algorithms to perceive topology.

[0045] Step S33: The weighted adjacency matrix and the spatiotemporal feature matrix are fused and encoded to form the feature representation of the distributed system.

[0046] In this embodiment, it is clarified that the final formation of the feature representation is the result of the combined action of the state matrix H and the structure matrix A. The fusion encoding is usually implemented by one or more layers of graph neural networks, and its output is a new node feature matrix H', in which the features of each node have included information about its neighbors, forming a high-level feature representation that contains both attributes and structure.

[0047] Step S20: Input the feature representation into the trained collaborative decision-making model, perform topology adaptation processing and long-range dependency modeling between devices through the collaborative decision-making model, and output the collaborative decision-making instructions corresponding to each physical device.

[0048] Understandably, this step is the core of the intelligent decision-making in the collaborative control method. This step inputs the previously constructed system feature representation into a pre-trained collaborative decision-making model (a graph attention-Transformer coupled engine). The model then performs two key technical processes: First, it performs topology adaptation, meaning the model can detect changes in the topology implicit in the feature representation in real time and dynamically adjust the internal information aggregation paths and interaction weights between nodes to ensure that the decision-making basis remains consistent with the current actual network conditions, thus solving the problem of fixed strategies failing in dynamic networks. Second, it models long-range dependencies between devices. Utilizing mechanisms such as graph attention, the model overcomes the limitations of traditional methods that only focus on direct neighbors, actively calculating and capturing indirect, multi-hop influence relationships between any two reachable nodes in the network, thereby predicting systemic risks and making collaborative decisions with a global perspective. Finally, the model integrates the above adaptation results to dynamic topology and the analysis of complex dependencies, outputting a set of precise collaborative decision-making instructions for each physical device, completing the crucial transformation from perception to decision-making.

[0049] like Figure 5As shown, the step of performing topology adaptation processing and modeling long-range dependencies between devices through the collaborative decision-making model, and outputting collaborative decision-making instructions corresponding to each physical device, specifically includes: Step S41: Detect the topological changes reflected by the feature representation through the graph attention-transformer coupling engine, and perform topological adaptive adjustment on the feature representation according to the topological changes to obtain the adaptively processed feature representation.

[0050] In this embodiment, the first step of the model processing is to adjust the input features in relation to topology changes and output a cleaned feature representation that is more adapted to the current network conditions.

[0051] Step S42: Based on the adaptively processed feature representation, model the long-range dependencies between devices, and based on the long-range dependencies and system-level resource constraints, output a set of collaborative decision instructions corresponding to each physical device.

[0052] In this embodiment, the profound influence relationships between modeling devices are clarified based on the adjusted features, and resource constraints are taken into account to finally generate instructions. This reveals that decision-making is a two-step process: first, ensuring that "the graph seen is correct," and then "making the optimal decision" on the correct graph.

[0053] like Figure 6 As shown, further, the step of detecting the topological changes reflected in the feature representation, and performing topological adaptive adjustment on the feature representation based on the topological changes to obtain an adaptively processed feature representation, specifically includes: Step S51: Calculate the fingerprint features of the current dynamic topology graph through topology invariance verification. When the change of the fingerprint features from the historical state exceeds a preset threshold, it is determined that the topology structure has changed.

[0054] In this embodiment, a specific method for "detecting topological changes" is defined: topological fingerprint comparison. The feature hash of the current topological graph (e.g., based on node degree sequences) is calculated and compared with historical fingerprints. Only changes exceeding a threshold are considered "significant changes." This is a lightweight and efficient change-aware strategy that avoids overreacting to momentary fluctuations.

[0055] Specifically, the process of topological invariance verification is represented as follows: ;in, This represents the hash value used to verify the topological invariance of graph G. This represents a cryptographic hash function (which maps input data of arbitrary length to a fixed-length (256-bit) hash value. Here, it is used to process the L2 norm calculated based on the node degree distribution to generate a unique and fixed-length hash value to ensure the reliability and security of verification). Represents all nodes in graph G The set of degrees, Describing the L2 norm, This represents the amount of change between the currently calculated topological hash value and the previously stored topological hash value. This represents a pre-defined threshold used to determine whether the change in the topological hash value is large enough. At that time, it was assumed that the topological structure of the graph had changed significantly. This indicates that the model is being retrained.

[0056] Step S52: When the topology changes, perform dynamic neighbor extraction and graph structure encoding on the feature representation to obtain the adaptively processed feature representation.

[0057] In this embodiment, two specific operations are defined to update the feature representation after a change is detected: dynamic neighbor extraction and graph structure encoding. It is also specified that the output of this step is an adaptively processed feature representation.

[0058] like Figure 7 As shown, further, the step of performing dynamic neighbor extraction and graph structure encoding on the feature representation specifically includes: Step S61: Extract the corresponding k-hop neighbor nodes for each node in the dynamic topology graph according to the preset or dynamically calculated number of hops, and construct a dynamic neighbor set for each node based on the k-hop neighbor nodes.

[0059] Understandably, this embodiment defines how to determine the currently valid interaction range for each node. "k-hop neighbors" means considering not only direct neighbors but also potentially indirect neighbors (such as friends of friends). "Constitutes a dynamic neighbor set" indicates that this range is calculated in real-time based on the current topology, rather than being fixed.

[0060] Specifically, the process of performing dynamic neighbor extraction on the feature representation can be represented as follows: ; in, , Represents the first in the network 1 node Represents the first in the network 1 node Represents a node The K-hop neighborhood set, Represents a node and nodes The shortest path length between them. This represents the dynamically calculated hop count used to determine the neighbor range. This indicates the maximum number of hops limit (which can be 3 in this example). This represents the function for rounding up. Represents a node The degree.

[0061] Step S62: Use a graph neural network to aggregate the individual features of each node and the set features of the corresponding dynamic neighbor set, and update the embedding representation of each node according to the individual features and the set features.

[0062] Understandably, this embodiment also defines how to update node features using the aforementioned neighbor set. By using graph neural network aggregation, neighbor features are converged to the central node through message passing, thus ensuring that each node's new features contain contextual information about its local topology.

[0063] Specifically, by using a graph neural network to aggregate the individual features of each node and the set features of the corresponding dynamic neighbor set, it can be represented as follows: ; ; in, This represents the initial graph structure encoding feature matrix. This represents a learnable embedding parameter matrix (used to map the original feature matrix to a feature space suitable for graph convolution operations, and by learning to adjust the parameters to better capture the relationships between device features). Indicates time Device feature matrix at time, Indicates the first After the layer graph convolution operation, the nodes Feature representation, Represents the ReLU activation function. Indicates a node All neighboring nodes Perform a summation operation. The elements in the adjacency matrix represent the graph structure. Indicates the first Learnable parameter moments of layer graph convolution, Indicates the first After the layer graph convolution operation, the nodes The characteristics are represented.

