Industrial internet of things based multi-source data fusion collection method for measurement and control equipment

Abstract

This invention discloses a method for multi-source data fusion acquisition of measurement and control equipment based on the Industrial Internet of Things (IIoT), relating to the fields of IIoT and data acquisition technology. It solves the technical problem of efficient, synchronous, and adaptive acquisition of multi-source measurement and control data under resource-constrained conditions in the IIoT. The method includes: constructing a digital twin model in the cloud to identify measurement and control points requiring synchronous acquisition; edge nodes calculating information gain based on predictive uncertainty and submitting bidding requests to the cloud with information gain per unit time as the utility function; the cloud solving the Nash equilibrium of the Stackelberg game with the goal of maximizing global information gain while satisfying synchronous acquisition coupling constraints, and dynamically allocating sampling time slot resources; edge nodes performing acquisition according to the allocated resources, fusing time-aligned multi-source data frames and feeding back the actual information gain; and the cloud updating the reputation value of the edge nodes based on the feedback and adjusting the relaxation of the synchronous acquisition coupling constraints.

Description

Technical Field

[0001] This invention belongs to the field of industrial Internet of Things and data acquisition technology, specifically relating to a method for multi-source data fusion acquisition of measurement and control equipment based on industrial Internet of Things. Background Technology

[0002] In the industrial Internet of Things (IIoT) environment, there are many types of measurement and control equipment, with varying sampling frequencies and inconsistent data formats.

[0003] The fusion acquisition of multi-source data faces the following technical challenges: First, edge nodes have limited communication bandwidth, computing power, and storage resources, while the number of measurement and control devices is large. Traditional periodic acquisition methods generate a large amount of redundant data, wasting scarce resources and potentially losing critical transient data. Second, different measurement and control devices have different sampling frequencies, start-up times, and transmission delay jitter, causing data to be misaligned on the time axis. Traditional linear interpolation methods cannot balance accuracy and efficiency and are difficult to adapt to dynamically changing network environments. Third, existing acquisition methods typically process data from each measurement and control device independently, ignoring the physical coupling between devices, resulting in a lack of physical consistency in the fused data. Fourth, traditional acquisition is passive and fixed-cycle, unable to dynamically adjust acquisition strategies based on data value and device status, and also difficult to predict future data fusion points that need attention.

[0004] Therefore, there is an urgent need for a method that can achieve efficient, synchronous, and adaptive fusion acquisition of multi-source measurement and control data under resource-constrained conditions. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a multi-source data fusion acquisition method for measurement and control equipment based on the Industrial Internet of Things, which is used to solve the technical problems of low acquisition efficiency, difficulty in synchronizing asynchronous multi-source data, lack of predictability and adaptability of acquisition strategies, and lack of physical consistency of fusion results due to independent processing of data from each device under resource-constrained conditions.

[0006] To address the above problems, this invention provides a method for multi-source data fusion acquisition of measurement and control equipment based on the Industrial Internet of Things, comprising the following steps: S1: The cloud server constructs a three-layer nested global digital twin model including a device layer, a coupling layer and a system layer. The system layer adopts a graph physical information neural network and introduces the residual of the physical conservation equation as a Laplace regularization term on the graph in message passing to correct prediction bias, so as to predict the state evolution trend of each measurement and control device in the future time window and identify the combination of measurement and control points that need to be synchronously collected. S2: Each edge node calculates the predicted information gain of each candidate sampling action based on the cognitive uncertainty component quantified in real time by the Monte Carlo Dropout method according to the local twin model, and submits a bid request containing the set of sampling actions and their coupling constraints to the cloud with the information gain within a unit sampling time window as the utility function. S3: The cloud server collects all bidding requests and solves the Nash equilibrium of the Stackelberg game with the goal of maximizing global information gain and satisfying the synchronous acquisition coupling constraint. It dynamically allocates sampling time slot resources, including the exclusive sampling right of each edge node for a specific measurement and control point at a specific time. S4: Each edge node performs multi-source data acquisition at a specified time according to the allocated sampling time slot resources, and fuses the multi-source data frames after the timestamps are aligned, and feeds back the actual information gain obtained. S5: The cloud updates the reputation value of each edge node based on the feedback information and adjusts the relaxation of the synchronous collection coupling constraint in the next round of the game.

[0007] Preferably, the construction of a three-layer nested global digital twin model comprising a device layer, a coupling layer, and a system layer includes the following steps: The global digital twin model is constructed as a three-layer nested architecture, including device layer twin, coupling layer twin, and system layer twin; The device-layer twin establishes an independent lightweight long short-term memory network model for each measurement and control device. The input is the historical acquisition time-series data of the measurement and control device, and the output is the state prediction value and prediction variance within the next time window. The coupling layer twin is constructed based on a graph attention network. The nodes are the state variables of each device-layer twin, and the edges are the potential physical coupling relationships between devices. The coupling strength is dynamically learned through attention coefficients, and the output is a node embedding matrix and the corresponding graph structure. Each row corresponds to the embedding vector of a measurement and control device, and the topology of the matrix is ​​defined by the physical connection relationships between devices. The system-layer twin takes the output of the coupling layer as input and uses a graph physical information neural network to directly perform message passing on the graph. The aggregation function of message passing introduces the residual of the physical conservation equation as a bias term. At the same time, the physical conservation equation is defined as a Laplace regularization term on the graph to constrain the state prediction of adjacent nodes to satisfy physical consistency. The output of the graph physical information neural network is the physical consistency state correction value of each node, which is used to correct the prediction deviation between the device layer and the coupling layer.

[0008] Preferably, the global digital twin model adopts a hierarchical training method: the device layer is trained locally at each edge node, and only the hash values ​​of the model parameters are uploaded to the cloud; the coupling layer and the system layer are trained by aggregating multi-source data frames with timestamp alignment in the cloud. During the training process, for nodes and their neighborhoods with a Laplace regularization term value > a preset regularization threshold, the system automatically marks them as high uncertainty regions, triggering the edge nodes to include the high uncertainty region as a candidate sampling action in the bidding request in the next round of the game in step S2.

[0009] Preferably, the global digital twin model adopts an incremental update mechanism driven by a data acquisition strategy, including the following steps: When the actual information gain fed back to the cloud by the edge node is greater than the preset gain threshold, a local incremental update of the model is triggered: only the subgraphs in the coupling layer that are related to the data with actual information gain greater than the preset gain threshold, and the graph physical information neural network parameters on the corresponding subgraphs in the system layer are retrained, while the twins in the device layer remain frozen; When the cumulative number of incremental updates reaches the preset round threshold, global retraining is triggered: the entire three-layer structure is retrained using compressed historical data frames stored in the cloud. Incremental updates and global retraining are facilitated by selective knowledge distillation driven by information gain. Both knowledge distillation and model retraining are executed asynchronously in independent background threads on the cloud and edge nodes. Read-write locks on model parameters are maintained to ensure that the distillation update process does not block the current game-theoretic sampling time slot resource allocation and real-time data acquisition process. Distillation is performed only on parameters of subgraphs related to high-gain data in the coupling layer and corresponding subgraphs in the system layer; device twins do not participate in the distillation process. The distillation is triggered when the actual information gain fed back by the edge node exceeds a preset gain threshold. An information gain weighting term is introduced into the distillation loss function.

