A method and system for generating linkage strategies for edge-distributed real-time platforms

Abstract

This application relates to the technical field of industrial internet. A method for generating a linkage strategy for an edge-distributed real-time platform includes: acquiring time-series signal data from edge device sensors and structured data from a production scenario to obtain time-series signal data feature vectors and structured data feature vectors; obtaining graph-level global features based on the time-series signal data feature vectors and structured data feature vectors; inputting the graph-level global features into a pre-trained target MDP decision model for inference, combining a hierarchical distributed decision architecture with a meta-learning adaptation method for collaborative scheduling optimization, generating multiple cross-production line collaborative execution strategies; performing resource modeling and real-time status synchronization based on the hardware and operational data of edge devices and cloud servers to obtain a unified cloud-edge resource view; and selecting the optimal solution from multiple cross-production line collaborative execution strategies to obtain the target cross-production line collaborative execution strategy.

Description

Technical Field

[0001] This application relates to the technical field of industrial internet, specifically to a method and system for generating linkage strategies for edge distributed real-time platforms. Background Technology

[0002] In the development of intelligent manufacturing and industrial internet technologies, cloud-edge collaborative architecture has become an important technical direction for industrial production resource management and equipment operation management. Industrial production scenarios have put forward higher practical requirements for the real-time, collaborative and intelligent levels of equipment linkage control, cloud-edge resource scheduling, and so on. There is an urgent need for an integrated management and control solution that adapts to the dynamic working conditions of industrial production in order to ensure production continuity and achieve optimal allocation of production resources.

[0003] In the current industrial production field, the control methods for equipment linkage and cloud-edge collaborative task scheduling are difficult to adapt efficiently to dynamically changing production conditions. They also lack a comprehensive design that takes into account multiple dimensions of production operation goals, and the collaborative decision-making capabilities across production lines and cloud-edge devices are insufficient, failing to fully match the dynamic changes in equipment status, production needs, and resource conditions in industrial production scenarios.

[0004] Existing technologies have poor connectivity in areas such as state perception of industrial equipment and cloud-edge resources, generation of linkage strategies, and optimization of scheduling schemes. The flexibility and adaptability of the control mode are insufficient, which makes it difficult to fully realize the overall effectiveness of the cloud-edge collaborative architecture. It cannot meet the actual needs of modern industrial production for efficient linkage and scheduling, and it also restricts the improvement of production efficiency and optimization of resource utilization efficiency, becoming a major obstacle to the implementation and application of industrial cloud-edge collaborative technology. Summary of the Invention

[0005] The main purpose of this application is to provide a method and system for generating linkage strategies for edge distributed real-time platforms, so as to solve the technical problems in the background art.

[0006] To achieve the above objectives, the first aspect of this application proposes a method for generating linkage strategies for edge-distributed real-time platforms, comprising: Acquire time-series signal data from edge device sensors and structured data from the production scenario to obtain time-series signal data feature vectors and structured data feature vectors. The time-series signals include time-series signal data of temperature, vibration, and current of the edge devices, and the structured data includes structured data of production work orders and environmental parameters. Based on the time-series signal data feature vector and the structured data feature vector, graph structure modeling is performed to form a dynamic production graph, which is then encoded through a time-series graph neural network to obtain graph-level global features. The graph-level global features are input into the pre-trained target MDP decision model for inference. The hierarchical distributed decision architecture and meta-learning adaptation method are combined for collaborative scheduling optimization to generate multiple cross-production line collaborative execution strategies. Based on the hardware and operational data of the edge devices and cloud servers, resource modeling and real-time status synchronization are performed to obtain a unified view of cloud-edge resources. Based on the unified view data of cloud-edge resources, the optimal solution is selected from multiple cross-production line collaborative execution strategies to obtain the target cross-production line collaborative execution strategy.

[0007] In some feasible methods, the step of acquiring time-series signal data from edge device sensors and structured data from the production scenario to obtain time-series signal data feature vectors and structured data feature vectors includes: The temperature, vibration, and current time-series signal data of the edge device sensor are used to extract local features using a Conv1D-CNN layer to obtain local features of the time-series signal. For the local features of the time series signal, long-range dependencies are captured by a BiLSTM network to obtain the feature vector of the time series signal data; The structured data of the production scenario is encoded sequentially through an embedding layer and a fully connected layer to obtain the structured data feature vector.

[0008] In some feasible methods, the step of performing graph structure modeling processing based on the time-series signal data feature vector and the structured data feature vector to form a dynamically generated graph, and encoding it through a time-series graph neural network to obtain graph-level global features, includes: Using edge devices as edge nodes, material flow, energy flow and information flow between edge devices as edges, and the time-series signal data feature vector as edge node features, combined with the interaction strength calculated by mutual information or Granger causality as edge features, a dynamic production graph is constructed. The dynamic production graph is updated by multi-layer iterative encoding using a temporal graph neural network (TGNN) layer to obtain the encoded features of each edge node; The encoded features of each edge node are aggregated to output the graph-level global features.

[0009] In some feasible approaches, the training steps of the target MDP decision model include: By integrating historical graph-level global features into the state space of a Markov decision process, a joint state space is constructed that includes the real-time state of edge devices, production scenarios, and environmental states. An action space for adjusting linkage rules is defined to form an initial MDP decision model. The historical graph-level global features are formed by using historical time-series signal data feature vectors and historical structured data feature vectors through dynamic production graph construction and time-series graph neural network encoding. The linkage rule adjustment includes enabling / disabling existing linkage rules, modifying existing linkage rule parameters, and generating new linkage rules. A multi-objective reward function is constructed with the optimization objectives of improving production efficiency, optimizing energy consumption, and reducing the lifespan loss of edge devices. The multi-objective reward function serves as the evaluation criterion for training the initial MDP decision model. Based on the initial MDP decision model and the multi-objective reward function, the initial MDP decision model is trained to obtain the converged target MDP decision model.

[0010] In some feasible approaches, the step of generating multiple cross-production line collaborative execution strategies by combining the target MDP decision model with a hierarchical distributed decision architecture and a meta-learning adaptation method for collaborative scheduling optimization includes: The target MDP decision model is optimized online by combining dual Q-learning with security constraint reinforcement learning to obtain an optimized MDP decision model that meets the security threshold of edge devices and ensures production continuity. Based on the MAML meta-learning framework and combined with adversarial training of the domain adaptation layer, the optimized MDP decision model is adapted across production lines to obtain an adapted MDP decision model that is adapted to the characteristics of different production lines. Based on a hierarchical distributed decision architecture of edge devices, production lines, and edge nodes, the adapted MDP decision model is coordinated and scheduled to generate multiple cross-production line collaborative execution strategies. The coordinated scheduling means that the edge nodes coordinate the linkage strategies output by the adapted MDP decision model, decompose them into sub-strategies for each production line, and further decompose them into operation instructions for each edge device. The instructions are issued in a hierarchical manner from edge nodes to production lines and from production lines to edge devices.

[0011] In some feasible methods, the step of performing resource modeling and real-time status synchronization based on the hardware and operational data of the edge device and the cloud server to obtain a unified view of cloud-edge resources includes: Based on the hardware and operational data of the edge devices, construct an edge device resource profile that includes the static attributes and dynamic operational status of the edge devices. Based on the hardware and operational data of the cloud server, a dynamic model of cloud resources, including available computing units, storage capacity, and service status, is constructed. The edge device resource profile and the cloud resource dynamic model are updated in real time to obtain a unified view of cloud and edge resources.

[0012] In some feasible methods, the step of selecting the optimal solution from multiple cross-production line collaborative execution strategies based on the unified cloud-edge resource view data to obtain the target cross-production line collaborative execution strategy includes: Based on the unified view data of cloud and edge resources, the resource demand characteristics of multiple cross-production line collaborative execution strategies and the computing power, load, energy consumption and reliability characteristics of cloud and edge resources are abstractly modeled to establish a strategy and resource matching constraint model. Using the strategy and resource matching constraint model as constraints, a multi-objective optimization evaluation of delay, cost, energy consumption, balance and reliability is performed on multiple cross-production line collaborative execution strategies. The cross-production line collaborative execution strategies are sorted according to Pareto dominance relationship. Dominated cross-production line collaborative execution strategies are eliminated. The optimal solution set composed of Pareto optimal strategies is obtained from multiple cross-production line collaborative execution strategies. Based on the real-time operating status of the current production system and actual operational needs, an adaptive decision-making process is performed on the optimal solution set to obtain the target cross-production line collaborative execution strategy.

