Data quality adaptive enhancement method and device for industrial edge gateway

Abstract

The application discloses a data quality adaptive enhancement method and device of an industrial edge gateway, comprising: protocol stripping and time sequence coarse alignment of an original data stream; root cause positioning and pollution propagation tracing based on an original data block sequence, a time sequence drift amount, a device health baseline and an industrial physical constraint atlas; construction of a multi-objective enhancement game model based on a physical violation point set, a trusted benchmark cluster and a pollution propagation path; edge side hierarchical enhancement execution based on the multi-objective enhancement game model and real-time resource state quantity; physical consistency closed loop verification based on the trusted benchmark cluster and a residual error tensor, to obtain a quality gain index and a strategy regret value; and dynamic transmission shaping of uplink data stream of the edge gateway according to the foregoing data. The application realizes physical consistency guarantee and multi-strategy collaborative optimization of edge side data quality, effectively improving the accuracy, resource utilization efficiency and transmission reliability of industrial edge side data quality enhancement.

Description

Technical Field

[0001] This application relates to the field of industrial IoT edge computing technology, and in particular to a method and apparatus for adaptive enhancement of data quality of industrial edge gateways. Background Technology

[0002] In the Industrial Internet of Things (IIoT) system, the edge gateway serves as a crucial hub connecting underlying sensing devices with the cloud data platform, undertaking core functions such as protocol conversion, data acquisition, preprocessing, and uploading. Industrial environments are complex, with diverse sensing node types and heterogeneous communication protocols. Furthermore, raw sensor data is often affected by factors such as electromagnetic interference, equipment aging, and network jitter, resulting in quality issues such as missing data, noise, outliers, and systematic biases. Uploading low-quality data directly to the cloud not only increases network transmission burden and cloud storage pressure but also distorts decision-making in upper-layer fault warnings and OEE analysis.

[0003] Existing data quality processing methods for industrial edge gateways typically employ single strategies such as fixed threshold filtering and linear interpolation completion. Each processing stage operates independently, lacking a deep understanding of the physical operating patterns of industrial equipment. This makes it difficult to distinguish between real physical anomalies and sensor data contamination, resulting in a persistently high false cleansing rate. Furthermore, various enhancement strategies, such as cleansing, reconstruction, and compensation, often compete and conflict on resource-constrained edge devices. Current technologies lack a collaborative optimization mechanism for the coupling relationships between these strategies, leading to an imbalance in edge computing resource allocation. Some critical data is not adequately enhanced, while secondary data is overprocessed. In addition, existing solutions often evaluate enhancement effects at the statistical residual level, failing to establish a consistency verification loop with industrial physical constraints. They cannot adaptively adjust strategy parameters based on enhancement effectiveness, and uplink transmission scheduling does not consider differences in data value density. This results in high-value data being delayed and lost due to bandwidth contention, while low-value data consumes a large amount of transmission resources. Summary of the Invention

[0004] The main objective of this application is to provide an adaptive data quality enhancement method for industrial edge gateways, which aims to solve the technical problems of poor data identification accuracy, unreasonable resource allocation, and poor transmission reliability in the process of data quality enhancement on the industrial edge side.

[0005] To achieve the above objectives, this application proposes an adaptive data quality enhancement method for industrial edge gateways, which includes: The raw data streams of heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway are obtained. The raw data streams are then stripped of protocols and coarsely aligned in time to obtain the raw data block sequence, time drift, equipment health baseline and industrial physical constraint map. Based on the original data block sequence, time-series drift, equipment health baseline, and industrial physical constraint map, the root cause of data quality defects and the traceability of pollution propagation are performed to obtain the set of physical violation points, the trusted benchmark cluster, and the pollution propagation path. A multi-objective augmented game model is constructed based on the physical violation point set, credible benchmark cluster, and pollution propagation path. The multi-objective augmented game model includes cleanup strategy participants, reconstruction strategy participants, and compensation strategy participants. The participants achieve strategy equilibrium through a Pareto collaborative optimization mechanism based on marginal utility evaluation. Based on the multi-objective enhanced game model and the real-time resource state quantity, the edge-side hierarchical enhanced execution is performed to obtain the enhanced data stream, residual error tensor and edge-side data value density distribution. Physical consistency closed-loop verification is performed based on the trusted benchmark cluster and the residual error tensor to obtain the quality gain index and the strategy regret value. The uplink data stream of the edge gateway is dynamically shaped based on the quality gain index, policy regret value, edge-side data value density distribution, and real-time resource status quantity to achieve adaptive enhancement of data quality at the edge of the industrial IoT.

