An industrial internet remote monitoring operation and maintenance management system and method

Abstract

The application discloses an industrial internet remote monitoring operation and maintenance management system and method, and belongs to the technical field of industrial internet. The method comprises the following steps: in response to access requests of multiple edge nodes, collecting real-time running state parameters of each edge node and performing heterogeneous data graph construction to obtain a dynamic attribute graph structure; performing federated transfer learning on the multiple edge nodes based on the dynamic attribute graph structure to obtain a cross-domain joint inference model; constructing a dynamic reconfigurable digital twin corresponding to each edge node, and performing closed-loop dynamic reconfiguration on the digital twin by using real-time feedback data; performing multi-step simulation deduction based on the reconfigured digital twin to generate an adaptive operation and maintenance strategy; and decomposing the adaptive operation and maintenance strategy into an atomic control instruction sequence, and delivering the atomic control instruction sequence to an execution unit through an asynchronous message queue to realize remote collaborative control.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, specifically to an industrial internet remote monitoring and maintenance management system and method. Background Technology

[0002] In the context of the Industrial Internet, edge nodes are numerous and widely distributed, and their operational status monitoring and remote operation and maintenance scheduling heavily rely on the collection and analysis of real-time operational data. Existing remote monitoring and operation and maintenance methods typically employ a centralized data aggregation architecture, uploading the operational parameters of each edge node to the cloud or a central server, where a unified status monitoring or fault prediction model is then trained. This centralized processing approach consumes significant network bandwidth during data transmission, limiting real-time performance, and the concentration of all node data in one location poses risks of data privacy leaks and single points of failure. On the other hand, some solutions employ isolated modeling to protect data privacy, training models using only local data from a single node. This results in models that struggle to capture common fault patterns and collaborative operational rules across nodes, leading to insufficient inference accuracy.

[0003] In the field of digital twins, most existing digital twins are static or quasi-static mappings constructed in one go based on an initial physical model. When the operating status of equipment drifts due to aging, sudden load changes, or environmental changes, the twin cannot be updated in a timely manner, resulting in a significant deviation between the simulation results and the actual state. At the same time, the generation of operation and maintenance strategies is mostly based on rule thresholds or reactive passive processing, lacking the ability to proactively predict and adaptively adjust the state evolution over multiple future time steps.

[0004] To address the aforementioned issues, it is necessary to solve how to integrate heterogeneous operational information from multiple edge nodes to obtain a high-precision inference model while avoiding the original data from leaving the domain. At the same time, it is necessary to solve how to enable the digital twin to perform closed-loop dynamic reconstruction in accordance with the real-time changes of the physical entity, and to perform multi-step forward-looking inference based on the reconstructed twin to generate an adaptive operation and maintenance strategy that can proactively adapt to future states. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for remote monitoring and operation and maintenance management of the industrial internet, so as to realize cross-domain joint reasoning of multiple nodes while ensuring the data privacy of edge nodes, and to construct a dynamic and reconfigurable digital twin that can follow the real-time evolution of the physical entity status. Through forward-looking inference, adaptive operation and maintenance strategies are generated to improve the accuracy and initiative of remote collaborative control.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] This invention provides a method for remote monitoring, operation and maintenance management of industrial internet, comprising:

[0008] In response to access requests from multiple edge nodes in the Industrial Internet, the real-time operating status parameters of each edge node are collected, and a heterogeneous data graph is constructed on the real-time operating status parameters to obtain a dynamic attribute graph structure representing the operating status of the edge nodes. By constructing the dynamic attribute graph structure, the multi-dimensional operating parameters of the edge nodes and the communication links between nodes are uniformly modeled as graph data, which significantly improves the representational completeness and semantic density of multi-source heterogeneous operating data.

[0009] Based on the dynamic attribute graph structure, federated transfer learning is performed on the multiple edge nodes to obtain a cross-domain joint inference model that integrates the local features of multiple edge nodes. The federated transfer learning mechanism completes cross-node knowledge aggregation and transfer without the need for centralized original data, which not only protects data privacy but also enhances the model's generalization inference ability under cross-domain conditions.

[0010] Based on the cross-domain joint reasoning model, a dynamically reconfigurable digital twin corresponding to each edge node is constructed, and the dynamically reconfigurable digital twin is dynamically reconstructed in a closed loop using real-time feedback data from the edge nodes to obtain the reconstructed digital twin. By introducing real-time feedback and incremental updates driven by state residuals, high-fidelity synchronization between the digital twin and the physical entity is maintained with low computational overhead, providing a reliable virtual mapping for simulation and deduction.

[0011] Based on the reconstructed digital twin, the future operating status of the corresponding edge nodes is simulated and extrapolated in multiple steps, and an adaptive operation and maintenance strategy for each edge node is generated based on the extrapolation results. Based on the forward-looking extrapolation of the digital twin, the operating trends and potential anomalies can be perceived in advance, thereby generating proactive operation and maintenance decisions and reducing the risk of unplanned downtime.

[0012] The adaptive operation and maintenance strategy is decomposed into atomic control instruction sequences, and these sequences are sent to the execution units of the corresponding edge nodes via asynchronous message queues to achieve remote collaborative control of the edge nodes. Through atomic decomposition and transaction message sending, the reliability, consistency, and orderly execution of remote control instructions are ensured, avoiding control risks caused by network jitter or instruction loss.

[0013] As a preferred technical solution of the present invention, the process of obtaining a dynamic attribute graph structure representing the operating state of edge nodes includes: performing time series alignment and missing value imputation on the real-time operating state parameters to obtain an aligned multidimensional parameter sequence; performing semantic type labeling on each parameter in the aligned multidimensional parameter sequence to obtain parameter attribute triples, wherein the parameter attribute triples include parameter name, parameter value, and parameter acquisition timestamp; using each edge node as a central node, using the parameter attribute triples as the attribute feature vector of the central node, and using the communication link status between different edge nodes as edge features to construct the dynamic attribute graph structure. This solution can integrate scattered heterogeneous parameters into structured graph data, accurately reflecting the cooperation and dependency relationships between edge nodes.

[0014] As a preferred technical solution of the present invention, the process of obtaining a cross-domain joint inference model that integrates local features of multiple edge nodes includes: distributing the dynamic attribute graph structure to each edge node participating in federated learning; using a local graph convolutional network on each edge node to extract local features from the dynamic attribute graph structure, obtaining local graph embedding parameters for each edge node; encrypting and uploading the local graph embedding parameters of each edge node to a federated aggregation server; performing weighted aggregation on the federated aggregation server on all received local graph embedding parameters to obtain global graph embedding parameters; and sending the global graph embedding parameters back to each edge node, whereby each edge node uses its stored local supervision data to perform transfer fine-tuning of the global graph embedding parameters, thereby obtaining the cross-domain joint inference model. This federated transfer learning design achieves effective aggregation and personalized adaptation of knowledge from multiple edge nodes under the premise of privacy protection, improving the inference accuracy of the model for local conditions.