[0013] Secondly, this application provides an edge distributed real-time platform linkage strategy generation system, applied to the aforementioned edge distributed real-time platform linkage strategy generation method, the system comprising: The acquisition unit is used to acquire time-series signal data from edge device sensors and structured data from the production scene to obtain time-series signal data feature vectors and structured data feature vectors. The time-series signals include time-series signal data of temperature, vibration and current of the edge device, and the structured data includes structured data of production work orders and environmental parameters. The feature processing unit is used to perform graph structure modeling processing based on the feature vector of the time-series signal data and the feature vector of the structured data to form a dynamic production graph, and to encode it through a time-series graph neural network to obtain graph-level global features; The optimization unit inputs the graph-level global features into the pre-trained target MDP decision model for inference, and performs collaborative scheduling optimization by combining the hierarchical distributed decision architecture and meta-learning adaptation method to generate multiple cross-production line collaborative execution strategies. The resource processing unit is used to perform resource modeling and real-time status synchronization based on the hardware and operating data of the edge device and the cloud server to obtain unified cloud-edge resource view data. The result unit is used to select the optimal solution from multiple cross-production line collaborative execution strategies based on the unified view data of cloud-edge resources, and obtain the target cross-production line collaborative execution strategy.

[0014] Thirdly, this application provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned method.

[0015] Fourthly, this application provides a computer program that, when executed by a processor, implements the steps of the aforementioned method.

[0016] The technical solutions provided by the embodiments of this application may include the following beneficial effects: This application presents a method for generating linkage strategies for an edge-distributed real-time platform. By extracting and fusing features from time-series signals from edge device sensors and structured data from the production scenario, and combining this with dynamic production graph construction and time-series graph neural network encoding, graph-level global features are obtained. This achieves close integration of industrial equipment with cloud-edge resource status perception, linkage strategy generation, and scheduling scheme optimization, improving the flexibility and adaptability of the control mode. The graph-level global features are then input into a pre-trained target MDP decision model for inference. Combined with a hierarchical distributed decision architecture and meta-learning adaptation methods, cross-production line collaborative execution strategies are generated, solving the problems of existing technologies. The problem of insufficient cross-production line and cloud-edge-device collaborative decision-making capabilities has been addressed by enabling linkage strategies to efficiently adapt to the changing characteristics of dynamic industrial production conditions. Simultaneously, by constructing a unified view of cloud-edge resources and selecting the optimal solution from cross-production line collaborative execution strategies, a comprehensive design for multi-dimensional production operation objectives has been achieved. This effectively improves the real-time performance, collaboration, and intelligence of equipment linkage control and cloud-edge resource scheduling in industrial production scenarios, ensuring production continuity and optimizing the allocation of production resources. It fully leverages the overall effectiveness of the cloud-edge collaborative architecture, thereby improving production efficiency and resource utilization efficiency, and overcoming obstacles to the practical application of industrial cloud-edge collaborative technology. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application. In the drawings: Figure 1 A flowchart illustrating a method for generating a linkage strategy for an edge-distributed real-time platform, as provided in this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0021] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in some cases to indicate a certain dependency or connection relationship. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0022] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linked," and "socketing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] The following explanations of some terms used in this application are provided to aid in understanding the technical solution of this application: The Conv1D-CNN layer, a one-dimensional convolution, is an adaptation of convolutional neural networks (CNNs) for one-dimensional temporal data. It uses the time step as the convolution dimension and performs convolution operations with the local window of the temporal signal through a sliding one-dimensional convolution kernel to extract local spatial (temporal) feature patterns. At the same time, it uses pooling layers to reduce dimensionality and retain key features.

[0024] BiLSTM networks, serving as long-range feature capture components in device state feature extractors, inherit the local feature outputs of Conv1D-CNN layers to mine long-term temporal dependencies in device time-series signals. BiLSTM (Bidirectional Long Short-Term Memory) is an extension of LSTM, consisting of a forward LSTM and a backward LSTM. The forward LSTM learns the forward dependencies of temporal data from the start to the end of the time step, while the backward LSTM learns the inverse dependencies from the end to the start of the time step. Finally, the outputs from both directions are concatenated to obtain a feature representation that simultaneously contains "past-future" temporal information.

[0025] The Embedding layer, as a fundamental component of the feature encoder in the production scenario, is responsible for converting discrete structured data in the production scenario into low-dimensional dense numerical feature vectors, solving the problem that discrete data (such as material type and process requirements) cannot directly participate in neural network calculations.

[0026] Fully connected layers (FC layers) are the most basic feature fusion and transformation components in neural networks. Each neuron in the layer is connected to all neurons in the previous layer, and feature integration, dimensionality reduction / upgrading, and standardization are achieved through linear transformation and nonlinear activation.

[0027] Mutual information is an indicator that quantifies the degree of interdependence between two random variables. It can break through the limitations of linear relationships and accurately characterize the close relationship between material flow, energy flow, and information flow between equipment in industrial scenarios.

[0028] Granger causality is a method from a time series perspective to determine whether the historical information of one variable can effectively predict the future value of another variable. It can quantify the causal relationship strength between the temporal states of industrial equipment, thereby determining the causal orientation of interactions between equipment.

[0029] The Temporal Graph Neural Network (TGNN) layer can encode the temporal dimension of a dynamic production graph with equipment as nodes and material / energy / information flow as edges. It is a dedicated graph neural network layer that integrates the features of the node itself and its neighboring nodes through a specific update formula, dynamically captures the changes in the graph over time, and outputs a graph-level feature representation.

[0030] Mean aggregation, a fundamental method for fusing node neighborhood features in graph neural networks, involves taking the arithmetic mean of the features of all neighboring nodes of a target node and using the result as the neighborhood aggregation feature of that node. It is suitable for fusing neighborhood features in industrial production scenarios where there is no obvious correlation between equipment nodes.

[0031] Attention mechanism aggregation, a node neighborhood feature fusion method with differentiated weights in graph neural networks, refers to assigning attention weights matching the association strength to each neighbor node feature of the target node, and obtaining neighborhood aggregation features through weighted summation, which can highlight the neighbor node features with strong interaction with the target equipment in industrial production scenarios.

[0032] Markov decision (Markov Decision Process, MDP) is a sequential decision model based on the Markov property. It is used to solve for the decision strategy that maximizes the cumulative reward in dynamic and uncertain environments by defining the state space, action space, state transition rules and reward function. It is the core modeling method for dynamic optimization of equipment linkage strategies in industrial production.

[0033] Dual Q-learning is an improved algorithm for reinforcement learning. By splitting a single value evaluation network into two independent Q-networks, which perform action selection and action value evaluation respectively, it effectively reduces the value overestimation bias of traditional Q-learning. In industrial production, it is used for dynamic optimization of equipment linkage strategies and improves the strategy optimization accuracy of MDP decision models.

[0034] Safety-constrained reinforcement learning is a variant of reinforcement learning that incorporates safety boundaries. During the iterative process of an agent learning the optimal decision-making strategy, strict safety thresholds are preset for all decision actions and real-time verification is performed to ensure that all learned and executed actions are within the safe operating range of the system. In industrial production, it is used to ensure the safety of equipment linkage operations and avoid production interruptions or equipment failures caused by strategy decisions.

[0035] The MAML meta-learning framework, or model-independent meta-learning framework, is a general meta-learning method that can pre-train a meta-model on a small amount of historical data from multiple similar production lines. For new production lines, only a small number of samples are needed to complete gradient updates, which can quickly adapt to the features of the new production line and achieve rapid model transfer across production lines.

[0036] The domain adaptation layer is a feature adaptation component used in conjunction with the Temporal Graph Neural Network (TGNN) and the MAML meta-learning framework to adapt to cross-production line scenarios in industrial applications. Deployed after the TGNN layer, its core function is to eliminate feature distribution differences between different production lines, improving the adaptability and generalization ability of the meta-model across production lines. A domain discriminator determines the production line domain of the graph-level features output by the TGNN. Then, through adversarial training, the TGNN learns to generate general features that can confuse the domain discriminator, ensuring that the extracted production line features are no longer affected by specific production line scenario differences, and making the feature distribution of different production lines more consistent. Finally, through the processing of the domain adaptation layer, the MAML-pretrained meta-model can quickly adapt to new production lines with only a small number of gradient updates.

[0037] Pareto dominance is a core criterion for evaluating the merits of schemes in the field of multi-objective optimization. It means that if a scheduling scheme performs no worse than another scheme on all optimization objectives (latency, cost, energy consumption, etc.) and performs better on at least one objective, then the scheme is said to Pareto dominate the other scheme. It is the core basis for the NSGA-III algorithm in this scheme to rank cloud-edge collaborative scheduling schemes and select high-quality schemes.

[0038] The NSGA-III algorithm is an improved non-dominated sorting genetic algorithm. It is an evolutionary algorithm for multi-objective optimization problems and is designed specifically to solve high-dimensional multi-objective optimization scenarios. Through population initialization, Pareto dominance sorting and elite retention in iterative evolution, it selects high-quality solutions under multi-objective dimensions and generates the optimal solution set.

[0039] like Figure 1 As shown, in a first aspect, this application provides a method for generating linkage strategies for an edge-distributed real-time platform, comprising: S100 acquires time-series signal data from edge device sensors and structured data from the production scenario, obtaining time-series signal data feature vectors and structured data feature vectors.