[0006] Furthermore, to achieve the above objectives, this application also proposes a data quality adaptive enhancement device for an industrial edge gateway, which includes: The acquisition module is used to acquire the raw data stream of heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway, and to perform protocol stripping and time-series coarse alignment on the raw data stream and the real-time resource status to obtain the raw data block sequence, time-series drift, equipment health baseline and industrial physical constraint map. The traceability module is used to locate the root cause of data quality defects and trace the spread of contamination based on the original data block sequence, time drift, equipment health baseline and industrial physical constraint map, so as to obtain the physical violation point set, credible benchmark cluster and contamination propagation path; The construction module is used to construct a multi-objective augmented game model based on the set of physical violation points, the cluster of credible benchmarks and the pollution propagation path. The multi-objective augmented game model includes cleansing strategy participants, reconstruction strategy participants and compensation strategy participants, and the participants achieve strategy equilibrium through a Pareto collaborative optimization mechanism based on marginal utility evaluation. The execution module is used to perform edge-side hierarchical enhancement execution based on the multi-objective enhanced game model and the real-time resource state quantity, so as to obtain the enhanced data stream, residual error tensor and edge-side data value density distribution. The verification module is used to perform physical consistency closed-loop verification based on the trusted benchmark cluster and the residual error tensor to obtain the quality gain index and the strategy regret value. The transmission module is used to dynamically shape the uplink data stream of the edge gateway according to the quality gain index, the policy regret value, the edge-side data value density distribution, and the real-time resource status quantity, so as to achieve adaptive enhancement of data quality at the edge side of the Industrial Internet of Things.

[0007] Furthermore, to achieve the above objectives, this application also proposes an adaptive data quality enhancement device for an industrial edge gateway, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the adaptive data quality enhancement method for an industrial edge gateway as described above.

[0008] In addition, to achieve the above objectives, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the data quality adaptive enhancement method for industrial edge gateways as described above.

[0009] The proposed technical solutions, including one or more, incorporate an industrial physical constraint map to embed the physical conservation relationships, parameter coupling boundaries, and state evolution laws of equipment operation into the data quality analysis process, thereby achieving root cause localization of data quality defects based on physical consistency. By constructing a multi-objective augmented game model, cleaning, reconstruction, and compensation strategies are modeled as game participants, and a Pareto collaborative optimization mechanism based on marginal utility evaluation is used to achieve resource equilibrium and conflict resolution among strategies. Through physical consistency closed-loop verification and strategy regret value calculation, adaptive correction of augmented strategies is achieved. Furthermore, by driving dynamic transmission shaping through edge-side data value density distribution, differentiated and precise allocation of edge computing resources and network bandwidth is achieved, significantly improving the accuracy, resource utilization efficiency, and transmission reliability of data quality enhancement at the industrial edge. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a flowchart illustrating an embodiment of the adaptive enhancement method for data quality of industrial edge gateways in this application. Figure 2 This is a schematic diagram of the module structure of the data quality adaptive enhancement device for an industrial edge gateway according to an embodiment of this application.

[0013] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0015] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0016] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a data quality adaptive enhancement device for an industrial edge gateway. The following description uses a data quality adaptive enhancement device for an industrial edge gateway as an example to illustrate this embodiment and the subsequent embodiments.

[0017] Based on this, embodiments of this application provide an adaptive data quality enhancement method for industrial edge gateways, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the data quality adaptive enhancement method for industrial edge gateways in this application.

[0018] In this embodiment, the data quality adaptive enhancement method for industrial edge gateways includes steps S10 to S60: Step S10: Obtain the raw data stream of heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway, and perform protocol stripping and coarse timing alignment on the raw data stream to obtain the raw data block sequence, timing drift, equipment health baseline and industrial physical constraint map.

