A power plant edge computing and 5g cooperative data transmission method

Abstract

The application discloses a power plant edge computing and 5G cooperative data transmission method, which comprises the following steps: obtaining original multi-modal data flow and analyzing task request, generating dynamic business feature vector, and selecting atomic processing unit to combine into an initial processing pipeline according to the feature vector; collecting 5G private network wireless channel state, backhaul link load, resource utilization rate of each edge computing node and interconnection state between nodes to generate a global resource state view; performing dynamic mapping and reconstruction on the initial processing pipeline based on the view and the feature vector, assigning appropriate nodes to each unit, planning transmission path and priority, controlling node execution and triggering online adaptive migration based on dynamic criticality level change; receiving intermediate results at the aggregation node, using attention mechanism to weight and fuse to generate final diagnostic results and output, and updating the global resource state view, so that dynamic cooperative optimization and adaptive scheduling of computing and transmission resources are realized.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and more specifically to a data transmission method for power plant edge computing and 5G collaboration. Background Technology

[0002] In the intelligent operation and maintenance of ultra-large coal-fired power plants, various types of sensors and cameras are typically deployed to collect multimodal data on equipment operating status to ensure safe and stable operation. With the development of 5G communication technology and edge computing, the industry has begun to explore the use of 5G private networks to achieve real-time transmission of massive monitoring data and perform preliminary processing at edge nodes close to the data source to reduce transmission latency and cloud load. Existing technical solutions typically optimize data transmission and computing tasks as relatively independent processes. For example, they may first plan the data transmission path based on network conditions and then allocate computing tasks based on node load, or adopt fixed task decomposition and node mapping strategies. While these solutions can improve local efficiency, they struggle to achieve global coordination and dynamic adaptation of computing resources and network transmission when facing dynamically changing business needs within the power plant (such as early warning of sudden faults) and complex network resource environments (such as wireless channel fluctuations and uneven node loads). Summary of the Invention

[0003] In view of the above problems, the present invention provides a data transmission method for power plant edge computing and 5G collaboration. By dynamically orchestrating and jointly optimizing the multimodal data processing pipeline based on dynamic service awareness and global resource status, the method realizes the collaborative scheduling and online adaptive migration of computing and transmission resources.

[0004] To achieve the above objectives, this application provides a data transmission method for power plant edge computing and 5G collaboration, comprising: Acquire raw multimodal data streams from heterogeneous sensing devices in the power plant and the corresponding analysis task requests; Based on the initial attributes of the analysis task request and the original multimodal data stream, a dynamic business feature vector is generated; Based on the business type in the dynamic business feature vector, select and combine from the pre-built atomic processing unit library to form the initial processing pipeline; Real-time collection of 5G private network wireless channel status, backhaul link load, resource utilization of each edge computing node and interconnection status between nodes, and fusion to generate a global resource status view; With the joint optimization goal of minimizing the overall task completion latency and maximizing resource utilization, dynamic mapping and reconstruction are performed on the initial processing pipeline based on the global resource status view and dynamic business feature vectors. Dynamic mapping and reconstruction include: Assign appropriate edge computing nodes as execution nodes to each atomic processing unit in the pipeline, and plan the transmission path and priority for the intermediate data flow between nodes based on the NSGA-II multi-objective optimization algorithm; Control each edge computing node to execute the corresponding atomic processing unit according to the mapping relationship, and trigger online adaptive migration of the transmission path and execution node according to the dynamic criticality level change in the dynamic business feature vector during the execution process; At the designated aggregation edge computing node, intermediate processing results from upstream nodes are received, and a weighted fusion analysis is performed using an attention mechanism to generate the final device status fusion diagnostic results. Output the fusion diagnostic results and update the global resource status view.

[0005] In some embodiments, a dynamic business feature vector is generated based on the initial attributes of the analysis task request and the original multimodal data stream, including: Parse and analyze the task request, extract the task identifier and expected output format, and generate initial business type labels based on the task identifier; Obtain the source device identifier, data format, and sampling rate of the raw multimodal data stream as initial attributes of the data stream; Based on the initial business type label and the initial data flow attributes, the initial latency tolerance limit and reliability requirements are determined by querying the predefined business type-service quality mapping table. Based on the initial business type labels and data format, the initial processing complexity estimate is calculated using a complexity prediction model. At the first edge computing node that receives the raw multimodal data stream, the pre-deployed pre-analysis atomic processing unit is invoked to perform real-time analysis on the data stream and obtain real-time analysis results. The real-time analysis results are matched with a pre-set abnormal pattern library, and the content criticality level is dynamically evaluated and output based on the matching confidence level. The initial business type label, initial latency tolerance limit, reliability requirements, initial processing complexity estimate, and content criticality level are vectorized and encapsulated to generate dynamic business feature vectors. Furthermore, during the execution of subsequent atomic processing units, if new real-time analysis results are received that cause a change in the content criticality level, a real-time update operation on the dynamic business feature vector is triggered.

[0006] In some embodiments, the real-time analysis results are matched against a preset abnormal pattern library, and the content criticality level is dynamically evaluated and output based on the matching confidence level, including: Extract structured feature descriptors from real-time analysis results. The structured feature descriptors include at least the anomaly type, spatial location, timestamp, and quantization parameters. The structured feature descriptor is input into a pre-built anomaly pattern library, which stores multi-dimensional feature templates of historical anomalies. Each feature template is associated with a preset key benchmark value. An attention-based feature matching algorithm is used to calculate the similarity between the structured feature descriptor and each feature template in the abnormal pattern library. Based on the similarity, the key benchmark values ​​associated with each feature template are weighted and fused to generate a preliminary key score. Obtain the data stream identifier corresponding to the real-time analysis results, and query the historical key score sequence of the data stream within a preset time window; The initial criticality score and the historical criticality score sequence are input into the criticality state transition model, which is built based on a hidden Markov chain and is used to smoothly evaluate and predict the evolution trend of criticality states. Based on the output of the critical state transition model, the final content criticality level at the current moment is determined. The content criticality level is used to characterize the value density and urgency of the information contained in the data stream for subsequent fusion decision-making tasks.

[0007] In some embodiments, the status of 5G private network wireless channels, backhaul link load, resource utilization of each edge computing node, and interconnection status between nodes are collected in real time and fused to generate a global resource status view, including: By using probes deployed on the 5G base station side, the reference signal received power and signal-to-interference-to-noise ratio of each area are periodically collected as the wireless channel status. By deploying link monitoring points between user plane functional network elements and edge computing nodes, the current uplink and downlink data throughput and transmission latency are collected, and the backhaul link load is calculated. By deploying resource monitoring agents on each edge computing node, the utilization rate of the central processing unit, the utilization rate of the graphics processing unit, and the memory usage rate are periodically collected as the node resource utilization rate. Based on node resource utilization and node hardware configuration information, a computing resource profile of each edge computing node is constructed. By running link layer discovery protocols or active probing mechanisms, the network topology connections between edge computing nodes are obtained, and the round-trip latency and packet loss rate of the links between nodes are measured to assess the interconnection status between nodes. The collected wireless channel status, backhaul link load, node resource utilization and inter-node interconnection status are cleaned and normalized to unify the heterogeneous indicators to the same numerical units. The normalized wireless channel state vector, backhaul link load scalar, computational resource profile vector of each node, and inter-node interconnection state matrix are encapsulated to generate a global resource state view.

[0008] In some embodiments, a computing resource profile for each edge computing node is constructed based on node resource utilization and node hardware configuration information, including: Read the hardware configuration information of the edge computing node. The hardware configuration information includes at least the CPU model and number of cores, the graphics processor model and video memory capacity, and the total memory capacity. Based on hardware configuration information, a predefined hardware performance benchmark scoring model is used to determine the benchmark computing performance scores of nodes in terms of central processing unit computing, graphics processing unit computing, and memory access. The CPU utilization, GPU utilization, and memory usage are periodically collected by the resource monitoring agent and used to construct a real-time utilization vector. Based on the utilization values ​​of each dimension in the real-time utilization vector and the corresponding benchmark computing performance scores, the real-time available computing power of the node in each dimension at the current moment is calculated; the real-time available computing power represents the processing capacity remaining after deducting the occupied resources that can be used for new tasks. The baseline computing performance score, real-time utilization vector, and real-time available computing power vector are combined and encapsulated to form a computing resource profile for the edge computing node.

[0009] In some embodiments, by running a link layer discovery protocol or an active probing mechanism, the network topology connections between edge computing nodes are obtained, and the round-trip latency and packet loss rate of the links between nodes are measured to evaluate the interconnection status between nodes, including: The resource monitoring agent deployed on each edge computing node integrates a network topology discovery module and a link quality detection module. The network topology discovery module periodically broadcasts link layer discovery protocol data frames in the local area network to which the edge computing node belongs, or sends topology query requests to a known list of neighboring node addresses. Receive link layer discovery protocol responses or topology information replies from other edge computing nodes, and parse the neighbor node identifiers, connection port information and hop counts contained in the response data; Based on the neighbor relationship information reported by all nodes, a dynamic edge node network topology is constructed and maintained in the intelligent decision-making center. The topology is based on nodes as vertices and physical or logical connections as edges. The link quality detection module periodically initiates bidirectional active detection for each link between nodes represented by an edge in the topology graph; active detection includes sending a sequence of probe data packets with timestamps. The receiving end records the arrival timestamp and sequence number of the probe data packets, and calculates the round-trip time and packet loss for a single probe. Based on all probe data packet samples collected during the probe period, after removing abnormal samples using a robust statistical algorithm, the average round-trip time and average packet loss rate of the link are calculated. The average round-trip time, average packet loss rate, and topological connectivity are collectively encapsulated into an inter-node interconnection state descriptor, which is used to characterize the inter-node interconnection state.

[0010] In some embodiments, a suitable edge computing node is assigned as an execution node to each atomic processing unit in the pipeline, and a transmission path and priority are planned for the intermediate data stream between nodes based on the NSGA-II multi-objective optimization algorithm, including: For each atomic processing unit in the initial processing pipeline, based on the node computing resource profile and the interconnection status between nodes in the global resource status view, edge computing nodes that meet its processing capabilities and communication requirements are selected to construct a candidate execution node set. With the optimization objectives of minimizing the overall task completion latency and maximizing the system resource load balance, a multi-objective optimization function is constructed. The calculation of the overall task completion latency requires a combination of the estimated execution latency of the atomic processing units on each candidate node and the estimated network latency of the intermediate data streams transmitted between nodes. The selection of candidate execution nodes for each atomic processing unit, as well as the selection of transmission paths between nodes, are collectively encoded as individual chromosomes of the optimization problem. Initialize the chromosome population by utilizing the content criticality levels in the global resource status view and dynamic business feature vectors; The NSGA-II multi-objective optimization algorithm is used to iteratively optimize the chromosome population, performing operations including non-dominated sorting, crowding calculation, selection, crossover and mutation, to approximate and obtain the Pareto optimal solution set; From the Pareto optimal solution set, based on the latency tolerance upper limit and reliability requirements in the dynamic service feature vector, the optimal solution is selected for decoding, including: The gene segments corresponding to the atomic processing units in the chromosome are decoded into defined edge computing nodes, which are then used as their execution nodes. Decode the gene segments corresponding to the inter-node transmission path into a defined path sequence and the transmission priority on that path; Output the mapping table between atomic processing units and execution nodes obtained from decoding, as well as the transmission path and priority planning table of intermediate data streams.

[0011] In some embodiments, a multi-objective optimization function is constructed with the optimization objectives of minimizing the overall task completion latency and maximizing the system resource load balancing, including: For each atomic processing unit in the initial processing pipeline, based on its processing complexity and the real-time available computing power in the candidate execution node computing resource profile, the estimated execution latency of the unit on each candidate node is calculated using a predefined execution latency prediction model. For each path that may transmit intermediate data streams, the estimated network latency of the path is calculated based on the transmission latency of the links traversed by the path and the current load status in the global resource status view, using a predefined network latency prediction model. The objective function for the overall task completion time is defined as follows: Starting from the atomic processing unit at the beginning of the pipeline, based on the decoded node mapping and path planning, the estimated execution time of each atomic processing unit and the estimated network time of the path through which its output data flows are accumulated in the order of execution, and the resulting total time value is used as the first optimization objective. The objective function for system resource load balancing is defined as follows: Based on the decoded node mapping relationship, estimate the future resource utilization rate of each edge computing node after undertaking the allocated atomic processing unit, calculate the variance of the future resource utilization rate of all nodes, and take the reciprocal of the variance as the second optimization objective. The multi-objective optimization function is composed of the objective function of the overall task completion delay and the objective function of system resource load balancing.