[0064] Furthermore, when performing dynamic neighbor extraction, the hop value used to constrain the neighbor range is dynamically determined based on the degree of the node. For nodes whose degree exceeds the first preset threshold, a hop value less than the second preset threshold is automatically used to limit the node's perception and communication range.

[0065] Specifically, this embodiment stipulates that the sensing radius (hop count k) of a node is inversely proportional to its number of connections (degree). For network hubs (high-degree nodes), their sensing range is automatically reduced. This is a key load balancing strategy to prevent computational and communication overload, ensuring the scalability of the algorithm in large-scale dense networks. Its core lies in introducing a resource-aware adaptive sensing range control mechanism based on node degree. This scheme explicitly stipulates that the hop count k used to constrain the neighbor range is not a globally fixed value or randomly selected, but is dynamically calculated and determined based on the degree of each node itself. Specifically, for "high-degree nodes" with a large number of connections in the network (e.g., gateways acting as data aggregation points in industrial IoT, or switches in core positions in communication networks), the system automatically assigns them a smaller k value, thereby proactively limiting the number of neighbor nodes they need to process during the topology sensing and information aggregation phases.

[0066] This design directly addresses the core engineering challenges of large-scale distributed collaborative systems—resource load balancing and scalability. High-density nodes naturally become performance bottlenecks in traditional graph computation: applying the same or even larger perception radius to them as to ordinary nodes leads to an exponential increase in neighbor information processing, causing a surge in computational latency, excessive memory consumption, and network congestion. By intelligently shrinking the "vision" of these critical nodes, excessive computational and communication loads are effectively prevented from accumulating on them, ensuring that the collaborative decision-making algorithm can still operate efficiently and stably in ultra-large-scale, densely connected network topologies. Furthermore, it embodies an inventive approach that deeply integrates system resource constraints into algorithm design, significantly improving the practicality and robustness of the entire method when facing complex, heterogeneous network structures in the real world.

[0067] like Figure 8 As shown, further, the step of modeling the long-range dependencies between devices based on the adaptively processed feature representation specifically includes: Step S71: Perform fusion and normalization processing on the node embeddings in the adaptively processed feature representation to generate enhanced node feature vectors.

[0068] Understandably, further normalization and transformation of the node embeddings in the adaptively processed feature representation are intended to improve the discriminative power and training stability of the features, thus preparing for attention computation.

[0069] Step S72: Input the enhanced node feature vector into the multi-head graph attention layer, and introduce an attention mask based on real-time topology connection relationship to calculate and capture the long-range dependency strength between any two reachable device nodes in the network, and output the context features representing the long-range dependency relationship.

[0070] This step is central to modeling long-range dependencies. By using a "real-time topology-based attention mask," attention is ensured to be computed only between topologically reachable node pairs, injecting physical constraints into the model. Its direct output is capturing the contextual features of long-range dependencies.

[0071] Specifically, the multi-head graph attention layer is used to execute the multi-head graph attention layer mechanism, which can be represented as: ; in, This represents the query matrix in the m-th attention head. The node feature matrix of the graph. This represents the learnable parameter matrix used to generate the query matrix in the m-th attention head. This represents the key matrix in the m-th attention head. Let represent the learnable parameter matrix used to generate the key matrix in the m-th attention head. This represents the value matrix in the m-th attention head. Let represent the learnable parameter matrix used to generate the value matrix in the m-th attention head. This represents the output matrix of the m-th attention head. Represents the normalized exponential function, The graph mask matrix representing the constrained attention. It represents the Hadamah accumulation. This represents the product of the transposes of the query matrix and the key matrix. Represents the dimension of the key vector The square root of.

[0072] Step S73: The context features are mapped to a value estimate representing the global state of the entire distributed system through a feedforward neural network. The value estimate is used to assist in generating collaborative decision-making instructions.

[0073] In this embodiment, the dependency information contained in each node is globally aggregated and mapped to a global state value estimate. This scalar value represents the overall quality of the system and is used to guide the generation of final device instructions.

[0074] Specifically, the context features are mapped to a value estimate representing the global state of the entire distributed system using a feedforward neural network, expressed as: ; in, The global state vector (is a vector obtained by aggregating the integrated results of the multi-head attention mechanism outputs, used to characterize the comprehensive state information of the entire graph, extracting key information from the outputs of multiple attention heads and fusing them to obtain a description of the overall features of the graph). Indicates the number of attention heads. This represents the output projection matrix of the m-th attention head. This represents the output matrix of the m-th attention head. This represents the final calculated global state value. This represents a weight vector used to perform a linear transformation on the result after the activation function. Represents the activation function of the Gaussian error linear unit. This represents the weight matrix of the first layer in a two-layer MLP.

[0075] The above formula describes the process from the output integration of the multi-head graph attention mechanism to the final global state value calculation. Through operations such as linear transformation and nonlinear activation, it makes full use of the graph's structural information and node features to obtain an evaluation of the graph's global state.

[0076] like Figure 9 As shown, further, before outputting the collaborative decision-making instructions corresponding to each physical device, the method also includes: Step S81: Using the dedicated computing unit of the collaborative decision-making model, counterfactual advantage calculation is performed based on the global state value estimate generated by the collaborative decision-making model, the probability distribution of the action strategies of each device, and the estimated immediate reward, to obtain the counterfactual advantage value of each device.

[0077] Understandably, this step defines the need to calculate a "counterfactual advantage value" for each device before executing a decision. This calculation utilizes the model's internal global value estimate, policy probability, and reward prediction, indicating a deeply coupled internal evaluation process within the decision-making model. The calculation is performed by a dedicated computational unit within the collaborative decision-making model and occurs before the final collaborative decision instruction is output. This suggests that the calculation is a necessary, embedded analytical step within the model's decision-making process, rather than a separate post-processing module.

[0078] Specifically, the process of performing counterfactual advantage calculations to obtain the counterfactual advantage values ​​for each device can be represented as follows: ; in, Indicates time At that time, individual The counterfactual advantage value, Indicates time Instant rewards at any time Indicates the discount factor. Indicates time Estimated state value at that time Indicates all possible actions Summation, Indicates the state of a given environment Next, take action The probability, Indicates time The environmental conditions at that time, Indicates time Take action The instant reward obtained afterward Indicates time Take action The estimated value of the new state that is reached later.

[0079] This formula is used for counterfactual advantage calculation, which calculates individual contribution through single-step branching and eliminates dimensional differences in industrial scenarios through cross-device normalization.

[0080] Step S82: Calculate the change in distributed system performance of each device based on the counterfactual advantage value, calculate the contribution of each device based on the change in distributed system performance, and update the parameters of the policy network of the collaborative decision-making model based on all the contribution values ​​to optimize subsequent decisions.