[0019] It should be noted that the heterogeneous sensing nodes in the industrial field of this embodiment include industrial equipment with data acquisition and communication capabilities, such as PLCs, sensors, smart meters, RFID readers, and CNC machine tools. Each node reports data to the edge gateway via Modbus, Profinet, OPC UA, MQTT, or a proprietary protocol. The raw data stream refers to the sequence of binary or character data packets generated by each sensing node without any quality processing. It includes at least the original sampled values, sampling timestamps, communication protocol types, and device topology identifiers of each sensing node. Real-time resource status measures refer to the currently available computing and communication resources of the edge gateway, including at least CPU utilization, memory availability, bandwidth availability, and edge cache queue depth. These parameters are collected in real time through the edge gateway's operating system resource monitoring interface and are key constraint inputs for subsequent hierarchical enhancement execution and dynamic transmission shaping. Protocol stripping refers to the process of removing the link layer, transport layer, and application layer protocol encapsulation headers from the raw data stream to extract the payload data. Coarse timing alignment refers to the preliminary time alignment process that unifies the data from sensing nodes in different clock domains to the local time reference of the edge gateway based on the sampling time stamp of each data frame. Its output includes the set of time-aligned data frames and the timing drift of each node relative to the unified time reference.

[0020] Understandably, industrial field sensing nodes often come from different vendors and are deployed on different network segments. Their communication protocols, sampling periods, and clock sources are independent, resulting in highly heterogeneous and fragmented raw data streams. Directly performing subsequent processing would lead to data association errors due to protocol parsing failures and timing discrepancies. Furthermore, the CPU, memory, and bandwidth resources of edge gateways fluctuate dynamically with field conditions. If enhancement strategies exceed resource capacity, it will cause gateway overload, data packet loss, and even control command failure. Therefore, protocol standardization, timing alignment, and resource status awareness must be completed at the initial stage of data processing to provide a unified data foundation and resource constraints for subsequent quality enhancement.

[0021] In one feasible implementation, step S10 may include: acquiring the raw data stream of heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway, wherein the raw data stream includes at least the raw sampled values, sampling timestamps, communication protocol types, and device topology identifiers of each sensing node, and the real-time resource status includes at least CPU utilization, memory margin, bandwidth margin, and edge cache queue depth; performing protocol stripping and frame structure standardization on the raw data stream, extracting payload data, and obtaining a standardized frame sequence; performing coarse temporal alignment on the standardized frame sequence based on the sampling timestamps, estimating the temporal drift of each sensing node relative to a unified time base, and obtaining a time-aligned data frame set; extracting historical statistical features from the time-aligned data frame set, establishing the data distribution characteristics of each sensing node under normal operating conditions, and obtaining a device health baseline; establishing an industrial physical constraint map characterizing the physical conservation relationships, parameter coupling boundaries, and state evolution laws of device operation based on the device topology identifier and a pre-set industrial equipment mechanism model library; and segmenting and aggregating the time-aligned data frame set according to a preset time window to obtain a raw data block sequence.

[0022] It should be noted that equipment topology identifiers are structured labels representing the physical connections and logical dependencies of sensing nodes in an industrial field. They typically include equipment type codes, installation location coordinates, parent node numbers, and communication interface identifiers. Equipment health baselines refer to the set of statistical distribution characteristics of the sampled data from each sensing node under normal operating conditions, including statistical quantities such as mean, variance, kurtosis, skewness, and quantiles, used for subsequent anomaly detection and reliable benchmark selection. Industrial physical constraint graphs are graph-based data structures where nodes represent industrial equipment, edges represent energy, mass, or signal transmission relationships between devices, node attributes include state variables and conservation constraints, and edge attributes include coupling weights and transmission delays, used to characterize the physical laws and topological relationships of industrial systems.

[0023] Understandably, the industrial equipment mechanism model library pre-constructs physical conservation equations for various types of industrial equipment, such as the power balance equation for motors, the energy conservation equation for heat exchangers, and the mass conservation equation for pipelines. By matching equipment topology identifiers with the mechanism model library, physical constraint maps of field equipment can be automatically generated, eliminating the need for manual configuration of each device and significantly improving the system's scalability and deployment efficiency.

[0024] In actual implementation, the edge gateway concurrently receives raw data streams from various sensing nodes through multi-protocol access ports, such as RS485, Ethernet, Wi-Fi, and 5G. When stripping the raw data streams from protocols, the protocol type field in the link layer frame header is first read. The corresponding protocol parser, such as a Modbus RTU parser or an OPC UA binary parser, is then invoked to perform layer-by-layer decapsulation and extract the payload data. The extracted payload is then standardized according to a unified frame structure, which includes at least: device ID, sampling time stamp, physical quantity measurement value, unit identifier, and quality flag.

[0025] Perform coarse timing alignment on the standardized frame sequence. Assume the local unified time base of the edge gateway is... The sampling timescale of the i-th sensing node in the k-th frame is: Then the time drift of this node relative to the unified time base Estimated using linear regression with a sliding window:

[0026] in, The sliding window length is used to select the most recent 50-100 data frames. This is the local timestamp when the edge gateway receives the k-th frame. Given the known sampling period of the i-th sensing node, this term is the sampling delay compensation amount, used to eliminate the deviation caused by random arrivals within the sampling period.