[0012] In some embodiments, the selection of candidate execution nodes corresponding to each atomic processing unit and the selection of inter-node transmission paths are jointly encoded as chromosome individuals of the optimization problem, including: The chromosome coding structure is determined. Chromosomes are composed of sequentially arranged gene segments, and each gene segment corresponds to an atomic processing unit in the initial processing pipeline. For each atomic processing unit, the corresponding gene segment is encoded using an integer encoding method. The value space of this integer is a list of index numbers of all nodes in its candidate execution node set. The gene value directly represents the execution node to which the atomic processing unit is assigned. In the chromosome, a path gene segment is added for each logical connection that needs to transmit intermediate data streams. The logical connection is determined by the execution node of the upstream atomic processing unit and the execution node of the downstream atomic processing unit. For each path segment, a variable-length integer sequence encoding method is used. Each integer in the sequence represents the index number of the edge computing node or network switching device that the path passes through. The entire sequence represents the complete transmission path from the upstream node to the downstream node. In the path gene segment, a priority coding bit is added. This priority coding bit adopts discrete value coding, and its value is associated with the content criticality level in the dynamic service feature vector. It is used to indicate the scheduling priority that the intermediate data stream should be assigned on the transmission path. The atomic processing unit gene segments, path gene segments, and priority encoding bits are spliced ​​together according to the execution logic order of the initial processing pipeline to form a complete chromosome individual. This chromosome individual fully represents the combination of the atomic processing unit execution node mapping scheme and the intermediate data stream transmission path and priority planning scheme.

[0013] In some embodiments, each edge computing node is controlled to execute the corresponding atomic processing unit according to the mapping relationship, and during the execution process, online adaptive migration of the transmission path and execution node is triggered based on the dynamic criticality level changes in the dynamic service feature vector, including: During the execution of the atomic processing unit, the dynamic business feature vector of the associated data stream is continuously monitored to detect whether the dynamic criticality level has jumped. If a dynamic criticality level jump is detected, the current execution node and transmission path are evaluated to determine whether they can still meet the quality of service constraints based on the latency and reliability requirements corresponding to the jump level. If the evaluation result is unsatisfactory, a migration decision process is initiated for the affected atomic processing units and their associated data streams based on the updated global resource status view, including: Using the improved quality of service constraints as a hard condition, the system quickly finds new execution nodes for the affected atomic processing units from the set of edge computing nodes that meet the conditions, and re-plans the transmission paths for their input and output data streams. Based on the new execution node and transmission path obtained through optimization, a migration instruction set containing atomic processing unit context migration and data stream switching timing is generated; Distribute the migration instruction set to relevant edge computing nodes and network forwarding devices; Coordinate relevant nodes and devices to perform operations such as migration of atomic processing unit computing state, transfer of intermediate data, and update of data stream forwarding path according to the timing specified by the instruction set, and complete online adaptive migration.

[0014] In some embodiments, using the upgraded quality of service constraint as a hard condition, from the set of edge computing nodes that meet the condition, a new execution node is quickly selected for the affected atomic processing unit, and its input and output data streams are re-planned for transmission, including: Based on the latency tolerance limit and reliability requirements corresponding to the upgraded dynamic criticality level, and combined with the updated global resource status view, nodes that meet the processing capacity requirements and the latency and reliability constraints in the computing resource profiles are selected from all edge computing nodes to form a set of feasible execution nodes. For each candidate node in the set of feasible execution nodes, evaluate the local cost of migrating the affected atomic processing unit to that node. The local cost includes at least the estimated time for migrating the atomic processing unit's computation context to that node, the estimated execution latency required for that node to process the unit, and the estimated network latency of the replanned input / output data flow path for that unit. While assessing the local costs, ensure that each link selected for the newly planned input / output data flow path meets the reliability requirements after the leap. From the set of feasible execution nodes, select the candidate node that minimizes the local cost and determine it as the new execution node; The input / output data flow paths planned for the new execution node during the evaluation process are determined as the replanned transmission paths.

[0015] In some embodiments, based on the optimized new execution node and transmission path, a migration instruction set containing atomic processing unit context migration and data stream switching timing is generated, including: The optimization results are analyzed to extract the new execution node identifier of the affected atomic processing unit, the new source node identifier and new transmission path of the input data stream, and the new target node identifier and new transmission path of the output data stream. Generate atomic processing unit context transition instructions. The atomic processing unit context transition instructions include the identifier of the computation state data to be transitioned, the source execution node address, the target execution node address, and the state transition triggering conditions. The computation state data includes at least the model parameters required for the atomic processing unit to run, the contents of the intermediate computation result buffer, and the session identifier. Generate a data stream switching preparation instruction. The data stream switching preparation instruction is sent to the current source node of the input data stream, instructing it to simultaneously copy and cache the data packets to be forwarded to the exit queues corresponding to the old and new transmission paths within a specified time window. A path update command is generated and sent to the network forwarding devices involved in the new transmission path. This command is used to pre-configure flow table rules, which specify that data packets matching the data flow identifier should be forwarded to the new execution node. Generate timing synchronization instructions, which specify the execution order and timestamps of context migration instructions, data stream switching preparation instructions, and path update instructions; timing synchronization instructions ensure that after the computation state is restored in the new execution node, the flow table rules on the new transmission path have taken effect, and the input data stream has started to be transmitted along the new path. The atomic processing unit context migration instructions, data stream switching preparation instructions, path update instructions, and timing synchronization instructions are encapsulated and serialized to form a migration instruction set.

[0016] In some embodiments, at a designated aggregation edge computing node, intermediate processing results from upstream nodes are received, and a weighted fusion analysis is performed using an attention mechanism to generate the final device state fusion diagnostic result, including: At the aggregated edge computing node, multiple intermediate processing results corresponding to the same monitored device are received from different upstream nodes; The received intermediate processing results are timestamped and their data formats are standardized to ensure data consistency in both time and semantic dimensions. Extract feature vectors representing the device state from each intermediate processing result, and concatenate all feature vectors to form a fused feature matrix; Meanwhile, metadata is extracted for each intermediate processing result. The metadata includes at least the confidence level of the atomic processing unit that generated the result and the real-time load status of the source edge computing node. The fused feature matrix and metadata are input into the attention weight calculation network. The attention weight calculation network calculates a dynamic attention weight for each intermediate processing result based on the correlation between feature vectors and the reliability of metadata representation. Using the calculated attention weights, the eigenvectors in the fusion feature matrix are summed in a weighted manner to obtain the weighted fusion feature vector. The weighted fusion feature vector is input into the pre-trained device status diagnosis classification model for inference analysis; The inference results of the output device status diagnosis classification model are used as the final device status fusion diagnosis results.

[0017] In some embodiments, the fused feature matrix and metadata are jointly input into the attention weight calculation network. Based on the correlation between feature vectors and the reliability of the metadata representation, the attention weight calculation network calculates a dynamic attention weight for each feature vector corresponding to an intermediate processing result, including: Each feature vector in the fused feature matrix is ​​linearly transformed through a shared fully connected layer to generate the corresponding query vector, key vector, and value vector, respectively. The metadata accompanying each intermediate processing result is encoded by an independent metadata encoder to generate a metadata embedding vector that characterizes the reliability of the intermediate processing result. The key vector is concatenated with the corresponding metadata embedding vector to form an enhanced key vector; The query vector is multiplied by all the enhanced key vectors to obtain an initial attention score, which represents the correlation between feature vectors. The initial attention score is input into a gated adjustment network based on metadata embedding vectors. The gated adjustment network outputs an adjustment factor, which is used to scale the initial attention score according to the reliability of the metadata representation, to obtain the scaled attention score. The scaled attention scores are then normalized to obtain the final attention weights.

[0018] In some embodiments, an initial processing pipeline is formed by selecting and combining atomic processing units from a pre-built library based on the service type in the dynamic service feature vector, including: Extract the business type from the dynamic business feature vector; Based on the parsed business type, query the predefined business type-pipeline template mapping relationship to determine the corresponding initial pipeline template. The initial pipeline template specifies the type and execution order constraints of the atomic processing units required to complete this type of business. Based on the initial pipeline template, select specific versions of atomic processing unit instances from the pre-built atomic processing unit library, taking into account the data format information contained in the dynamic business feature vector when selecting; Obtain the input / output interface definitions of the selected atomic processing unit instance, and construct a data dependency graph between atomic processing units based on the data types and semantics of the interface definitions; The data dependency graph is a directed acyclic graph structure, where nodes represent instances of atomic processing units and directed edges represent data flow directions. Based on the data dependency graph and the execution order constraints in the initial pipeline template, an initial processing pipeline with a clear execution order and data transfer relationship is generated. The initial processing pipeline is described using a graph structure that includes a sequence of nodes and a set of edges.

[0019] In some embodiments, based on an initial pipeline template, a specific version of an atomic processing unit instance is selected from a pre-built atomic processing unit library. The selection process references data format information contained in the dynamic business feature vector, including: Parse the initial pipeline template to obtain the type identifier of each atomic processing unit specified in the initial pipeline template; Based on the type identifier of each atomic processing unit, a search is performed in the pre-built atomic processing unit library to obtain a list of all candidate atomic processing unit instances that match the type identifier; Extract the data format information of the original multimodal data stream from the dynamic business feature vector. The data format information includes at least the encoding format, resolution, sampling rate, and quantization accuracy. For each candidate atomic processing unit instance, read its predefined input data format requirements; The matching degree is calculated by comparing the data format information of the original multimodal data stream with the input data format requirements of the candidate atomic processing unit instances. The matching degree calculation includes checking the encoding format compatibility, resolution scaling feasibility, and sampling rate adaptability. Based on the matching degree calculation results, a format adaptation score is generated for each candidate atomic processing unit instance; Combining the format fit score with the version information of the atomic processing unit instance, the atomic processing unit instance with the highest format fit score and the version that meets the preset requirements is selected from the candidate list and used as the final atomic processing unit instance corresponding to this type identifier.

[0020] Unlike existing technologies, the above technical solution provides a data transmission method for power plant edge computing and 5G collaboration. It acquires raw multimodal data streams and analyzes task requests to generate dynamic service feature vectors, which are then used to select atomic processing units to form an initial processing pipeline. Real-time acquisition of 5G private network wireless channel status, backhaul link load, resource utilization of each edge computing node, and interconnection status between nodes generates a global resource status view. With the goal of minimizing latency and maximizing resource utilization, dynamic mapping and reconstruction are performed on the initial processing pipeline based on this view and feature vectors. Appropriate nodes are allocated to each unit, and transmission paths and priorities are planned. Node execution is controlled, and online adaptive migration is triggered based on dynamic criticality level changes. Intermediate results are received at the aggregation node, and a weighted fusion using an attention mechanism is employed to generate and output the final diagnostic results, updating the global resource status view. This achieves dynamic collaborative optimization and adaptive scheduling of computing and transmission resources.

[0021] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0022] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0023] In the accompanying drawings of the instruction manual: Figure 1 This is a schematic diagram illustrating steps S101 to S105 of the method described in the specific implementation embodiment; Figure 2 This is a schematic diagram illustrating steps S201 to S207 of the method described in a specific implementation. Figure 3 The diagram illustrates steps S301 to S306 of the method described in the specific implementation. Detailed Implementation

[0024] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0025] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0026] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0027] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0028] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0029] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0030] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0031] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0032] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0033] The computer program involved in the embodiments can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributed and stored in multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device, or can be connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0034] Please see Figure 1 This embodiment provides a data transmission method for power plant edge computing and 5G collaboration, including: S101. Obtain the raw multimodal data stream from the heterogeneous sensing devices in the power plant and the corresponding analysis task request; S102. Based on the initial attributes of the analysis task request and the original multimodal data stream, generate a dynamic business feature vector; S103. Based on the business type in the dynamic business feature vector, select and combine from the pre-built atomic processing unit library to form the initial processing pipeline; S104. Real-time collection of 5G private network wireless channel status, backhaul link load, resource utilization of each edge computing node and interconnection status between nodes, and fusion to generate a global resource status view. S105. With the joint optimization goal of minimizing the overall task completion delay and maximizing resource utilization, dynamic mapping and reconstruction are performed on the initial processing pipeline based on the global resource status view and dynamic business feature vector. Dynamic mapping and reconstruction include: Assign appropriate edge computing nodes as execution nodes to each atomic processing unit in the pipeline, and plan the transmission path and priority for the intermediate data flow between nodes based on the NSGA-II multi-objective optimization algorithm; Control each edge computing node to execute the corresponding atomic processing unit according to the mapping relationship, and trigger online adaptive migration of the transmission path and execution node according to the dynamic criticality level change in the dynamic business feature vector during the execution process; At the designated aggregation edge computing node, intermediate processing results from upstream nodes are received, and a weighted fusion analysis is performed using an attention mechanism to generate the final device status fusion diagnostic results. Output the fusion diagnostic results and update the global resource status view.

[0035] In step S101, the heterogeneous sensing devices in the power plant include visible light cameras, infrared thermal imagers, vibration sensors, and acoustic sensors deployed in the equipment area. The raw multimodal data stream refers to the unprocessed raw monitoring data output by the aforementioned devices through their network interfaces, such as H.264 video streams, vibration waveforms transmitted via Modbus / TCP protocol, or infrared thermal images. The corresponding analysis task request is issued by the operation and maintenance platform deployed in the plant-level monitoring center via a RESTful API. The request encapsulates the target device identifier, the desired analysis type, and a unique task identifier. This step completes the acquisition of data and requests through the data receiving interface of the 5G private network base station or edge gateway, providing the data source and task context for subsequent processing.