[0010] Preferably, in step S2, the calculation of the predicted information gain includes the following steps: in, For the first One candidate sampling action, For edge node local twin models performing sampling actions The previous state of knowledge, The prediction uncertainty entropy under the aforementioned knowledge state. To perform sampling actions as assumed And obtain sampling data Then, update the conditional uncertainty entropy of the model. For mathematical expectation operators; The predicted uncertainty entropy is quantified in real time using the Monte Carlo Dropout method, including a cognitive uncertainty component and a random uncertainty component. Only the information gain corresponding to the cognitive uncertainty component is used for utility function calculation in the bidding request. The random uncertainty component is used to evaluate the sensor noise level and as a threshold for data pre-filtering. The sampling action... The information gain per unit time is: ,in, For sampling action The sampling duration.

[0011] Preferably, in step S3, synchronously acquiring coupling constraints includes the following steps: Suppose a combination of monitoring and control points identified by the cloud that requires synchronous data collection. ,in, For the first One monitoring and control point, Let be the total number of control points within the group, and let be the sampling time allocated to each control point within the group. ,in, For monitoring and control points The assigned sampling time; the synchronous acquisition coupling constraint requires that the maximum time difference between the sampling times of all measurement and control points within the combination be ≤ a preset time alignment error threshold, i.e.: ,in, For monitoring and control points and The absolute value of the difference between sampling times. This is the maximum value operator. This is the preset time alignment error threshold.

[0012] Preferably, in step S3, solving for the Nash equilibrium of the Stackelberg game uses a distributed iterative algorithm, including the following steps: S301: Cloud Leader Initializes Resource Price Vector And the resource quota limit for each edge node; S302: Each edge node follower adjusts its price based on the current price. Solve the resource demand optimization problem for each: ,in, For edge nodes The requested sampling time slot resource vector, wherein the sampling time slot resources are allocated in units of time slices. For edge nodes Applying for resources Information gain per unit time under the given conditions For the first The inner product of the transpose of the price vector and the resource vector in each iteration. This is a vector transpose operation; S303: Collect all demand data in the cloud. The deviation between total demand and total resource capacity is calculated, and the price is updated accordingly. Specifically: ,in, To adaptively update the step size, This is the total resource capacity vector. This represents the total number of edge nodes. S304: Repeat steps S302 to S303 until the relative rate of change of the price vectors of two adjacent rounds is less than or equal to the preset convergence threshold. The conflict resolution rule for exclusive sampling rights is as follows: when multiple edge nodes submit bid requests for the same monitoring and control point at the same time, the game equilibrium result will grant the sampling time slot to the node with the highest overall priority; the overall priority is calculated as follows: ,in, For edge nodes For monitoring and control points At any moment Overall priority For edge nodes For monitoring and control points At any moment Information gain per unit time For edge nodes Reputation value The urgency factor is determined by the edge nodes. Local twin model predicts control points At any moment The state value is obtained after normalizing its proximity to the preset danger threshold. ,and These are the weighting coefficients.

[0013] Preferably, in step S4, the fusion of the collected timestamp-aligned multi-source data frames includes the following steps: S401: Each edge node collects raw data from each measurement and control point at a specified time according to the allocated sampling time slot resources, and adds a hardware timestamp to each collected data point; S402: Perform timestamp alignment on the data collected by each control point within the same synchronous acquisition group, taking the group as the unit: Assume there are m control points within the same group, and the original data collected by each control point is: ,in, For the first The measured values ​​of each control point For the first Hardware timestamps at each monitoring and control point This timestamp is generated by hardware; a reference timeline is selected for reference time. For the target time point, based on the reputation value of the edge node performing the alignment operation. Dynamically select interpolation method: when ≥ At that time, the cubic spline interpolation method was used to calculate the values ​​of each control point. alignment value ;when < At that time, the linear interpolation method was used to calculate the values ​​of each control point. alignment value The aligned data from each control point are organized into multi-source data frames with timestamp alignment according to the control point index order. ; S403: Perform a fusion operation on the multi-source data frames obtained in step S402. The fusion weights are determined based on the attention coefficients output by the graph attention network to obtain the fusion feature vector. S404: For data frames marked as high uncertainty regions, a robust fusion method is used to replace the weighted fusion in step S403 to obtain a robust fusion feature vector. The robust fusion method is median fusion. S405: Pack the fused feature vector or robust fused feature vector and its corresponding actual information gain and feed them back to the cloud.

[0014] Preferably, in step S5, adjusting the relaxation of the synchronous acquisition coupling constraint refers to dynamically adjusting the preset time alignment error threshold value based on the reputation value of the edge node, including the following steps: The adjustment rules are as follows: set edge nodes Reputation value The first reputation threshold is The second reputation threshold is The reference time alignment error threshold is ; when At that time, the time alignment error threshold of the synchronous acquisition combination to which the node belongs. The calculation is as follows: ,in, This is the relaxation gain coefficient. The maximum time alignment error threshold. This is the minimum value operator; when At that time, the time alignment error threshold of the synchronous acquisition combination to which the node belongs. The calculation is as follows: ,in, This is the compression gain coefficient. The minimum time alignment error threshold, This is the maximum value operator; when At that time, keep constant.

[0015] Preferably, in step S5, updating the reputation value of each edge node based on the feedback information includes the following steps: The process of updating the reputation value of each edge node is as follows: ,in, For edge nodes Reputation value before update For edge nodes Updated reputation value Historical weighting factor For the accuracy of the bid, To improve resource utilization efficiency, Honesty factor; The cloud implements tiered penalty measures for edge nodes whose reputation value is less than the penalty threshold for K consecutive rounds.