[0015] As a preferred embodiment of the present invention, the process of obtaining the reconstructed digital twin includes: performing component-level decomposition on the physical entity of each edge node to obtain the component topology relationship of each physical entity, and assigning initial state variables to each component according to the cross-domain joint inference model; constructing an initial digital twin in virtual space that maps one-to-one with the physical entity according to the component topology relationship and the initial state variables, the initial digital twin including a component state matrix and a component connection matrix; when receiving real-time feedback data from the edge node, extracting the state residual between the real-time feedback data and the initial digital twin, and using the state residual to incrementally update the component state matrix of the initial digital twin to obtain the reconstructed digital twin. Preferably, when the norm of the state residual exceeds a preset reconstruction threshold, a closed-loop dynamic reconstruction of the initial digital twin is triggered, thereby timely capturing the state mutation of the physical entity and preventing the simulation model from becoming distorted.

[0016] As a further preferred embodiment of the present invention, incrementally updating the component state matrix of the initial digital twin using the state residual includes: calculating the deviation between the measured value of each monitoring point in the real-time feedback data and the corresponding simulated value in the initial digital twin, obtaining a deviation vector; inputting the deviation vector into a pre-trained state correction filter, the state correction filter outputting a state correction amount for each component; vector-superimposing the state correction amount with the current state value of the corresponding component in the component state matrix to obtain an updated component state matrix, and writing the updated component state matrix back to the initial digital twin. This method can achieve rapid calibration of the digital twin with relatively low computational cost, ensuring its real-time consistency.

[0017] As a preferred technical solution of the present invention, the process of generating an adaptive operation and maintenance strategy for each edge node based on the simulation results includes: taking the current state snapshot of the reconstructed digital twin as the simulation starting point, using a time-series prediction network to perform forward simulation of the reconstructed digital twin for multiple prediction steps to obtain a prediction state sequence corresponding to each prediction step; performing abnormal mode detection on each state in the prediction state sequence, and when an abnormal mode is detected, extracting the key component set and key parameter set that trigger the abnormal mode; and retrieving matching candidate operation and maintenance actions from a predefined operation and maintenance strategy library based on the key component set and key parameter set, and performing conflict resolution and priority sorting on the candidate operation and maintenance actions to generate the adaptive operation and maintenance strategy. Preferably, during the multi-step simulation simulation of the reconstructed digital twin, a confidence score is calculated for the prediction state sequence of each prediction step, and the simulation is terminated early and the system reverts to the previous known safe state when the confidence score is lower than the safety threshold, thereby avoiding erroneous operations caused by incorrect simulation and improving the reliability of the operation and maintenance strategy.

[0018] When retrieving matching candidate operation and maintenance actions, the key component set and key parameter set are combined into a query feature vector, and the cosine similarity between the query feature vector and each strategy feature vector in the operation and maintenance strategy library is calculated. The operation and maintenance actions corresponding to all strategy feature vectors with a cosine similarity exceeding a preset threshold are selected as the preliminary candidate set. For each operation and maintenance action in the preliminary candidate set, the set of preconditions for the operation and maintenance action is obtained. When all conditions in the set of preconditions are satisfied by the current edge node state, the operation and maintenance action is added to the final candidate operation and maintenance action set, thereby achieving accurate matching of operation and maintenance strategies.

[0019] As a preferred technical solution of the present invention, the process of realizing remote collaborative control of edge nodes includes: parsing the adaptive operation and maintenance strategy, decomposing the adaptive operation and maintenance strategy into multiple strategy primitives, each strategy primitive corresponding to an indivisible operation and maintenance operation; mapping each strategy primitive to a control instruction recognizable by the corresponding edge node execution unit, obtaining the atomic control instruction sequence, wherein the instructions in the atomic control instruction sequence are topologically ordered according to execution dependencies; encapsulating the topologically ordered atomic control instruction sequence into a transaction message, and pushing the transaction message to a designated topic partition of a distributed message queue, whereby the execution unit of the corresponding edge node pulls the transaction message from the designated topic partition and executes it sequentially. Preferably, an integrity verification field is appended to the end of the atomic control instruction sequence, and the execution unit performs integrity verification on the received instruction sequence before execution, further ensuring the integrity of the instructions and the security of execution.

[0020] As a further preferred embodiment of the present invention, the process of encapsulating the topologically sorted atomic control instruction sequence into a transaction message includes: generating a globally unique transaction identifier for the atomic control instruction sequence and writing the transaction identifier into the header of the transaction message; calculating the cyclic redundancy check (CRC) value of the atomic control instruction sequence and appending the CRC value to the tail of the transaction message; merging the transaction message carrying the transaction identifier and the CRC value with the atomic control instruction sequence to form a verifiable transaction message body. Through the transaction mechanism and the verification mechanism, reliable issuance and consistency assurance of control instructions are achieved.

[0021] As a preferred embodiment of the present invention, the process of collecting real-time operating status parameters of each edge node includes: deploying a lightweight monitoring agent on each edge node; collecting CPU utilization, memory usage, network throughput, and disk I / O latency of the edge node at a non-uniform sampling period through the lightweight monitoring agent to obtain a raw parameter stream; performing outlier removal on the raw parameter stream using a sliding window to obtain a cleaned parameter sequence; and performing max-min normalization on the cleaned parameter sequence to map the normalized parameter values ​​to a uniform numerical range to obtain the real-time operating status parameters. The above collection and preprocessing process can obtain high-quality operating parameters in a low-intrusive manner, providing an accurate data foundation for subsequent analysis.

[0022] This invention also provides an industrial internet remote monitoring and maintenance management system, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the aforementioned industrial internet remote monitoring and maintenance management method. This system enables intelligent remote monitoring and proactive maintenance of edge nodes, improving the overall operational reliability of the industrial internet.

[0023] The beneficial effects of this invention are:

[0024] By constructing a dynamic attribute graph structure and conducting federated transfer learning on multiple edge nodes, each edge node can participate in joint model training without uploading its original operating parameters. Each edge node uses a local graph convolutional network to extract local features from the dynamic attribute graph representing the node's operating state, and only uploads the encrypted graph embedding parameters to the federated aggregation server. The server then performs weighted aggregation to obtain the global graph embedding parameters and sends them back to each node for transfer fine-tuning, forming a cross-domain joint inference model. This mechanism, while strictly protecting data privacy, integrates the differentiated operating characteristics distributed across different nodes, enabling the model to learn common fault modes and collaborative state evolution patterns across nodes, significantly improving the accuracy and generalization ability of inference about edge node states.