[0081] This step clarifies the role of counterfactual advantage values: firstly, they quantify individual contributions; secondly, they are directly used to update policy network parameters. This solves the credit allocation problem of "who gets the credit" in multi-agent collaboration, guiding the model to learn more cooperative policies.

[0082] Specifically, this step clarifies how counterfactual advantage values ​​can be used to directly drive model evolution. Contribution quantification: The counterfactual advantage value calculated in the first section is itself defined as a quantified value of the individual contribution of a device. A positive value indicates that the action has an above-average positive contribution to the system, while a negative value indicates that its performance is below its average. This transforms contribution evaluation from a vague concept into a precisely calculable and comparable scalar signal. Policy update role: This contribution signal is directly used to guide the parameter updates of the policy network. In the policy gradient algorithm, this advantage value is used as a weight coefficient for updating the gradient. Specifically, during training, the model adjusts the internal parameters of its policy network in a direction that amplifies the probability of high-contribution actions and suppresses the probability of low-contribution actions. This allows the model to clearly identify the specific merits and demerits of each device's action to the team's success, thereby achieving precise and efficient gradient updates. This drives the entire collaborative strategy to continuously optimize towards maximizing overall effectiveness, effectively avoiding the problem of ambiguous merit distribution in multi-agent systems.

[0083] The process of calculating the distributed system performance of each device based on the counterfactual advantage value can be expressed as follows: ; in, This represents the change in the performance of a distributed system. This represents the weighting coefficient of the control effectiveness term. This represents the objective function for pruning proxies in near-end strategy optimization. Describe the objective function of the pruning proxy. The partial derivatives with respect to the policy network parameters θ, This represents the weighting coefficient for the energy cost item. Control input vector, Represents the power consumption constraint matrix. Represents a scalar value, indicating the value at a given control input. and power consumption constraint matrix A certain power consumption-related metric, This represents the partial derivative of the aforementioned power consumption-related metrics with respect to the policy network parameter θ. This represents the weighting coefficient of the counterfactual advantage contribution item. Indicates the number of devices. Indicates equipment In time The counterfactual advantage value, Represents the normalization function. Indicates equipment The industrial importance weight.

[0084] Step S30: Distribute the collaborative decision-making instructions to the corresponding physical devices for execution, and collect system-level performance feedback generated when the collaborative decision-making instructions are executed.

[0085] Understandably, this step is crucial for realizing the transition from intelligent decision-making to physical execution and forming a learning loop. First, the instruction set output by the collaborative decision-making model is distributed to each physical device based on network reliability and target address, using strategies such as efficient broadcasting or robust relaying to drive their execution. This transforms the optimization strategy in the digital space into actual control actions in the physical world. Simultaneously, the system collects performance data at two levels: first, the micro-states such as the local execution results and resource consumption of each device; and second, macro-interaction indicators such as communication latency and link quality between devices. These data are fused and calculated to generate system-level performance feedback characterizing the overall effect of this collaborative action. This step not only completes the delivery of control commands but, more importantly, captures the "consequences" of the decision's execution in the real physical environment. This provides indispensable real-world evidence for subsequent dynamic model optimization, thus forming a complete closed loop of decision-making, execution, and evaluation.

[0086] like Figure 10 As shown, distributing the collaborative decision-making instructions to the corresponding physical devices for execution specifically includes: Step S91: Obtain the current network conditions. If the current network conditions are a reliable network, then use a broadcast matrix to uniformly distribute the instructions to the corresponding physical devices for execution.

[0087] It is understood that the reliable network specified in this embodiment typically refers to wired industrial Ethernet (such as EtherCAT, PROFINET IRT) or a high-quality, stable wireless private network. In this mode, the system uses a broadcast matrix (or multicast) for unified distribution. The technical principle is to leverage the high determinism and low packet loss rate of the underlying network, allowing the control node to send command packets to all relevant devices at once, or to cover all nodes through a single network flood. The advantage of this approach is extremely low communication latency and minimal network overhead, making it suitable for periodic control scenarios with stringent real-time requirements (such as synchronous motion control).

[0088] Step S92: If the current network conditions are a communication-restricted or unreliable network, a tree-like relay transmission strategy is adopted to forward the instructions to the corresponding physical devices for execution in a hierarchical manner.

[0089] Understandably, communication may be restricted or unreliable, encompassing common wireless IoT scenarios (such as the use of Wi-Fi and Zigbee in complex environments), or wired networks with unstable links. In such cases, the system switches to a tree-based relay transmission strategy. The technical principle is to organize the instruction transmission path into a logical tree with the control node as the root and the target device as the leaves. Instructions are forwarded along this pre-defined or dynamically generated "tree" in multiple hops, with each hop ensuring delivery to the next level's parent or child node. This approach sacrifices some end-to-end latency, but by decomposing long-distance, unstable direct communication into multiple short-distance, more reliable links, and leveraging the store-and-forward capabilities of intermediate nodes, it significantly improves the final delivery rate (robustness) of instructions under adverse network conditions.

[0090] In summary, this step is a key design element ensuring that collaborative control commands are reliably and robustly delivered to each target physical device in real, unreliable industrial network environments. It defines an adaptive dual-mode command distribution strategy, the core of which is to intelligently switch or select between two different communication protocol modes based on real-time sensing or assessment of current network conditions, thereby achieving an optimal trade-off between efficiency and reliability. It does not prescribe a fixed communication method, but rather protects a complete set of dynamic transport layer adaptation mechanisms based on network state awareness. This directly addresses the core engineering challenge of non-ideal and variable network conditions in industrial settings, ensuring that upper-layer intelligent collaborative decisions, regardless of the network environment, can reliably operate in the physical world, thus guaranteeing the integrity and practicality of the entire method from "decision-making" to "execution."

[0091] like Figure 11 As shown, further, the collection of system-level performance feedback generated during the execution of the collaborative decision-making instructions specifically includes: Step S101: Collect real-time operating status data from each physical device after executing the collaborative decision-making instruction. The data includes the device's own performance indicators and resource consumption.

[0092] Understandably, once an instruction is issued, the system needs to retrieve the performance metrics (such as task completion rate, computational load, and accuracy error) and resource consumption (such as instantaneous power consumption, memory usage, and network I / O) of each executing device locally. These data directly reflect the direct and local impact of the decision on individual devices.

[0093] Step S102: Collect actual interaction data of communication links between devices in the distributed system. The actual interaction data includes communication latency and link reliability indicators.

[0094] In this embodiment, the system needs to measure the actual interaction data of the communication link between devices, with the core being communication latency (the delay from sending an instruction or data to acknowledging receipt) and link reliability (such as packet loss rate and signal strength). These indicators reveal the quality and efficiency of the underlying network communication during the execution of this collaborative task, and they directly affect the overall smoothness and timeliness of the collaboration.