[0016] The beneficial effects of this invention are: This invention constructs a global digital twin model as a three-layer nested architecture: a device layer, a coupling layer, and a system layer. The device layer uses an LSTM network to predict the state of each device; the coupling layer uses a graph attention network to dynamically learn the physical coupling relationships between devices; and the system layer uses a graph physical information neural network with a Laplace regularization term to constrain the physical conservation equations. This solves the problem of existing technologies neglecting the physical coupling relationships between devices, leading to a lack of physical consistency in the fused data. It ensures that the state predictions of adjacent nodes satisfy the physical conservation laws of energy, mass, and momentum, effectively correcting prediction biases. This invention uses the Monte Carlo Dropout method to quantify and predict uncertainty entropy in real time, decomposing it into cognitive uncertainty (insufficient model knowledge) and random uncertainty (sensor noise). Only the information gain corresponding to cognitive uncertainty is used for bidding decisions. This solves the problem that traditional periodic data acquisition cannot distinguish data value, allowing acquisition resources to be prioritized for high-value data areas that the model "understands the least". At the same time, random uncertainty is used for sensor noise assessment and data pre-filtering, improving the signal-to-noise ratio of the acquired data. This invention uses information gain per unit time as the utility function, with edge nodes acting as followers submitting bidding requests and the cloud acting as the leader solving the Nash equilibrium to dynamically allocate sampling time slot resources; when multiple nodes conflict over the same measurement and control point, the decision is made according to the comprehensive priority; this invention solves the problem of resource conflicts caused by edge nodes acting independently, and maximizes global information gain under the conditions of limited communication bandwidth and computing resources, while ensuring rapid response in emergency situations; This invention dynamically selects the timestamp alignment method based on the reputation value of edge nodes: high-reputation nodes use high-precision cubic spline interpolation, while low-reputation nodes use low-complexity linear interpolation. Simultaneously, it dynamically adjusts the relaxation of the synchronous acquisition coupling constraint based on the reputation value, allowing high-reputation nodes to have a larger time alignment error threshold, while forcing low-reputation nodes to use a stricter threshold. This solves the problem that fixed alignment methods cannot balance accuracy and efficiency, enabling high-reputation nodes to achieve higher alignment accuracy and saving computational resources for low-reputation nodes, thus achieving adaptive optimization of accuracy and efficiency. This invention introduces the attention coefficients output by the graph attention network into the determination of fusion weights, so that measurement and control points with high coupling strength can obtain higher weights during fusion; median fusion is used instead of weighted fusion in high uncertainty regions to enhance robustness; it solves the problem that traditional weighted fusion ignores the physical relationship between devices, so that the fusion result can reflect the physical relationship between devices such as thermal coupling and vibration transmission, while effectively suppressing the influence of outliers in high uncertainty regions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

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

[0019] Please see Figure 1 As shown, this invention is a multi-source data fusion acquisition method for measurement and control equipment based on the Industrial Internet of Things, comprising the following steps: S1: The cloud server constructs a three-layer nested global digital twin model including a device layer, a coupling layer and a system layer. The system layer adopts a graph physical information neural network and introduces the residual of the physical conservation equation as a Laplace regularization term on the graph in message passing to correct prediction bias, so as to predict the state evolution trend of each measurement and control device in the future time window and identify the combination of measurement and control points that need to be synchronously collected. S2: Each edge node calculates the predicted information gain of each candidate sampling action based on the cognitive uncertainty component quantified in real time by the Monte Carlo Dropout method according to the local twin model, and submits a bid request containing the set of sampling actions and their coupling constraints to the cloud with the information gain within a unit sampling time window as the utility function. S3: The cloud server collects all bidding requests and solves the Nash equilibrium of the Stackelberg game with the goal of maximizing global information gain and satisfying the synchronous acquisition coupling constraint. It dynamically allocates sampling time slot resources, including the exclusive sampling right of each edge node for a specific measurement and control point at a specific time. S4: Each edge node performs multi-source data acquisition at a specified time according to the allocated sampling time slot resources, and fuses the multi-source data frames after the timestamps are aligned, and feeds back the actual information gain obtained. S5: The cloud updates the reputation value of each edge node based on the feedback information and adjusts the relaxation of the synchronous collection coupling constraint in the next round of the game.

[0020] In one embodiment of the present invention, the construction of a three-layer nested global digital twin model comprising a device layer, a coupling layer, and a system layer includes the following steps: The global digital twin model is constructed as a three-layer nested architecture, including device layer twin, coupling layer twin, and system layer twin; The device-layer twin establishes an independent lightweight long short-term memory network model for each measurement and control device. The input is the historical acquisition time-series data of the measurement and control device, and the output is the state prediction value and prediction variance within the next time window. The coupling layer twin is constructed based on a graph attention network. The nodes are the state variables of each device-layer twin, and the edges are the potential physical coupling relationships between devices. The coupling strength is dynamically learned through attention coefficients, and the output is a node embedding matrix and the corresponding graph structure. Each row corresponds to the embedding vector of a measurement and control device, and the topology of the matrix is ​​defined by the physical connection relationships between devices. The system-layer twin takes the output of the coupling layer as input and uses a graph physical information neural network to directly perform message passing on the graph. The aggregation function of message passing introduces the residual of the physical conservation equation as a bias term. At the same time, the physical conservation equation is defined as a Laplace regularization term on the graph to constrain the state prediction of adjacent nodes to satisfy physical consistency. The output of the graph physical information neural network is the physical consistency state correction value of each node, which is used to correct the prediction deviation between the device layer and the coupling layer.