[0025] In the construction of a digital twin, the physical entity of each edge node is decomposed at the component level, and component topology relationships and initial state variables are established to form an initial digital twin. When real-time feedback data is received, the state residual between the measured values ​​and the simulated values ​​is continuously extracted, and the state correction amount of each component is calculated through a state correction filter. The component state matrix is ​​adjusted in an incremental update manner to achieve closed-loop dynamic reconstruction of the physical entity state by the digital twin. This reconstruction method enables the digital twin to synchronize in real time with the physical entity's operational drift, performance degradation, or environmental changes, always maintaining high fidelity. Based on the reconstructed digital twin, multi-step simulation is performed starting from the current state snapshot, and anomaly pattern detection is performed on the simulated state sequence. When an anomaly is detected, key components and parameters are extracted, and matching actions are retrieved from the operation and maintenance strategy library to resolve conflicts and generate an adaptive operation and maintenance strategy. Through the real-time feedback-driven reconstruction mechanism and multi-step simulation, the operation and maintenance strategy can proactively anticipate potential faults and generate time-specific operation and maintenance actions before anomalies occur, effectively reducing the risk of unplanned downtime and improving the accuracy and autonomous decision-making level of remote operation and maintenance. Attached Figure Description

[0026] The invention will now be further described with reference to the accompanying drawings.

[0027] Figure 1 This is a flowchart of the industrial internet remote monitoring and operation and maintenance management method;

[0028] Figure 2 This is a flowchart of the construction process of the dynamic attribute graph structure of edge nodes;

[0029] Figure 3 This is a schematic diagram of the closed-loop dynamic reconstruction process of a digital twin;

[0030] Figure 4 This is a flowchart of the adaptive operation and maintenance strategy processing. Detailed Implementation

[0031] 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.

[0032] See Figure 1 This invention provides a method for remote monitoring and operation and maintenance management of industrial internet, comprising: responding to access requests from multiple edge nodes in the industrial internet, collecting real-time operating status parameters of each edge node, and constructing a heterogeneous data graph of the real-time operating status parameters to obtain a dynamic attribute graph structure characterizing the operating status of the edge nodes; based on the dynamic attribute graph structure, performing federated transfer learning on the multiple edge nodes to obtain a cross-domain joint inference model that integrates local features of multiple edge nodes; constructing a dynamically reconfigurable digital twin corresponding to each edge node according to the cross-domain joint inference model, and performing closed-loop dynamic reconstruction of the dynamically reconfigurable digital twin using real-time feedback data from the edge nodes to obtain a reconfigured digital twin; performing multi-step simulation and deduction of the future operating status of the corresponding edge nodes based on the reconfigured digital twin, and generating an adaptive operation and maintenance strategy for each edge node based on the deduction results; decomposing the adaptive operation and maintenance strategy into atomic control instruction sequences, and sending the atomic control instruction sequences to the execution units of the corresponding edge nodes through an asynchronous message queue to realize remote collaborative control of the edge nodes.

[0033] In specific implementation, please refer to Figure 2 A lightweight monitoring agent is deployed on each edge node. This agent collects CPU utilization, memory usage, network throughput, and disk I / O latency data at each edge node using a non-uniform sampling period. The non-uniform sampling period is determined as follows: when the rate of change of CPU utilization exceeds a preset threshold, the sampling period is shortened; when the rate of change of CPU utilization is below the preset threshold, the sampling period is extended. The lower limit of the sampling period is 100 milliseconds, and the upper limit is 10 seconds. The collected CPU utilization, memory usage, network throughput, and disk I / O latency data are arranged according to the collection timestamp to form a raw parameter stream.

[0034] A sliding window outlier removal operation is performed on the original parameter stream. The sliding window size is set to 100 sampling points, and the sliding step size is set to 1 sampling point. For each parameter sequence within the sliding window, the median of the parameter sequence and the absolute deviation of each parameter value from the median are calculated. Then, the median of absolute deviations is calculated based on all absolute deviations. Parameter values ​​with an absolute deviation greater than 5 times the median of absolute deviations are identified as outliers and removed from the original parameter stream. The outlier positions are marked as empty. After removing outliers, the cleaned parameter sequence is obtained.

[0035] The cleaned parameter sequence is then subjected to max-min normalization, mapping the normalized parameter values ​​to a uniform numerical range [0,1] to obtain the real-time operating status parameters. The calculation formula for max-min normalization is as follows:

[0036]

[0037] in: This represents the normalized real-time running status parameter value. This represents a single parameter value in the cleaned parameter sequence. This represents the minimum value of the parameter within a preset historical time window in the cleaned parameter sequence. This indicates the maximum value of the parameter within a preset historical time window in the cleaned parameter sequence. The preset historical time window length is 600 seconds. This represents the lower limit of the normalization target value range, with a value of 0. This represents the upper limit of the normalization target value range, and its value is 1.

[0038] The real-time running status parameters undergo time series alignment, using a 1-second time granularity to align the timestamps of all parameters to the nearest whole second. If multiple parameter values ​​exist at the same whole second, the average of these values ​​is taken as the aligned parameter value. After time series alignment, missing time points are imputed using linear interpolation: for a missing parameter value between two known parameter values, a weighted average is calculated based on the time distance between the two known parameter values ​​before and after the missing value to obtain the imputed value. After imputed missing values, the aligned multidimensional parameter sequence is obtained.

[0039] Each parameter in the aligned multidimensional parameter sequence is semantically labeled, and a parameter semantic type mapping table is pre-established, storing the correspondence between parameter names and semantic types. Semantic types include CPU load, memory usage, network throughput, and disk I / O. Each parameter in the aligned multidimensional parameter sequence is traversed, and its corresponding semantic type is looked up in the parameter semantic type mapping table based on its name. The parameter name, parameter value, parameter acquisition timestamp, and the retrieved semantic type are combined to form a parameter attribute triple. The specific format of the parameter attribute triple is: {Parameter name: CPU utilization, Parameter value: 0.75, Parameter acquisition timestamp: 2025-01-01-12:00:01, Semantic type: CPU load}.

[0040] Each edge node is used as a central node, and all parameter attribute triples generated on the edge nodes are used as the attribute feature vector of the central node. The attribute feature vector is organized using a key-value pair list structure, where each key-value pair stores a parameter name and its corresponding normalized parameter value. The communication link status between different edge nodes is used as edge features, including link throughput, link latency, and link packet loss rate. A dynamic attribute graph structure is constructed, defined as G=(V,E,A,X), where V represents the node set, which contains the central nodes corresponding to all edge nodes; E represents the edge set, where each edge corresponds to a communication link between two edge nodes; A represents the adjacency matrix, where the element values ​​are calculated by weighting the corresponding communication link status; and X represents the node attribute matrix, where each row stores the attribute feature vector of a central node. The dynamic attribute graph structure is updated over time. Whenever a new parameter attribute triple is generated, the value of the same parameter name in the corresponding central node's attribute feature vector is overwritten, and the edge feature values ​​in the adjacency matrix are recalculated, completing the update of the dynamic attribute graph structure.