[0095] Step S103: Based on the collected real-time operating status data and the actual interaction data, calculate and fuse them according to the preset global optimization target to generate system-level performance feedback as the basis for subsequent optimization.

[0096] Understandably, after collecting the real-time operational status data and the actual interaction data, the system does not simply report the raw data. Instead, it needs to perform calculations and integration according to preset global optimization goals (such as minimum total energy consumption, shortest overall task completion time, and highest system stability). This is usually achieved through a predefined performance function, which weights, aggregates, or normalizes all the collected data, and finally outputs one or a set of scalars or vectors representing the overall performance of this collaborative action, i.e., the system-level performance feedback.

[0097] This step upgrades traditional outcome feedback, which may focus only on a single objective (such as whether a task is completed), into a comprehensive evaluation system that integrates individual status, group interaction, and global goals. The feedback signals generated in this way not only let the model know whether it performed well or poorly, but also suggest why it did or didn't, thus providing a much richer and more instructive basis for subsequent precise dynamic optimization than a single reward signal. It ensures that the optimization process not only pursues apparent success but also delves into optimizing the system's intrinsic health and resource efficiency.

[0098] Step S40: Based on the performance feedback and preset system-level resource constraints, dynamically optimize the collaborative decision-making model to continuously adapt to the topology and state changes of the distributed system.

[0099] Understandably, this step is the core optimization loop driving the entire collaborative control system to achieve continuous self-evolution and lifelong learning. This step uses collected system-level performance feedback and preset system-level resource constraints as input to dynamically optimize the collaborative decision-making model. Its technical essence lies in the fact that it does not perform a one-time parameter tuning, but rather establishes a continuous, goal-oriented online learning mechanism for the model. The optimization process calculates the direction the model should adjust using specific algorithms (such as policy gradient methods), enabling the model to simultaneously pursue multiple key objectives in subsequent decisions: first, maximizing the immediate task performance (such as efficiency and success rate) represented by system-level performance feedback; second, strictly meeting resource constraints (such as power consumption and bandwidth limits) to prevent excessive consumption; and third, improving the long-term robustness of decisions by incorporating mechanisms such as long-term risk assessment. This step allows the system not only to adapt to immediate topology and state changes, but also to abstract experience and optimize strategies from historical interaction data, ensuring the sustainability and growth of intelligent collaborative capabilities in complex dynamic environments.

[0100] like Figure 12 As shown, the step of dynamically optimizing the collaborative decision-making model based on the performance feedback and preset system-level resource constraints specifically includes: Step S111: Dynamically update the power consumption constraint parameters in the system-level resource constraints based on the instruction execution data in the performance feedback.

[0101] In this embodiment, the core is to transform static, preset "system-level resource constraints" (such as "total power consumption not exceeding 1000W") into dynamic, adjustable internal parameters of the model. It utilizes instruction execution data from performance feedback to adjust these constraint parameters in real time. This allows the constraints to flexibly scale according to the actual system load: when collaborative tasks are heavy and instruction intensity is high, constraints can be appropriately relaxed to avoid system stagnation; when the load is light, constraints automatically tighten to improve energy efficiency. This resolves the contradiction between hard constraints and dynamic task requirements during the optimization process.

[0102] Step S112: Based on the updated resource constraints, the policy gradient is weighted according to the degree of network nodes to generate a policy update direction that adapts to different network densities.

[0103] Understandably, this step addresses the "overfitting" problem that causes model performance degradation in new network topologies. Its core operation involves weighting the policy gradient based on the degree of network nodes. During training, different weights or learning rates are assigned to the policy gradients of high-degree nodes (network hubs) and low-degree nodes (edge ​​devices). This essentially introduces prior knowledge of the network structure into the optimization process, guiding the model to learn collaborative policies that differentiate node roles, thereby significantly improving the generalization and transfer capabilities of the trained policy across networks of different sizes and connection densities.

[0104] Step S113: Calculate the long-term reward based on the system state prediction data in the performance feedback. The long-term reward is used to enhance the optimization objective's consideration of long-term risks.

[0105] Understandably, this embodiment breaks through the limitations of traditional reinforcement learning, which mostly optimizes immediate rewards, and aims to prevent future systemic risks. Based on the model's ability to predict system states, it calculates a long-term reward signal (e.g., a long-term reward signal can be quantified as a negative estimate of the risk of cascading failures occurring within a certain future period). During optimization, this long-term reward is incorporated into the objective function along with the immediate reward. This forces the model to not only pursue the immediate benefits of the current action but also to weigh the potential impact of the current decision on the long-term stability of the system, thereby making a more forward-looking and proactively avoidable catastrophic state-based collaborative strategy.

[0106] Step S114: Combine the strategy update direction adapted to different network densities with the enhancement optimization objective that incorporates long-term risk considerations to drive the parameter update of the collaborative decision-making model and realize the dynamic optimization of the collaborative decision-making model.

[0107] In this embodiment, this step combines the outputs generated in the first three steps—namely, the policy update directions adapted to different network densities and the enhancement optimization objective incorporating long-term risk considerations—to jointly calculate a final comprehensive policy gradient. This gradient is then used to drive the parameter updates of the collaborative decision-making model. In this way, each iteration of the model is simultaneously guided by three signals: immediate resource constraints, structural generalization guidance, and long-term risk warning, thereby achieving efficient and balanced collaborative evolution in a complex objective space.

[0108] Furthermore, the device collaborative control method based on distributed systems is deployed on industrial IoT edge computing nodes for real-time collaborative control of physical devices in smart energy networks, smart manufacturing production lines, or distributed communication networks.

[0109] Understandably, this content firmly anchors the methods described above to specific application areas with significant industrial value, thereby clearly defining its scope of protection.

[0110] Specifically, this method can be directly deployed on edge computing nodes in the Industrial Internet of Things (IIoT) to achieve real-time collaborative control of physical devices in smart energy networks, intelligent manufacturing production lines, and distributed communication networks. In smart energy networks, by sensing real-time changes in grid topology and quantifying the implicit impacts between non-adjacent nodes, the method can dynamically adjust loads and quickly isolate high-risk nodes, effectively predicting and preventing the cross-regional propagation of cascading faults. In intelligent manufacturing production lines, the system dynamically abstracts heterogeneous devices such as robotic arms, automated guided vehicles (AGVs), and sensors into network nodes. When production line configurations change or equipment states change, the method can adaptively replan collaborative tasks between devices and issue control commands through reliable communication strategies, ensuring the continuity and efficiency of flexible production. In distributed communication networks (such as 5G edge computing clusters), the method models the complex dependencies between nodes such as base stations and user equipment to achieve precise and dynamic allocation of wireless resources and computing tasks, thereby optimizing overall network throughput and quality of service. These applications demonstrate the core value of this method in addressing dynamic topologies, achieving precise collaboration, and improving overall system performance.