[0040] The timing signals include timing signal data of temperature, vibration and current of edge devices, and the structured data includes structured data of production work orders and environmental parameters.

[0041] Temperature, vibration, and current sensors are deployed on each edge device. Each sensor has a preset acquisition frequency (e.g., 10Hz for high-frequency operation and 5Hz for normal operation) based on the actual needs of the industrial production scenario. These sensors continuously collect temperature, vibration, and current signals during edge device operation, obtaining raw time-series signal data. Simultaneously, structured raw data from the existing industrial production management system and on-site environmental monitoring module are acquired. This includes production work order data containing information such as material type, process requirements, and delivery deadlines, and environmental parameter data including temperature, humidity, and air pressure. The sensors used can be existing sensors in the industrial production scenario, or additional sensors of corresponding types can be added as needed for actual monitoring.

[0042] Next, the acquired raw time-series signal data is preprocessed: random noise in the signal is removed using the moving average method, and min-max normalization is used to map the values ​​of all signals to the [0,1] interval to eliminate the dimensional differences between signals from different sensors. The structured raw data is then preprocessed: missing values ​​of discrete features are filled using the mean-filling method, and missing values ​​of continuous features are filled using linear interpolation. Simultaneously, normalization is performed on the continuous structured data to ensure the effectiveness of subsequent feature encoding. After the above preprocessing, standardized time-series signal data and structured data are obtained, respectively.

[0043] Specifically, obtaining the feature vectors of time-series signal data and structured data may include the following steps: S101, local feature extraction is performed on the temperature, vibration and current time-series signal data of the edge device sensor using a Conv1D-CNN layer to obtain local features of the time-series signal.

[0044] Specifically, the standardized temperature, vibration, and current time-series signal data are concatenated into three-dimensional input data along the time dimension (dimension: time step × number of sensor types × single sensor signal dimension), and then input into the Conv1D-CNN layer for local feature extraction.

[0045] The Conv1D-CNN layer is based on the local receptive field principle of one-dimensional convolutional kernels. It performs sliding convolution operations on the time dimension of the temporal signal using a pre-defined one-dimensional convolutional kernel. Within each sliding time window, the kernel performs feature-weighted summation on the temperature, vibration, and current signals within that window, capturing the local temporal features of single-type sensor signals (such as signal peaks, valleys, and short-term trends) as well as the local coupling features between multiple types of sensor signals (such as the local correlation between temperature increases and current increases). The feature map output from the convolution operation undergoes non-linear activation processing to enhance its non-linear representation capability. Then, a pooling layer performs dimensionality reduction and feature filtering on the activated feature map, removing redundant local feature information and retaining effective local features that characterize the device's operating state. Finally, the local features of the temporal signal are obtained; these features are a sequence of local features in the temporal dimension, accurately reflecting the changes in the operating state of the edge device over a short period.

[0046] S102, for the local features of the time series signal, long-range dependencies are captured by a BiLSTM network to obtain the feature vector of the time series signal data.

[0047] Specifically, the local features of the time-series signal obtained by S101 are input into the BiLSTM network in chronological order. By utilizing the core principle of bidirectional time-series learning of the BiLSTM network, the long-range dependencies of temperature, vibration, and current time-series signals can be captured.

[0048] The BiLSTM network consists of two sub-networks: a forward LSTM and a backward LSTM. The forward LSTM learns the forward time series dependency of the signal sequentially from the start to the end, capturing the influence of earlier signal changes on later signals during equipment operation. The backward LSTM learns the backward time series dependency of the signal sequentially from the end to the start, capturing the feedback characteristics of later signals on earlier signals and the inverse correlation characteristics of signals. Through bidirectional learning, this BiLSTM network can fully capture the long-range dependency characteristics of time series signals across the entire time dimension, such as the periodic variation characteristics of equipment vibration signals and the slow cumulative variation characteristics of temperature signals.

[0049] The bidirectional long-range features learned by the forward LSTM and backward LSTM are fused and concatenated to obtain a fused time-series feature sequence. This sequence is then input into a fully connected layer, where linear transformations are used to unify the feature dimensions. Meanwhile, nonlinear activation is used to enhance the representational capability of the features. Finally, a time-series signal data feature vector, i.e., an edge device-level feature vector, is output. The value of each dimension corresponds to a feature dimension of the edge device's operating state, which can completely represent the overall operating state of the edge device.

[0050] S103, the structured data of the production scenario is encoded sequentially through the Embedding layer and the fully connected layer to obtain the structured data feature vector.

[0051] Specifically, the pre-processed structured data such as production work orders and environmental parameters in S100 are first subjected to feature pre-coding. For example, discrete features in the structured data (such as material type and process requirements) are mapped to unique discrete integer identifiers to eliminate the non-numerical influence of text-based discrete features; continuous features in the structured data (such as temperature and humidity, and the remaining time of delivery deadline) are kept numerical to maintain consistency with the pre-processed results.

[0052] The pre-encoded structured data is input into the Embedding layer. Utilizing the principle of low-dimensional dense mapping in the Embedding layer, the discrete integer identifiers are mapped into low-dimensional dense vectors, solving the sparsity problem of discrete structured features while preserving semantic relationships between features (such as the relationship between different material types and different production process requirements). The discrete feature dense vector output from the Embedding layer is concatenated with the preprocessed continuous feature vector along the feature dimension to obtain a fused structured feature vector, which is then input into the fully connected layer.

[0053] The fully connected layer achieves deep integration of multi-dimensional structured features of the production scenario through linear transformation of the weight matrix. Then, it mines the intrinsic correlation features between production work orders and environmental parameters (such as the constraint features of high temperature environment on the production process of specific materials) through nonlinear activation. Finally, it outputs a structured data feature vector, namely the production scenario feature vector, to ensure the dimensional matching between node features and scenario features in the subsequent graph structure modeling process. The value of each dimension corresponds to a feature dimension of the production scenario, which can completely represent the overall working condition of the production scenario.

[0054] S200: Based on the time-series signal data feature vector and the structured data feature vector, graph structure modeling is performed to form a dynamic production graph, which is then encoded through a time-series graph neural network to obtain graph-level global features.

[0055] In this process, the time-series signal data feature vectors and structured data feature vectors obtained in the aforementioned steps are concatenated and fused according to their feature dimensions. Then, through graph structure modeling, multi-layer iterative encoding of temporal graph neural networks (TGNN), and global aggregation of edge node features, a graph-level global feature that can characterize the operating status of edge devices, production scene conditions, and inter-device interactions in the entire industrial production scenario is obtained. This feature provides core global feature input for the subsequent inference of the pre-trained target MDP decision model, and the feature dimension is consistent with the edge device and scene feature vector dimensions in the early stage, ensuring the feature adaptability of the subsequent model inference.

[0056] Specifically, obtaining map-level global features may include the following steps: S201, using edge devices as edge nodes, material flow, energy flow, and information flow between the edge devices as edges, the time-series signal data feature vector as edge node features, and the interaction strength calculated by mutual information or Granger causality as edge features, a dynamic production graph is constructed.

[0057] Specifically, the formal representation of the dynamic production graph is as follows: In this embodiment, the dynamic production graph is formally represented as Gt=(Vt,Et), where t is the time step, which can be preset according to the update requirements of industrial production conditions (e.g., 1s / 5s / 10s is a time step). The graph will complete the real-time update of nodes, edges and corresponding features with time step t, accurately matching the dynamic changes of the production scenario.

[0058] Edge node set Vt and node features: Each edge device in the industrial production scenario is regarded as an edge node in the dynamic production map. All edge devices in the production scenario at time t constitute the node set Vt. If edge devices are added or deactivated in the production scenario, the node set Vt will be dynamically adjusted synchronously. The feature vector of the time-series signal data obtained by S100 is dimensionally fused with the feature vector of the structured data as the core feature of each edge node. It retains the operating status features of the edge device itself, such as temperature, vibration, and current, and also incorporates the operating condition features of the production scenario, such as production work orders and environmental parameters, so that the node features can completely represent the overall state of the edge device under specific production conditions.

[0059] Edge set Et and the physical meaning of edges: The actual material flow, energy flow, and information flow between edge devices are used as edges in the dynamic production graph. At time t, all effective flow connections between edge devices in the production scenario constitute edge set Et. Edges are not constructed between edge devices without actual flow connections. Among them, material flow refers to the production material transmission relationship between edge devices, energy flow refers to the energy supply relationship such as electricity and power between edge devices, and information flow refers to the interactive relationship such as operation status data and production control commands between edge devices. If the flow connection between devices changes such as being added or disconnected, edge set Et will be dynamically adjusted synchronously.