[0041] In practice, the dynamic attribute graph structure is distributed to each edge node participating in federated learning. The distribution process is as follows: the federated aggregation server serializes the latest version of the dynamic attribute graph structure using a protocol buffer format, encrypts it through a secure transport layer protocol, and then broadcasts it to all edge nodes participating in federated learning. After receiving the dynamic attribute graph structure, each edge node performs a deserialization operation to recover the node attribute matrix and adjacency matrix.

[0042] A local graph convolutional network (GCNN) is used at each edge node to extract local features from the dynamic attribute graph structure. The GCNN structure employs a two-layer stacked GCNN design. The first GCNN layer receives the node attribute matrix and adjacency matrix as input, has 128 neurons, and uses a rectified linear function (CLM) as its activation function. The second GCNN layer takes the output of the first GCNN layer as input, has 64 neurons, and also uses a CLM. Following the second GCNN layer is a fully connected layer with a 32-dimensional output, which represents the local feature vector extracted by the GCNN for each edge node.

[0043] The training process of this map convolutional network employs supervised learning. At each edge node, each center node in the dynamic attribute graph structure is used as a training sample, with the attribute feature vector of the center node serving as the input feature. The operational state label for each center node is provided using the local supervised data stored within the edge nodes themselves, including both normal and abnormal states. The loss function of this map convolutional network is the cross-entropy loss function, and the optimizer is an adaptive moment estimation optimizer with an initial learning rate of 0.001. This map convolutional network is trained locally for a preset number of epochs, set to 50 epochs. Each epoch uses full-batch gradient descent. After completing the preset number of epochs, all weight and bias parameters of the local map convolutional network are saved as local map embedding parameters.

[0044] The local graph embedding parameters of each edge node are encrypted and uploaded to the federated aggregation server. The encryption upload process uses a homomorphic encryption algorithm. Each edge node generates a public-private key pair, uses the public key to encrypt the local graph embedding parameters, and uploads the encrypted local graph embedding parameters along with the public key to the federated aggregation server. After receiving the encrypted local graph embedding parameters uploaded from all participating edge nodes, the federated aggregation server performs a weighted aggregation of all local graph embedding parameters within the encrypted domain.

[0045] The formula for weighted aggregation performed by the federated aggregation server is:

[0046]

[0047] in: Indicates global graph embedding parameters. This represents the total number of edge nodes participating in federated learning. This represents the index of the edge node participating in federated learning. Indicates the index number is The number of local supervision data samples stored on the edge nodes. This represents the total number of locally supervised data samples stored on all edge nodes participating in federated learning. By all Summing them up gives us the result. Indicates the index number is The edge nodes upload local graph embedding parameters. The weighted aggregation operation is implemented in the encrypted domain using the homomorphic addition and homomorphic scalar multiplication properties of the homomorphic encryption algorithm. The federated aggregation server calculates the encrypted global graph embedding parameters without decrypting any local graph embedding parameters.

[0048] The federated aggregation server sends encrypted global graph embedding parameters back to each edge node. Each edge node decrypts the received encrypted global graph embedding parameters using its stored private key to obtain the plaintext global graph embedding parameters. Each edge node then performs transfer fine-tuning of the global graph embedding parameters using its stored local supervision data. The transfer fine-tuning process involves each edge node initializing all weight and bias parameters of its local graph convolutional network to the global graph embedding parameters, and then fine-tuning the local graph convolutional network using its stored local supervision data. The optimizer used for fine-tuning is an adaptive moment estimation optimizer, with a learning rate of 0.0001 and 10 epochs. After fine-tuning, the fine-tuned local graph convolutional network is saved as a cross-domain joint inference model. This cross-domain joint inference model is used to output the local feature vector of each edge node based on any input dynamic attribute graph structure, thereby completing subsequent inference tasks.

[0049] In specific implementation, please refer to Figure 3 The physical entity of each edge node is decomposed at the component level. A physical entity refers to the industrial field equipment corresponding to the edge node. Component-level decomposition is based on the hardware module division and functional unit division of the physical entity. Hardware module division includes central processing unit modules, memory modules, network interface modules, disk storage modules, and power supply modules; functional unit division includes data processing units, communication management units, and sensing and acquisition units. A component list is established for each physical entity, recording the component identifier, component name, component type, and component functional description for each component. Based on the hardware connection relationships and signal transmission relationships of the physical entities, the connection relationships between components are determined, generating component topology relationships. Component topology relationships are represented in the form of a directed acyclic graph (DAG). Vertices in the DAG represent components, and directed edges represent data or control flow transmission between components, with the direction of the directed edges indicating the direction of data or control flow transmission.

[0050] Initial state variables are assigned to each component based on the cross-domain joint inference model. The cross-domain joint inference model is a fine-tuned local graph convolutional network (Graph Convolutional Network), consisting of two graph convolutional layers and one fully connected layer. The dynamic attribute graph structure is input into the cross-domain joint inference model, which outputs a 32-dimensional local feature vector for each edge node. A linear mapping is performed on the local feature vectors, transforming the 32-dimensional vectors into initial state variable vectors equal in number to the number of components in the component list. This linear mapping uses a trainable weight matrix with a shape equal to the number of components multiplied by 32. The values ​​of the weight matrix are synchronously optimized and updated during the fine-tuning training of the cross-domain joint inference model. The optimization method involves adding a state variable prediction error term to the loss function, calculated using the mean squared error loss. Each component is assigned an initial state variable, the value of which is the element value at the corresponding position in the linearly mapped initial state variable vector.

[0051] An initial digital twin, mapped one-to-one with the physical entity, is constructed in a virtual space. The virtual space serves as the digital twin's runtime engine, running on cloud computing nodes. This engine is built upon a physical engine and component simulation models. The initial digital twin comprises a component state matrix and a component connection matrix. The component state matrix has the shape of the number of component rows multiplied by the number of state dimension columns, with three columns representing the component's operational health, load level, and energy efficiency index. The first column of the i-th row of the component state matrix stores the initial value of the i-th component's operational health, the second column of the i-th row stores the initial value of the i-th component's load level, and the third column of the i-th row stores the initial value of the i-th component's energy efficiency index. The initial state variables of the i-th component are directly assigned to the three state dimension columns of the i-th row of the component state matrix: the initial value of operational health is the value of the initial state variable, the initial value of load level is the value of the initial state variable multiplied by 0.5, and the initial value of energy efficiency index is the value of the initial state variable multiplied by 0.8. The component connection matrix is ​​an adjacency matrix generated based on the topological relationships of components. The shape of the component connection matrix is ​​the number of rows of components multiplied by the number of columns of components. If there is a directed edge between the i-th component and the j-th component, the element in the i-th row and j-th column of the component connection matrix is ​​1; otherwise, the element in the i-th row and j-th column of the component connection matrix is ​​0.