[0111] Furthermore, such as Figure 13 As shown, based on the above-described device collaborative control method based on a distributed system, the present invention also provides a device collaborative control system based on a distributed system, wherein the device collaborative control system based on a distributed system includes: The perception construction module 50 is used to acquire the operating status data and communication link information of multiple physical devices in real time, map each physical device to a node in a dynamic topology graph, and construct a feature representation of the distributed system based on the operating status data and the communication link information. The adaptive decision module 60 is used to input the feature representation into the trained collaborative decision model, perform topology adaptive processing and model long-range dependencies between devices through the collaborative decision model, and output collaborative decision instructions corresponding to each physical device. The execution and feedback module 70 is used to distribute the collaborative decision-making instructions to the corresponding physical devices for execution, and to collect system-level performance feedback generated when the collaborative decision-making instructions are executed; The dynamic optimization module 80 is used to dynamically optimize the collaborative decision-making model based on the performance feedback and preset system-level resource constraints, so as to continuously adapt to the topology and state changes of the distributed system.

[0112] like Figure 14 As shown, in the device collaborative control system based on a distributed system in this embodiment of the invention, the perception construction module 50 includes: a data perception unit 501 and a feature construction unit 502; The data sensing unit 501 is used to read the operating status data of multiple physical devices in real time by parsing the communication messages of the industrial fieldbus protocol or industrial Ethernet protocol, map each addressable physical device as a network node in the dynamic topology diagram, assign a unique node identifier to each physical device, measure or define the communication links between nodes based on the actual communication connection between devices or the preset network architecture, and assign a weight value representing the connection strength or communication quality to each link to generate communication link information. The feature construction unit 502 is used to organize the operating status data of each physical device into a node state vector, aggregate the state vectors of all nodes into a spatiotemporal feature matrix that represents the overall real-time status of the system, generate a dynamic weighted adjacency matrix based on the communication link information, wherein the matrix elements of the weighted adjacency matrix represent the connection relationship and communication attributes between the corresponding physical device nodes, and fuse and encode the weighted adjacency matrix with the spatiotemporal feature matrix to form the feature representation of the distributed system.

[0113] like Figure 15 As shown, in the device collaborative control system based on a distributed system in this embodiment of the invention, the adaptive decision module 60 includes a topology adaptive adjustment unit 601 and a long-range dependency modeling unit 602. The topology adaptive adjustment unit 601 is used to detect the topological structure changes reflected by the feature representation through the graph attention-Transformer coupling engine, and perform topology adaptive adjustment on the feature representation according to the topological structure changes to obtain the adaptively processed feature representation. The long-range dependency modeling unit 602 is used to model the long-range dependencies between devices based on the adaptively processed feature representation, and output a set of collaborative decision instructions corresponding to each physical device based on the long-range dependencies and system-level resource constraints.

[0114] like Figure 16 As shown, in the device collaborative control system based on a distributed system in this embodiment of the invention, the execution and feedback module 70 includes an instruction execution unit 701 and an efficiency feedback unit 702. The instruction execution unit 701 is used to obtain the current network conditions. If the current network conditions are a reliable network, the instruction is uniformly distributed to the corresponding physical device for execution using a broadcast matrix. If the current network conditions are a communication-restricted or unreliable network, the instruction is forwarded to the corresponding physical device for execution using a tree-like relay transmission strategy. The performance feedback unit 702 is used to collect real-time operating status data from each physical device after executing the collaborative decision-making instruction. The data includes the device's own performance indicators and resource consumption. It also collects actual interaction data of the communication links between devices in the distributed system. The actual interaction data includes communication latency and link reliability indicators. Based on the collected real-time operating status data and the actual interaction data, it calculates and integrates the data according to a preset global optimization goal to generate system-level performance feedback as the basis for subsequent optimization.

[0115] like Figure 17 As shown, in the device collaborative control system based on a distributed system in this embodiment of the invention, the dynamic optimization module 80 includes a resource constraint update unit 801, a strategy update unit 802, a long-range reward calculation unit 803, and a dynamic optimization unit 804. The resource constraint update unit 801 is used to dynamically update the power consumption constraint parameters in the system-level resource constraints based on the instruction execution data in the performance feedback. The policy update unit 802 is used to generate a policy update direction that adapts to different network densities by weighting the policy gradient according to the network node degree based on the updated resource constraints. The long-term reward calculation unit 803 is used to calculate long-term rewards based on the system state prediction data in the performance feedback. The long-term rewards are used to enhance the optimization target's consideration of long-term risks. The dynamic optimization unit 804 is used to combine the strategy update direction adapted to different network densities with the enhancement optimization objective that incorporates long-term risk considerations, thereby driving the parameter update of the collaborative decision-making model and realizing the dynamic optimization of the collaborative decision-making model.

[0116] The core advantage of the method proposed in this invention lies in achieving precise, adaptive, and continuous optimization of multi-device collaborative control in dynamically changing physical networks. Specifically: (1) This invention realizes adaptive collaborative decision-making for dynamic topology: Traditional methods are prone to failure when facing device online / offline or link quality fluctuations. This method, by sensing and quantifying changes in network connectivity in real time, enables the decision model to dynamically adjust its information aggregation path and decision basis, thereby maintaining the effectiveness and stability of collaborative control without manual intervention or retraining when the network structure changes.

[0117] (2) This invention overcomes the challenge of long-range dependency modeling and enhances the global perspective and predictability of decision-making: Unlike traditional solutions that only focus on direct device interactions, this method can proactively analyze and quantify the indirect and multi-hop impact relationships between any two reachable nodes in the network. This enables the system not only to optimize the current task but also to predict and avoid potential cascading risks (such as fault propagation in the power grid and congestion and collapse in the communication network), thereby significantly enhancing the overall robustness of complex systems.

[0118] (3) This invention achieves continuous self-optimization under multi-objective collaboration: the system integrates multiple objectives such as resource consumption, network generalization ability, and long-term risks into the optimization process through a closed-loop learning mechanism. This enables the collaborative strategy to not only pursue immediate performance optimization, but also automatically balance efficiency, energy consumption, and long-term reliability, and adapt to different network sizes and densities, ultimately achieving autonomous evolution and continuous improvement in the system during operation.

[0119] Furthermore, such as Figure 18 As shown, based on the above-mentioned device collaborative control system and device collaborative control method based on distributed system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 18 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0120] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a distributed system-based device collaborative control program 40, which can be executed by the processor 10 to implement the distributed system-based device collaborative control method of this application.

[0121] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the device collaborative control method based on a distributed system.