[0060] Calculating edge features (interaction strength between edge devices): The interaction strength obtained through mutual information or Granger causality calculation is used as the core feature of the edge to characterize the strength of the flow connection between any two edge nodes. When using mutual information calculation, the correlation between the fused features of two edge nodes is measured to obtain a value in the range of 0 to 1. The larger the value, the stronger the feature correlation and the tighter the flow connection between the devices. When using Granger causality calculation, the corresponding causal correlation value is obtained by determining whether the state change of one edge node will have a causal impact on the state of another edge node. The larger the correlation value, the more significant the causal impact between the devices. This is used as the edge feature to characterize the interaction strength between devices.

[0061] Real-time updates of the dynamic production graph: At each time step t, the fusion features of all edge nodes and the interaction strength of all edges are recalculated. If changes occur in edge devices / stream connections in the production scenario, the node set Vt and edge set Et are adjusted synchronously to ensure accurate matching between the dynamic production graph and the actual industrial production conditions.

[0062] S202, the dynamic production map is updated by multi-layer iterative encoding using a temporal graph neural network (TGNN) layer to obtain the encoding features of each edge node.

[0063] Specifically, the edge node input features are initialized as follows: the fused features of each edge node in the dynamically generated graph Gt=(Vt,Et) constructed in S201 are used as the initial input features of the TGNN layer. All edge nodes that have flow connections with the edge nodes are defined as their neighbor edge node set, which changes synchronously with the update of the graph edge set.

[0064] The number of iteration layers in the TGNN is set according to the complexity of the industrial production scenario. 2-3 layers are used for a single production line, and 4-5 layers are used for complex production scenarios involving multiple production lines. The number of TGNN layers is determined based on the principle of fully integrating the features of the edge nodes themselves with the global relational features. The TGNN layers encode and update the features of each edge node sequentially from layer k to layer k+1. Each layer update is based on the core principle of neighbor feature aggregation and node self-feature transformation. First, the features of the edge node itself are aggregated and fused with the features of the neighboring edge node sets. Then, non-linear activation is used to achieve non-linear transformation and deep encoding of the features, allowing the edge node features to gradually integrate their own state, the relational state of neighboring edge nodes, and the global relational information of the entire graph in the iterative process.

[0065] After completing the iterative encoding of all preset layers, the last layer feature of each edge node is used as the final encoded feature of that node. This feature integrates the edge node's own operation and scene features, the association features of neighboring edge nodes, and the global association information of the entire dynamic production graph, which can accurately represent the overall interaction and operation status of the edge node in the entire production scene.

[0066] S203, aggregate the encoded features of each edge node and output the graph-level global features.

[0067] Specifically, the core purpose of the S203 step aggregation is to integrate the scattered coding features of all edge nodes in the graph into global features that can represent the entire production scenario. During the aggregation process, feature integrity is guaranteed, and all core global information on edge device operation, scenario conditions, and interactions between edge devices is retained.

[0068] Based on the differences in functional importance of various edge devices in the production scenario, feature fusion is achieved by choosing between mean aggregation or attention mechanism aggregation. If the functional importance of each device in the scenario is similar, the mean of the final encoded features of all nodes is calculated dimension by dimension, and the resulting mean vector is the aggregated global feature. This method can evenly integrate the feature information of all devices without feature bias. If there is a clear distinction between core production equipment, key control equipment, and auxiliary equipment in the scenario, different attention weights are assigned to different nodes, with higher weights for core and key equipment and lower weights for auxiliary equipment. The weight values ​​can be preset according to the actual importance of the equipment and optimized during offline pre-training of the model. Then, the final encoded features of the nodes are weighted and summed based on the weights, and the resulting weighted vector is the aggregated global feature. This method can highlight the feature information of core equipment and fit the needs of actual production conditions.

[0069] A linear transformation is performed on the aggregated global feature vector to ensure that its feature dimensions are consistent with those of the previous time-series signal data feature vectors and structured data feature vectors. This guarantees the feature dimension adaptability when inputting into the pre-trained target MDP decision model for inference. The global feature vector with unified dimensions is the graph-level global feature. This feature fully integrates global information such as the operating status of all equipment, the strength of interaction between equipment, and the overall working conditions of the production scenario. It provides a core global feature foundation for subsequent model inference and the generation of cross-production line collaborative execution strategies.

[0070] S300, based on the target MDP decision model, combined with a hierarchical distributed decision architecture and meta-learning adaptation method, collaborative scheduling optimization is performed to generate multiple cross-production line collaborative execution strategies.

[0071] The S300 step is based on the offline-trained and converged target MDP decision model. For the online real-time working conditions of industrial production, it sequentially completes the online optimization of the target MDP decision model and cross-production line feature adaptation. Then, it relies on the three-layer hierarchical distributed decision architecture of edge devices, production lines, and edge nodes to carry out hierarchical collaborative scheduling, and finally generates multiple sets of cross-production line collaborative execution strategies adapted to different production scenarios.

[0072] Specifically, generating multiple cross-production line collaborative execution strategies may include the following steps: S301, the target MDP decision model is optimized online by combining dual Q-learning with security constraint reinforcement learning to obtain an optimized MDP decision model that meets the security threshold of edge devices and ensures production continuity.

[0073] The target MDP decision model is optimized online by combining dual Q-learning with security-constrained reinforcement learning to obtain an optimized MDP decision model that meets the safety threshold of edge devices and ensures production continuity. Specifically, the offline-trained target MDP decision model learns the linkage patterns of edge devices based on historical production conditions. However, the actual online production process is characterized by real-time dynamic changes in operating conditions. Therefore, it is necessary to optimize the model online by combining dual Q-learning with security-constrained reinforcement learning to allow the model to quickly adapt to the current real-time production state.

[0074] Specifically, the core optimization principle of dual Q-learning involves setting up two independent evaluation networks to bidirectionally verify and score the effect of the target MDP decision model after executing linked decisions. This avoids decision bias caused by a single evaluation network, allowing the target MDP decision model to more accurately learn the optimal linkage strategy under real-time operating conditions during online updates, thereby improving the stability and accuracy of the target MDP decision model's decisions. Simultaneously, a safety constraint reinforcement learning mechanism is introduced to set clear safety thresholds for all linked operations of edge devices. These thresholds cover core indicators such as device operating parameter thresholds, load thresholds, and energy consumption thresholds, ensuring that all linked decision actions output by the target MDP decision model after online optimization are within the safe operating range of the edge devices. This mitigates the risks of equipment failure and production interruption from the decision-making level, ensuring the continuity of industrial production.

[0075] After the above online optimization, the resulting optimized MDP decision model retains the general edge device linkage rules learned in the offline training phase, accurately adapts to online real-time production conditions, and all decisions meet the safety operation requirements of edge devices, laying the model foundation for subsequent cross-production line adaptation.

[0076] S302, Based on the MAML meta-learning framework and combined with adversarial training of the domain adaptation layer, the optimized MDP decision model is adapted across production lines to obtain an adapted MDP decision model that is adapted to the characteristics of different production lines.

[0077] Specifically, in industrial production scenarios, there are differences in process requirements, equipment configurations, and production tasks among different production lines. In order to enable the optimized MDP decision model to have cross-production line adaptability, this step is to carry out cross-production line pre-training based on the MAML meta-learning framework, and combine it with adversarial training of the domain adaptation layer to reduce the feature distribution differences between different production lines.

[0078] First, relying on the MAML meta-learning framework, a small amount of historical operating condition data from multiple similar production lines in an industrial production scenario is selected to pre-train the optimized MDP decision model. This allows the optimized MDP decision model to learn the common features and adaptation logic of different production lines. For feature adaptation to new production lines, only a small number of new production line samples are needed to perform parameter gradient updates on the pre-trained optimized MDP decision model. This quickly yields the parameters of the optimized MDP decision model adapted to the new production line. Specifically, the update logic is based on the original parameters of the pre-trained optimized MDP decision model. The learning rate of the optimized MDP decision model's gradient update is used, combined with the gradient operator of the optimized MDP decision model parameters, to calculate the operating condition loss function of the new production line. The original parameters are then adjusted and updated through gradient descent, eliminating the need for retraining with a large amount of data and achieving rapid adaptation to new production lines.

[0079] Simultaneously, a domain discriminator is added after the Temporal Graph Neural Network (TGNN), and adversarial training of the domain adaptation layer is conducted. This allows the optimized MDP decision model to automatically reduce the feature distribution differences between different production lines during the feature encoding stage, further improving the cross-production line adaptation capability of the optimized MDP decision model. After pre-training with the MAML meta-learning framework and adversarial training with the domain adaptation layer, the adapted MDP decision model is obtained. This adapted MDP decision model can simultaneously adapt to multiple production lines with different characteristics, meeting the needs of cross-production line collaborative decision-making.

[0080] S303, based on a hierarchical distributed decision architecture of edge devices, production lines, and edge nodes, performs collaborative scheduling on the adapted MDP decision model to generate multiple cross-production line collaborative execution strategies.

[0081] The collaborative scheduling refers to the coordinated linkage strategy output by the adapted MDP decision model, which is organized by the edge node, broken down into sub-strategies for each production line and further broken down into operation instructions for each edge device, and the instructions are issued in a hierarchical manner from edge node to production line and from production line to edge device.