[0027] get Then, correct all sampling timescales of the i-th node to This yields a time-aligned set of data frames.

[0028] Historical statistical features are extracted from the time-aligned data frame set. For each sensing node, the most recent M time windows are selected, such as the normal operating data of the past 24 hours, to determine the sample mean of each physical quantity measurement. Sample standard deviation Sample kurtosis and sample skewness Sample mean Sample standard deviation As shown in the following formula: ;

[0029] in, Let M be the sequence of measurement values ​​of the i-th sensing node within the m-th time window, where M is the total number of historical time windows.

[0030] The above statistics will be used as the device health baseline for this sensing node. It is used for subsequent multi-node consensus voting and anomaly detection.

[0031] Construct a set of device nodes based on device topology identifiers. Each node It includes equipment type, set of state variables, and operational constraints. It retrieves the conservation equations corresponding to each equipment type from a pre-built industrial equipment mechanism model library; for example, for a motor node, it establishes the electric power... Mechanical power and heat loss power Energy conservation constraint: Based on the parent node number and communication interface information in the device topology identifier, the parameter coupling relationship between each device node is determined, and a set of directed coupling edges representing energy transfer, mass transfer, and signal transfer is established. The device topology network G=(V,E) is obtained. Conservation constraints and coupling relationships are integrated to construct an industrial physical constraint map. ,in Let be the set of node conservation constraints.

[0032] Finally, the time-aligned data frame set is divided into a preset time window T. win For example, the data frames are segmented and aggregated at 100ms or 1s. All data frames within the same window are sorted by device ID and packaged to obtain the original data block sequence, where each block contains the aligned data frames of all sensing nodes within the time window.

[0033] Optionally, to improve the accuracy and adaptability of protocol stripping, the step of performing protocol stripping and frame structure standardization on the original data stream, extracting payload data, and obtaining a standardized frame sequence includes: performing adaptive protocol identification on the original data stream based on information entropy distribution and a protocol feature fingerprint database to obtain protocol type confidence, wherein the protocol feature fingerprint database includes the message entropy range, feature code position, and payload offset of each industrial communication protocol; filtering the original data stream according to the protocol type confidence, marking the original data stream with a protocol type confidence below a preset confidence threshold as a data stream to be verified, and marking the original data stream with a protocol type confidence above or equal to the preset confidence threshold as a reliable data stream; performing frame structure standardization and payload extraction on the reliable data stream to obtain a standardized frame sequence; and attaching the protocol type confidence as a data confidence label to the frame header metadata of the standardized frame sequence to obtain a standardized frame sequence with a confidence label.

[0034] It should be noted that information entropy distribution refers to the measure of uncertainty in the value of each byte in a data packet. Different industrial protocols exhibit significant differences in packet entropy values ​​due to their fixed frame structures, address fields, and function code patterns. The protocol feature fingerprint database is a pre-built collection containing typical packet features of various industrial communication protocols. Each fingerprint record includes at least the protocol name and the packet entropy value range. Feature code position offset Feature code byte sequence and effective load starting offset .

[0035] Understandably, proprietary protocols or protocol variants often exist in industrial environments, and traditional fixed port mapping methods cannot cover all protocol types. By using dual verification of information entropy and signature codes, adaptive protocol identification can be achieved even without prior port information, significantly improving the protocol compatibility of edge gateways.

[0036] In actual execution, protocol adaptive identification is performed on the original data stream. Assume the original data packet is a byte sequence. Calculate its information entropy:

[0037] in, Let j be the frequency of byte value j in the message, and L be the total number of bytes in the message.

[0038] The calculated Entropy range in each protocol fingerprint database Perform matching to filter out a set of candidate protocols. For each protocol m in the candidate protocol set, at the position offset... Extraction length is Byte fragments, and signatures Hamming distance comparison is performed to determine the protocol type confidence level. The protocol type with the highest confidence level is selected. If the highest confidence level is greater than or equal to a preset confidence threshold (0.85), the message is marked as a trusted data stream; otherwise, it is marked as a data stream to be verified, triggering manual review or deep packet inspection. For trusted data streams, the payload is extracted based on the payload offset corresponding to the protocol type and standardized according to a unified frame structure. Simultaneously, the confidence level corresponding to the protocol type is written as a data confidence tag into the frame header metadata, resulting in a standardized frame sequence with confidence tags.