[0036] In step S102, the initial attributes of the original multimodal data stream are obtained by parsing the data packet header or querying device metadata, and at least include the data format, sampling rate, and geographical location of the data source. The analysis task request implicitly imposes constraints on end-to-end latency and result reliability for the business. By parsing the business type in the task request and combining it with the processing complexity implied by the initial data attributes, a structured feature label is formed, which is the dynamic business feature vector. This vector dynamically encapsulates the business type, latency tolerance limit, reliability requirements, processing complexity estimate, and a dynamic criticality level determined by the real-time analysis results of the data content. This step transforms specific operation and maintenance instructions into quantitative features that can be directly processed by optimization algorithms, establishing a bridge between business needs and resource scheduling.

[0037] In step S103, the service type is determined based on the analysis objective defined in the analysis task request, such as video anomaly detection or vibration spectrum diagnosis. A pre-built atomic processing unit library stores standardized software modules that implement various basic analysis functions, such as YOLO-based target detection units or units that perform Fast Fourier Transform. Based on the service type in the dynamic service feature vector, a predefined service-pipeline template mapping relationship is queried to obtain the corresponding atomic processing unit type and execution order. According to this template, specific instances with compatible data formats are selected from the atomic processing unit library and combined into an initial processing pipeline according to data dependencies. Preferably, the initial processing pipeline is described as a directed acyclic graph, where nodes are atomic processing units and edges define the flow direction of intermediate data. This step achieves the decomposability and standardized orchestration of complex analysis tasks.

[0038] In step S104, the 5G private network wireless channel status is obtained by periodically collecting reference signal received power and signal-to-interference-plus-noise ratio (SINR) using network probes on the base station side. Backhaul link load is calculated by monitoring the throughput and queue depth of the links between user plane functional network elements and edge computing nodes. Resource utilization of each edge computing node is collected by a resource monitoring agent deployed on the node's operating system, including CPU utilization, memory usage, and GPU utilization. Inter-node interconnection status is measured by running link-layer discovery protocols between nodes and actively sending probe packets to measure round-trip latency and packet loss rate. After timestamp alignment, missing value imputation, and normalization, the aforementioned status data from different sources and with different dimensions are encapsulated into a unified data structure to obtain a global resource status view. This view provides a real-time, quantitative snapshot of the entire edge computing and 5G network infrastructure.

[0039] In step S105, dynamic mapping and reconstruction are performed based on the global resource status view and dynamic service feature vectors. Each atomic processing unit in the pipeline is assigned a suitable edge computing node as an execution node, specifically based on a comprehensive evaluation of the node's real-time available computing power, load, and network conditions between the node and the data source. Simultaneously, transmission paths and priorities are planned for the intermediate data streams that need to be transmitted between nodes, specifically based on the path's real-time latency, bandwidth, and criticality level in the dynamic service feature vectors. By employing the NSGA-II multi-objective optimization algorithm, node allocation and path planning are encoded as the same optimization problem, jointly solving for the objectives of minimizing the overall task completion latency and maximizing resource utilization, outputting the optimal mapping and routing scheme in one go. Preferably, the NSGA-II multi-objective optimization algorithm balances these two potentially conflicting objectives by maintaining a Pareto optimal solution set.

[0040] The system controls each edge computing node to execute the corresponding atomic processing unit according to the mapping relationship. The intelligent decision center issues execution instructions containing atomic processing unit identifiers, input data source addresses, and startup parameters to each node. When the real-time pre-analysis results of the data stream (such as detecting an abnormal temperature) cause its dynamic criticality level to jump from "normal" to "urgent", the system immediately re-evaluates the affected atomic processing unit and selects a better execution node based on the latest global resource status view. At the same time, it re-plans a low-latency, high-reliability transmission path for its data stream and coordinates relevant nodes and network devices to complete the seamless switching of computing context and data stream.

[0041] At a designated aggregation edge computing node, intermediate processing results from upstream nodes are received. This node is preferably one that facilitates the aggregation of multiple data streams in its topology and possesses sufficient fusion computing capabilities. An attention mechanism is employed to dynamically calculate different weights for the received feature vectors based on meta-information such as the confidence level of their source processing units and the load status of the generating node. The weighted features are then integrated, and finally, a diagnostic model is used to generate a fusion diagnostic result for the device status.

[0042] The diagnostic results are pushed to the operation and maintenance platform. At the same time, based on the actual node computing power and network bandwidth consumed in this task execution, the node resource utilization and link load data in the resource status view are corrected and updated to form a closed-loop feedback.

[0043] This embodiment quantifies the latency, reliability, and dynamic criticality level of multimodal services using dynamic service feature vectors. It unifies the perception and modeling of 5G private network wireless channels, backhaul links, edge computing node resources, and inter-node interconnection status through a global resource status view. Based on this, with the objectives of minimizing overall task completion latency and maximizing resource utilization, the NSGA-II multi-objective optimization algorithm is used to dynamically map and reconstruct the initial processing pipeline. This achieves simultaneous joint optimization of allocating suitable edge computing nodes to atomic processing units and planning transmission paths and priorities for intermediate data streams. Furthermore, by monitoring dynamic criticality level changes during execution and triggering online adaptive migration, the system ensures real-time resource reconfiguration for sudden high-value services. Finally, an attention mechanism is used at the aggregated edge computing nodes to weightedly fuse multi-source intermediate processing results, generating accurate device status fusion diagnostic results and updating the global resource status view to form an optimization closed loop. This embodiment enables dynamic adjustment of resource allocation based on the intrinsic value density of services in complex power plant environments. This significantly improves the overall resource utilization efficiency and system resilience of 5G private networks and edge computing clusters while strictly ensuring low latency and high reliability of critical services.

[0044] Please see Figure 2In some embodiments, a dynamic business feature vector is generated based on the initial attributes of the analysis task request and the original multimodal data stream, including: S201. Parse and analyze the task request, extract the task identifier and expected output format, and generate an initial business type label based on the task identifier; S202. Obtain the source device identifier, data format, and sampling rate of the original multimodal data stream as initial attributes of the data stream; S203. Based on the initial business type label and the initial data flow attributes, determine the initial latency tolerance limit and reliability requirements by querying the predefined business type-service quality mapping table; S204. Based on the initial business type label and data format, the initial processing complexity estimate is calculated using the complexity prediction model. S205. At the first edge computing node that receives the original multimodal data stream, the pre-deployed pre-analysis atomic processing unit is invoked to perform real-time analysis on the data stream and obtain real-time analysis results. S206. Match the real-time analysis results with the preset abnormal pattern library, dynamically evaluate and output the content criticality level based on the matching confidence level; S207. Vectorize and encapsulate the initial business type label, initial latency tolerance limit, reliability requirements, initial processing complexity estimate, and content criticality level to generate a dynamic business feature vector. Furthermore, during the execution of subsequent atomic processing units, if new real-time analysis results are received that cause a change in the content criticality level, a real-time update operation on the dynamic business feature vector is triggered.

[0045] In step S201, the analysis task request is encapsulated in a structured data format. Parsing the request involves extracting predefined fields from its message: the task identifier is a string or code used to uniquely distinguish the task; the expected output format field specifies the data structure type that the diagnostic results should return. An initial business type label is generated based on the task identifier, which can be achieved by querying a pre-configured mapping relationship within the system. This mapping relationship associates different task identifiers with preset business type enumeration values.

[0046] In step S202, the source device identifier of the original multimodal data stream is obtained by parsing the corresponding field defined in the network layer or application layer protocol header of the data packet. The data format is determined by identifying the encoding protocol and structure of the data payload. The sampling rate is obtained from protocol information or device metadata based on the type of data stream; for video streams, it corresponds to the frame rate, and for time-series data, it corresponds to the number of sampling points per unit time. These parameters together constitute the initial attributes of the data stream.

[0047] In step S203, the service type-service quality mapping table is a pre-configured query table that stores the corresponding latency tolerance upper limit and reliability requirement value using the service type as the key. The latency tolerance upper limit is measured in time units, and the reliability requirement is expressed as a success probability percentage. By querying this table, the initial latency tolerance upper limit and reliability requirement bound to the initial service type label can be directly obtained.

[0048] In step S204, the complexity prediction model is a prediction model trained based on historical task performance data. Its inputs are business type and data format characteristics, and its output is an initial processing complexity estimate representing the processing overhead. This estimate can be quantified as an estimated amount of computation or processing time. The model training process uses the actual execution time and resource consumption data of historical tasks as supervision labels.

[0049] In step S205, the first edge computing node receiving the raw multimodal data stream is deployed with a pre-analysis atomic processing unit. This unit is a lightweight data analysis program used to perform preliminary real-time analysis on the incoming raw data, such as motion detection, amplitude statistics, or threshold comparison. Calling this unit to process the data stream yields a structured, simplified descriptive information as the real-time analysis result.

[0050] In step S206, a preset anomaly pattern library stores feature templates of historical anomaly events. The real-time analysis results are matched against this library, i.e., the similarity between their features is calculated. Based on the calculated matching confidence, it is mapped to a discrete or continuous content criticality level using preset evaluation rules. Preferably, the matching process can be implemented by calculating the distance or similarity between feature vectors.

[0051] In step S207, vectorization encapsulation refers to combining the data from five dimensions—initial business type label, initial latency tolerance limit, reliability requirements, initial processing complexity estimate, and content criticality level—in a predetermined order and format into a complete data structure, thereby generating a dynamic business feature vector. The real-time update operation of the dynamic business feature vector includes reassessing the criticality level based on the new analysis results, replacing the original values ​​in the vector with the new level values, and simultaneously sending a notification to the scheduling module that the vector has been updated.

[0052] This embodiment generates a comprehensive quantitative feature that represents business needs—a dynamic business feature vector—by combining static attributes from task requests, constraint parameters derived from mapping tables and models, and content value assessments derived from real-time data stream analysis. This vector not only includes fixed service quality requirements but also embeds real-time perception of data content value through dynamic criticality levels. It provides accurate and timely decision input for subsequent dynamic resource scheduling and forms the basis for achieving business perception and adaptive resource collaboration.

[0053] In some embodiments, the real-time analysis results are matched against a preset abnormal pattern library, and the content criticality level is dynamically evaluated and output based on the matching confidence level, including: Extract structured feature descriptors from real-time analysis results. The structured feature descriptors include at least the anomaly type, spatial location, timestamp, and quantization parameters. The structured feature descriptor is input into a pre-built anomaly pattern library, which stores multi-dimensional feature templates of historical anomalies. Each feature template is associated with a preset key benchmark value. An attention-based feature matching algorithm is used to calculate the similarity between the structured feature descriptor and each feature template in the abnormal pattern library. Based on the similarity, the key benchmark values ​​associated with each feature template are weighted and fused to generate a preliminary key score. Obtain the data stream identifier corresponding to the real-time analysis results, and query the historical key score sequence of the data stream within a preset time window; The initial criticality score and the historical criticality score sequence are input into the criticality state transition model, which is built based on a hidden Markov chain and is used to smoothly evaluate and predict the evolution trend of criticality states. Based on the output of the critical state transition model, the final content criticality level at the current moment is determined. The content criticality level is used to characterize the value density and urgency of the information contained in the data stream for subsequent fusion decision-making tasks.

[0054] In this embodiment, extracting structured feature descriptors from real-time analysis results refers to parsing and formatting the brief description information output in step S205. The anomaly type field is obtained by matching keywords in the description information with a predefined anomaly category dictionary. The spatial location field is extracted from the coordinate data or associated device topology identifiers contained in the description information. The timestamp field directly uses the time of generation of the real-time analysis results. The quantization parameter field extracts specific measured values ​​from the description information, such as excessive temperature values ​​or vibration amplitudes. These fields together constitute a standardized data structure, namely, the structured feature descriptor.

[0055] The pre-built anomaly pattern library is derived from the analysis and feature extraction of a large number of historical failure event cases. The multi-dimensional feature template of each historical anomaly event stored in the library is the result of standardizing the structured feature descriptor of that historical event. Each feature template is associated with a preset key benchmark value, which is pre-set based on the severity of the actual operational impact caused by the historical event, such as being assigned by expert experience or calculated based on the downtime caused by the event.

[0056] A feature matching algorithm based on an attention mechanism is employed. First, the structured feature descriptor and each feature template in the anomaly pattern library are converted into high-dimensional feature vectors. The attention mechanism calculates the correlation score between the descriptor vector and each template vector, obtaining a set of attention weights, which represent the similarity between the descriptor and each template. The similarity between the structured feature descriptor and each feature template in the anomaly pattern library is calculated, specifically by calculating the cosine similarity or dot product between the converted vectors and normalizing it. Based on the similarity-weighted fusion of key benchmark values ​​associated with each feature template, a preliminary key score is generated. This involves using the calculated similarity scores as weights to perform a weighted sum of the key benchmark values ​​for the corresponding templates; the resulting weighted sum is the preliminary key score.

[0057] The data stream identifier corresponding to the real-time analysis result is determined in step S202 as part of or associated with the source device identifier. Querying the historical critical score sequence of the data stream within a preset time window means retrieving all historical critical scores generated by the data stream within a recent preset time period (e.g., the past few minutes) from the time-series database maintained by the system, based on the data stream identifier, and arranging them into a sequence in chronological order.