[0052] During the operation of the digital twin runtime engine, real-time feedback data is received from edge nodes. This real-time feedback data is collected by lightweight monitoring agents on the edge nodes and includes sensor readings such as CPU utilization, memory usage, network throughput, disk I / O latency, and temperature and power consumption of various components. The digital twin runtime engine compares the measured value of each monitoring point in the real-time feedback data with the corresponding simulated value in the initial digital twin. The corresponding simulated value is calculated by the initial digital twin based on the internal component simulation model at the current moment. The state residual between the real-time feedback data and the initial digital twin is extracted. The state residual is calculated as follows: for each monitoring point, the measured value is subtracted from the corresponding simulated value to obtain a difference. The differences of all monitoring points are aggregated into a state residual vector. The dimension of the state residual vector is equal to the total number of monitoring points.

[0053] The norm of the state residual vector is calculated using the L2 norm, which is the square root of the sum of the squares of all components in the state residual vector. When the L2 norm of the state residual exceeds a preset reconstruction threshold, a closed-loop dynamic reconstruction of the initial digital twin is triggered. The preset reconstruction threshold is set by collecting historical data of the state residual vectors from edge nodes during normal operation for 72 consecutive hours, calculating the L2 norm values ​​of all state residual vectors, and then calculating the mean of the L2 norm values. and standard deviation , set the preset reconstruction threshold Set as The triggering condition is expressed as follows:

[0054]

[0055] in: Represents the state residual vector. Represents the state residual vector L2 norm, This indicates the preset reconstruction threshold. This represents the mean of the historical L2 norm values. The standard deviation of the historical L2 norm values ​​is represented.

[0056] After triggering closed-loop dynamic reconstruction, the component state matrix of the initial digital twin is incrementally updated using the state residuals. The first step of the incremental update is to calculate the deviation between the measured value of each monitoring point in the real-time feedback data and the corresponding simulated value in the initial digital twin, thus obtaining a deviation vector. The deviation vector and the state residual vector are numerically identical, and each element in the deviation vector corresponds to the difference between the measured value and the simulated value of a monitoring point.

[0057] The second step of incremental updates involves inputting the bias vector into a pre-trained state correction filter. The state correction filter is a pre-trained feedforward neural network with an architecture consisting of an input layer, three hidden layers, and an output layer connected sequentially. The number of neurons in the input layer is equal to the dimension of the bias vector. The first hidden layer has 128 neurons, activated by a linear rectified function (RCF). The second hidden layer has 64 neurons, also activated by an RCF. The third hidden layer has 32 neurons, activated by an RCF. The output layer has three neurons, equal to the number of components, and does not use an activation function. The output vector is transformed to obtain a matrix with the number of rows multiplied by three columns; this matrix represents the state correction matrix for each component. The training process for the state correction filter involves collecting multiple sets of historical incremental update data. Each set contains a historical bias vector and its corresponding manually labeled true state correction values. These true state correction values ​​are obtained by engineers based on the historical bias vectors, analyzing the actual degradation level or state offset of each component. The historical deviation vector is used as input, and the true value of the state correction is used as the supervision signal. The filter is trained using the mean squared error loss function and an adaptive moment estimation optimizer. The batch size is 32, the initial learning rate is 0.001, and the pre-trained state correction filter is obtained after 100 training epochs. After training, the weight and bias parameters of the state correction filter are fixed and saved.

[0058] The third step of incremental updates is to vector-superimpose the state correction value of each component in the state correction matrix with the current state value of the corresponding component in the component state matrix. The vector superposition is performed as follows: for the i-th component, the three state dimension columns of the i-th row of the component state matrix correspond to operational health, load level, and energy efficiency index, respectively; the three values ​​of the i-th row of the state correction matrix correspond to the operational health correction, load level correction, and energy efficiency index correction, respectively. The current operational health value is added to the operational health correction to obtain the updated operational health value; the current load level value is added to the load level correction to obtain the updated load level value; and the current energy efficiency index value is added to the energy efficiency index correction to obtain the updated energy efficiency index value. After traversing all components, the updated component state matrix is ​​obtained.

[0059] The updated component state matrix is ​​written back to the initial digital twin, overwriting the original component state matrix in the initial digital twin, resulting in the reconstructed digital twin. The reconstructed digital twin continues to perform runtime simulations to reflect the latest operational state of the physical entities of the edge nodes.

[0060] In practice, a snapshot of the current state of the reconstructed digital twin is obtained as the starting point for simulation. The reconstructed digital twin includes a component state matrix and a component connection matrix, and the current state snapshot is a complete copy of the component state matrix at this simulation moment. The current state snapshot is converted into a simulation starting state vector, which is formed by expanding the component state matrix row by row and concatenating the vectors. The dimension is equal to the number of components multiplied by the number of state dimension columns.

[0061] A temporal prediction network is used to perform forward extrapolation of the reconstructed digital twin with multiple prediction steps. The temporal prediction network adopts an autoregressive Transformer model architecture, which consists of an encoder and a decoder. The encoder contains six identical encoder layers, each containing a multi-head self-attention sublayer and a fully connected feedforward sublayer. The multi-head self-attention sublayer has eight attention heads, each with a dimension of 64. The fully connected feedforward sublayer contains two linear transformation layers: the first has an output dimension of 2048, and the second has an output dimension of 512. The activation function is a linear rectified function. Each sublayer is followed by a residual connection and a layer normalization operation. The decoder contains six identical decoder layers, each containing a masked multi-head self-attention sublayer, an encoder-decoder multi-head attention sublayer, and a fully connected feedforward sublayer. The masked multi-head self-attention sublayer has eight attention heads, each with a dimension of 64. The masking operation ensures that the prediction at position i depends only on the outputs before position i. The encoder-decoder multi-head attention sublayer has 8 attention heads, each with a dimension of 64. The query vector comes from the output of the previous layer of the decoder, while the key and value vectors come from the output of the encoder. The structure of the fully connected feedforward sublayer is the same as that in the encoder.