[0082] Specifically, step S303 uses the adapted MDP decision model as the core decision-making basis and relies on a three-layered distributed decision architecture of edge devices, production lines, and edge nodes to carry out hierarchical collaborative scheduling. This architecture is divided into three core decision-making levels: the first layer is a single device-level adaptive control strategy, responsible for the local linkage decision of a single edge device; the second layer is a production line-level device group collaborative optimization strategy, responsible for the linkage and resource allocation of all devices within a single production line; the third layer is a cross-production line resource scheduling strategy between edge nodes, which, as the highest decision-making level, is responsible for coordinating the cross-production line collaboration and global resource allocation of all production lines. The three levels coordinate step by step, with the upper layer coordinating the lower layer, and the lower layer can feed back the production status to the upper layer in real time, forming a closed-loop collaborative decision-making logic.

[0083] With edge nodes as the highest level of coordination, they receive the global cross-production line linkage strategy output by the adapted MDP decision model. This global cross-production line linkage strategy is formulated based on real-time production conditions across all scenarios, covering core aspects such as cross-production line equipment linkage, resource allocation, and production task collaboration. The edge nodes first break down the global linkage strategy into exclusive sub-strategies for each production line according to their process requirements, production tasks, and equipment configuration characteristics. The production line sub-strategies then perform localized adaptation and adjustments to the global strategy to ensure compatibility with the actual production conditions of each production line.

[0084] Each production line then further breaks down its sub-strategies into specific operational instructions for each edge device within the production line. These instructions clearly define the adjustment requirements for the linkage rules of the corresponding edge devices, including specific actions such as enabling / disabling existing linkage rules, modifying core rule parameters, and generating new linkage rules. Finally, the operational instructions are distributed hierarchically from edge node to production line to edge device, ensuring the coordinated consistency of cross-production line strategy execution.

[0085] During the aforementioned hierarchical collaborative scheduling process, multiple different cross-production line collaborative execution strategies can be generated based on different production conditions, resource allocation priorities, and production task objectives. All cross-production line collaborative execution strategies are practically feasible and can be used to select the optimal solution by combining cloud-edge resource unified view data.

[0086] S400: Based on the hardware and operational data of the edge device and the cloud server, resource modeling and real-time status synchronization are performed to obtain unified cloud-edge resource view data.

[0087] The S400 step takes the hardware basic data and real-time operation data of edge devices and cloud servers as core inputs to complete the resource modeling of the edge side and the cloud side respectively. Then, the two types of model data are integrated through a real-time status synchronization mechanism to finally form a unified cloud-edge resource view data that can comprehensively represent the resource status of the entire cloud-edge domain. This data provides the core reference basis for the optimal selection of subsequent cross-production line collaborative execution strategies in terms of resource dimensions.

[0088] Specifically, the steps to obtain a unified view of cloud-edge resources may include the following: S401, Based on the hardware and operating data of the edge device, construct an edge device resource profile that includes the static attributes and dynamic operating status of the edge device.

[0089] Specifically, step S401 constructs a unique and complete edge device resource profile for each edge device in an industrial production scenario, based on the full amount of hardware and operational data of that edge device. This edge device resource profile covers both static attribute dimensions and dynamic operational status dimensions, achieving a full-dimensional characterization of the edge device resource status.

[0090] Among them, the static attribute dimension includes the inherent hardware information of the edge device, including basic attributes that do not change with the operating status, such as device model, factory parameters, hardware configuration (number of CPU cores, memory capacity, storage specifications, interface type), deployment location, production line, and preset functional positioning. The data in the static attribute dimension will be initialized and entered when the device is deployed to the network, and will only be updated in scenarios such as device hardware upgrades and deployment location adjustments.

[0091] The dynamic operating status dimension records the real-time operating data of the edge device, including CPU utilization, memory usage, remaining storage capacity, network transmission rate, current device load, energy consumption data, fault alarm status, linkage rule execution status, production task processing progress, and other operating indicators that change dynamically with the production process. The data in the dynamic operating status dimension is collected in real time by the edge device's local acquisition module at a preset acquisition frequency (e.g., once per second) to ensure that the data in this dimension can accurately reflect the current operating status of the edge device.

[0092] After integrating the dimensional data of static attributes and dynamic operating status, a complete resource profile of a single edge device is formed. The resource profiles of all edge devices are classified and stored according to the hierarchical structure of "production line-device number", laying the edge-side data foundation for subsequent cloud-edge resource integration.

[0093] S402, Based on the hardware and operating data of the cloud server, construct a dynamic model of cloud resources that includes available computing units, storage capacity, and service status.

[0094] Specifically, step S402 targets cloud server clusters and, based on the hardware configuration data and real-time operating data of the cloud server clusters, constructs a dynamic cloud resource model that can dynamically reflect the usage status of cloud resources. This dynamic cloud resource model covers three core resource dimensions and comprehensively represents the resource supply capacity of the cloud.

[0095] The first category is the available computing unit dimension. This dimension collects core data such as the number of CPU cores, GPU computing power, computing power scheduling priority, occupied computing resources, and remaining available computing resources of all nodes in the cloud server cluster. By integrating the scattered computing resources into a unified computing resource pool through the principle of resource pooling, the overall available computing power of the computing resource pool is output in real time, providing a basis for cloud-edge collaborative computing power scheduling.

[0096] The second category is the storage capacity dimension. This dimension counts the total storage capacity, used storage capacity, remaining available storage capacity, storage data types (historical data of edge devices, production work order data, model training data, etc.), storage read and write rates, and other data of the cloud server cluster. Storage partitions are divided according to data importance, and the remaining available space of this storage capacity dimension is updated in real time to ensure sufficient storage resources for cloud-edge data interaction.

[0097] The third category is the service status dimension. This dimension collects statistics on the running status, response latency, service availability, and concurrent connection count of various services (data storage service, model training service, policy distribution service, status monitoring service, etc.) provided by the cloud server cluster. It marks abnormal service nodes in real time and triggers an automatic switching mechanism to ensure that the dynamic model of cloud resources can accurately reflect the availability status of cloud services.

[0098] S403, perform real-time status synchronization updates on the edge device resource profile and the cloud resource dynamic model to obtain unified cloud-edge resource view data.

[0099] Specifically, step S403 establishes a two-way real-time synchronization channel between the cloud and the edge. Based on the principle of incremental data synchronization, it updates the status data of all edge device resource profiles and the dynamic model of the cloud resource in real time, and finally integrates them to form a unified view of cloud and edge resources.

[0100] The synchronization mechanism is divided into two directions: First, edge-to-cloud synchronization, where all edge devices synchronize incremental change data (such as CPU utilization increasing from 50% to 60%, new fault alarms, etc.) in the dynamic operating status dimension of the edge device resource profile to the cloud at a preset synchronization frequency (consistent with the dynamic operating status collection frequency); Second, cloud-to-edge synchronization, where incremental change data (such as a decrease in the remaining computing power of the computing resource pool, an increase in the response latency of a certain type of service) in the cloud resource dynamic model are synchronized to each edge node to ensure that the resource status data on both sides of the cloud and the edge are consistent in real time.

[0101] After data synchronization is completed, a unified resource visualization modeling method is used to integrate the scattered edge device resource profile data and the cloud resource dynamic model data into a standardized unified cloud-edge resource view data. This unified cloud-edge resource view data is organized in a hierarchical structure of "cloud resource pool - edge node - production line - edge device", and displays the core resource status of computing power, storage, services, and device operation in the entire cloud-edge domain in real time. It supports filtering and querying by production line, device type, resource type and other dimensions, and provides a panoramic resource status reference for the optimal selection of subsequent cross-production line collaborative execution strategies.

[0102] S500: Based on the unified view data of cloud-edge resources, the optimal solution is selected from multiple cross-production line collaborative execution strategies to obtain the target cross-production line collaborative execution strategy.

[0103] Step S500 uses the unified cloud-edge resource view data as the resource basis and multiple cross-production line collaborative execution strategies as candidate solutions. Through resource matching modeling, multi-objective optimization evaluation and adaptive decision selection, the target cross-production line collaborative execution strategy that best suits the current cloud-edge resources and production conditions is selected from the multiple cross-production line collaborative execution strategies.

[0104] Specifically, obtaining the target cross-production line collaborative execution strategy may include the following steps: S501, based on the unified view data of cloud and edge resources, abstract and model the resource demand characteristics of multiple cross-production line collaborative execution strategies and the computing power, load, energy consumption and reliability characteristics of cloud and edge resources respectively, and establish a strategy and resource matching constraint model.