[0021] Specifically, the global digital twin model is constructed as a three-layer nested architecture, including device layer twin, coupling layer twin, and system layer twin; The device-level twin establishes an independent lightweight long short-term memory network model for each measurement and control device. The input of this model is the historical time-series data acquired by the measurement and control device, and the output is the predicted state value and prediction variance for the next time window: for the first... For a measurement and control device, the input sequence of its LSTM model is: ,in, For the first Each monitoring and control device is at a constant time The collected values, The time window length represents the number of historical data points used for prediction. For the position index in the sequence, Corresponding to the earliest time , Corresponding to the current moment The LSTM model controls the flow of information through three gating units: the forget gate, the input gate, and the output gate. For the first If the hidden state is at time 1, then the model output is: ,in, The model outputs the predicted state values ​​for the next time window; it also outputs the prediction variance. This is achieved through the Monte Carlo Dropout method: maintaining Dropout activation during the inference phase. The second forward propagation calculates the variance of the predicted values ​​as an estimate of the uncertainty. ,in, For the first The predicted value of the second forward propagation, for The mean of the predicted values; the loss function for device-level twins is negative log-likelihood loss: ,in, The total number of measurement and control devices. For the true value, For predicted values, To predict variance; The coupling layer twin is constructed based on a graph attention network. The nodes of this network represent the state variables output by each device layer twin, and the edges represent the potential physical coupling relationships between devices, including thermal coupling, vibration transmission, and electromagnetic interference. The graph structure is defined as follows: ,in, It is a node set, where the number of nodes equals the total number of measurement and control devices. For the edge set, the initial edges are defined based on the physical connections between devices, and subsequently dynamically adjusted using attention coefficients; for nodes... and nodes Attention coefficient The calculation formula is: ,in, and For nodes and nodes The feature vector (from the output of the device layer). This is the weight matrix. For attention weight vectors, It is the transpose vector. The activation function is used; the attention coefficients are normalized using the softmax function. ,in, For nodes The set of neighboring nodes; nodes The updated feature vector is: ,in, The activation function is nonlinear, using the ReLU function; the output of the coupling layer is a node embedding matrix with a graph structure. Its form is: ,in, This represents the total number of measurement and control devices, i.e., the number of rows in the matrix. The embedding vector dimension for each node, i.e., the number of columns in the matrix. For the first The first monitoring and control device in the first Embedded values ​​in each dimension , The first of the matrix OK Corresponding to the The embedding vector of a measurement and control device, denoted as The topology of the matrix is ​​defined by the physical connections between devices: if the first... The first monitoring and control device and the first If there is a physical coupling relationship between the measurement and control devices (e.g., thermal coupling, vibration transmission, or electromagnetic interference), then in the diagram structure... There is a connection node in it. With nodes The loss function of the coupling layer twin is the mean square error between the predicted coupling state and the measured coupling state. ,in, The size of the edge set is the total number of edges in the graph. The system-level twin takes the graph structure and node embedding matrix output by the coupling layer as input and employs a graph-physical information neural network. The message passing process of the graph-physical information neural network is defined as follows: for a node... , No. The message passed from the layer to this node is: ,in, For message functions, The edge features are used; the node state update formula is: ,in, For the update function, These are the physical constraint weighting coefficients. For the physical regularization term with respect to the node state The gradient; simultaneously, the physical conservation equation is defined as a Laplace regularization term on the graph, for nodes Its physical conservation equation residual Among them, considering the discreteness of graph structures, the divergence operator in continuous space Approximation is made using the spatial difference operator on the graph, i.e. ,in, For nodes With nodes The actual physical distance or equivalent thermal resistance between the corresponding measurement and control equipment; For nodes The physical quantity, For flow rate, For source terms, The time partial derivative operator; the Laplace regularization term is defined as: ,in, These are the attention coefficients output by the coupled layer graph attention network. and For nodes With nodes The physical conservation equation residuals; the output of the graph physical information neural network is the physical consistency state correction value for each node, used to correct prediction biases in the device layer and coupling layer. The total loss function for system-level twins is: ,in, The mean square error between the predicted and actual values. This represents the weighting coefficient for physical constraints, with a value ranging from 0.01 to 0.1. This is the Laplace regularization term.

[0022] In one embodiment of the present invention, the global digital twin model adopts a hierarchical training method: the device layer is trained locally at each edge node, and only the hash values ​​of the model parameters are uploaded to the cloud; the coupling layer and the system layer are trained by aggregating multi-source data frames with timestamp alignment in the cloud. During the training process, for nodes and their neighborhoods with a Laplace regularization term value > a preset regularization threshold, the system automatically marks them as high uncertainty regions, triggering the edge nodes to include the high uncertainty region as a candidate sampling action in the bidding request in step S2 of the next round of the game.

[0023] Specifically, the preset regularization threshold is calculated by taking the historical values ​​of the Laplacian regularization term of each node in the most recent 100 training rounds, calculating its average and standard deviation, and setting the threshold as the average plus twice the standard deviation. The preset regularization threshold is recalculated every 50 training rounds to adapt to model evolution. For nodes and their neighborhoods where the Laplacian regularization term value is greater than the preset regularization threshold, the system automatically marks them as high uncertainty regions. After marking, the cloud automatically generates a trigger signal and sends it to the edge node that governs the region. After receiving the trigger signal, the edge node will include the high uncertainty region as a candidate sampling action and submit it to the cloud in the next round of bidding requests, requesting the allocation of additional sampling time slot resources to collect multi-source data in the region.

[0024] In one embodiment of the present invention, the global digital twin model adopts an incremental update mechanism driven by a data acquisition strategy, including the following steps: When the actual information gain fed back to the cloud by the edge node is greater than the preset gain threshold, a local incremental update of the model is triggered: only the subgraphs in the coupling layer that are related to the data with actual information gain greater than the preset gain threshold, and the graph physical information neural network parameters on the corresponding subgraphs in the system layer are retrained, while the twins in the device layer remain frozen; When the cumulative number of incremental updates reaches the preset round threshold, global retraining is triggered: the entire three-layer structure is retrained using compressed historical data frames stored in the cloud. Incremental updates and global retraining are facilitated by selective knowledge distillation driven by information gain. Both knowledge distillation and model retraining are executed asynchronously in independent background threads on the cloud and edge nodes. Read-write locks on model parameters are maintained to ensure that the distillation update process does not block the current game-theoretic sampling time slot resource allocation and real-time data acquisition process. Distillation is performed only on parameters of subgraphs related to high-gain data in the coupling layer and corresponding subgraphs in the system layer; device twins do not participate in the distillation process. The distillation is triggered when the actual information gain fed back by the edge node exceeds a preset gain threshold. An information gain weighting term is introduced into the distillation loss function.

[0025] Specifically, the preset gain threshold is calculated by taking the actual information gain values ​​from the most recent 100 edge node feedbacks, calculating their average and standard deviation, setting the initial threshold to be equal to the average plus 1.5 times the standard deviation, and recalculating the preset gain threshold every 50 feedbacks for adaptive adjustment; The preset round threshold represents the upper limit of the cumulative number of local incremental updates, which is dynamically set according to the system's computing resources: the preset round threshold is larger when resources are sufficient, and smaller when resources are limited. Among them, subgraphs with actual information gain greater than a preset gain threshold are marked as high-gain data-related subgraphs; The scope of local incremental updates is limited to: Coupled layer: Only the subgraphs related to high-gain data are retrained; specifically, the control points corresponding to the high-gain data are... The subgraph contains nodes. and its first-order neighbor nodes And the edges between these nodes; System layer: only retrain the neural network parameters of the graph physical information on the above subgraphs; Device layer: keep frozen and do not participate in the update; The loss function for local incremental update is: ,in, Let be the set of nodes in the subgraph. For the Laplacian regularization term on the subgraph; When the cumulative number of local incremental updates reaches a preset round threshold, global retraining is triggered. The cloud uses the stored compressed historical data frames to retrain the entire three-layer structure (device layer, coupling layer, and system layer). The loss function for global retraining is: ,in, , and These are the loss functions for the device layer, coupling layer, and system layer, respectively; after global retraining is completed, the cumulative incremental update count is reset to zero and the count is restarted. Incremental updates and global retraining are facilitated by selective knowledge distillation driven by information gain. The distillation trigger condition is the same as for local incremental updates: distillation is triggered when the actual information gain from edge nodes exceeds a preset gain threshold. Distillation is performed only on parameters of subgraphs in the coupling layer related to high-gain data and corresponding subgraphs in the system layer; device-level twins do not participate in the distillation process. The distillation loss function incorporates a weighted term based on the information gain, specifically defined as: ,in, , The information gain weighting coefficient. The actual information gain fed back by the edge nodes. The preset gain threshold, For knowledge distillation loss, a teacher-student architecture is adopted: the globally retrained model serves as the teacher model, and the model before incremental updates serves as the student model. The distillation loss is the KL divergence of the output probability distributions of the two models. ,in, For Kullback-Leibler divergence operators, The probability distribution output by the teacher model. The probability distribution output by the student model is used to ensure that the direction of knowledge transfer in the model is consistent with the goal of maximizing the value of the collected data. The higher the information gain of the data, the greater the corresponding distillation weight, and the more the model tends to retain knowledge related to high-value data.