[0039] Step S20: Based on the original data block sequence, time-series drift, equipment health baseline, and industrial physical constraint map, locate the root cause of data quality defects and trace the spread of contamination to obtain a set of physical violation points, a reliable benchmark cluster, and a contamination propagation path.

[0040] It should be noted that root cause localization of data quality defects refers to identifying the underlying causes of data anomalies from the original data block sequence, such as sensor failure, clock synchronization failure, communication interference, or physical process anomalies, and locating them to specific data frames or sensing nodes. Contamination propagation tracing refers to analyzing the path, direction, and depth of impact of identified quality defects propagating along the equipment topology network to related nodes, thereby determining the scope of the defect's impact. The physical violation point set refers to the set of data frames that violate energy conservation, parameter coupling boundaries, or state evolution continuity after verification using an industrial physical constraint map. The trusted baseline cluster refers to the set of high-confidence sensing node data selected through multi-node consensus voting, whose deviation from the equipment health baseline is within a preset threshold. The contamination propagation path refers to the set of directed weighted paths, represented in graph structure or matrix form, that propagate quality defects from the anomaly source node to related nodes.

[0041] Understandably, data anomalies in industrial settings often have a cascading effect. A single sensor malfunction or clock outage not only distorts its own data but can also affect the data performance of related devices through control loop feedback or physical coupling. Simply cleaning isolated data points cannot break the chain of defect propagation and may even spread contamination to previously functioning nodes during the reconstruction process. Therefore, root cause analysis and propagation tracing must be completed before enhancement implementation to identify which data is unreliable, where the defects originate, and which nodes will be affected, providing precise input for subsequent strategy formulation in the game theory model.

[0042] Step S30: Construct a multi-objective augmented game model based on the set of physical violation points, the credible benchmark cluster, and the pollution propagation path. The multi-objective augmented game model includes participants in the cleaning strategy, participants in the reconstruction strategy, and participants in the compensation strategy. The participants achieve strategy equilibrium through a Pareto collaborative optimization mechanism based on marginal utility evaluation.

[0043] It's important to note that the multi-objective reinforcement game model is a mathematical framework that abstracts the data quality enhancement process into a non-cooperative game among multiple strategic participants. Each participant represents an enhancement strategy—cleansing, reconstructing, or compensating—and aims to achieve a Nash equilibrium or Pareto optimal strategy combination by optimizing their own utility function and considering the influence of other participants' strategies. Cleaning strategy participants are responsible for reducing data noise and outliers through filtering and elimination; reconstructing strategy participants are responsible for repairing missing and outlier segments through imputation and prediction; and compensating strategy participants are responsible for eliminating systematic biases through residual correction. The Pareto collaborative optimization mechanism refers to finding a set of non-dominated strategies in multi-objective optimization that improves the utility of at least one participant without harming the utility of other participants, and achieving resource allocation equilibrium among strategies through marginal utility evaluation.

[0044] It is understandable that edge gateways have limited resources. Allocating a large amount of CPU resources to data cleaning will lead to insufficient resources for reconstruction and compensation. Conversely, excessive reconstruction of missing data may introduce false patterns, increasing the difficulty of compensation. Traditional serial fixed strategies—cleaning first, then reconstruction, and finally compensation—do not consider resource competition and synergy between strategies, making it difficult to achieve global optimum under dynamic resource constraints. This implementation explicitly uses a game theory model to make the interaction of the three strategies explicit, using Pareto fronts to guide resource allocation, achieving a globally optimal balance between enhancement effects and resource consumption.

[0045] In one feasible implementation, step S30 may include: determining the initial strategy spaces of the participants in the cleaning strategy, reconstruction strategy, and compensation strategy based on the defect type distribution and severity level of the physical violation point set; determining the strategy influence domain of each participant based on the contamination propagation path, wherein the strategy influence domain is used to limit the action boundary of each participant when performing enhancement operations; determining the reference benchmark domain of the reconstruction strategy participant based on the trusted benchmark cluster, wherein the reference benchmark domain is used to provide neighborhood trusted reference values ​​for missing or anomalous data segments; constructing the marginal utility function of each participant, wherein the marginal utility function characterizes the amount of data quality improvement brought about by unit edge computing resource investment; solving the non-dominated strategy set through the Pareto collaborative optimization mechanism based on the initial strategy space, strategy influence domain, reference benchmark domain, and marginal utility function to obtain the Pareto optimal strategy combination; and verifying the resource feasibility of the Pareto optimal strategy combination according to the real-time resource state quantity to obtain a multi-objective enhancement game model.