[0058] The critical state transition model is built upon Hidden Markov Chains. This model defines different levels of content criticality as hidden states. The training process uses historical criticality score sequences as observation sequences, and employs unsupervised learning methods such as the Baum-Welch algorithm to estimate the state transition probability matrix and observation probability distribution. The initial criticality score and historical criticality score sequences are input into the model. Using a forward-backward algorithm or a Viterbi algorithm, the model calculates the most probable hidden state probability distribution at the current moment based on the latest observations (initial scores) and past observation sequences (historical sequences), thus achieving smooth evaluation and predicting the evolution trend of the state at the next moment based on the state transition probabilities.

[0059] Based on the output of the critical state transition model, the final content criticality level at the current moment is determined. Typically, the hidden state with the highest probability calculated by the model is selected as the final level. The content criticality level is used to characterize the value density and urgency of the information contained in the data stream for subsequent fusion decision-making tasks. The higher the level, the more likely the information carried by the data stream is to indicate device malfunction, and the more critical it is to generating accurate fusion diagnostic results. Therefore, it should receive higher priority in resource scheduling.

[0060] This embodiment generates a preliminary criticality score by structuring real-time analysis results and performing intelligent matching based on an attention mechanism with a historical anomaly pattern library. Then, by combining the historical score sequence of the data stream, a hidden Markov chain model is used to perform temporal smoothing and prediction of critical states, ultimately outputting a stable content criticality level. This overcomes the limitations of simple threshold judgment or static rule matching, taking into account both the similarity between current anomaly features and historical patterns, and incorporating the continuity of critical states over time. This makes the evaluation results more stable and accurate, effectively filtering out instantaneous noise interference, and truly reflecting the continuous value and urgency of the data stream, providing a reliable basis for triggering precise online adaptive migration of resources.

[0061] In some embodiments, the status of 5G private network wireless channels, backhaul link load, resource utilization of each edge computing node, and interconnection status between nodes are collected in real time and fused to generate a global resource status view, including: By using probes deployed on the 5G base station side, the reference signal received power and signal-to-interference-to-noise ratio of each area are periodically collected as the wireless channel status. By deploying link monitoring points between user plane functional network elements and edge computing nodes, the current uplink and downlink data throughput and transmission latency are collected, and the backhaul link load is calculated. By deploying resource monitoring agents on each edge computing node, the utilization rate of the central processing unit, the utilization rate of the graphics processing unit, and the memory usage rate are periodically collected as the node resource utilization rate. Based on node resource utilization and node hardware configuration information, a computing resource profile of each edge computing node is constructed. By running link layer discovery protocols or active probing mechanisms, the network topology connections between edge computing nodes are obtained, and the round-trip latency and packet loss rate of the links between nodes are measured to assess the interconnection status between nodes. The collected wireless channel status, backhaul link load, node resource utilization and inter-node interconnection status are cleaned and normalized to unify the heterogeneous indicators to the same numerical units. The normalized wireless channel state vector, backhaul link load scalar, computational resource profile vector of each node, and inter-node interconnection state matrix are encapsulated to generate a global resource state view.

[0062] In this embodiment, the probe deployed on the 5G base station side is a software agent or a dedicated hardware device. It reads the reference signal received power and signal interference noise ratio values ​​at fixed time intervals by calling the application programming interface of the base station baseband processing unit. These values ​​reflect the signal strength and quality of the wireless link in a specific geographical area.

[0063] Link monitoring points deployed between user plane functional network elements and edge computing nodes acquire network data packets flowing through the link using traffic mirroring or port mirroring technology, and calculate the total amount of data per unit time to determine throughput. Simultaneously, they calculate end-to-end transmission latency by recording the timestamps of the data packets. The backhaul link load is calculated based on the collected throughput and latency data, combined with the link's nominal bandwidth, to arrive at a comprehensive indicator characterizing the link congestion level through weighted calculation.

[0064] The resource monitoring agent deployed on each edge computing node is a background service process running on the node's operating system. It periodically obtains the average utilization of the CPU across all cores, the utilization of the graphics processing unit, and the current percentage of system memory usage by calling the operating system's performance monitoring interface.

[0065] When constructing a computing resource profile for each edge computing node, hardware configuration information includes static parameters such as processor model and number of cores, graphics processor model and video memory capacity, and total memory capacity. The construction process combines static hardware capability benchmarks with dynamic resource utilization to calculate a multi-dimensional vector characterizing the node's real-time processing capability. Preferably, this profile may include dimensions such as benchmark computing power score and real-time available computing power.

[0066] Link-layer discovery protocols discover neighbors by periodically broadcasting messages containing information about the current node within the local area network; active probing mechanisms confirm connectivity by sending probe packets to known or potential neighbor node addresses and waiting for responses. The round-trip time and packet loss rate of inter-node links are measured by analyzing the time difference between sending and receiving probe packets and packet loss, thereby assessing the interconnection status between nodes.

[0067] Data cleaning is performed on the collected wireless channel status, backhaul link load, node resource utilization, and inter-node interconnection status, including handling instantaneous outliers or missing values ​​caused by network jitter. Normalization processing uses methods such as linear scaling to transform raw data from different sources and with different dimensions into the same numerical range, such as between zero and one, in order to eliminate the influence of dimensions and facilitate comprehensive processing.

[0068] This embodiment systematically acquires multi-dimensional raw data, such as wireless channel quality, backhaul link load, node computing resource utilization, and inter-node network performance, by deploying acquisition probes on various types of network and computing entities. After cleaning, normalization, and structured encapsulation, the data forms a unified global resource status view, realizing unified measurement and real-time perception of distributed heterogeneous resources. This provides accurate and consistent decision-making basis for dynamic mapping and reconstruction based on multi-objective optimization, ensuring that the resource scheduling scheme fits the actual network and computing environment.

[0069] In some embodiments, a computing resource profile for each edge computing node is constructed based on node resource utilization and node hardware configuration information, including: Read the hardware configuration information of the edge computing node. The hardware configuration information includes at least the CPU model and number of cores, the graphics processor model and video memory capacity, and the total memory capacity. Based on hardware configuration information, a predefined hardware performance benchmark scoring model is used to determine the benchmark computing performance scores of nodes in terms of central processing unit computing, graphics processing unit computing, and memory access. The CPU utilization, GPU utilization, and memory usage are periodically collected by the resource monitoring agent and used to construct a real-time utilization vector. Based on the utilization values ​​of each dimension in the real-time utilization vector and the corresponding benchmark computing performance scores, the real-time available computing power of the node in each dimension at the current moment is calculated; the real-time available computing power represents the processing capacity remaining after deducting the occupied resources that can be used for new tasks. The baseline computing performance score, real-time utilization vector, and real-time available computing power vector are combined and encapsulated to form a computing resource profile for the edge computing node.

[0070] In this embodiment, the hardware configuration information of the edge computing node is read by calling the system information query interface provided by the operating system, such as reading files like / proc / cpuinfo in the Linux system, or obtaining detailed specifications of the graphics processor and memory through the hardware management library.

[0071] A predefined hardware performance benchmark scoring model is a function or lookup table that maps hardware specifications to standardized performance scores. This model is pre-built based on publicly available hardware performance benchmark data. For example, it uses a lookup table to obtain the theoretical floating-point performance score based on the CPU model and core count, determines the parallel computing performance score based on the graphics processor model and video memory capacity, and directly or proportionally converts the total memory capacity into a memory access performance score.

[0072] The CPU utilization, GPU utilization, and memory usage collected periodically by the resource monitoring agent are arranged in a fixed order to form a real-time utilization vector.

[0073] When calculating the real-time available computing power of a computing node in each dimension at the current moment, for each dimension, the real-time available computing power is equal to the difference between the baseline computing performance score for that dimension and the utilization rate for that dimension. For example, if the CPU baseline score is 100 and the current utilization rate is 0.3, then the CPU's real-time available computing power is 100*(1-0.3)=70. This calculation is performed separately for the central processing unit, graphics processing unit, and memory dimensions, resulting in a real-time available computing power vector containing three values.

[0074] The benchmark computing performance score, real-time utilization vector, and real-time available computing power vector are organized and serialized according to a predefined data structure and encapsulated to form a computing resource profile of the edge computing node. This profile is a composite data object containing multi-dimensional information such as static performance benchmark, dynamic load status, and real-time availability.

[0075] This embodiment quantifies static hardware capabilities into a benchmark score and combines it with real-time resource utilization to derive real-time available computing power that reflects the instantaneous processing capability of a node. Finally, it encapsulates this into a structured computing resource profile, which describes the inherent performance of a node and dynamically reflects its current load and remaining capacity. This provides a refined quantitative basis for accurately assessing the execution cost and potential of a node during task scheduling, which is conducive to achieving load balancing and efficient resource utilization.

[0076] In some embodiments, by running a link layer discovery protocol or an active probing mechanism, the network topology connections between edge computing nodes are obtained, and the round-trip latency and packet loss rate of the links between nodes are measured to evaluate the interconnection status between nodes, including: The resource monitoring agent deployed on each edge computing node integrates a network topology discovery module and a link quality detection module. The network topology discovery module periodically broadcasts link layer discovery protocol data frames in the local area network to which the edge computing node belongs, or sends topology query requests to a known list of neighboring node addresses. Receive link layer discovery protocol responses or topology information replies from other edge computing nodes, and parse the neighbor node identifiers, connection port information and hop counts contained in the response data; Based on the neighbor relationship information reported by all nodes, a dynamic edge node network topology is constructed and maintained in the intelligent decision-making center. The topology is based on nodes as vertices and physical or logical connections as edges. The link quality detection module periodically initiates bidirectional active detection for each link between nodes represented by an edge in the topology graph; active detection includes sending a sequence of probe data packets with timestamps. The receiving end records the arrival timestamp and sequence number of the probe data packets, and calculates the round-trip time and packet loss for a single probe. Based on all probe data packet samples collected during the probe period, after removing abnormal samples using a robust statistical algorithm, the average round-trip time and average packet loss rate of the link are calculated. The average round-trip time, average packet loss rate, and topological connectivity are collectively encapsulated into an inter-node interconnection state descriptor, which is used to characterize the inter-node interconnection state.

[0077] In this embodiment, the network topology discovery module and the link quality detection module are integrated into the program code of the resource monitoring agent as software libraries or plugins, and are uniformly scheduled and executed by the agent. Preferably, the network topology discovery module broadcasts link layer discovery protocol data frames in the local area network to which the node belongs by calling the operating system's raw socket interface. The known list of neighboring node addresses comes from the system's initial configuration or the cache of historical topology discovery results.

[0078] After receiving responses from other edge computing nodes, the parsing process extracts fields such as the sending node's device identifier, sending port identifier, and protocol-defined time to live from the data frame payload according to the standard format of the link layer discovery protocol.

[0079] The intelligent decision-making center collects neighbor relationship information reported by all nodes, constructs a graph data structure in memory with edge computing nodes as vertices and physical or logical connections between nodes as edges, and continuously updates this graph with periodic discoveries to maintain its dynamism.

[0080] Based on the current topology, the link quality detection module periodically sends a set of User Datagram Protocol (UDP) probe packets containing a high-precision transmission timestamp and an incrementing sequence number to the network address of each neighboring node. Upon receiving the probe packet, the receiving end immediately records its arrival timestamp and confirms whether any packets have been lost by comparing the sequence number. The round-trip time is calculated by subtracting the transmission timestamp carried in the packet from the arrival timestamp.

[0081] Based on all samples collected within a detection period, an abnormal delay value that deviates significantly from the normal range is identified and removed using a method based on the median absolute deviation. Then, the arithmetic mean of the remaining delay samples is calculated as the average round-trip delay of the link, and the average packet loss rate is calculated based on the ratio of the number of lost packets to the total number of sent packets.

[0082] The calculated average round-trip time, average packet loss rate, and node identification information at both ends of the link are organized into a structured data record, namely the inter-node interconnection state descriptor. This descriptor quantitatively characterizes the transmission performance and reliability of the inter-node link.

[0083] This embodiment achieves automatic network topology discovery and proactive link quality measurement through integrated modules. It employs robust statistical methods to handle noise caused by network jitter and ultimately generates quantified inter-node interconnection state descriptors. This enables the global resource state view to contain accurate and stable network performance information, providing key input for evaluating data transmission costs in subsequent optimization algorithms and ensuring the accuracy of path planning decisions.