[0122] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminal's processor 10, memory 20, and display 30 communicate with each other via a system bus.

[0123] In one embodiment, when the processor 10 executes the device coordination control program 40 based on a distributed system in the memory 20, the following steps are performed: The system acquires real-time operational status data and communication link information of multiple physical devices, maps each physical device to a node in a dynamic topology graph, and constructs a feature representation of the distributed system based on the operational status data and the communication link information. The feature representation is input into the trained collaborative decision-making model, and the collaborative decision-making model is used to perform topology adaptation processing and modeling of long-range dependencies between devices, and outputs collaborative decision-making instructions corresponding to each physical device. The collaborative decision-making instructions are distributed to the corresponding physical devices for execution, and system-level performance feedback generated during the execution of the collaborative decision-making instructions is collected. Based on the performance feedback and preset system-level resource constraints, the collaborative decision-making model is dynamically optimized to continuously adapt to the topology and state changes of the distributed system.

[0124] Specifically, the real-time acquisition of operational status data and communication link information of multiple physical devices includes: By parsing communication messages of industrial fieldbus protocols or industrial Ethernet protocols, the operating status data of multiple physical devices can be read in real time. Each addressable physical device is mapped to a network node in a dynamic topology graph, and each physical device is assigned a unique node identifier. Based on the actual communication connections between devices or the preset network architecture, the communication links between nodes are measured or defined, and each link is assigned a weight value that characterizes the connection strength or communication quality to generate communication link information.

[0125] Specifically, mapping each physical device to a node in a dynamic topology graph and constructing a feature representation of the distributed system based on the operational status data and the communication link information includes: The operational status data of each physical device is organized into node status vectors, and the status vectors of all nodes are aggregated into a spatiotemporal feature matrix that represents the overall real-time status of the system. Based on the communication link information, a dynamic weighted adjacency matrix is ​​generated, and the matrix elements of the weighted adjacency matrix represent the connection relationship and communication attributes between the corresponding physical device nodes. The weighted adjacency matrix and the spatiotemporal feature matrix are fused and encoded to form the feature representation of the distributed system.

[0126] The collaborative decision-making model is a graph attention-transformer coupled engine. The process of performing topology adaptation processing and modeling long-range dependencies between devices through the collaborative decision-making model, and outputting collaborative decision-making instructions corresponding to each physical device, specifically includes: The graph attention-transformer coupling engine detects the topological changes reflected in the feature representation, and performs topological adaptive adjustment on the feature representation based on the topological changes to obtain the adaptively processed feature representation. Based on the adaptively processed feature representation, the long-range dependencies between devices are modeled, and based on the long-range dependencies and system-level resource constraints, a set of collaborative decision-making instructions corresponding to each physical device is output.

[0127] Specifically, detecting the topological changes reflected in the feature representation and performing topological adaptive adjustment on the feature representation based on the topological changes to obtain an adaptively processed feature representation includes: By verifying topological invariance, the fingerprint features of the current dynamic topology are calculated. When the change of the fingerprint features from the historical state exceeds a preset threshold, it is determined that the topology has changed. When the topology changes, dynamic neighbor extraction and graph structure encoding are performed on the feature representation to obtain an adaptively processed feature representation.

[0128] Specifically, performing dynamic neighbor extraction and graph structure encoding on the feature representation includes: Based on a preset or dynamically calculated number of hops, extract the corresponding k-hop neighbor nodes for each node in the dynamic topology graph, and construct a dynamic neighbor set for each node based on the k-hop neighbor nodes. The embedding representation of each node is updated based on its own features and the set features of its corresponding dynamic neighbor set by using a graph neural network.

[0129] When performing dynamic neighbor extraction, the hop value used to constrain the neighbor range is dynamically determined based on the degree of the node. For nodes whose degree exceeds the first preset threshold, a hop value less than the second preset threshold is automatically used to limit the node's perception and communication range.

[0130] Specifically, modeling the long-range dependencies between devices based on the adaptively processed feature representation includes: The node embeddings in the adaptively processed feature representation are fused and normalized to generate an enhanced node feature vector. The enhanced node feature vector is input into the multi-head graph attention layer, and an attention mask based on real-time topology connection is introduced to calculate and capture the long-range dependency strength between any two reachable device nodes in the network, and output the context features representing the long-range dependency relationship. The contextual features are mapped to a value estimate representing the global state of the entire distributed system through a feedforward neural network, and the value estimate is used to assist in generating collaborative decision-making instructions.

[0131] Before outputting the collaborative decision-making instructions corresponding to each physical device, the method further includes: The dedicated computing unit of the collaborative decision-making model performs counterfactual advantage calculation based on the global state value estimate generated by the collaborative decision-making model, the probability distribution of the action strategies of each device, and the estimated immediate reward, to obtain the counterfactual advantage value of each device. The change in distributed system performance of each device is calculated based on the counterfactual advantage value. The contribution of each device is calculated based on the change in distributed system performance. Based on all the contribution values, the parameters of the policy network of the collaborative decision-making model are updated to optimize subsequent decisions.

[0132] Specifically, distributing the collaborative decision-making instructions to the corresponding physical devices for execution includes: Obtain the current network conditions. If the current network conditions are a reliable network, then use a broadcast matrix to uniformly distribute the instructions to the corresponding physical devices for execution. If the current network conditions are a communication-restricted or unreliable network, a tree-like relay transmission strategy is adopted to forward the instructions to the corresponding physical devices for execution in a hierarchical manner.

[0133] Specifically, the collection of system-level performance feedback generated during the execution of the collaborative decision-making instructions includes: Real-time operational status data is collected from each physical device after the execution of the collaborative decision-making instructions. The data includes the device's own performance indicators and resource consumption. Collect actual interaction data of communication links between devices in the distributed system, including communication latency and link reliability indicators; Based on the collected real-time operating status data and the actual interaction data, calculations are performed and fused according to the preset global optimization goals to generate system-level performance feedback as the basis for subsequent optimization.

[0134] The step of dynamically optimizing the collaborative decision-making model based on the performance feedback and preset system-level resource constraints specifically includes: Based on the instruction execution data in the performance feedback, the power consumption constraint parameters in the system-level resource constraints are dynamically updated; Based on the updated resource constraints, the policy gradient is weighted according to the degree of network nodes to generate a policy update direction that adapts to different network densities. Based on the system state prediction data in the performance feedback, a long-term reward is calculated, which is used to enhance the optimization objective's consideration of long-term risks. By combining the strategy update direction adapted to different network densities with the enhancement optimization objective incorporating long-term risk considerations, the parameter update of the collaborative decision-making model is driven, thereby achieving dynamic optimization of the collaborative decision-making model.