[0026] In one embodiment of the present invention, step S2, the calculation of the predicted information gain, includes the following steps: in, For the first One candidate sampling action, For edge node local twin models performing sampling actions The previous state of knowledge, The prediction uncertainty entropy under the aforementioned knowledge state. To perform sampling actions as assumed And obtain sampling data Then, update the conditional uncertainty entropy of the model. For mathematical expectation operators; The predicted uncertainty entropy is quantified in real time using the Monte Carlo Dropout method, including a cognitive uncertainty component and a random uncertainty component. Only the information gain corresponding to the cognitive uncertainty component is used for utility function calculation in the bidding request. The random uncertainty component is used to evaluate the sensor noise level and as a threshold for data pre-filtering. The sampling action... The information gain per unit time is: ,in, For sampling action The sampling duration.

[0027] Specifically, the edge node runs a lightweight local twin model (i.e., a lightweight version of the device-level twin in the cloud-based global model) after INT8 quantization and channel pruning compression. It then uses the Monte Carlo Dropout method to quantize and predict uncertainty entropy in real time. Specifically, during the model inference phase, the Dropout layer remains active, and the edge-side NPU (Neural Processing Unit) performs M forward propagations on the same input in parallel batch processing mode to meet real-time requirements with limited power consumption, resulting in M ​​predictions. It is assumed that under multiple forward propagations of Monte Carlo Dropout, the model's prediction output for the same input approximately follows a Gaussian distribution. Based on this Gaussian distribution assumption, the formula for calculating the prediction uncertainty entropy is: ;in, Let M be the variance of the prediction results. It is a natural constant; The prediction uncertainty entropy is decomposed into cognitive uncertainty and random uncertainty components: Cognitive uncertainty is caused by insufficient model knowledge and can be reduced by collecting more data. It is calculated as the variance of the M prediction results. Random uncertainty is caused by sensor noise and cannot be reduced by collecting data. It is calculated as the average variance of the M prediction results. Among them, only the cognitive uncertainty component The corresponding information gain is used for utility function calculation in the bid request, that is, the information gain actually used for bidding is: ; Random uncertainty component Used to evaluate the sensor noise level and as the threshold for data pre-filtering, the pre-filtering threshold is obtained by taking the random uncertainty values ​​of the most recent 100 sampling points, calculating their average and standard deviation, and setting the filtering threshold to be equal to the average plus 3 times the standard deviation. When the noise estimate of the sensor data exceeds the filtering threshold, the data is subjected to smoothing filtering. Among them, the unit time information gain serves as the utility function when the edge node submits a bidding request to the cloud, and is used to quantify the resource input-output ratio of different sampling actions.

[0028] In one embodiment of the present invention, step S3, synchronously acquiring coupling constraints, includes the following steps: Suppose a combination of monitoring and control points identified by the cloud that requires synchronous data collection. ,in, For the first One monitoring and control point, Let be the total number of control points within the group, and let be the sampling time allocated to each control point within the group. ,in, For monitoring and control points The assigned sampling time; the synchronous acquisition coupling constraint requires that the maximum time difference between the sampling times of all measurement and control points within the combination be ≤ a preset time alignment error threshold, i.e.: ,in, For monitoring and control points and The absolute value of the difference between sampling times. This is the maximum value operator. This is the preset time alignment error threshold.

[0029] Specifically, the identification criteria include: the physical coupling relationship between devices (such as thermal coupling, vibration transmission), the high uncertainty area predicted by the twin model, and the strong correlation shown in historical data; The preset time alignment error threshold is set as follows: for high-precision measurement and control scenarios (such as vibration monitoring and high-speed rotating equipment monitoring), the time alignment error threshold is 1 millisecond; for conventional measurement and control scenarios (such as temperature monitoring and pressure monitoring), the time alignment error threshold is 10 milliseconds; for low-speed change scenarios (such as environmental monitoring), the time alignment error threshold is 100 milliseconds. The threshold is pre-configured by the operation and maintenance personnel according to the actual measurement and control needs, and dynamically adjusted in step S5 according to the reputation value of the edge node.

[0030] In one embodiment of the present invention, step S3, solving for the Nash equilibrium of the Stackelberg game, employs a distributed iterative algorithm and includes the following steps: S301: Cloud Leader Initializes Resource Price Vector And the resource quota limit for each edge node; S302: Each edge node follower adjusts its price based on the current price. Solve the resource demand optimization problem for each: ,in, For edge nodes The requested sampling time slot resource vector, wherein the sampling time slot resources are allocated in units of time slices. For edge nodes Applying for resources Information gain per unit time under the given conditions For the first The inner product of the transpose of the price vector and the resource vector in each iteration. This is a vector transpose operation; S303: Collect all demand data in the cloud. The deviation between total demand and total resource capacity is calculated, and the price is updated accordingly. Specifically: ,in, To adaptively update the step size, This is the total resource capacity vector. This represents the total number of edge nodes. S304: Repeat steps S302 to S303 until the relative rate of change of the price vectors of two adjacent rounds is less than or equal to the preset convergence threshold. The conflict resolution rule for exclusive sampling rights is as follows: when multiple edge nodes submit bid requests for the same monitoring and control point at the same time, the game equilibrium result will grant the sampling time slot to the node with the highest overall priority; the overall priority is calculated as follows: ,in, For edge nodes For monitoring and control points At any moment Overall priority For edge nodes For monitoring and control points At any moment Information gain per unit time For edge nodes Reputation value The urgency factor is determined by the edge nodes. Local twin model predicts control points At any moment The state value is obtained after normalizing its proximity to the preset danger threshold. , and These are the weighting coefficients.