[0062] The input to the temporal prediction network is the superposition of the simulation start-state vector and the prediction position encoding. The dimension of the simulation start-state vector is the number of components multiplied by 3, and the prediction position encoding uses a sinusoidal position encoding method. The encoder encodes the simulation start-state vector into a context hiding representation sequence. The length of the context hiding representation sequence is equal to the length of the simulation start-state vector, and the dimension of each element in the context hiding representation sequence is 512. The decoder generates the prediction state vector for the next prediction step in an autoregressive manner. When generating the t-th prediction step, the decoder takes the prediction state vectors of the 1st to t-1th prediction steps as input and receives the context hiding representation sequence output by the encoder, outputting the prediction state vector for the t-th prediction step. The dimension of the prediction state vector is the same as that of the simulation start-state vector, i.e., the number of components multiplied by 3. The prediction state vector corresponding to each prediction step is converted into a component state matrix form, resulting in the prediction state sequence corresponding to that prediction step. The prediction state sequence contains the set of component state matrices for all prediction steps.

[0063] The temporal prediction network is trained using supervised learning. Training samples are constructed using historical data. Each training sample contains a sequence of component state matrices for a continuous time window, with a time window length of 96 time steps. The component state matrix sequences for the first 48 time steps are used as input, and the component state matrix sequences for the last 48 time steps are used as output supervision signals. The loss function is the mean squared error loss, which is the average of the sum of squared differences between all predicted and true state vectors in the last 48 time steps. The optimizer is an adaptive moment estimation optimizer, with an initial learning rate of 0.0005, a batch size of 16, and 200 training epochs.

[0064] During multi-step simulation, a confidence score is calculated for the predicted state sequence at each prediction step. The confidence score is calculated based on the predicted probability distribution of the output layer of the temporal prediction network. When outputting each predicted state vector, the decoder of the temporal prediction network simultaneously outputs a variance vector, with the same dimension as the predicted state vector. The value of each component in the variance vector represents the degree of uncertainty of the corresponding predicted state component; a higher value indicates greater prediction uncertainty. For the t-th prediction step, the confidence score is calculated as follows: The calculation formula is:

[0065]

[0066] in, This represents the confidence score for the t-th prediction step. This represents the dimension of the predicted state vector. It equals the number of components multiplied by 3. Indicates the index of a component in the predicted state vector. Let represent the variance of the l-th component in the predicted state vector at the t-th prediction step. This represents the preset upper limit of variance. The value of is determined based on the 99.9 quantile of the variance output by the time series prediction network on the validation set. It is set to the 99.9 quantile of the variance of all prediction step sizes on the validation set, ensuring that the variance of the vast majority of normal predictions is below this value. If the confidence score... If the score falls below the safety threshold, the simulation terminates prematurely and reverts to the previous known safe state. The safety threshold is set to 0.7, based on the following criteria: on the validation set of the time series prediction network, when the confidence score is below 0.7, more than 30% of the components in the predicted state vector deviate from the true state vector beyond the preset accuracy tolerance, making the simulation result unreliable.

[0067] Anomaly pattern detection is performed on each state in the predicted state sequence. Anomaly pattern detection uses a predefined anomaly pattern rule base for matching. This rule base stores multiple anomaly pattern rules, each containing a conditional expression and a corresponding anomaly type. The conditional expression is composed of logical operations on the element values ​​in the component state matrix; for example, "operational health less than 0.3 and load level greater than 0.9" corresponds to the overload degradation anomaly type. For each state in the predicted state sequence, the component state matrix is ​​traversed through all conditional expressions in the anomaly pattern rule base. When a conditional expression is satisfied by a value in the component state matrix, an anomaly pattern is detected, and the corresponding anomaly type is recorded. When an anomaly pattern is detected, the key component set and key parameter set that triggered the anomaly pattern are extracted. The key component set is a list of component identifiers referenced in the conditional expression, and the key parameter set is a list of state dimension names referenced in the conditional expression, including operational health, load level, and energy efficiency index.

[0068] Retrieve matching candidate operations and maintenance (O&M) actions from a predefined O&M strategy library. The O&M strategy library is a pre-built database storing multiple O&M action records. Each O&M action record includes an O&M action identifier, O&M action name, strategy feature vector, set of preconditions, and action execution body. The strategy feature vector is a fixed-length numerical vector. It is constructed by representing the target components and control parameters involved in the O&M action using pre-trained component embedding vectors and parameter embedding vectors. The component embedding vectors and parameter embedding vectors are obtained through embedding matrix lookup. The component embedding vector has a dimension of 64, and the parameter embedding vector has a dimension of 32. The component embedding vectors and parameter embedding vectors are concatenated and mapped to a 128-dimensional strategy feature vector through a fully connected layer.

[0069] During retrieval, the key component set and key parameter set are combined into a query feature vector. The query feature vector is constructed in the same way as the strategy feature vector: for each component identifier in the key component set, the component embedding vector is obtained through an embedding matrix lookup; for each parameter name in the key parameter set, the parameter embedding vector is obtained through an embedding matrix lookup. All retrieved component embedding vectors and parameter embedding vectors are averaged and pooled to obtain an average component embedding vector and an average parameter embedding vector. These two vectors are then concatenated and mapped to a 128-dimensional query feature vector through the same fully connected layer. The cosine similarity between the query feature vector and each strategy feature vector in the operation and maintenance strategy library is calculated. The cosine similarity range is [-1, 1]. Operation and maintenance actions corresponding to all strategy feature vectors with cosine similarity exceeding a preset threshold are selected as the initial candidate set. The preset threshold is set to 0.8, based on the following principle: in offline evaluation, when the cosine similarity is greater than 0.8, the matching rate between the retrieved operation and maintenance actions and the manually labeled expected operation and maintenance actions reaches over 95%.

[0070] For each operation and maintenance action in the initial candidate set, obtain the set of preconditions for that action. The precondition set consists of multiple precondition expressions, each used to determine whether the current edge node state meets the execution prerequisites for the operation and maintenance action. The current edge node state is provided by the component state matrix of the reconstructed digital twin. When all precondition expressions in the precondition set are satisfied by the current edge node state, the operation and maintenance action is added to the final candidate operation and maintenance action set.

[0071] The candidate O&M actions in the final candidate O&M action set undergo conflict resolution and priority ranking. Conflict resolution involves checking if any two candidate O&M actions modify the same component state dimension. If modifications to the same component state dimension exist, a conflict is identified, the higher-priority action is retained, and the lower-priority action is removed. Priority ranking is based on: each O&M action is pre-assigned an initial priority value, determined by domain experts based on the urgency and impact of the O&M action. The initial priority value ranges from 1 to 10, with higher values ​​indicating higher priority. For O&M actions of the same priority level, a secondary ranking is performed according to the order in which the abnormal mode was triggered, with the O&M action corresponding to the earlier triggered abnormal mode appearing first. After conflict resolution and priority ranking, an ordered adaptive O&M strategy list is obtained, containing the O&M action identifiers to be executed in sequence.