[0084] Please see Figure 3In some embodiments, appropriate edge computing nodes are assigned as execution nodes to each atomic processing unit in the pipeline, and transmission paths and priorities for intermediate data streams between nodes are planned based on the NSGA-II multi-objective optimization algorithm, including: S301. For each atomic processing unit in the initial processing pipeline, based on the node computing resource profile and the interconnection status between nodes in the global resource status view, select edge computing nodes that meet its processing capabilities and communication requirements, and construct a candidate execution node set. S302. With the optimization objectives of minimizing the overall task completion delay and maximizing the system resource load balance, a multi-objective optimization function is constructed. The calculation of the overall task completion delay needs to take into account the estimated execution delay of the atomic processing unit on each candidate node and the estimated network delay of the intermediate data stream transmitted between nodes. S303. The selection of candidate execution nodes corresponding to each atomic processing unit, as well as the selection of transmission paths between nodes, are jointly encoded into the chromosome of the optimization problem. S304. Initialize the chromosome population by utilizing the content criticality levels in the global resource status view and dynamic business feature vectors; S305. The NSGA-II multi-objective optimization algorithm is used to iteratively optimize the chromosome population, performing operations including non-dominated sorting, crowding calculation, selection, crossover and mutation, to approximate and obtain the Pareto optimal solution set. S306. From the Pareto optimal solution set, based on the latency tolerance upper limit and reliability requirements in the dynamic service feature vector, select the optimal solution for decoding, including: The gene segments corresponding to the atomic processing units in the chromosome are decoded into defined edge computing nodes, which are then used as their execution nodes. Decode the gene segments corresponding to the inter-node transmission path into a defined path sequence and the transmission priority on that path; Output the mapping table between atomic processing units and execution nodes obtained from decoding, as well as the transmission path and priority planning table of intermediate data streams.

[0085] In step S301, the node computing resource profile and the interconnection status between nodes in the global resource status view are used for screening. Specifically, for an atomic processing unit, if the real-time available computing power of an edge computing node is not lower than the estimated processing complexity of the unit, and the link latency and packet loss rate between the node and the input data source node required by the unit meet the basic communication requirements, then the node is included in its candidate execution node set.

[0086] In step S302, when constructing the multi-objective optimization function, the overall task completion delay objective function needs to be based on the pipeline execution logic, accumulating the estimated execution delay of each atomic processing unit on its selected node and the estimated network delay of the intermediate data flow between units along the selected path. The system resource load balancing objective function needs to be based on the node allocation scheme of all atomic processing units, estimating the future resource utilization rate of each node after it bears the load, and calculating the dispersion of these utilization rates. The estimated execution delay can be estimated based on the unit processing complexity and the real-time available computing power of the node, and the estimated network delay can be estimated based on the delay status of the links traversed by the path and the data flow size.

[0087] In step S303, the candidate execution node selection corresponding to each atomic processing unit is encoded as a fixed-length gene segment in the chromosome, where the gene value represents the index of the selected node in the candidate set. The selection of the inter-node transmission path is encoded as another variable-length gene sequence in the chromosome, where the values ​​in the sequence represent the node identifiers traversed by the path. In this way, a single chromosome fully represents a complete set of node mapping and path planning schemes.

[0088] In step S304, when initializing the chromosome population, some chromosome individuals are generated based on the content criticality level in the dynamic business feature vector. For example, for processing units associated with high-criticality data streams, nodes and paths with better computing resource profiles and better interconnection link quality are preferentially allocated to construct a high-quality initial solution.

[0089] In step S305, the non-dominated sorting divides chromosome individuals into multiple non-dominated levels based on their performance on the two optimization objectives. Crowding density calculation assesses the distribution density of individuals within the same non-dominated level in the target space. The selection operation, based on non-dominated level and crowding density, prioritizes retaining individuals with higher levels and sparser distributions for the next generation. Crossover and mutation operations simulate biological evolution, exchanging information and randomly perturbing chromosome genes to explore new feasible solution spaces. Through multiple iterations, the population gradually approaches a solution set composed of numerous Pareto optimal solutions.

[0090] In step S306, when selecting the optimal solution from the Pareto optimal solution set, solutions whose overall task completion latency exceeds the latency tolerance upper limit in the dynamic business feature vector are first filtered out. Then, the solution with the optimal system resource load balancing is selected from these solutions. The decoding process proceeds in reverse according to chromosome encoding rules: the gene segment value corresponding to the atomic processing unit is mapped back to the specific edge computing node identifier, which serves as its execution node; the value of the path gene segment is parsed into a sequence of node identifiers from the upstream node to the downstream node, which serves as the transmission path; the transmission priority can be determined based on the criticality level of the data flow or specific encoded bits in the path gene segment. Finally, a structured mapping table and path planning table are output.

[0091] This embodiment models the complex node allocation and path planning problem as a multi-objective optimization problem and uses the NSGA-II algorithm for efficient solution, achieving an optimal trade-off between the conflicting objectives of minimizing task latency and maximizing resource balance. This method jointly optimizes the two types of decision variables through unified chromosome encoding, avoiding local optima caused by step-by-step optimization. It can output a globally coordinated scheduling and routing scheme in one go, providing a systematic solution for the coordinated optimization of service performance and resource efficiency in power plant edge computing scenarios.

[0092] In some embodiments, a multi-objective optimization function is constructed with the optimization objectives of minimizing the overall task completion latency and maximizing the system resource load balancing, including: For each atomic processing unit in the initial processing pipeline, based on its processing complexity and the real-time available computing power in the candidate execution node computing resource profile, the estimated execution latency of the unit on each candidate node is calculated using a predefined execution latency prediction model. For each path that may transmit intermediate data streams, the estimated network latency of the path is calculated based on the transmission latency of the links traversed by the path and the current load status in the global resource status view, using a predefined network latency prediction model. The objective function for the overall task completion time is defined as follows: Starting from the atomic processing unit at the beginning of the pipeline, based on the decoded node mapping and path planning, the estimated execution time of each atomic processing unit and the estimated network time of the path through which its output data flows are accumulated in the order of execution, and the resulting total time value is used as the first optimization objective. The objective function for system resource load balancing is defined as follows: Based on the decoded node mapping relationship, estimate the future resource utilization rate of each edge computing node after undertaking the allocated atomic processing unit, calculate the variance of the future resource utilization rate of all nodes, and take the reciprocal of the variance as the second optimization objective. The multi-objective optimization function is composed of the objective function of the overall task completion delay and the objective function of system resource load balancing.

[0093] In this embodiment, the predefined execution latency estimation model is a function that maps processing complexity and the real-time available computing power of a node to latency. Specifically, it can be expressed as the estimated execution latency equals the processing complexity divided by the real-time available computing power of the node. The processing complexity is derived from the estimated value in the dynamic business feature vector, and the real-time available computing power is obtained from the computing resource profile of the candidate execution nodes. Using this model, the estimated execution latency of each atomic processing unit on each candidate node can be calculated.

[0094] A predefined network latency prediction model is used to estimate the time required for a data stream to travel along a path. The model's inputs include the transmission latency of each link along the path, the current available bandwidth of each link, and the size of the data stream to be transmitted. Transmission latency and link load status are obtained from the inter-node interconnection status in the global resource status view. The predicted network latency can be calculated as the sum of the transmission latency of each link on the path, plus the quotient of the data stream size divided by the available bandwidth of the bottleneck link on the path, taking into account a certain queuing latency margin.

[0095] The calculation of the objective function for the overall task completion delay requires, for a specific node mapping and path planning scheme (i.e., a chromosome individual), first determining the critical path of the initial processing pipeline. The estimated execution delay of each atomic processing unit on the critical path at its mapped node, and the estimated network delay of the data stream produced by that unit transmitted along the mapped path, are then summed sequentially. The sum obtained is the overall task completion delay value for this scheme.

[0096] The calculation of the system resource load balancing objective function, based on the same node mapping scheme, specifically includes: estimating the future resource utilization of the node after executing these tasks based on the sum of the processing complexity of all atomic processing units allocated to each edge computing node, the node's baseline computing performance score, and its current resource utilization. One optional estimation method is to divide the sum of the estimated processing complexity of all atomic processing units allocated to the node by the node's baseline computing performance score to obtain an incremental utilization rate; then add this incremental rate to the node's current resource utilization rate to obtain the estimated future resource utilization rate; finally, calculate the variance of the future resource utilization rates of all nodes. The system resource load balancing objective function value is defined as the reciprocal of this variance; the smaller the variance, the more balanced the load, and the larger the reciprocal.

[0097] The multi-objective optimization function is composed of the objective function of the overall task completion delay and the objective function of system resource load balancing. During the optimization process, each individual chromosome needs to calculate these two objective values ​​to evaluate its merits.

[0098] This embodiment establishes a prediction model for execution latency and network latency, and uses pipeline critical path latency as a performance indicator. By predicting the node load after task allocation and calculating its inverse variance as a balance indicator, the abstract optimization objective is transformed into a computable mathematical function, providing an accurate fitness evaluation basis for the NSGA-II algorithm and ensuring the actual effectiveness of the optimization results in both reducing business latency and improving resource utilization.

[0099] In some embodiments, the selection of candidate execution nodes corresponding to each atomic processing unit and the selection of inter-node transmission paths are jointly encoded as chromosome individuals of the optimization problem, including: The chromosome coding structure is determined. Chromosomes are composed of sequentially arranged gene segments, and each gene segment corresponds to an atomic processing unit in the initial processing pipeline. For each atomic processing unit, the corresponding gene segment is encoded using an integer encoding method. The value space of this integer is a list of index numbers of all nodes in its candidate execution node set. The gene value directly represents the execution node to which the atomic processing unit is assigned. In the chromosome, a path gene segment is added for each logical connection that needs to transmit intermediate data streams. The logical connection is determined by the execution node of the upstream atomic processing unit and the execution node of the downstream atomic processing unit. For each path segment, a variable-length integer sequence encoding method is used. Each integer in the sequence represents the index number of the edge computing node or network switching device that the path passes through. The entire sequence represents the complete transmission path from the upstream node to the downstream node. In the path gene segment, a priority coding bit is added. This priority coding bit adopts discrete value coding, and its value is associated with the content criticality level in the dynamic service feature vector. It is used to indicate the scheduling priority that the intermediate data stream should be assigned on the transmission path. The atomic processing unit gene segments, path gene segments, and priority encoding bits are spliced ​​together according to the execution logic order of the initial processing pipeline to form a complete chromosome individual. This chromosome individual fully represents the combination of the atomic processing unit execution node mapping scheme and the intermediate data stream transmission path and priority planning scheme.

[0100] In this embodiment, when determining the chromosome coding structure, the chromosome is pre-set to contain the same number of atomic processing unit gene segments as the total number of atomic processing units in the initial processing pipeline. The arrangement order of these gene segments in the chromosome is consistent with the execution order of the atomic processing units in the pipeline.

[0101] For each atomic processing unit, the corresponding gene segment is encoded using an integer encoding method. The value space of this integer, i.e., the list of index numbers of all nodes in the candidate execution node set, is generated for this unit when constructing the candidate set in step S301. Each index number in the list uniquely corresponds to a candidate edge computing node. The gene value is directly taken from this list, indicating that the atomic processing unit has been assigned to the node corresponding to the index number for execution.

[0102] In a chromosome, a path gene segment is added for each logical connection that requires the transmission of intermediate data streams. Logical connections are determined based on the data dependencies between atomic processing units in the initial processing pipeline; that is, if the output of one atomic processing unit serves as the input of another, a logical connection exists between them. Path gene segments are arranged in the chromosome immediately following all atomic processing unit gene segments.

[0103] For each path segment, a variable-length integer sequence encoding method is used. Each integer in the sequence represents a globally unique index of a network entity traversed by the path, including edge computing nodes and network switching devices. Global indices are assigned to all known network devices during system initialization. The entire sequence lists all entity indices traversed by the path in order from upstream to downstream nodes, thus fully representing a transmission path.

[0104] A priority encoding bit is added to the path gene segment. This priority encoding bit uses discrete value encoding, such as integers like 0, 1, and 2. Its value is associated with the content criticality level in the dynamic service feature vector by querying a preset mapping table. For example, "high" criticality level is mapped to a priority value of 2, "medium" to 1, and "low" to 0. This value is used to indicate the priority level that network devices should give to this data stream during scheduling.

[0105] The atomic processing unit gene segments, path gene segments, and priority encoding bits are concatenated according to the execution logic order of the initial processing pipeline. Specifically, according to the execution sequence of the atomic processing units, their node selection gene segments are placed sequentially, and the corresponding path gene segment (which already contains priority encoding bits) is inserted between every two units with data dependencies, forming a complete chromosome individual. This individual encodes the complete atomic processing unit execution node mapping scheme and the transmission paths and priority planning schemes of all intermediate data streams in a linear data structure.

[0106] This embodiment employs a hybrid encoding scheme, organically combining integer encoding for node selection with variable-length sequence encoding for path planning, and embedding business-driven priority information. This allows a single chromosome to completely and unambiguously represent a complex scheduling and routing joint scheme. This encoding method directly reflects the problem structure, facilitating efficient exploration of the solution space by genetic operations (crossover, mutation), which is a key foundation for the NSGA-II algorithm to successfully solve this multi-objective optimization problem.

[0107] In some embodiments, each edge computing node is controlled to execute the corresponding atomic processing unit according to the mapping relationship, and during the execution process, online adaptive migration of the transmission path and execution node is triggered based on the dynamic criticality level changes in the dynamic service feature vector, including: During the execution of the atomic processing unit, the dynamic business feature vector of the associated data stream is continuously monitored to detect whether the dynamic criticality level has jumped. If a dynamic criticality level jump is detected, the current execution node and transmission path are evaluated to determine whether they can still meet the quality of service constraints based on the latency and reliability requirements corresponding to the jump level. If the evaluation result is unsatisfactory, a migration decision process is initiated for the affected atomic processing units and their associated data streams based on the updated global resource status view, including: Using the improved quality of service constraints as a hard condition, the system quickly finds new execution nodes for the affected atomic processing units from the set of edge computing nodes that meet the conditions, and re-plans the transmission paths for their input and output data streams. Based on the new execution node and transmission path obtained through optimization, a migration instruction set containing atomic processing unit context migration and data stream switching timing is generated; Distribute the migration instruction set to relevant edge computing nodes and network forwarding devices; Coordinate relevant nodes and devices to perform operations such as migration of atomic processing unit computing state, transfer of intermediate data, and update of data stream forwarding path according to the timing specified by the instruction set, and complete online adaptive migration.