[0031] Specifically, the initial value of the resource price vector Each component is set to 0.1. The resource quota limit for each edge node is pre-allocated based on its historical performance and computing power, with all nodes having the same quota initially. The sampling time slot resource (i.e., the bandwidth window through which edge nodes are allowed to report high-frequency feature data to the cloud) is allocated in time slices as the smallest unit, with each time slice ranging from 100 milliseconds to 1 millisecond in length. The Nash equilibrium of the Stackelberg game is solved. Data acquisition by the underlying hardware is still continuously executed by the edge nodes at a microsecond frequency based on their local real-time clock and stored in a local circular buffer. The game result only controls the reporting rights and cloud fusion computing rights of high-value data blocks within this buffer, thereby masking the latency of the game algorithm and network communication. For the first In each iteration, the transpose of the price vector and the inner product of the resource vector represent the cost of requesting resources. The resource demand optimization problem represents how each edge node, given a price, selects the resource request quantity that maximizes its net benefit (information gain minus resource cost). Through the design of the local model architecture, the information gain per unit time is ensured. It's about the amount of resources requested. The strictly concave function (i.e., exhibiting diminishing marginal returns) ensures that the distributed iterative algorithm for the Stark game can stably converge to the unique Nash equilibrium. The adaptive update step size is calculated as follows: , This is the step size adjustment coefficient, with a value of 0.1. This represents the average value of all components of the total resource capacity vector, expressed in milliseconds. In the price update formula, the price increases when total demand exceeds total capacity and decreases when total demand is less than total capacity. The formula for calculating the relative rate of change is: ,in, The preset convergence threshold is set to 0.01. The specific calculation method for the urgency factor is as follows: ,in, For edge nodes Control points predicted by the local twin model At any moment The state value, The preset danger threshold is obtained by analyzing historical data; in, , and These are the weighting coefficients; in this embodiment, the corresponding values ​​are determined using the analytic hierarchy process (AHP). The value is 0.5. The value is 0.3. The value is 0.2.

[0032] In one embodiment of the present invention, step S4 involves fusing the collected timestamp-aligned multi-source data frames, including the following steps: S401: Each edge node collects raw data from each measurement and control point at a specified time according to the allocated sampling time slot resources, and adds a hardware timestamp to each collected data point; S402: Perform timestamp alignment on the data collected by each control point within the same synchronous acquisition group, taking the group as the unit: Assume there are m control points within the same group, and the original data collected by each control point is: ,in, For the first The measured values ​​of each control point For the first Hardware timestamps at each monitoring and control point This timestamp is generated by hardware; a reference timeline is selected for reference time. For the target time point, based on the reputation value of the edge node performing the alignment operation. Dynamically select interpolation method: when ≥ At that time, the cubic spline interpolation method was used to calculate the values ​​of each control point. alignment value ;when < At that time, the linear interpolation method was used to calculate the values ​​of each control point. alignment value The aligned data from each control point are organized into multi-source data frames with timestamp alignment according to the control point index order. ; S403: Perform a fusion operation on the multi-source data frames obtained in step S402. The fusion weights are determined based on the attention coefficients output by the graph attention network to obtain the fusion feature vector. S404: For data frames marked as high uncertainty regions, a robust fusion method is used to replace the weighted fusion in step S403 to obtain a robust fusion feature vector. The robust fusion method is median fusion. S405: Pack the fused feature vector or robust fused feature vector and its corresponding actual information gain and feed them back to the cloud.

[0033] Specifically, each edge node collects raw data from each monitoring and control point at a specified time according to the allocated sampling time slot resources, and adds a hardware timestamp to each collected data point. The hardware timestamp is generated by the local clock of the edge node, with an accuracy of microseconds. Each collected data point is represented as... The binary tuple, where, For measured values, Hardware timestamp; Among them, reference time The median or mean of the timestamps of each hardware device is usually chosen. The specific implementation of cubic spline interpolation is as follows: a cubic polynomial is constructed between adjacent hardware timestamps to ensure that the interpolation curve is continuous and smooth at the nodes. In this process, a fusion operation is performed on the multi-source data frames obtained in step S402. The fusion weights are determined based on the attention coefficients output by the graph attention network: Let the nodes output by the graph attention network... With nodes The attention coefficient between them is Then the first The coupling strength of a single monitoring and control point is defined as: ;No. The fusion weight of each monitoring and control point The calculation of the fused feature vector is as follows: ; For data frames marked as high uncertainty regions, a robust fusion method is used to replace the weighted fusion in step S403. This robust fusion method is median fusion, and the calculation formula is as follows: ,in, The median fusion operator takes the median of the aligned values ​​of all monitoring and control points as the fusion result. Median fusion can effectively suppress the influence of outliers (such as sensor failures and communication errors) on the fusion result and enhance the robustness of the system. The fused feature vector obtained in step S403 or the robust fused feature vector obtained in step S404, together with the information gain actually obtained during the acquisition process, are packaged and fed back to the cloud server for reputation value update and relaxation adjustment in step S5.

[0034] In one embodiment of the present invention, step S5, adjusting the relaxation of the synchronous acquisition coupling constraint, refers to dynamically adjusting the value of a preset time alignment error threshold based on the reputation value of the edge node, including the following steps: The adjustment rules are as follows: set edge nodes Reputation value The first reputation threshold is The second reputation threshold is The reference time alignment error threshold is ; when At that time, the time alignment error threshold of the synchronous acquisition combination to which the node belongs. The calculation is as follows: ,in, This is the relaxation gain coefficient. The maximum time alignment error threshold. This is the minimum value operator; when At that time, the time alignment error threshold of the synchronous acquisition combination to which the node belongs. The calculation is as follows: ,in, This is the compression gain coefficient. The minimum time alignment error threshold, This is the maximum value operator; when At that time, keep constant.

[0035] Specifically, the first reputation threshold is set to 0.7, the second reputation threshold is set to 0.3, and the initial reputation value is set to 0.5, falling between the two; wherein, the maximum time alignment error threshold is set to... The minimum time alignment error threshold is set to ; The meaning of the adjustment rule is: nodes with higher reputation values ​​are allowed a larger time alignment error threshold (i.e., greater slack) because their historical performance is reliable; nodes with lower reputation values ​​are forced to use a stricter time alignment error threshold (i.e., smaller slack) to prevent their unreliable behavior from affecting the quality of synchronous acquisition.

[0036] In one embodiment of the present invention, step S5, updating the reputation value of each edge node based on feedback information, includes the following steps: The process of updating the reputation value of each edge node is as follows: ,in, For edge nodes Reputation value before update For edge nodes Updated reputation value Historical weighting factor For the accuracy of the bid, To improve resource utilization efficiency, Honesty factor; The cloud implements tiered penalty measures for edge nodes whose reputation value is less than the penalty threshold for K consecutive rounds.