[0105] Specifically, step S501 first extracts real-time computing power characteristics, real-time load characteristics, real-time energy consumption characteristics, and real-time reliability characteristics of cloud servers and edge devices based on the unified cloud-edge resource view data, forming a cloud-edge resource feature set corresponding to the unified cloud-edge resource view data. Simultaneously, resource requirement analysis is performed on each cross-production line collaborative execution strategy, abstracting the computing power requirement, load capacity, energy consumption limit, and reliability requirements required during the execution of the cross-production line collaborative execution strategy, forming the resource requirement characteristics corresponding to that cross-production line collaborative execution strategy. The cloud-edge resource feature set is then matched with the resource requirement characteristics of the cross-production line collaborative execution strategy, using computing power matching, load matching, energy consumption matching, and reliability matching as constraints to establish a strategy-resource matching constraint model. This strategy-resource matching constraint model is used to determine whether each cross-production line collaborative execution strategy meets the actual carrying capacity of the current cloud-edge resources, ensuring that the subsequently selected target cross-production line collaborative execution strategy can be executed stably.

[0106] S502, using the strategy and resource matching constraint model as constraints, perform multi-objective optimization evaluation on multiple cross-production line collaborative execution strategies in terms of delay, cost, energy consumption, balance and reliability, sort the cross-production line collaborative execution strategies according to Pareto dominance relationship, eliminate the dominated cross-production line collaborative execution strategies, and select the optimal solution set composed of Pareto optimal strategies from multiple cross-production line collaborative execution strategies.

[0107] Specifically, under the constraints of the strategy and resource matching constraint model, multiple cross-production line collaborative execution strategies that meet the resource matching conditions are evaluated using multi-objective optimization. The evaluation dimensions include execution latency, operating cost, overall energy consumption, cloud-edge resource balance, and system operational reliability.

[0108] Based on the Pareto dominance principle, multiple cross-production line collaborative execution strategies are comprehensively ranked. Dominated strategies that are not superior to other strategies in all evaluation dimensions are eliminated, retaining only the Pareto optimal strategy that performs best in at least one evaluation dimension and cannot be completely surpassed by any other strategy. All retained Pareto optimal strategies are integrated to form an optimal solution set composed of Pareto optimal strategies. Each cross-production line collaborative execution strategy in this optimal solution set is a candidate strategy that meets cloud-edge resource constraints and has superior overall performance.

[0109] S503, based on the real-time operating status of the current production system and actual operational needs, adaptive decision-making is performed on the optimal solution set to obtain the target cross-production line collaborative execution strategy.

[0110] Specifically, step S503 acquires the real-time operating status of the current production system, including the urgency of production tasks, equipment health status, production line operating load, remaining cloud-edge resources, and actual production and operational needs, such as prioritizing energy consumption reduction, shortening execution latency, and ensuring production continuity. Based on the real-time operating status of the current production system and the actual operational needs, an optimal Pareto strategy that best matches the current scenario is adaptively selected from the set of optimal solutions. This selected cross-production line collaborative execution strategy is determined as the target cross-production line collaborative execution strategy. This target cross-production line collaborative execution strategy satisfies the resource constraints corresponding to the unified cloud-edge resource view data, meets multi-objective optimization requirements, and adapts to the real-time operating conditions and operational needs of the current production system. It can be directly deployed to edge nodes and edge devices for execution.

[0111] In addition, if the status of edge devices / production lines changes frequently in a localized manner, the target strategy is only fine-tuned locally within the corresponding level (such as modifying the linkage parameters of a single device), without the need to re-select the overall strategy across production lines. The overall target strategy is only reselected when the global status of edge nodes (such as cloud-edge overall resources or cross-production line tasks) changes frequently, thus achieving a layered response of "local fine-tuning and global reselection".

[0112] In one embodiment, the training steps of the target MDP decision model may include: S601 integrates historical graph-level global features into the state space of the Markov decision process, constructs a joint state space that includes the real-time state of edge devices, production scenarios, and environmental states, and defines the action space for adjusting linkage rules to form an initial MDP decision model.

[0113] The historical graph-level global features are formed by the feature vectors of historical time-series signal data and the feature vectors of historical structured data through dynamic graph construction and time-series graph neural network encoding. The linkage rule adjustment includes enabling / disabling existing linkage rules, modifying existing linkage rule parameters, and generating new linkage rules.

[0114] Specifically, the historical graph-level global features obtained through dynamic production graph construction and temporal graph neural network encoding are deeply integrated into the state space construction stage of the Markov decision process. Based on the historical operating features of edge devices, historical operating conditions of production scenarios, and historical interaction and correlation features between devices integrated into these features, a three-dimensional joint state space is constructed. This joint state space is the core state reference for training the initial MDP decision model. The historical state dimension of edge devices covers core information such as device operating parameters, health status, and load status; the production scenario dimension covers key content such as production work order progress, process execution requirements, and product quality indicators; and the environmental state dimension covers relevant information such as production environment temperature and humidity, and energy supply status. These three dimensions together constitute a complete state input, enabling the model to perceive the production status across the entire scenario.

[0115] While constructing the joint state space, an action space for the Markov decision process is defined. All actions within this action space are adjustments to the edge device linkage rules and are adapted to the dynamic changes in the production scenario. The action space specifically includes three types of adjustment actions: the first is enabling / disabling existing linkage rules. For pre-configured device linkage rules in the production line, the initial MDP decision model can choose to enable or disable the corresponding rules based on changes in production status to adapt to basic production needs. The second is modifying the parameters of existing linkage rules. For enabled linkage rules, the initial MDP decision model can adjust their core parameters such as threshold, priority, and validity period to achieve refined adaptation of the linkage rules. The third is generating new linkage rules. The initial MDP decision model can obtain new device linkage schemes from the strategy generation module based on production conditions and implement them as executable new linkage rules to meet diverse production decision-making needs.

[0116] S602, construct a multi-objective reward function with the optimization objectives of improving production efficiency, optimizing energy consumption, and reducing the lifespan loss of edge devices.

[0117] The multi-objective reward function serves as the evaluation criterion for training the initial MDP decision model.

[0118] Specifically, a multi-objective reward function is constructed around three core industrial production goals: improving production efficiency, optimizing energy consumption, and reducing the lifespan of edge devices. The core design logic of this function is to achieve synergistic optimization of the three goals, avoiding production imbalance caused by optimization of a single goal. At the same time, this multi-objective reward function serves as the core evaluation criterion for training the initial MDP decision model. During the initial MDP decision model training process, the output of this function will guide the learning of the optimal equipment linkage decision strategy.

[0119] The formula for the multi-objective reward function in this embodiment is: ,in This is the final output value of the multi-objective reward function, i.e., the reward score after the model executes a certain linked decision; , and The weighting coefficients for the three optimization objectives—production efficiency, energy consumption optimization, and equipment life loss—are respectively, and they must add up to 1. The values ​​of each weighting coefficient can be dynamically adjusted according to the core needs of actual production. Production efficiency, energy consumption optimization, and equipment life loss are all normalized characteristic values. Production efficiency represents production efficiency indicators such as work order completion rate and capacity achievement rate. Energy consumption optimization represents energy consumption indicators such as unit product energy consumption reduction rate and energy utilization rate. Equipment life loss represents equipment life-related indicators such as wear rate of key equipment components and changes in equipment maintenance cycle.

[0120] S603, Based on the initial MDP decision model and the multi-objective reward function, the initial MDP decision model is trained to obtain the converged target MDP decision model.

[0121] Specifically, the joint state space constructed in step S601 serves as the state input basis for training the initial MDP decision model. The defined linkage rule adjustment action space serves as the decision action range of the initial MDP decision model. The multi-objective reward function constructed in step S602 serves as the evaluation criterion for model decision-making. Offline training of the initial MDP decision model is then conducted. During the training process, a safety constraint reinforcement learning mechanism is introduced to set safety thresholds for all linkage rule adjustment actions of the initial MDP decision model. This ensures that all decision actions learned by the initial MDP decision model are within the safe operating range of the edge device, fundamentally guaranteeing the continuity of industrial production and avoiding production interruptions or equipment failures caused by model decisions.

[0122] The core logic of the initial MDP decision model training is as follows: Under different production states in the joint state space, the initial MDP decision model autonomously selects linkage rules in the action space to adjust actions. The execution effect of the action is scored through a multi-objective reward function. If the action can achieve synergistic optimization of improved production efficiency, reduced energy consumption, and reduced equipment lifespan loss, it will receive a higher reward value. The initial MDP decision model will iteratively update its own parameters based on the result to strengthen the optimal decision in that state. If the action cannot achieve multi-objective synergistic optimization, or even causes problems such as decreased production efficiency and increased equipment loss, it will receive a lower reward value. The initial MDP decision model will then correct the decision logic in that state.

[0123] The training process continues iteratively until the output values ​​of the multi-objective reward function corresponding to the decision actions of the initial MDP decision model stabilize under all historical production conditions, and the decision results of the initial MDP decision model no longer show significant changes. At this point, the initial MDP decision model is considered to have reached training convergence, and the trained initial MDP decision model is the target MDP decision model. This target MDP decision model can be deployed to an edge distributed real-time platform. In the subsequent online inference stage, it receives real-time graph-level global feature input and quickly generates cross-production line collaborative execution strategies adapted to the current production conditions.