[0135] The device collaborative control method based on distributed systems is deployed on industrial IoT edge computing nodes and is used to perform real-time collaborative control of physical devices in smart energy networks, smart manufacturing production lines or distributed communication networks.

[0136] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a device cooperative control program based on a distributed system, and the device cooperative control program based on a distributed system, when executed by a processor, implements the steps of the device cooperative control method based on a distributed system as described above.

[0137] In summary, this invention provides a device collaborative control method, system, terminal, and storage medium based on a distributed system. The method includes: acquiring real-time operating status data and communication link information of multiple physical devices; mapping each physical device to a node in a dynamic topology graph; and constructing a feature representation of the distributed system based on the operating status data and the communication link information; inputting the feature representation into a trained collaborative decision-making model; performing topology adaptation processing and modeling long-range dependencies between devices through the collaborative decision-making model; and outputting collaborative decision-making instructions corresponding to each physical device; distributing the collaborative decision-making instructions to the corresponding physical devices for execution; and collecting system-level performance feedback generated during the execution of the collaborative decision-making instructions; and dynamically optimizing the collaborative decision-making model based on the performance feedback and preset system-level resource constraints to continuously adapt to changes in the topology and state of the distributed system. This invention achieves adaptation to dynamically changing network topologies, enabling accurate collaborative decision-making and efficient resource control.

[0138] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0139] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0140] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A device collaborative control method based on a distributed system, characterized in that, The device collaborative control method based on a distributed system includes: The system acquires real-time operational status data and communication link information of multiple physical devices, maps each physical device to a node in a dynamic topology graph, and constructs a feature representation of the distributed system based on the operational status data and the communication link information. The feature representation is input into the trained collaborative decision-making model, and the collaborative decision-making model is used to perform topology adaptation processing and modeling of long-range dependencies between devices, and outputs collaborative decision-making instructions corresponding to each physical device. The collaborative decision-making instructions are distributed to the corresponding physical devices for execution, and system-level performance feedback generated during the execution of the collaborative decision-making instructions is collected. Based on the performance feedback and preset system-level resource constraints, the collaborative decision-making model is dynamically optimized to continuously adapt to the topology and state changes of the distributed system.

2. The device collaborative control method based on a distributed system according to claim 1, characterized in that, The real-time acquisition of operational status data and communication link information of multiple physical devices specifically includes: By parsing communication messages of industrial fieldbus protocols or industrial Ethernet protocols, the operating status data of multiple physical devices can be read in real time. Each addressable physical device is mapped to a network node in a dynamic topology graph, and each physical device is assigned a unique node identifier. Based on the actual communication connections between devices or the preset network architecture, the communication links between nodes are measured or defined, and each link is assigned a weight value that characterizes the connection strength or communication quality to generate communication link information.

3. The device collaborative control method based on a distributed system according to claim 1, characterized in that, The process of mapping each physical device to a node in a dynamic topology graph and constructing a feature representation of the distributed system based on the operational status data and the communication link information specifically includes: The operational status data of each physical device is organized into node status vectors, and the status vectors of all nodes are aggregated into a spatiotemporal feature matrix that represents the overall real-time status of the system. Based on the communication link information, a dynamic weighted adjacency matrix is ​​generated, and the matrix elements of the weighted adjacency matrix represent the connection relationship and communication attributes between the corresponding physical device nodes. The weighted adjacency matrix and the spatiotemporal feature matrix are fused and encoded to form the feature representation of the distributed system.

4. The device collaborative control method based on a distributed system according to claim 1, characterized in that, The collaborative decision-making model is a graph attention-Transformer coupled engine; The process of performing topology adaptation processing and modeling long-range dependencies between devices through the collaborative decision-making model, and outputting collaborative decision-making instructions corresponding to each physical device, specifically includes: The graph attention-transformer coupling engine detects the topological changes reflected in the feature representation, and performs topological adaptive adjustment on the feature representation based on the topological changes to obtain the adaptively processed feature representation. Based on the adaptively processed feature representation, the long-range dependencies between devices are modeled, and based on the long-range dependencies and system-level resource constraints, a set of collaborative decision-making instructions corresponding to each physical device is output.

5. The device collaborative control method based on a distributed system according to claim 4, characterized in that, The process of detecting the topological changes reflected in the feature representation, and then performing topological adaptive adjustment on the feature representation based on these changes to obtain an adaptively processed feature representation, specifically includes: By verifying topological invariance, the fingerprint features of the current dynamic topology are calculated. When the change of the fingerprint features from the historical state exceeds a preset threshold, it is determined that the topology has changed. When the topology changes, dynamic neighbor extraction and graph structure encoding are performed on the feature representation to obtain an adaptively processed feature representation.

6. The device cooperative control method based on a distributed system according to claim 5, characterized in that, The process of performing dynamic neighbor extraction and graph structure encoding on the feature representation specifically includes: Based on a preset or dynamically calculated number of hops, extract the corresponding k-hop neighbor nodes for each node in the dynamic topology graph, and construct a dynamic neighbor set for each node based on the k-hop neighbor nodes. The embedding representation of each node is updated based on its own features and the set features of its corresponding dynamic neighbor set by using a graph neural network.

7. The device cooperative control method based on a distributed system according to claim 6, characterized in that, When performing dynamic neighbor extraction, the hop value used to constrain the neighbor range is dynamically determined based on the degree of the node. For nodes whose degree exceeds the first preset threshold, a hop value less than the second preset threshold is automatically used to limit the node's perception and communication range.

8. The device collaborative control method based on a distributed system according to claim 4, characterized in that, The step of modeling long-range dependencies between devices based on the adaptively processed feature representation specifically includes: The node embeddings in the adaptively processed feature representation are fused and normalized to generate an enhanced node feature vector. The enhanced node feature vector is input into the multi-head graph attention layer, and an attention mask based on real-time topology connection is introduced to calculate and capture the long-range dependency strength between any two reachable device nodes in the network, and output the context features representing the long-range dependency relationship. The contextual features are mapped to a value estimate representing the global state of the entire distributed system through a feedforward neural network, and the value estimate is used to assist in generating collaborative decision-making instructions.

9. The device cooperative control method based on a distributed system according to claim 1, characterized in that, Before outputting the collaborative decision-making instructions corresponding to each physical device, the following is also included: The dedicated computing unit of the collaborative decision-making model performs counterfactual advantage calculation based on the global state value estimate generated by the collaborative decision-making model, the probability distribution of the action strategies of each device, and the estimated immediate reward, to obtain the counterfactual advantage value of each device. The change in distributed system performance of each device is calculated based on the counterfactual advantage value. The contribution of each device is calculated based on the change in distributed system performance. Based on all the contribution values, the parameters of the policy network of the collaborative decision-making model are updated to optimize subsequent decisions.