[0037] Specifically, In this embodiment, the historical weighting factor is used. The value is 0.7; To measure bidding accuracy, reflecting the deviation between the predicted information gain declared by edge nodes during bidding and the actual information gain obtained, the calculation formula is as follows: ,in, For edge nodes The actual information gain of the feedback, For edge nodes The predicted information gain declared during the bidding process, It is a very small positive number, with a value of 10. -6 When the reported value is exactly the same as the actual value, When the deviation is greater than or equal to the declared value, Resource utilization efficiency ,in, For edge nodes The total amount of sampling time slot resources actually used, in milliseconds, and the resource utilization efficiency. The larger the value, the higher the information gain the node achieves with fewer resources; honesty factor The redundancy and consistency of data collected by multiple edge nodes from the same monitoring and control point are calculated through cloud-based cross-validation. Specifically, the cloud selects data collected by multiple edge nodes from the same monitoring and control point in the most recent 10 rounds of the game, calculates the correlation coefficient between the data, and sets the edge nodes... The average correlation coefficient between the data and the data from other nodes is ,but When a node is detected intentionally falsifying data (e.g., data that deviates significantly from that of other nodes), It will significantly reduce, The value is 0.5, otherwise, The value of is 1; Specifically, the cloud-based system implements tiered penalties for edge nodes whose reputation values ​​fall below the penalty threshold for K consecutive rounds (K being a preset positive integer ranging from 2 to 5): Level 1 penalty (first trigger): Reduces the node's resource quota limit for the next round of the game to 50% of the original quota and issues a warning notification; Level 2 penalty (two consecutive triggers): Based on Level 1 penalty, the approval delay for the node's bid request is increased by one game cycle (i.e., the node's bid request is processed last in the next round, reducing the reputation value weight in its overall priority calculation); Level 3 penalty (three or more consecutive triggers): Based on Level 2 penalty, the node is added to the list of nodes whose relaxation of synchronous acquisition coupling constraints is reduced, and the minimum time alignment error threshold is forcibly used. That is, the time alignment error threshold of the synchronous acquisition combination to which the node belongs is fixed at the minimum value and is no longer dynamically adjusted according to the reputation value. If it is triggered five times consecutively, the node is marked as an untrusted node, its bidding qualification is suspended, and manual intervention is required to restore it. The penalty removal conditions are as follows: when the reputation value of an edge node recovers to above the penalty threshold for K consecutive rounds, the penalty measures are gradually lifted. At the same time, to prevent nodes from falling into a vicious cycle of continuous reputation deterioration due to network jitter, a reputation circuit breaker and calibration mechanism is set up: when a node is in a level 3 penalty state, it is allowed to actively apply to the cloud to perform a round of low-frequency, high-latency tolerance benchmark calibration collection task. If the data in the calibration task meets the consistency standard with the cloud fusion prediction value, the node's reputation value is directly reset to the benchmark initial value of 0.5, and all penalty measures are lifted.

[0038] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for multi-source data fusion acquisition of measurement and control equipment based on the Industrial Internet of Things, characterized in that, Includes the following steps: S1: The cloud server constructs a three-layer nested global digital twin model including a device layer, a coupling layer and a system layer. The system layer adopts a graph physical information neural network and introduces the residual of the physical conservation equation as a Laplace regularization term on the graph in message passing to correct prediction bias, so as to predict the state evolution trend of each measurement and control device in the future time window and identify the combination of measurement and control points that need to be synchronously collected. S2: Each edge node calculates the predicted information gain of each candidate sampling action based on the cognitive uncertainty component quantified in real time by the Monte Carlo Dropout method according to the local twin model, and submits a bid request containing the set of sampling actions and their coupling constraints to the cloud with the information gain within a unit sampling time window as the utility function. S3: The cloud server collects all bidding requests and solves the Nash equilibrium of the Stackelberg game with the goal of maximizing global information gain and satisfying the synchronous acquisition coupling constraint. It dynamically allocates sampling time slot resources, including the exclusive sampling right of each edge node for a specific measurement and control point at a specific time. S4: Each edge node performs multi-source data acquisition at a specified time according to the allocated sampling time slot resources, and fuses the multi-source data frames after the timestamps are aligned, and feeds back the actual information gain obtained. S5: The cloud updates the reputation value of each edge node based on the feedback information and adjusts the relaxation of the synchronous collection coupling constraint in the next round of the game.

2. The multi-source data fusion acquisition method for measurement and control equipment based on the Industrial Internet of Things as described in claim 1, characterized in that, The construction of a three-layer nested global digital twin model comprising a device layer, a coupling layer, and a system layer includes the following steps: The global digital twin model is constructed as a three-layer nested architecture, including device layer twin, coupling layer twin, and system layer twin; The device-layer twin establishes an independent lightweight long short-term memory network model for each measurement and control device. The input is the historical acquisition time-series data of the measurement and control device, and the output is the state prediction value and prediction variance within the next time window. The coupling layer twin is constructed based on a graph attention network. The nodes are the state variables of each device-layer twin, and the edges are the potential physical coupling relationships between devices. The coupling strength is dynamically learned through attention coefficients, and the output is a node embedding matrix and the corresponding graph structure. Each row corresponds to the embedding vector of a measurement and control device, and the topology of the matrix is ​​defined by the physical connection relationships between devices. The system-layer twin takes the output of the coupling layer as input and uses a graph physical information neural network to directly perform message passing on the graph. The aggregation function of message passing introduces the residual of the physical conservation equation as a bias term. At the same time, the physical conservation equation is defined as a Laplace regularization term on the graph to constrain the state prediction of adjacent nodes to satisfy physical consistency. The output of the graph physical information neural network is the physical consistency state correction value of each node, which is used to correct the prediction deviation between the device layer and the coupling layer.

3. The method for multi-source data fusion acquisition of measurement and control equipment based on industrial Internet of Things according to claim 2, characterized in that, The global digital twin model adopts a hierarchical training method: the device layer is trained locally at each edge node, and only the hash values ​​of the model parameters are uploaded to the cloud; the coupling layer and the system layer are trained by aggregating multi-source data frames with timestamp alignment in the cloud. During the training process, for nodes and their neighborhoods with a Laplace regularization term value greater than the preset regularization threshold, the system automatically marks them as high uncertainty regions, triggering the edge nodes to include the high uncertainty region as a candidate sampling action in the bidding request in step S2 of the next round of the game.

4. The method for multi-source data fusion acquisition of measurement and control equipment based on industrial Internet of Things according to claim 2, characterized in that, The global digital twin model adopts an incremental update mechanism driven by a data acquisition strategy, including the following steps: When the actual information gain fed back to the cloud by the edge node is greater than the preset gain threshold, a local incremental update of the model is triggered: only the subgraphs in the coupling layer that are related to the data with actual information gain greater than the preset gain threshold, and the graph physical information neural network parameters on the corresponding subgraphs in the system layer are retrained, while the twins in the device layer remain frozen; When the cumulative number of incremental updates reaches the preset round threshold, global retraining is triggered: the entire three-layer structure is retrained using compressed historical data frames stored in the cloud. Incremental updates and global retraining are facilitated by selective knowledge distillation driven by information gain. Both knowledge distillation and model retraining are executed asynchronously in independent background threads on the cloud and edge nodes. Read-write locks on model parameters are maintained to ensure that the distillation update process does not block the current game-theoretic sampling time slot resource allocation and real-time data acquisition process. Distillation is performed only on parameters of subgraphs related to high-gain data in the coupling layer and corresponding subgraphs in the system layer; device twins do not participate in the distillation process. The distillation is triggered when the actual information gain fed back by the edge node exceeds a preset gain threshold. An information gain weighting term is introduced into the distillation loss function.