[0124] In this embodiment, the training of the target MDP decision model is an offline training process. The entire process uses historical graph-level global features as the core input and relies on the modeling logic of Markov decision process to allow the initial MDP decision model to learn the equipment linkage rules and multi-objective optimization decision logic in industrial production scenarios. Finally, a converged target MDP decision model is obtained, which provides model support for generating cross-production line collaborative execution strategies in the subsequent online inference stage.

[0125] In one embodiment, after step S302 and before step S302, multimodal federated learning is also included for collaborative optimization of cross-production line models.

[0126] Specifically, to further enhance the cross-production line generalization capability of the adapted MDP decision model while protecting production line data privacy, a multimodal federated learning cross-production line model collaborative optimization step is included after S302 (cross-production line adaptation) and before S303 (collaborative scheduling). This step, based on the "adapted MDP decision model" output from S302, optimizes model performance without leaking the original data through multimodal data expansion and federated learning collaborative training. Specifically, it may include the following steps: S3021, divide the federated learning nodes and expand the preprocessed multimodal data to obtain encrypted cross-production line multimodal features.

[0127] Specifically, this step aims to build the node architecture for federated learning and expand the data dimensions to enrich the model input features, while ensuring data privacy through encryption.

[0128] Federated learning node partitioning: Each production line in the industrial production scenario is treated as an independent federated learning node. Each node retains only the right to store and process local data and does not transmit raw data to other nodes or the cloud. The cloud platform acts as the federated coordinator and is only responsible for receiving the model parameter gradients of each node and performing aggregation updates. It does not store any raw data of any production line, thus blocking the path of data privacy leakage from the architecture.

[0129] Multimodal data acquisition expansion: Building upon the existing "edge device timing signal data (temperature, vibration, current) + structured data from production scenarios (production work orders, environmental parameters)," multimodal data acquisition is expanded. Equipment appearance image data: Industrial cameras deployed at key locations on the production line are used to collect images of the appearance status of edge equipment (such as wear on the equipment casing, loose parts, indicator light status, etc.), with the acquisition frequency consistent with the timing signal (e.g., 1 second / time). Equipment operating audio data: Acoustic sensors are used to collect operating noise from edge devices (such as motor rotation noise, mechanical friction noise, etc.) to capture the audio characteristics of the equipment's operating status; Infrared thermal imaging data: By collecting thermal imaging data on the temperature distribution of edge devices through infrared sensors, abnormal conditions such as local overheating of the devices can be accurately captured.

[0130] Multimodal data preprocessing: Image data preprocessing: The acquired equipment appearance image is subjected to mean filtering to remove noise, irrelevant background areas are cropped, the image size is uniformly scaled to a fixed size (e.g., 224×224 pixels), and then pixel value normalization (mapping pixel values ​​to the [0,1] interval) is performed to convert it into numerical features that can be recognized by the model. Audio data preprocessing: The audio signal is converted into the spectral features of the time-frequency domain (by using the short-time Fourier transform principle, the continuous audio is divided into multiple short time periods, and the frequency distribution features of each time period are extracted), and then the spectral features are normalized to eliminate the dimensional differences between different sensors. Infrared data preprocessing: Infrared thermal imaging data is converted into grayscale images, the temperature distribution matrix in the image is extracted, and the matrix values ​​are standardized to ensure that they are consistent with the numerical range of other modal features. Multimodal feature fusion: The preprocessed image features, audio features, and infrared features are concatenated with the existing time-series signal data feature vectors and structured data feature vectors according to their dimensions to form a cross-production line multimodal feature vector, ensuring that the feature dimensions match the input dimensions of the adapted MDP decision model.

[0131] Data encryption processing: Homomorphic encryption principle is used to encrypt the fused cross-production line multimodal features. The encrypted data can be directly used for calculation operations in model training without decrypting the original data to obtain effective calculation results. This ensures that the multimodal features of each federated node are not leaked during transmission and use, and finally the encrypted cross-production line multimodal features are obtained.

[0132] S3022: Each federated node trains the model locally and uploads the gradients, which are then aggregated and updated in the cloud to obtain a cross-production line collaborative optimization model.

[0133] Specifically, this step achieves cross-production line model collaborative optimization through the "local training + cloud aggregation" model of federated learning, avoiding the privacy risks caused by the centralized transmission of the original dataset.

[0134] Local model training initialization: Each federated node (each production line) uses the "adapted MDP decision model" output by S302 as its local initial model, loads the encrypted cross-production line multimodal features obtained by S3021, and conducts local model training. During training, the model only uses local data to update parameters and does not exchange any raw data or intermediate calculation results with other nodes.

[0135] Local model parameter gradient calculation: Based on the training process of local multimodal features, the gradient change values ​​(i.e., the direction and magnitude of parameter updates) of each parameter of the adapted MDP decision model are calculated. The core principle of gradient calculation is: by comparing the difference between the model's predicted output and the actual labels of local data (such as the label of normal equipment operation status and the label of production efficiency target), the influence of each parameter on the difference is deduced in reverse to obtain the gradient value.

[0136] Gradient encryption upload: Each federated node uploads the calculated parameter gradient to the cloud coordinator through a secure encrypted channel (a transmission channel based on the SSL protocol). During the upload process, the gradient data is encrypted again to prevent data leakage during transmission.

[0137] After receiving the encrypted gradients from all federated nodes, the cloud coordinator integrates the gradients using a weighted average aggregation method.

[0138] The cloud distributes the aggregated global gradient to each federated node. Each node uses the global gradient to update its local model parameters, completing one round of collaborative training. The process of "local training → gradient calculation → encrypted upload → cloud aggregation → parameter update" is repeated for n rounds until the training loss value of the model tends to stabilize (the fluctuation range of the loss value in 3 consecutive iterations is ≤3%). The iteration is then stopped, and the cross-production line collaborative optimization model is obtained.

[0139] S3023 uses MAML meta-learning to perform lightweight fine-tuning on the new production line, resulting in a collaborative optimization model adapted to the new production line.

[0140] Specifically, to address the issues of limited data and long adaptation cycles for new production lines, this step leverages the rapid adaptation capabilities of MAML meta-learning to achieve efficient model adaptation for new production lines.

[0141] New production line data preparation: For new production lines, only a small number (50-100) of multimodal samples are collected, including equipment timing signals, structured data, images, audio, and infrared data. Feature extraction and encryption are completed according to the S3021 preprocessing standard to form a new production line multimodal feature set.

[0142] Meta-model initialization: The cross-production line collaborative optimization model obtained in S3022 is used as the meta-model. This meta-model has learned the common features and adaptation logic of multiple existing production lines and has the basic ability to generalize across production lines.

[0143] Lightweight MAML fine-tuning: Utilizing a small set of multimodal features from the new production line, the meta-model undergoes 1-2 steps of gradient updates to achieve rapid adaptation. The core principle of fine-tuning is that the meta-model already possesses a general cross-production line parameter base, eliminating the need for extensive retraining with large amounts of data. Only a small number of samples are needed to adjust local parameters to adapt to the unique characteristics of the new production line (such as process differences and equipment configuration differences). During fine-tuning, the learning rate follows the MAML meta-learning setting in S302 (e.g., 0.01) to ensure the stability of parameter updates.

[0144] S3024 integrates cross-production line and new production line adaptation results and outputs a federated collaborative optimization MDP decision model.

[0145] Specifically, this step aims to integrate the results of existing production line collaborative optimization and new production line adaptation to form a unified optimized model, providing a basis for decision-making in subsequent collaborative scheduling.

[0146] Model parameter integration: The cross-production line collaborative optimization model parameters obtained in S3022 are integrated with the adapted model parameters of each new production line obtained in S3043. The core general parameters of the cross-production line collaborative optimization model are retained (to ensure generalization ability for most production lines), and the adaptation parameters of each new production line are embedded (to ensure adaptability to specific new production lines), forming a model parameter system that takes into account both "general generalization" and "personalized adaptation".

[0147] Model consistency verification: Perform consistency verification on the integrated model to ensure that the output logic of the model is consistent in different production line scenarios (such as the unified judgment standard for safety thresholds and the unified logic for adjusting linkage rules), and avoid decision conflicts caused by parameter integration.

[0148] Output the final model: After successful verification, define the model as the "Federated Collaborative Optimized MDP Decision Model". This model inherits the cross-production line basic capabilities of the adapted MDP decision model, adds the ability to represent multimodal features, the privacy protection features brought by federated learning, and the ability to quickly adapt to new production lines. It can be directly input into step S303 as the core decision basis of the hierarchical distributed decision architecture.

[0149] It should be noted that in the implementation of this application, the design methods and algorithms of Conv1D-CNN layers, BiLSTM networks, Embedding layers and fully connected layers are all adopted in the conventional way used in the industrial field. This application does not improve the above core architecture and usage logic, but only reasonably integrates them and applies them to the technical solution of this application to achieve the technical effect to be achieved by this application.