[0072] In specific implementation, please refer to Figure 4An adaptive operation and maintenance (O&M) strategy consists of an ordered list of O&M action identifiers, each corresponding to an O&M action record in the O&M strategy library. During syntax parsing of the adaptive O&M strategy, a predefined strategy syntax rule table is used, which stores the mapping relationship between O&M action identifiers and strategy primitive templates. A strategy primitive template is predefined as a five-tuple structure, containing an operation object field, an operation type field, an operation parameter field, a precondition field, and a postcondition field. The operation object field specifies the component identifier targeted by the operation; the operation type field specifies the operation type number; the operation parameter field specifies the list of parameter values ​​required for the operation; the precondition field specifies the state constraint expression that must be satisfied before executing the operation; and the postcondition field specifies the expected state assertion after the operation is executed. The syntax parsing process traverses each O&M action identifier in the adaptive O&M strategy, searches for the corresponding strategy primitive template in the strategy syntax rule table, fills the actual parameter values ​​carried by the O&M action identifier into the corresponding position of the strategy primitive template, and instantiates the complete strategy primitive. After parsing, the adaptive operation and maintenance strategy is broken down into multiple strategy primitives, each of which corresponds to an indivisible operation and maintenance operation.

[0073] After decomposing the policy primitives, each policy primitive is mapped to a control instruction recognizable by the corresponding edge node execution unit. The edge node execution unit has a pre-set control instruction set, where each control instruction contains an opcode and operands. A pre-set mapping table from policy primitives to control instructions is used, with the combination of the policy primitive's operation type field and operation object field as the lookup key, mapping to the specific opcode in the control instruction set. The mapping process is as follows: extract the operation type field and operation object field values ​​of the policy primitive, perform an exact match lookup in the mapping table to obtain the corresponding opcode; fill the sequence of parameter values ​​from the policy primitive's operation parameter field into the instruction operand part in sequence to form a complete control instruction. After mapping all policy primitives, an atomic control instruction sequence is obtained, and each control instruction in the atomic control instruction sequence can be directly parsed and executed by the corresponding edge node execution unit.

[0074] Instructions in an atomic control instruction sequence have execution dependencies, which include data dependencies and control dependencies. Data dependencies mean that the output operand of one instruction is used as the input operand of another instruction. Control dependencies mean that the execution of one instruction depends on the execution result of another instruction. A directed acyclic graph (DAG) reflecting these dependencies is constructed, where the vertices of the DAG represent the instructions in the atomic control instruction sequence, and the directed edges in the DAG represent dependencies from the dependent instruction to the dependent instruction. A topological sort is then performed on the DAG using the Kahn algorithm, repeatedly performing the following operations: selecting a vertex with an in-degree of zero from the DAG, outputting the instruction corresponding to that vertex into the sorted sequence, and deleting that vertex and all its outgoing edges. This process continues until all vertices have been output, resulting in the topologically sorted atomic control instruction sequence.

[0075] The topologically sorted sequence of atomic control instructions is encapsulated into a transaction message. The first step in the encapsulation process is to generate a globally unique transaction identifier for the topologically sorted sequence of atomic control instructions. The transaction identifier is generated using the Universally Unique Identifier Version 4 (UUUID 4) algorithm. A random number generator produces a 128-bit random number, which is then formatted into a 36-bit string conforming to the UUUID 4 format, serving as the transaction identifier. The transaction identifier is written into the transaction message header. The transaction message header is a fixed-length key-value pair structure, where the key is named "transaction_id" and the value is the transaction identifier string.

[0076] The second step in the encapsulation process is to calculate the cyclic redundancy check (CRC) value of the topologically sorted atomic control instruction sequence. The CRC-32 algorithm is used to calculate the CRC value, and the generator polynomial is: The topologically sorted sequence of atomic control instructions is treated as a continuous byte stream. Each byte is input into the CRC-32 calculation logic, and after shifting and XOR operations, a 32-bit cyclic redundancy check (CRC) value is obtained. This 32-bit CRC value is appended to the end of the transaction message, which occupies 4 bytes.

[0077] The third step in the encapsulation process is to merge the transaction message header carrying the transaction identifier, the topologically sorted sequence of atomic control instructions, and the cyclic redundancy check value appended to the tail to form a verifiable transaction message body. The structure of the verifiable transaction message body, from beginning to end, is as follows: transaction message header, sequence of atomic control instructions, and cyclic redundancy check value.

[0078] An integrity check field is appended to the end of the atomic control instruction sequence. The integrity check field is calculated by performing a sequential XOR operation on all bytes in the topologically sorted atomic control instruction sequence. The formula for calculating the integrity check field is:

[0079]

[0080] in, This represents the integrity verification field. This represents the bitwise XOR operator. This represents the sequential index from the first byte to the last byte in the topologically sorted sequence of atomic control instructions. This represents the total number of bytes in the topologically sorted sequence of atomic control instructions. This indicates that the index of the atomic control instruction sequence after topological sorting is... The integrity check field is 1 byte (8 bits) long and ranges from 0x00 to 0xFF. The calculated integrity check field is appended to the end of the topology-sorted atomic control command sequence. This integrity check field is used by the edge node execution unit to perform integrity checks on the received atomic control command sequence before execution. The encapsulated transaction message is pushed to the specified topic partition of the distributed message queue. The distributed message queue is built using the Apache Kafka distributed stream processing platform. The naming convention for the specified topic partition is: the topic name is "edge-control-commands", and the partition number is obtained by taking the hash value of the edge node's node identifier modulo the total number of partitions. After the transaction message is serialized into a byte stream, it is sent to the corresponding topic partition through the Kafka producer client.

[0081] The execution unit on the corresponding edge node pulls transaction messages from the specified topic partition. Each edge node execution unit runs a Kafka consumer client, which subscribes to the specified topic partition and pulls transaction messages from the distributed message queue using long polling. After pulling the transaction messages, the execution unit parses the verifiable transaction message body, extracting the transaction message header, atomic control instruction sequence, and cyclic redundancy check (CRC) value. The execution unit recalculates the CRC value based on the extracted atomic control instruction sequence and compares the recalculated CRC value with the CRC value appended to the end of the transaction message. If they do not match, the transaction message is discarded and a CRC failure alarm is reported.

[0082] After the cyclic redundancy check passes, the execution unit uses the integrity check field to perform integrity checks on the atomic control instruction sequence. The integrity check process is as follows: extract the integrity check field appended to the end of the atomic control instruction sequence, perform a re-XOR operation on all bytes of the extracted atomic control instruction sequence in sequence, compare the result with the extracted integrity check field, if they match, the integrity check is considered to have passed; if they do not match, the atomic control instruction sequence is discarded and an integrity check failure alarm is reported.