10. The device cooperative control method based on a distributed system according to claim 1, characterized in that, The step of distributing the collaborative decision-making instructions to the corresponding physical devices for execution specifically includes: Obtain the current network conditions. If the current network conditions are a reliable network, then use a broadcast matrix to uniformly distribute the instructions to the corresponding physical devices for execution. If the current network conditions are a communication-restricted or unreliable network, a tree-like relay transmission strategy is adopted to forward the instructions to the corresponding physical devices for execution in a hierarchical manner.

11. The device cooperative control method based on a distributed system according to claim 1, characterized in that, The collection of system-level performance feedback generated during the execution of the collaborative decision-making instructions specifically includes: Real-time operational status data after the execution of the collaborative decision-making instructions is collected from each physical device, including the device's own performance indicators and resource consumption. Collect actual interaction data of communication links between devices in the distributed system, including communication latency and link reliability indicators; Based on the collected real-time operating status data and the actual interaction data, calculations are performed and fused according to the preset global optimization goals to generate system-level performance feedback as the basis for subsequent optimization.

12. The device cooperative control method based on a distributed system according to claim 1, characterized in that, The step of dynamically optimizing the collaborative decision-making model based on the performance feedback and preset system-level resource constraints specifically includes: Based on the instruction execution data in the performance feedback, the power consumption constraint parameters in the system-level resource constraints are dynamically updated; Based on the updated resource constraints, the policy gradient is weighted according to the degree of network nodes to generate a policy update direction that adapts to different network densities. Based on the system state prediction data in the performance feedback, a long-term reward is calculated, which is used to enhance the optimization objective's consideration of long-term risks. By combining the strategy update direction adapted to different network densities with the enhancement optimization objective incorporating long-term risk considerations, the parameter update of the collaborative decision-making model is driven, thereby achieving dynamic optimization of the collaborative decision-making model.

13. The device cooperative control method based on a distributed system according to claim 1, characterized in that, The device collaborative control method based on distributed systems is deployed on industrial IoT edge computing nodes and is used to perform real-time collaborative control of physical devices in smart energy networks, smart manufacturing production lines or distributed communication networks.

14. A device collaborative control system based on a distributed system, characterized in that, The distributed system-based device collaborative control system includes: The perception construction module is used to acquire the operating status data and communication link information of multiple physical devices in real time, map each physical device as a node in a dynamic topology graph, and construct a feature representation of the distributed system based on the operating status data and the communication link information. An adaptive decision module is used to input the feature representation into a trained collaborative decision model, perform topology adaptive processing and model long-range dependencies between devices through the collaborative decision model, and output collaborative decision instructions corresponding to each physical device. The execution and feedback module is used to distribute the collaborative decision-making instructions to the corresponding physical devices for execution, and to collect system-level performance feedback generated when the collaborative decision-making instructions are executed. The dynamic optimization module is used to dynamically optimize the collaborative decision-making model based on the performance feedback and preset system-level resource constraints, so as to continuously adapt to the topology and state changes of the distributed system.

15. The device collaborative control system based on a distributed system according to claim 14, characterized in that, The perception construction module includes a data perception unit and a feature construction unit; The data sensing unit is used to read the operating status data of multiple physical devices in real time by parsing the communication messages of the industrial fieldbus protocol or industrial Ethernet protocol, map each addressable physical device as a network node in the dynamic topology, assign a unique node identifier to each physical device, measure or define the communication links between nodes based on the actual communication connection between devices or the preset network architecture, and assign a weight value representing the connection strength or communication quality to each link to generate communication link information. The feature construction unit is used to organize the operating status data of each physical device into a node state vector, aggregate all node state vectors into a spatiotemporal feature matrix that represents the overall real-time status of the system, generate a dynamic weighted adjacency matrix based on the communication link information, wherein the matrix elements of the weighted adjacency matrix represent the connection relationship and communication attributes between the corresponding physical device nodes, and fuse and encode the weighted adjacency matrix with the spatiotemporal feature matrix to form the feature representation of the distributed system.

16. The device collaborative control system based on a distributed system according to claim 14, characterized in that, The adaptive decision-making module includes a topology adaptive adjustment unit and a long-range dependency modeling unit; The topology adaptive adjustment unit is used to detect the topological changes reflected by the feature representation through the graph attention-transformer coupling engine, and to perform topology adaptive adjustment on the feature representation according to the topological changes to obtain the adaptively processed feature representation. The long-range dependency modeling unit is used to model the long-range dependencies between devices based on the adaptively processed feature representation, and output a set of collaborative decision instructions corresponding to each physical device based on the long-range dependencies and system-level resource constraints.

17. The device collaborative control system based on a distributed system according to claim 14, characterized in that, The execution and feedback module includes an instruction execution unit and a performance feedback unit; The instruction execution unit is used to obtain the current network conditions. If the current network conditions are a reliable network, the instruction is uniformly distributed to the corresponding physical device for execution using a broadcast matrix. If the current network conditions are a communication-restricted or unreliable network, the instruction is forwarded to the corresponding physical device for execution using a tree-like relay transmission strategy. The performance feedback unit is used to collect real-time operating status data from each physical device after executing the collaborative decision-making command. The data includes the device's own performance indicators and resource consumption. It also collects actual interaction data of the communication links between devices in the distributed system. The actual interaction data includes communication latency and link reliability indicators. Based on the collected real-time operating status data and the actual interaction data, it calculates and integrates the data according to the preset global optimization target to generate system-level performance feedback as the basis for subsequent optimization.

18. The device collaborative control system based on a distributed system according to claim 14, characterized in that, The dynamic optimization module includes a resource constraint update unit, a strategy update unit, a long-range reward calculation unit, and a dynamic optimization unit. The resource constraint update unit is used to dynamically update the power consumption constraint parameters in the system-level resource constraints based on the instruction execution data in the performance feedback; The policy update unit is used to generate policy update directions that adapt to different network densities by weighting the policy gradient according to the network node degree based on the updated resource constraints. The long-term reward calculation unit is used to calculate long-term rewards based on the system state prediction data in the performance feedback. The long-term rewards are used to enhance the optimization target's consideration of long-term risks. The dynamic optimization unit is used to combine the strategy update direction adapted to different network densities with the enhancement optimization objective that incorporates long-term risk considerations, thereby driving the parameter update of the collaborative decision-making model and realizing the dynamic optimization of the collaborative decision-making model.

19. A terminal, characterized in that, The terminal includes: a memory, a processor, and a distributed system-based device collaborative control program stored in the memory and executable on the processor. When the distributed system-based device collaborative control program is executed by the processor, it implements the steps of the distributed system-based device collaborative control method as described in any one of claims 1-13.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a device cooperative control program based on a distributed system, which, when executed by a processor, implements the steps of the device cooperative control method based on a distributed system as described in any one of claims 1-13.

Images