[0108] In this embodiment, the intelligent decision-making center is responsible for continuously monitoring the dynamic business feature vector of the associated data stream, which is achieved by subscribing to the real-time update event stream of this vector. Detecting whether a dynamic criticality level has jumped is done by comparing the current content criticality level field value in the vector with the previous stable state value of the data stream cached in the system. If the current level value is higher than the historical value and the difference exceeds a preset sensitivity threshold, a jump is determined. This sensitivity threshold is pre-determined through experimental calibration based on the system's response requirements to sudden business changes. It is important to note that the "jump" detected here specifically refers to the criticality level being raised to a level requiring higher service quality (such as shorter latency and higher reliability), serving as a threshold for triggering resource reallocation decisions and avoiding unnecessary migration overhead due to slight fluctuations or declines in the criticality level.

[0109] When evaluating latency and reliability requirements based on the upgraded criticality level, these requirements are obtained by querying a pre-defined mapping table between criticality level and quality of service constraints. This mapping table is predefined before system deployment based on the urgency and importance of fault handling in different business scenarios; for example, a "high" criticality level is mapped to a millisecond-level latency cap and near-100% reliability requirements. The evaluation process, based on the updated global resource status view, calculates the estimated latency required by the current execution node to process the atomic processing unit, as well as the end-to-end latency and packet loss rate of the current transmission path, and compares these values ​​with the new constraint values ​​retrieved from the mapping table.

[0110] If the evaluation results are unsatisfactory, the migration decision process is initiated. This process uses the upgraded quality of service constraints retrieved from the aforementioned mapping table as a hard screening criterion, quickly selecting a set of nodes from all edge computing nodes whose computing resource profiles meet the processing capacity requirements and whose interconnection status indicates that their network connectivity meets the new latency and reliability requirements. Within this set, a new execution node is selected for the affected atomic processing unit, and a new transmission path that meets the constraints is replanned for its input and output data streams based on the current network topology. Preferably, migration overhead and expected performance are considered comprehensively during the selection process.

[0111] Based on the optimized new execution node and transmission path, the migration instruction set needs to include instructions to migrate the current computing state of the atomic processing unit from the original node to the new node, instructions to instruct network devices to switch the data flow from the old path to the new path, and timing control instructions to coordinate the execution order of these two types of operations. The specific format and content of the instructions are predefined according to the system's internal communication protocol.

[0112] The migration instruction set is distributed to relevant edge computing nodes and network forwarding devices, and different instruction subsets are sent to the corresponding execution entities through reliable control channels, such as message queues or remote procedure calls.

[0113] The system coordinates relevant nodes and devices to perform operations according to the timing specified in the instruction set. Upon receiving the instruction, each node is required to synchronously or sequentially perform computational state transmission, memory data transfer, and flow table entry update according to the time point specified in the instruction or by waiting for a specific signal. Ultimately, this enables the atomic processing unit to resume execution on the new node, and the data flow to be transmitted along the new path, thereby completing the online adaptive migration.

[0114] This embodiment achieves dynamic resource assurance for high-value, sudden business operations by monitoring critical changes in real time and triggering rapid local re-optimization and collaborative migration when constraints are not met. This mechanism avoids the high overhead and latency caused by global rescheduling, and can quickly improve the service quality of critical businesses with minimal disturbance, significantly enhancing the system's adaptability and resilience in dynamically changing environments.

[0115] In some embodiments, using the upgraded quality of service constraint as a hard condition, from the set of edge computing nodes that meet the condition, a new execution node is quickly selected for the affected atomic processing unit, and its input and output data streams are re-planned for transmission, including: Based on the latency tolerance limit and reliability requirements corresponding to the upgraded dynamic criticality level, and combined with the updated global resource status view, nodes that meet the processing capacity requirements and the latency and reliability constraints in the computing resource profiles are selected from all edge computing nodes to form a set of feasible execution nodes. For each candidate node in the set of feasible execution nodes, evaluate the local cost of migrating the affected atomic processing unit to that node. The local cost includes at least the estimated time for migrating the atomic processing unit's computation context to that node, the estimated execution latency required for that node to process the unit, and the estimated network latency of the replanned input / output data flow path for that unit. While assessing the local costs, ensure that each link selected for the newly planned input / output data flow path meets the reliability requirements after the leap. From the set of feasible execution nodes, select the candidate node that minimizes the local cost and determine it as the new execution node; The input and output data flow paths planned for the new execution node during the evaluation process are determined as the replanned transmission paths.

[0116] In this embodiment, the computing resource profile satisfies the processing capacity, meaning that the real-time available computing power of the candidate node is not lower than the estimated processing complexity of the affected atomic processing unit; the network connectivity satisfies the constraints, meaning that the latency and packet loss rate indicators recorded in the global resource status view of the links between the candidate node and the data source node and the data destination node are lower than the packet loss rate thresholds calculated from the latency tolerance upper limit and reliability requirements after the jump.

[0117] When evaluating the local cost for each candidate node in the set of feasible execution nodes, the estimated time for the atomic processing unit to migrate its computation context to that node can be estimated based on the size of the computational state data to be migrated and the currently available network bandwidth. The estimated execution latency required for that node to process the unit is obtained by dividing the estimated processing complexity of the unit by the real-time available computing power of the node. The estimated network latency of the replanned input / output data flow path for the unit is estimated by accumulating the transmission latency of each link on the new path and adding the data flow size divided by the transmission time of the path bottleneck bandwidth. The local cost is the sum of the above three times.

[0118] While assessing the local costs, ensure that each link selected for the newly planned input and output data flow path meets the reliability requirements after the jump. Specifically, this is achieved by checking whether the real-time packet loss rate of the link in the global resource status view is lower than the packet loss rate threshold preset according to the reliability requirements.

[0119] From the set of feasible execution nodes, select the candidate node that minimizes the local cost and designate it as the new execution node for the affected atomic processing unit. The input / output data flow paths planned for this new execution node during the evaluation process, and whose reliability requirements have been verified for all links, are directly determined as the replanned transmission paths.

[0120] This embodiment constructs a feasible node set and evaluates local costs centered on migration overhead, processing latency, and transmission latency, thus realizing a fast and efficient local re-optimization mechanism. Under the premise of strictly meeting the quality of service constraints after the leap, it aims to minimize the total migration and execution latency, intelligently selects the optimal new execution node and transmission path for sudden high-critical business, ensuring the timeliness and economy of online adaptive migration, and avoiding complex global rescheduling.

[0121] In some embodiments, based on the optimized new execution node and transmission path, a migration instruction set containing atomic processing unit context migration and data stream switching timing is generated, including: The optimization results are analyzed to extract the new execution node identifier of the affected atomic processing unit, the new source node identifier and new transmission path of the input data stream, and the new target node identifier and new transmission path of the output data stream. Generate atomic processing unit context transition instructions. The atomic processing unit context transition instructions include the identifier of the computation state data to be transitioned, the source execution node address, the target execution node address, and the state transition triggering conditions. The computation state data includes at least the model parameters required for the atomic processing unit to run, the contents of the intermediate computation result buffer, and the session identifier. Generate a data stream switching preparation instruction. The data stream switching preparation instruction is sent to the current source node of the input data stream, instructing it to simultaneously copy and cache the data packets to be forwarded to the exit queues corresponding to the old and new transmission paths within a specified time window. A path update command is generated and sent to the network forwarding devices involved in the new transmission path. This command is used to pre-configure flow table rules, which specify that data packets matching the data flow identifier should be forwarded to the new execution node. Generate timing synchronization instructions, which specify the execution order and timestamps of context migration instructions, data stream switching preparation instructions, and path update instructions; timing synchronization instructions ensure that after the computation state is restored in the new execution node, the flow table rules on the new transmission path have taken effect, and the input data stream has started to be transmitted along the new path. The atomic processing unit context migration instructions, data stream switching preparation instructions, path update instructions, and timing synchronization instructions are encapsulated and serialized to form a migration instruction set.

[0122] In this embodiment, parsing the optimization result refers to reading the data structure output by the fast optimization process. This structure explicitly records the new execution node identifier and the new transmission path of the input and output data streams, described in the form of a node sequence. From the node sequence of the new transmission path, the starting node identifier is extracted as the new source node identifier, and the ending node identifier is extracted as the new target node identifier.

[0123] When generating an atomic processing unit context migration instruction, the identifier of the computational state data to be migrated is generated by the resource monitoring agent on the original execution node according to predefined rules. This identifier uniquely identifies the runtime state of the atomic processing unit instance. The source and target execution node addresses are obtained by querying the node information table in the global resource state view. The state migration trigger condition is typically set to receiving a specific start signal issued by a timing synchronization instruction.

[0124] When generating the data stream switching preparation instruction, the specified time window is determined based on the estimated completion time of the atomic processing unit context migration and the network transmission delay. This instruction requires the current source node of the input data stream to immediately perform a copy operation on subsequent data packets matching the data stream identifier upon receiving the instruction, and temporarily store the copies in the exit queue buffer pointing to the new execution node, while the original data packets continue to be sent along the old path.

[0125] The flow table rules encapsulated in the path update command contain packet characteristics (such as source and destination addresses, port numbers, and protocol types) for accurately matching data flows, as well as the corresponding forwarding actions (i.e., the next-hop address). These rules are pre-distributed to network switching devices along the new transmission path, but are set to a dormant state, awaiting activation.

[0126] The timing synchronization instructions explicitly state: First, ensure that the path update instructions have been issued and pre-configured; then issue the data flow switching preparation instructions; after the atomic processing unit context migration instructions have been executed and the new node confirms that the state has been restored, send an activation signal to the relevant network devices to enable the new flow table, and at the same time notify the source node to start sending data from the cache queue of the new path.

[0127] The generated atomic processing unit context migration instructions, data stream switching preparation instructions, path update instructions, and timing synchronization instructions are assembled and binary serialized according to the message format agreed upon within the system, ultimately forming a complete migration instruction set data packet that can be independently transmitted and parsed.

[0128] This embodiment generates a set of precise and time-controlled control instructions, transforming migration decisions into specific operational steps that can be executed collaboratively by distributed nodes and network devices. This ensures the atomicity and consistency of computing state migration and data flow path switching, providing technical support for achieving uninterrupted online adaptive migration of services.

[0129] In some embodiments, at a designated aggregation edge computing node, intermediate processing results from upstream nodes are received, and a weighted fusion analysis is performed using an attention mechanism to generate the final device state fusion diagnostic result, including: At the aggregated edge computing node, multiple intermediate processing results corresponding to the same monitored device are received from different upstream nodes; The received intermediate processing results are timestamped and their data formats are standardized to ensure data consistency in both time and semantic dimensions. Extract feature vectors representing the device state from each intermediate processing result, and concatenate all feature vectors to form a fused feature matrix; Meanwhile, metadata is extracted for each intermediate processing result. The metadata includes at least the confidence level of the atomic processing unit that generated the result and the real-time load status of the source edge computing node. The fused feature matrix and metadata are input into the attention weight calculation network. The attention weight calculation network calculates a dynamic attention weight for each intermediate processing result based on the correlation between feature vectors and the reliability of metadata representation. Using the calculated attention weights, the eigenvectors in the fusion feature matrix are summed in a weighted manner to obtain the weighted fusion feature vector. The weighted fusion feature vector is input into the pre-trained device status diagnosis classification model for inference analysis; The inference results of the output device status diagnosis classification model are used as the final device status fusion diagnosis results.

[0130] In this embodiment, when receiving intermediate processing results from different upstream nodes, the aggregation edge computing node groups multiple results belonging to the same monitored device based on the data source device identifier encapsulated in each result. The received intermediate processing results are then timestamped using a linear interpolation method based on a reference time point to align the feature data of each result to a unified time. Data format unification is achieved by calling predefined data transformation functions to map feature data from different sources and dimensions to the same vector space and numerical range.

[0131] Feature vectors representing device status are extracted from each intermediate processing result. If the result itself is structured data, key fields are extracted using a preset feature selection or projection layer to construct a vector; if the result is already in vector form, it is used directly. All feature vectors are concatenated row-wise to form a fused feature matrix. When extracting metadata accompanying each intermediate processing result, the confidence level of the atomic processing unit is calculated and appended by the unit itself when outputting the result, such as classification probability or error estimation based on its internal algorithm. The real-time load status of the source edge computing node is obtained from the CPU utilization and memory usage fields of that node in the latest global resource status view.