[0150] Secondly, this application provides an edge distributed real-time platform linkage strategy generation system, applied to the aforementioned edge distributed real-time platform linkage strategy generation method, the system comprising: The acquisition unit is used to acquire time-series signal data from edge device sensors and structured data from the production scene to obtain time-series signal data feature vectors and structured data feature vectors. The time-series signals include time-series signal data of temperature, vibration and current of the edge device, and the structured data includes structured data of production work orders and environmental parameters. The feature processing unit is used to perform graph structure modeling processing based on the feature vector of the time-series signal data and the feature vector of the structured data to form a dynamic production graph, and to encode it through a time-series graph neural network to obtain graph-level global features; The optimization unit inputs the graph-level global features into the pre-trained target MDP decision model for inference, and performs collaborative scheduling optimization by combining the hierarchical distributed decision architecture and meta-learning adaptation method to generate multiple cross-production line collaborative execution strategies. The resource processing unit is used to perform resource modeling and real-time status synchronization based on the hardware and operating data of the edge device and the cloud server to obtain unified cloud-edge resource view data. The result unit is used to select the optimal solution from multiple cross-production line collaborative execution strategies based on the unified view data of cloud-edge resources, and obtain the target cross-production line collaborative execution strategy.

[0151] Thirdly, this application provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned method.

[0152] Fourthly, this application provides a computer program that, when executed by a processor, implements the steps of the aforementioned method.

[0153] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0154] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0155] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An edge distributed real-time platform linkage strategy generation method, characterized in that, include: Acquire time-series signal data from edge device sensors and structured data from the production scenario to obtain time-series signal data feature vectors and structured data feature vectors. The time-series signals include time-series signal data of temperature, vibration, and current of the edge devices, and the structured data includes structured data of production work orders and environmental parameters. Based on the time-series signal data feature vector and the structured data feature vector, graph structure modeling is performed to form a dynamic production graph, which is then encoded through a time-series graph neural network to obtain graph-level global features. The graph-level global features are input into the pre-trained target MDP decision model for inference. The hierarchical distributed decision architecture and meta-learning adaptation method are combined for collaborative scheduling optimization to generate multiple cross-production line collaborative execution strategies. Based on the hardware and operational data of the edge devices and cloud servers, resource modeling and real-time status synchronization are performed to obtain a unified view of cloud-edge resources. Based on the unified view data of cloud-edge resources, the optimal solution is selected from multiple cross-production line collaborative execution strategies to obtain the target cross-production line collaborative execution strategy.

2. The edge distributed real-time platform federating strategy generation method of claim 1, wherein, The step of acquiring time-series signal data from edge device sensors and structured data from the production scenario to obtain time-series signal data feature vectors and structured data feature vectors includes: The temperature, vibration, and current time-series signal data of the edge device sensor are used to extract local features using a Conv1D-CNN layer to obtain local features of the time-series signal. For the local features of the time series signal, long-range dependencies are captured by a BiLSTM network to obtain the feature vector of the time series signal data; The structured data of the production scenario is encoded sequentially through an embedding layer and a fully connected layer to obtain the structured data feature vector.

3. The edge distributed real-time platform federating strategy generation method of claim 1, wherein, The step of performing graph structure modeling based on the time-series signal data feature vector and the structured data feature vector to form a dynamic production graph, and encoding it through a time-series graph neural network to obtain graph-level global features includes: Using edge devices as edge nodes, material flow, energy flow and information flow between edge devices as edges, and the time-series signal data feature vector as edge node features, combined with the interaction strength calculated by mutual information or Granger causality as edge features, a dynamic production graph is constructed. The dynamic production graph is updated by multi-layer iterative encoding using a temporal graph neural network (TGNN) layer to obtain the encoded features of each edge node; The encoded features of each edge node are aggregated to output the graph-level global features.

4. The edge distributed real-time platform federating strategy generation method of claim 1, wherein, The training steps of the target MDP decision model include: By integrating historical graph-level global features into the state space of a Markov decision process, a joint state space is constructed that includes the real-time state of edge devices, production scenarios, and environmental states. An action space for adjusting linkage rules is defined to form an initial MDP decision model. The historical graph-level global features are formed by using historical time-series signal data feature vectors and historical structured data feature vectors through dynamic production graph construction and time-series graph neural network encoding. The linkage rule adjustment includes enabling / disabling existing linkage rules, modifying existing linkage rule parameters, and generating new linkage rules. A multi-objective reward function is constructed with the optimization objectives of improving production efficiency, optimizing energy consumption, and reducing the lifespan loss of edge devices. The multi-objective reward function serves as the evaluation criterion for training the initial MDP decision model. Based on the initial MDP decision model and the multi-objective reward function, the initial MDP decision model is trained to obtain the converged target MDP decision model.

5. The edge distributed real-time platform linkage strategy generation method as described in claim 1, characterized in that, The step of generating multiple cross-production line collaborative execution strategies by combining the target MDP decision model with a hierarchical distributed decision architecture and meta-learning adaptation method for collaborative scheduling optimization includes: The target MDP decision model is optimized online by combining dual Q-learning with security constraint reinforcement learning to obtain an optimized MDP decision model that meets the security threshold of edge devices and ensures production continuity. Based on the MAML meta-learning framework and combined with adversarial training of the domain adaptation layer, the optimized MDP decision model is adapted across production lines to obtain an adapted MDP decision model that is adapted to the characteristics of different production lines. Based on a hierarchical distributed decision architecture of edge devices, production lines, and edge nodes, the adapted MDP decision model is coordinated and scheduled to generate multiple cross-production line collaborative execution strategies. The coordinated scheduling means that the edge nodes coordinate the linkage strategies output by the adapted MDP decision model, decompose them into sub-strategies for each production line, and further decompose them into operation instructions for each edge device. The instructions are issued in a hierarchical manner from edge nodes to production lines and from production lines to edge devices.

6. The edge distributed real-time platform linkage strategy generation method as described in claim 1, characterized in that, The step of performing resource modeling and real-time status synchronization based on the hardware and operational data of the edge device and the cloud server to obtain a unified view of cloud-edge resources includes: Based on the hardware and operational data of the edge devices, construct an edge device resource profile that includes the static attributes and dynamic operational status of the edge devices. Based on the hardware and operational data of the cloud server, a dynamic model of cloud resources, including available computing units, storage capacity, and service status, is constructed. The edge device resource profile and the cloud resource dynamic model are updated in real time to obtain a unified view of cloud and edge resources.

7. The edge distributed real-time platform linkage strategy generation method as described in claim 1, characterized in that, The step of selecting the optimal solution from multiple cross-production line collaborative execution strategies based on the unified cloud-edge resource view data to obtain the target cross-production line collaborative execution strategy includes: Based on the unified view data of cloud and edge resources, the resource demand characteristics of multiple cross-production line collaborative execution strategies and the computing power, load, energy consumption and reliability characteristics of cloud and edge resources are abstractly modeled to establish a strategy and resource matching constraint model. Using the strategy and resource matching constraint model as constraints, a multi-objective optimization evaluation of delay, cost, energy consumption, balance and reliability is performed on multiple cross-production line collaborative execution strategies. The cross-production line collaborative execution strategies are sorted according to Pareto dominance relationship. Dominated cross-production line collaborative execution strategies are eliminated. The optimal solution set composed of Pareto optimal strategies is obtained from multiple cross-production line collaborative execution strategies. Based on the real-time operating status of the current production system and actual operational needs, an adaptive decision-making process is performed on the optimal solution set to obtain the target cross-production line collaborative execution strategy.

8. A system for generating linkage strategies for an edge-distributed real-time platform, characterized in that, The method for generating linkage strategies for edge distributed real-time platforms according to any one of claims 1-7, the system comprising: The acquisition unit is used to acquire time-series signal data from edge device sensors and structured data from the production scene to obtain time-series signal data feature vectors and structured data feature vectors. The time-series signals include time-series signal data of temperature, vibration and current of the edge device, and the structured data includes structured data of production work orders and environmental parameters. The feature processing unit is used to perform graph structure modeling processing based on the feature vector of the time-series signal data and the feature vector of the structured data to form a dynamic production graph, and to encode it through a time-series graph neural network to obtain graph-level global features; The optimization unit inputs the graph-level global features into the pre-trained target MDP decision model for inference, and performs collaborative scheduling optimization by combining the hierarchical distributed decision architecture and meta-learning adaptation method to generate multiple cross-production line collaborative execution strategies. The resource processing unit is used to perform resource modeling and real-time status synchronization based on the hardware and operating data of the edge device and the cloud server to obtain unified cloud-edge resource view data. The result unit is used to select the optimal solution from multiple cross-production line collaborative execution strategies based on the unified view data of cloud-edge resources, and obtain the target cross-production line collaborative execution strategy.

9. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images