[0083] After the integrity check passes, the edge node execution unit executes each control instruction sequentially according to the atomic control instruction sequence. Each control instruction is decoded by the instruction decoder inside the execution unit. The instruction decoder looks up the micro-operation sequence based on the instruction opcode and distributes the micro-operation sequence to the corresponding hardware execution unit to complete the operation. After all control instructions have been executed sequentially, the edge node execution unit reports an execution completion confirmation message to the industrial internet remote monitoring and maintenance management system.

[0084] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for remote monitoring, operation and maintenance management of industrial internet, characterized in that, include: In response to access requests from multiple edge nodes in the Industrial Internet, real-time operating status parameters of each edge node are collected, and heterogeneous data graphs are constructed from these real-time operating status parameters to obtain a dynamic attribute graph structure representing the operating status of the edge nodes. Based on the dynamic attribute graph structure, federated transfer learning is performed on the multiple edge nodes to obtain a cross-domain joint inference model that integrates the local features of multiple edge nodes. Based on the cross-domain joint reasoning model, a dynamically reconfigurable digital twin corresponding to each edge node is constructed, and the dynamically reconfigurable digital twin is dynamically reconfigured in a closed loop using real-time feedback data from the edge nodes to obtain the reconfigured digital twin. Based on the reconstructed digital twin, the future operating status of the corresponding edge nodes is simulated and deduced in multiple steps, and an adaptive operation and maintenance strategy for each edge node is generated based on the simulation results. The adaptive operation and maintenance strategy is decomposed into a sequence of atomic control instructions, and the sequence of atomic control instructions is sent to the execution unit of the corresponding edge node through an asynchronous message queue to realize remote collaborative control of the edge node.

2. The industrial internet remote monitoring and maintenance management method according to claim 1, characterized in that, The dynamic attribute graph structure representing the operating state of the edge nodes is obtained, including: The real-time operating status parameters are aligned by time series and missing values ​​are imputed to obtain an aligned multidimensional parameter sequence. Each parameter in the aligned multidimensional parameter sequence is semantically labeled to obtain a parameter attribute triplet, which includes the parameter name, parameter value and parameter acquisition timestamp. Using each edge node as the center node, the parameter attribute triplet is used as the attribute feature vector of the center node, and the communication link state between different edge nodes is used as the edge feature to construct the dynamic attribute graph structure.

3. The industrial internet remote monitoring and maintenance management method according to claim 1, characterized in that, The resulting cross-domain joint inference model that integrates local features from multiple edge nodes includes: The dynamic attribute graph structure is distributed to each edge node participating in federated learning. Local graph convolutional networks are used on each edge node to extract local features from the dynamic attribute graph structure, thereby obtaining the local graph embedding parameters of each edge node. The local graph embedding parameters of each edge node are encrypted and uploaded to the federated aggregation server. All the received local graph embedding parameters are weighted and aggregated on the federated aggregation server to obtain the global graph embedding parameters. The global graph embedding parameters are fed back to each edge node, and each edge node uses its own stored local supervision data to perform migration fine-tuning of the global graph embedding parameters to obtain the cross-domain joint inference model.

4. The industrial internet remote monitoring and maintenance management method according to claim 1, characterized in that, The reconstructed digital twin includes: The physical entity of each edge node is decomposed at the component level to obtain the component topology relationship of each physical entity, and initial state variables are assigned to each component according to the cross-domain joint inference model. Based on the component topology and the initial state variables, an initial digital twin is constructed in the virtual space that maps one-to-one with the physical entity. The initial digital twin includes a component state matrix and a component connection matrix. When real-time feedback data from edge nodes is received, the state residual between the real-time feedback data and the initial digital twin is extracted, and the component state matrix of the initial digital twin is incrementally updated using the state residual to obtain the reconstructed digital twin.

5. The industrial internet remote monitoring and maintenance management method according to claim 4, characterized in that, When the norm of the state residual exceeds a preset reconstruction threshold, a closed-loop dynamic reconstruction of the initial digital twin is triggered.

6. The industrial internet remote monitoring and maintenance management method according to claim 4, characterized in that, Incremental updates to the component state matrix of the initial digital twin using the state residuals include: The deviation vector is obtained by calculating the deviation between the measured value of each monitoring point in the real-time feedback data and the corresponding simulated value in the initial digital twin; The deviation vector is input into a pre-trained state correction filter, and the state correction filter outputs the state correction amount for each component. The state correction amount is vector-superimposed with the current state value of the corresponding component in the component state matrix to obtain the updated component state matrix, and the updated component state matrix is ​​written back to the initial digital twin.

7. The industrial internet remote monitoring and maintenance management method according to claim 1, characterized in that, Based on the simulation results, adaptive operation and maintenance strategies are generated for each edge node, including: Using the current state snapshot of the reconstructed digital twin as the simulation starting point, a time-series prediction network is used to perform forward extrapolation of the reconstructed digital twin for multiple prediction steps to obtain the prediction state sequence corresponding to each prediction step. For each state in the predicted state sequence, perform abnormal pattern detection. When an abnormal pattern is detected, extract the set of key components and the set of key parameters that trigger the abnormal pattern. Based on the set of key components and the set of key parameters, matching candidate operation and maintenance actions are retrieved from the predefined operation and maintenance strategy library, and conflict resolution and priority sorting are performed on the candidate operation and maintenance actions to generate the adaptive operation and maintenance strategy.

8. The industrial internet remote monitoring and maintenance management method according to claim 7, characterized in that, During the multi-step simulation of the reconstructed digital twin, a confidence score is calculated for the predicted state sequence for each prediction step. When the confidence score is lower than the safety threshold, the simulation is terminated early and the system reverts to the previous known safe state.

9. The industrial internet remote monitoring and maintenance management method according to claim 1, characterized in that, To achieve remote collaborative control of edge nodes, including: The adaptive operation and maintenance strategy is parsed and decomposed into multiple strategy primitives, each of which corresponds to an indivisible operation and maintenance operation. Each policy primitive is mapped to a control instruction that can be recognized by the corresponding edge node execution unit to obtain the atomic control instruction sequence. The instructions in the atomic control instruction sequence are topologically sorted according to the execution dependency relationship. The atomic control instruction sequence after topological sorting is encapsulated into a transaction message, and the transaction message is pushed to a designated topic partition of the distributed message queue. The execution unit of the corresponding edge node pulls the transaction message from the designated topic partition and executes it sequentially.

10. The industrial internet remote monitoring and maintenance management method according to claim 9, characterized in that, An integrity check field is appended to the end of the atomic control instruction sequence, and the execution unit performs integrity check on the received instruction sequence before execution.

Images