[0132] The attention weight calculation network learns the interaction relationships between feature vectors and the reliability differences indicated by metadata, outputting a scalar weight value for each feature vector in the fused feature matrix. The higher the weight value, the greater the contribution of that vector to the final diagnosis. Using the calculated attention weights, each feature vector in the fused feature matrix is ​​multiplied by its corresponding weight, and all vectors are summed to obtain a comprehensive weighted fused feature vector.

[0133] The pre-trained equipment status diagnosis classification model is a multi-classifier based on a deep neural network. Its training process uses a large amount of historical case data. The input is a feature vector that has been aligned and weighted, and the label is the actual equipment status confirmed manually or by rules. By inputting the weighted fused feature vector into this model, the current equipment status category and confidence level can be inferred.

[0134] This embodiment processes multi-source asynchronous data through time alignment and format standardization, dynamically fuses features with different reliability and correlation using an attention mechanism, and finally performs accurate diagnosis through a pre-trained model, effectively improving the accuracy and robustness of device status assessment in a multimodal, distributed processing environment.

[0135] In some embodiments, the fused feature matrix and metadata are jointly input into the attention weight calculation network. Based on the correlation between feature vectors and the reliability of the metadata representation, the attention weight calculation network calculates a dynamic attention weight for each feature vector corresponding to an intermediate processing result, including: Each feature vector in the fused feature matrix is ​​linearly transformed through a shared fully connected layer to generate the corresponding query vector, key vector, and value vector, respectively. The metadata accompanying each intermediate processing result is encoded by an independent metadata encoder to generate a metadata embedding vector that characterizes the reliability of the intermediate processing result. The key vector is concatenated with the corresponding metadata embedding vector to form an enhanced key vector; The query vector is multiplied by all the enhanced key vectors to obtain an initial attention score, which represents the correlation between feature vectors. The initial attention score is input into a gated adjustment network based on metadata embedding vectors. The gated adjustment network outputs an adjustment factor, which is used to scale the initial attention score according to the reliability of the metadata representation, to obtain the scaled attention score. The scaled attention scores are then normalized to obtain the final attention weights.

[0136] In this embodiment, the shared fully connected layer is a neural network layer with a fixed weight matrix. This layer performs three independent matrix multiplication operations on each feature vector in the fused feature matrix, thereby generating a query vector, key vector, and value vector of the same dimension for each feature vector. These three vectors are used for subsequent attention matching, matching, and information aggregation, respectively.

[0137] The independent metadata encoder is a small, fully connected neural network whose input is a numerical vector consisting of metadata accompanying each intermediate processing result. This encoder maps the metadata into a low-dimensional, dense metadata embedding vector through a nonlinear transformation. This vector compresses and characterizes the reliability information of the intermediate processing result.

[0138] The key vector and its corresponding metadata embedding vector are concatenated end-to-end along their feature dimensions to form a new, higher-dimensional enhanced key vector. This operation fuses feature content information with source reliability information at the representation level.

[0139] The query vector is multiplied by the dot product of all the enhanced key vectors. For each query vector, its inner product with each enhanced key vector is calculated, resulting in a set of scalar values, which are the initial attention scores. This score reflects the combined correlation between the features that the current query vector focuses on and the features represented by each enhanced key vector.

[0140] The gating modulation network based on metadata embedding vectors takes metadata embedding vectors as input, passes them through a fully connected layer with a sigmoid activation function, and outputs a modulation factor between 0 and a positive number. This factor is used to scale the initial attention score element-wise. If the metadata indicates high reliability, a factor greater than 1 is output to amplify the score; otherwise, a factor less than 1 is output to suppress the score.

[0141] The scaled attention scores are normalized, usually using the Softmax function, which converts the scaled scores into a probability distribution. That is, after all scores are exponentially operated and normalized, their sum is 1. The result is the final attention weight.

[0142] This embodiment incorporates a metadata-aware attention mechanism, explicitly considering the reliability of data sources when calculating feature relevance, and dynamically adjusts weight allocation through a gating network. This allows the fusion process to place greater trust in analysis results from high-confidence, low-load nodes, thereby improving the accuracy and robustness of the final device status diagnosis.

[0143] In some embodiments, an initial processing pipeline is formed by selecting and combining atomic processing units from a pre-built library based on the service type in the dynamic service feature vector, including: Extract the business type from the dynamic business feature vector; Based on the parsed business type, query the predefined business type-pipeline template mapping relationship to determine the corresponding initial pipeline template. The initial pipeline template specifies the type and execution order constraints of the atomic processing units required to complete this type of business. Based on the initial pipeline template, select specific versions of atomic processing unit instances from the pre-built atomic processing unit library, taking into account the data format information contained in the dynamic business feature vector when selecting; Obtain the input / output interface definitions of the selected atomic processing unit instance, and construct a data dependency graph between atomic processing units based on the data types and semantics of the interface definitions; The data dependency graph is a directed acyclic graph structure, where nodes represent instances of atomic processing units and directed edges represent data flow directions. Based on the data dependency graph and the execution order constraints in the initial pipeline template, an initial processing pipeline with a clear execution order and data transfer relationship is generated. The initial processing pipeline is described using a graph structure that includes a sequence of nodes and a set of edges.

[0144] In this embodiment, the field value storing business type information in the dynamic business feature vector is read. Based on the parsed business type, the predefined business type-pipeline template mapping relationship is queried. This mapping relationship is pre-configured based on domain knowledge during system deployment. For example, the "video inspection" business type is mapped to a template that includes atomic processing unit types and sequences such as "video decoding", "object detection", and "result encapsulation".

[0145] The atomic processing unit library is stored as a container image or function package. Each instance comes with metadata describing its functional type, version, and supported data formats. The selection process first filters candidate instances based on the unit type required by the template, and then selects versions with compatible input interfaces from the candidate instances based on the data format information in the dynamic business feature vector.

[0146] The system retrieves the input / output interface definitions of the selected atomic processing unit instance. These definitions, read from the instance's metadata file, specify the data structure, semantics, and output data type required by the instance. Based on these interface definitions, a data dependency graph between atomic processing units is automatically constructed by matching the output type of upstream instances with the input type of downstream instances. This data dependency graph is a directed acyclic graph (DAG), whose acyclicity ensures that the pipeline is free from deadlocks or circular waits and can be scheduled for execution. Nodes represent selected atomic processing unit instances, and directed edges point from instances providing data to instances consuming data, explicitly representing the producer-consumer relationship of intermediate data.

[0147] Based on the data dependency graph and the execution order constraints in the initial pipeline template, a graph topology sorting algorithm is used to generate a linear or parallel branching sequence of node executions. This sequence satisfies both data dependencies and template constraints, thus forming an initial processing pipeline with a clear execution order and data transfer relationships. This pipeline is described and maintained in the system using a graph data structure containing information about all node objects and their connecting edges. The initial processing pipeline is described by a graph structure containing a node sequence and an edge set. The node sequence is the list of atomic processing unit instances after topology sorting, and the edge set is all directed edges in the data dependency graph, fully defining the logical topology of the processing task.

[0148] This embodiment transforms business requirements into a processing logic framework through predefined templates and instantiates specific processing units based on data format compatibility, ultimately constructing an executable task dependency graph. This achieves automated and standardized transformation from abstract business types to specific, distributed execution pipelines, laying the foundation for subsequent optimization and scheduling.

[0149] In some embodiments, based on an initial pipeline template, a specific version of an atomic processing unit instance is selected from a pre-built atomic processing unit library. The selection process references data format information contained in the dynamic business feature vector, including: Parse the initial pipeline template to obtain the type identifier of each atomic processing unit specified in the initial pipeline template; Based on the type identifier of each atomic processing unit, a search is performed in the pre-built atomic processing unit library to obtain a list of all candidate atomic processing unit instances that match the type identifier; Extract the data format information of the original multimodal data stream from the dynamic business feature vector. The data format information includes at least the encoding format, resolution, sampling rate, and quantization accuracy. For each candidate atomic processing unit instance, read its predefined input data format requirements; The matching degree is calculated by comparing the data format information of the original multimodal data stream with the input data format requirements of the candidate atomic processing unit instances. The matching degree calculation includes checking the encoding format compatibility, resolution scaling feasibility, and sampling rate adaptability. Based on the matching degree calculation results, a format adaptation score is generated for each candidate atomic processing unit instance; Combining the format fit score with the version information of the atomic processing unit instance, the atomic processing unit instance with the highest format fit score and the version that meets the preset requirements is selected from the candidate list and used as the final atomic processing unit instance corresponding to this type identifier.

[0150] In this embodiment, parsing the initial pipeline template involves reading the atomic processing unit type identifier string defined for each processing step in the template data structure. Based on this type identifier, the pre-built atomic processing unit library is queried using an index table with the type identifier as the key to retrieve all instances in the metadata that declare the same type identifier, forming a candidate list.

[0151] The data format information of the original multimodal data stream is a component of the dynamic business feature vector and is directly read from the corresponding fields of the dynamic business feature vector, such as the encoding format field, resolution field, and sampling rate field.

[0152] For each candidate atomic processing unit instance, predefined input data format requirements are read from its accompanying metadata description file. These requirements explicitly list the set of input encoding formats supported by the instance, the acceptable resolution range or list, and the supported sampling rate values.

[0153] The matching degree is calculated between the data format information of the original multimodal data stream and the input data format requirements of the candidate atomic processing unit instance. Specifically, the encoding format compatibility check is performed to see if the encoding format of the data stream exists in the set of formats supported by the instance; the resolution scaling feasibility check is performed to see if the resolution of the data stream falls within the acceptable range declared by the instance, or if the instance is marked as having internal scaling capabilities; and the sampling rate adaptability check is performed to see if the sampling rate of the data stream is consistent with the sampling rate supported by the instance, or if they are integer multiples of each other to facilitate resampling.

[0154] The format fit score is generated based on the matching degree calculation results. Specifically, if the encoding format is incompatible, the score is zero. If it is compatible, sub-scores are assigned according to the degree of resolution matching (e.g., perfect match, within range, scaling required) and the degree of sampling rate matching (e.g., perfect match, resampling is possible). These sub-scores are then added together according to preset weights to obtain the final format fit score.

[0155] The selection process combines the format fit score with the version information of the atomic processing unit instance. First, instances with version numbers lower than the minimum system requirement are excluded. Then, the instance with the highest format fit score from the remaining instances is selected as the final atomic processing unit instance corresponding to that type identifier. If multiple instances have the same score, the instance with the higher version number is selected first.

[0156] This embodiment achieves automatic selection of the instance that best matches the current data stream from multiple atomic processing units with the same function through fine-grained data format compatibility checks and quantitative scoring. This ensures that each processing step in the pipeline can process the input data efficiently and correctly, avoids processing failures or performance degradation caused by format mismatches, and improves the robustness and processing efficiency of the system.

[0157] By adopting the above technical solutions, this invention differs from existing technologies and possesses the following beneficial effects: It quantifies the latency, reliability, and dynamic criticality levels of multimodal services through dynamic service feature vectors; it unifies the perception and modeling of 5G private network wireless channels, backhaul links, edge computing node resources, and inter-node interconnection status through a global resource status view, providing accurate input for subsequent optimization. Based on this, with the objectives of minimizing overall task completion latency and maximizing resource utilization, the NSGA-II multi-objective optimization algorithm is used to dynamically map and reconstruct the initial processing pipeline. This achieves simultaneous joint optimization of allocating suitable edge computing nodes to atomic processing units and planning transmission paths and priorities for intermediate data streams, thereby significantly improving overall resource utilization efficiency while ensuring service performance. Furthermore, by monitoring dynamic criticality level changes and triggering online adaptive migration during execution, real-time resource reconfiguration can be performed for sudden high-value services, ensuring low latency and high reliability of critical services. Finally, an attention mechanism is used at the aggregated edge computing nodes to weightedly fuse multi-source intermediate processing results, generating accurate device status fusion diagnostic results and updating the global resource status view to form an optimization closed loop. This invention enables dynamic adjustment of resource allocation based on the intrinsic value density of services in complex power plant environments, thereby significantly improving the overall resource utilization efficiency and system resilience of 5G private networks and edge computing clusters while strictly ensuring the quality of critical business services.

[0158] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A data transmission method for power plant edge computing and 5G collaboration, characterized in that, include: Acquire raw multimodal data streams from heterogeneous sensing devices in the power plant and the corresponding analysis task requests; Based on the initial attributes of the analysis task request and the original multimodal data stream, a dynamic business feature vector is generated; Based on the business type in the dynamic business feature vector, select and combine from the pre-built atomic processing unit library to form an initial processing pipeline; Real-time collection of 5G private network wireless channel status, backhaul link load, resource utilization of each edge computing node and interconnection status between nodes, and fusion to generate a global resource status view; With the joint optimization goal of minimizing the overall task completion latency and maximizing resource utilization, dynamic mapping and reconstruction are performed on the initial processing pipeline based on the global resource status view and the dynamic business feature vector; The dynamic mapping and reconstruction include: Assign appropriate edge computing nodes as execution nodes to each atomic processing unit in the pipeline, and plan the transmission path and priority for the intermediate data flow between nodes based on the NSGA-II multi-objective optimization algorithm; Each edge computing node is controlled to execute the corresponding atomic processing unit according to the mapping relationship, and during the execution process, the transmission path and execution node are triggered to adapt online based on the dynamic criticality level change in the dynamic service feature vector. At the designated aggregation edge computing node, intermediate processing results from upstream nodes are received, and a weighted fusion analysis is performed using an attention mechanism to generate the final device status fusion diagnostic results. Output the fusion diagnostic results and update the global resource status view.