5. The multi-source data fusion acquisition method for measurement and control equipment based on the Industrial Internet of Things as described in claim 1, characterized in that, In step S2, the calculation of the predicted information gain includes the following steps: in, For the first One candidate sampling action, For edge node local twin models performing sampling actions The previous state of knowledge, The prediction uncertainty entropy under the aforementioned knowledge state. To perform sampling actions as assumed And obtain sampling data Then, update the conditional uncertainty entropy of the model. For mathematical expectation operators; The predicted uncertainty entropy is quantified in real time using the Monte Carlo Dropout method, including a cognitive uncertainty component and a random uncertainty component. Only the information gain corresponding to the cognitive uncertainty component is used for utility function calculation in the bidding request. The random uncertainty component is used to evaluate the sensor noise level and as a threshold for data pre-filtering. The sampling action... The information gain per unit time is: ,in, For sampling action The sampling duration.

6. The multi-source data fusion acquisition method for measurement and control equipment based on the Industrial Internet of Things as described in claim 1, characterized in that, In step S3, the synchronous acquisition of coupling constraints includes the following steps: Suppose a combination of monitoring and control points identified by the cloud that requires synchronous data collection. ,in, For the first One monitoring and control point, Let be the total number of control points within the group, and let be the sampling time allocated to each control point within the group. ,in, For monitoring and control points The assigned sampling time; the synchronous acquisition coupling constraint requires that the maximum time difference between the sampling times of all measurement and control points within the combination be ≤ a preset time alignment error threshold, i.e.: ,in, For monitoring and control points and The absolute value of the difference between sampling times. This is the maximum value operator. This is the preset time alignment error threshold.

7. The method for multi-source data fusion acquisition of measurement and control equipment based on industrial Internet of Things according to claim 1, characterized in that, In step S3, the Nash equilibrium of the Stackelberg game is solved using a distributed iterative algorithm, including the following steps: S301: Cloud Leader Initializes Resource Price Vector And the resource quota limit for each edge node; S302: Each edge node follower adjusts its price based on the current price. Solve the resource demand optimization problem for each: ,in, For edge nodes The requested sampling time slot resource vector, wherein the sampling time slot resources are allocated in units of time slices. For edge nodes Applying for resources Information gain per unit time under the given conditions For the first The inner product of the transpose of the price vector and the resource vector in each iteration. This is a vector transpose operation; S303: Collect all demand data in the cloud. The deviation between total demand and total resource capacity is calculated, and the price is updated accordingly. Specifically: ,in, To adaptively update the step size, This is the total resource capacity vector. This represents the total number of edge nodes. S304: Repeat steps S302 to S303 until the relative rate of change of the price vectors of two adjacent rounds is less than or equal to the preset convergence threshold. The conflict resolution rule for exclusive sampling rights is as follows: when multiple edge nodes submit bid requests for the same monitoring and control point at the same time, the game equilibrium result will grant the sampling time slot to the node with the highest overall priority; the overall priority is calculated as follows: ,in, For edge nodes For monitoring and control points At any moment Overall priority For edge nodes For monitoring and control points At any moment Information gain per unit time For edge nodes Reputation value The urgency factor is determined by the edge nodes. Local twin model predicts control points At any moment The state value is obtained after normalizing its proximity to the preset danger threshold. , and These are the weighting coefficients.

8. The method for multi-source data fusion acquisition of measurement and control equipment based on industrial Internet of Things according to claim 1, characterized in that, In step S4, the collected timestamp-aligned multi-source data frames are fused, including the following steps: S401: Each edge node collects raw data from each measurement and control point at a specified time according to the allocated sampling time slot resources, and adds a hardware timestamp to each collected data point; S402: Perform timestamp alignment on the data collected by each control point within the same synchronous acquisition group, taking the group as the unit: Assume there are m control points within the same group, and the original data collected by each control point is: ,in, For the first The measured values ​​of each control point For the first Hardware timestamps at each monitoring and control point This timestamp is generated by hardware; a reference timeline is selected for reference time. For the target time point, based on the reputation value of the edge node performing the alignment operation. Dynamically select interpolation method: when ≥ At that time, the cubic spline interpolation method was used to calculate the values ​​of each control point. alignment value ;when < At that time, the linear interpolation method was used to calculate the values ​​of each control point. alignment value The aligned data from each control point are organized into multi-source data frames with timestamp alignment according to the control point index order. ; S403: Perform a fusion operation on the multi-source data frames obtained in step S402. The fusion weights are determined based on the attention coefficients output by the graph attention network to obtain the fusion feature vector. S404: For data frames marked as high uncertainty regions, a robust fusion method is used to replace the weighted fusion in step S403 to obtain a robust fusion feature vector. The robust fusion method is median fusion. S405: Pack the fused feature vector or robust fused feature vector and its corresponding actual information gain and feed them back to the cloud.

9. The method for multi-source data fusion acquisition of measurement and control equipment based on industrial Internet of Things according to claim 1, characterized in that, In step S5, adjusting the relaxation of the synchronous acquisition coupling constraint refers to dynamically adjusting the preset time alignment error threshold value based on the reputation value of the edge node, including the following steps: The adjustment rules are as follows: set edge nodes Reputation value The first reputation threshold is The second reputation threshold is The reference time alignment error threshold is ; when At that time, the time alignment error threshold of the synchronous acquisition combination to which the node belongs. The calculation is as follows: ,in, This is the relaxation gain coefficient. The maximum time alignment error threshold. This is the minimum value operator; when At that time, the time alignment error threshold of the synchronous acquisition combination to which the node belongs. The calculation is as follows: ,in, This is the compression gain coefficient. The minimum time alignment error threshold, This is the maximum value operator; when At that time, keep constant.

10. The multi-source data fusion acquisition method for measurement and control equipment based on the Industrial Internet of Things according to claim 1, characterized in that, In step S5, updating the reputation value of each edge node based on the feedback information includes the following steps: The process of updating the reputation value of each edge node is as follows: ,in, For edge nodes Reputation value before update For edge nodes Updated reputation value Historical weighting factor For the accuracy of the bid, To improve resource utilization efficiency, Honesty factor; The cloud implements tiered penalty measures for edge nodes whose reputation value is less than the penalty threshold for K consecutive rounds.

Images