2. The data transmission method for power plant edge computing and 5G collaboration according to claim 1, characterized in that, Based on the analysis task request and the initial attributes of the original multimodal data stream, a dynamic business feature vector is generated, including: Parse the analysis task request, extract the task identifier and expected output format, and generate an initial business type label based on the task identifier; Obtain the source device identifier, data format, and sampling rate of the original multimodal data stream as initial attributes of the data stream; Based on the initial business type label and the initial data flow attributes, the initial latency tolerance limit and reliability requirements are determined by querying the predefined business type-service quality mapping table. Based on the initial business type label and data format, the initial processing complexity estimate is calculated using the complexity prediction model. At the first edge computing node that receives the original multimodal data stream, a pre-deployed pre-analysis atomic processing unit is invoked to perform real-time analysis on the data stream and obtain real-time analysis results. The real-time analysis results are matched with a preset abnormal pattern library, and the content criticality level is dynamically evaluated and output based on the matching confidence level. The initial business type label, initial latency tolerance limit, reliability requirements, initial processing complexity estimate, and content criticality level are vectorized and encapsulated to generate the dynamic business feature vector. Furthermore, during the execution of subsequent atomic processing units, if new real-time analysis results are received that cause a change in the content criticality level, a real-time update operation on the dynamic business feature vector is triggered.

3. The data transmission method for power plant edge computing and 5G collaboration according to claim 2, characterized in that, The real-time analysis results are matched against a pre-defined anomaly pattern library. Based on the matching confidence level, the criticality level of the content is dynamically evaluated and output, including: The structured feature descriptor is extracted from the real-time analysis results. The structured feature descriptor includes at least the anomaly type, spatial location, timestamp, and quantization parameters. The structured feature descriptor is input into a pre-built anomaly pattern library, which stores multi-dimensional feature templates of historical anomaly events, and each feature template is associated with a preset key benchmark value. An attention-based feature matching algorithm is used to calculate the similarity between the structured feature descriptor and each feature template in the abnormal pattern library. Based on the similarity, the key benchmark values ​​associated with each feature template are weighted and fused to generate a preliminary key score. Obtain the data stream identifier corresponding to the real-time analysis result, and query the historical key score sequence of the data stream within a preset time window; The preliminary criticality score and the historical criticality score sequence are input into the criticality state transition model, which is constructed based on a hidden Markov chain and is used to smoothly evaluate and predict the evolution trend of criticality states. Based on the output of the critical state transition model, the final content criticality level at the current moment is determined. The content criticality level is used to characterize the value density and urgency of the information contained in the data stream for subsequent fusion decision-making tasks.

4. The data transmission method for power plant edge computing and 5G collaboration according to claim 1, characterized in that, Real-time data collection of 5G private network wireless channel status, backhaul link load, resource utilization of each edge computing node, and interconnection status between nodes are used to generate a global resource status view, including: By using probes deployed on the 5G base station side, the reference signal received power and signal-to-interference-to-noise ratio of each area are periodically collected as the wireless channel status. By deploying link monitoring points between user plane functional network elements and edge computing nodes, the current uplink and downlink data throughput and transmission latency are collected, and the backhaul link load is calculated. By deploying resource monitoring agents on each edge computing node, the utilization rate of the central processing unit, the utilization rate of the graphics processing unit, and the memory usage rate are periodically collected as the node resource utilization rate. Based on node resource utilization and node hardware configuration information, a computing resource profile of each edge computing node is constructed. By running link layer discovery protocols or active probing mechanisms, the network topology connections between edge computing nodes are obtained, and the round-trip latency and packet loss rate of links between nodes are measured to assess the interconnection status between nodes. The collected wireless channel status, backhaul link load, node resource utilization and inter-node interconnection status are cleaned and normalized to unify the heterogeneous indicators to the same numerical units. The normalized wireless channel state vector, backhaul link load scalar, computational resource profile vector of each node, and inter-node interconnection state matrix are encapsulated to generate a global resource state view.

5. The data transmission method for power plant edge computing and 5G collaboration according to claim 4, characterized in that, Based on node resource utilization and node hardware configuration information, a computing resource profile for each edge computing node is constructed, including: Read the hardware configuration information of the edge computing node, which includes at least the central processing unit model and number of cores, the graphics processor model and video memory capacity, and the total memory capacity; Based on the hardware configuration information, the benchmark computing performance scores of the node in terms of central processing unit computing, graphics processing unit computing, and memory access are determined through a predefined hardware performance benchmark scoring model. The CPU utilization, GPU utilization, and memory usage are periodically collected by the resource monitoring agent and used to construct a real-time utilization vector. Based on the utilization values ​​of each dimension in the real-time utilization vector and the corresponding benchmark computing performance scores, the real-time available computing power of the node in each dimension at the current moment is calculated; the real-time available computing power represents the processing capacity remaining after deducting the occupied resources that can be used for new tasks. The benchmark computing performance score, real-time utilization vector, and real-time available computing power vector are combined and encapsulated to form a computing resource profile of the edge computing node.

6. The data transmission method for power plant edge computing and 5G collaboration according to claim 1, characterized in that, Assign appropriate edge computing nodes as execution nodes to each atomic processing unit in the pipeline, and plan transmission paths and priorities for intermediate data streams between nodes based on the NSGA-II multi-objective optimization algorithm, including: For each atomic processing unit in the initial processing pipeline, based on the node computing resource profile and the interconnection status between nodes in the global resource status view, edge computing nodes that meet its processing capabilities and communication requirements are selected to construct a candidate execution node set. With the optimization objectives of minimizing the overall task completion latency and maximizing the system resource load balance, a multi-objective optimization function is constructed. The calculation of the overall task completion latency requires a combination of the estimated execution latency of the atomic processing units on each candidate node and the estimated network latency of the intermediate data streams transmitted between nodes. The selection of candidate execution nodes for each atomic processing unit, as well as the selection of transmission paths between nodes, are collectively encoded as individual chromosomes of the optimization problem. Initialize the chromosome population by utilizing the content criticality levels in the global resource status view and dynamic business feature vectors; The NSGA-II multi-objective optimization algorithm is used to iteratively optimize the chromosome population, performing operations including non-dominated sorting, crowding calculation, selection, crossover and mutation, to approximate and obtain the Pareto optimal solution set. From the Pareto optimal solution set, based on the latency tolerance upper limit and reliability requirements in the dynamic service feature vector, the optimal solution is selected for decoding, including: The gene segments corresponding to the atomic processing units in the chromosome are decoded into defined edge computing nodes, which are then used as their execution nodes. Decode the gene segments corresponding to the inter-node transmission path into a defined path sequence and the transmission priority on that path; Output the mapping table between atomic processing units and execution nodes obtained from decoding, as well as the transmission path and priority planning table of intermediate data streams.

7. The data transmission method for power plant edge computing and 5G collaboration according to claim 6, characterized in that, With the optimization objectives of minimizing the overall task completion latency and maximizing the system resource load balancing, a multi-objective optimization function is constructed, including: For each atomic processing unit in the initial processing pipeline, based on its processing complexity and the real-time available computing power in the candidate execution node computing resource profile, the estimated execution latency of the unit on each candidate node is calculated using a predefined execution latency prediction model. For each path that may transmit intermediate data streams, the estimated network latency of the path is calculated based on the transmission latency of the links traversed by the path and the current load status in the global resource status view, using a predefined network latency prediction model. The objective function for the overall task completion time is defined as follows: Starting from the atomic processing unit at the beginning of the pipeline, based on the decoded node mapping and path planning, the estimated execution time of each atomic processing unit and the estimated network time of the path through which its output data flows are accumulated in the order of execution, and the resulting total time value is used as the first optimization objective. The objective function for system resource load balancing is defined as follows: Based on the decoded node mapping relationship, estimate the future resource utilization rate of each edge computing node after undertaking the allocated atomic processing unit, calculate the variance of the future resource utilization rate of all nodes, and take the reciprocal of the variance as the second optimization objective. The multi-objective optimization function is composed of the overall task completion delay objective function and the system resource load balancing objective function.

8. The data transmission method for power plant edge computing and 5G collaboration according to claim 1, characterized in that, Control each edge computing node to execute the corresponding atomic processing unit according to the mapping relationship, and trigger online adaptive migration of the transmission path and execution node based on the dynamic criticality level change in the dynamic service feature vector during the execution process, including: During the execution of the atomic processing unit, the dynamic business feature vector of the associated data stream is continuously monitored to detect whether the dynamic criticality level has jumped. If a dynamic criticality level jump is detected, the current execution node and transmission path are evaluated to determine whether they can still meet the quality of service constraints based on the latency and reliability requirements corresponding to the jump level. If the evaluation result is unsatisfactory, a migration decision process is initiated for the affected atomic processing units and their associated data streams based on the updated global resource status view, including: Using the improved quality of service constraints as a hard condition, the system quickly finds new execution nodes for the affected atomic processing units from the set of edge computing nodes that meet the conditions, and re-plans the transmission paths for their input and output data streams. Based on the new execution node and transmission path obtained through optimization, a migration instruction set containing atomic processing unit context migration and data stream switching timing is generated; Distribute the migration instruction set to relevant edge computing nodes and network forwarding devices; Coordinate relevant nodes and devices to perform operations such as migration of atomic processing unit computing state, transfer of intermediate data, and update of data stream forwarding path according to the timing specified by the instruction set, and complete online adaptive migration.

9. The data transmission method for power plant edge computing and 5G collaboration according to claim 8, characterized in that, Using the improved quality of service constraints as a hard condition, from the set of edge computing nodes that meet the conditions, a new execution node is quickly selected for the affected atomic processing units, and the transmission paths for their input and output data streams are replanned, including: Based on the latency tolerance limit and reliability requirements corresponding to the upgraded dynamic criticality level, and combined with the updated global resource status view, nodes that meet the processing capacity requirements and the latency and reliability constraints in the computing resource profiles are selected from all edge computing nodes to form a set of feasible execution nodes. For each candidate node in the set of feasible execution nodes, evaluate the local cost of migrating the affected atomic processing unit to that node. The local cost includes at least the estimated time for migrating the atomic processing unit's computation context to that node, the estimated execution latency required for that node to process the unit, and the estimated network latency of the replanned input / output data flow path for that unit. While assessing the local costs, ensure that each link selected for the newly planned input / output data flow path meets the reliability requirements after the leap. From the set of feasible execution nodes, select the candidate node that minimizes the local cost and determine it as the new execution node; The input and output data flow paths planned for the new execution node during the evaluation process are determined as the replanned transmission paths.

10. The data transmission method for power plant edge computing and 5G collaboration according to claim 1, characterized in that, At the designated aggregation edge computing node, intermediate processing results from upstream nodes are received, and a weighted fusion analysis is performed using an attention mechanism to generate the final device status fusion diagnostic results, including: At the aggregated edge computing node, multiple intermediate processing results corresponding to the same monitored device are received from different upstream nodes; The received intermediate processing results are timestamped and their data formats are standardized to ensure data consistency in both time and semantic dimensions. Extract feature vectors representing the device state from each intermediate processing result, and concatenate all feature vectors to form a fused feature matrix; Simultaneously, metadata is extracted from each intermediate processing result, including at least the confidence level of the atomic processing unit that generated the result and the real-time load status of the source edge computing node. The fused feature matrix and metadata are input together into the attention weight calculation network. The attention weight calculation network calculates a dynamic attention weight for each intermediate processing result based on the correlation between feature vectors and the reliability of metadata representation. Using the calculated attention weights, the eigenvectors in the fusion feature matrix are summed in a weighted manner to obtain a weighted fusion feature vector. The weighted fusion feature vector is input into a pre-trained device status diagnosis classification model for inference analysis; The inference results of the equipment status diagnosis classification model are output as the final equipment status fusion diagnosis results.

11. The data transmission method for power plant edge computing and 5G collaboration according to claim 1, characterized in that, Based on the service type in the dynamic service feature vector, an initial processing pipeline is formed by selecting and combining units from a pre-built atomic processing unit library, including: Extract the business type from the dynamic business feature vector; Based on the parsed business type, query the predefined business type-pipeline template mapping relationship to determine the corresponding initial pipeline template. The initial pipeline template specifies the type and execution order constraints of the atomic processing units required to complete this type of business. Based on the initial pipeline template, select specific versions of atomic processing unit instances from the pre-built atomic processing unit library, taking into account the data format information contained in the dynamic business feature vector when selecting; Obtain the input / output interface definitions of the selected atomic processing unit instance, and construct a data dependency graph between atomic processing units based on the data types and semantics of the interface definitions; The data dependency graph is a directed acyclic graph structure, where nodes represent atomic processing unit instances and directed edges represent data flow directions. Based on the data dependency graph and the execution order constraints in the initial pipeline template, an initial processing pipeline with a clear execution order and data transfer relationship is generated. The initial processing pipeline is described using a graph structure that includes a sequence of nodes and a set of edges.

Images