[0046] It should be noted that the initial strategy space refers to the set of specific operational parameters that each participant can choose, such as the filter window length and threshold coefficient in the cleaning strategy, the interpolation method type and neighborhood radius in the reconstruction strategy, and the compensation gain and update frequency in the compensation strategy. The strategy influence domain refers to the range of nodes and data intervals affected by each participant's enhancement operations, determined based on the contamination propagation path, used to prevent strategy overreach operations from causing the misprocessing of normal data. The reference benchmark domain refers to a subset of high-confidence data selected from the trusted benchmark cluster that is closest in physical distance or logical correlation to the missing interval to be reconstructed, providing a spatiotemporal correlation reference for reconstruction. The marginal utility function is a function characterizing the amount of data quality improvement brought about by a unit of resource input; it typically decreases with increasing resource input, conforming to the law of diminishing marginal utility in economics.

[0047] Understandably, different defect types require significantly different reinforcement strategies: random noise necessitates cleaning strategies, large-scale defects require reconstruction strategies, and systematic drift requires compensation strategies. Initializing the strategy space based on defect type distribution and severity levels can avoid ineffective searches and improve game-solving efficiency.

[0048] In actual implementation, the physical violation point set is first classified by defect type and its severity is quantified. Let the physical violation point set be... For each violation point Extract its feature vector ,in Defect type The severity level is derived from the constraint violation level mapping, such as levels 1 to 5. The duration of the defect is considered. The distribution of the number of defects of each type is statistically analyzed, and the weighted sum of their severity is used to determine the initial strategy space for the three participants. For example, if noise-related defects account for a high proportion and are severe, then the scope should be expanded. The range of values ​​for the length of the filtering window.

[0049] The influence domain of each participant's strategy is determined based on the pollution propagation path R. The pollution propagation path R is represented by the propagation path weight matrix R = [r ij ] indicates that r ij Let be the impact weight of the defect propagating from node i to node j. Let the threshold for the role of participants in the cleaning strategy be... Then its strategy influence domain is This means that cleaning is performed only on associated nodes with high propagation impact weights. Similarly, [the following is determined]. and .

[0050] From the trusted benchmark cluster T, take the node v containing the missing interval to be reconstructed. target Centered on, select a physical topological distance d topo ≤D max And the deviation δ from the health baseline dev ≤θ trust The nodes constitute the reference datum domain B recon ⊆T, where D max θ is the maximum topological neighborhood radius. trust This is the confidence threshold.

[0051] Construct the marginal utility function for each participant. Let u be the resource input of each participant in the cleaning strategy. cpu Measured by the increase in CPU utilization, its marginal utility function is:

[0052] in, For investment The number of physical violations eliminated by post-cleaning The total number of physical violations. This is a scaling factor for cleaning utility, used to map the rate of elimination of physical violations to a dimensionless utility value. This represents a slight increase in resource input.

[0053] Similarly, the marginal utility function of the participants in the restructuring strategy is:

[0054] in, For memory usage increments, The length of missing data that has been reconstructed and filled. Total missing length, This is the scaling factor for the reconstructed utility.

[0055] The marginal utility function of the participants in the compensation strategy is:

[0056] in, This is the bandwidth usage increment used for transmitting compensation reference data or downloading model parameters. The energy of the systematic error that has been compensated and eliminated. The total systematic error energy, This is a scaling factor for compensating for utility.

[0057] Based on the above factors, a multi-objective optimization problem is constructed, and the Pareto optimal policy combination is solved using the NSGA-II algorithm based on fast non-dominated sorting or the MOEA / D algorithm based on decomposition. Where P is the number of nondominated solutions on the Pareto front. Based on real-time resource state variables... By performing resource feasibility verification and eliminating solutions that exceed the hard constraints of resources, the final multi-objective augmented game model is obtained.

[0058] Step S40: Perform edge-side hierarchical enhancement execution based on the multi-objective enhanced game model and the real-time resource state quantity to obtain the enhanced data stream, residual error tensor, and edge-side data value density distribution.

[0059] It should be noted that edge-side hierarchical enhancement execution refers to the process of hierarchical resource quota allocation and enhancement operation execution based on the Pareto optimal strategy combination output by the multi-objective enhancement game model, according to the priority order of cleaning, reconstruction, and compensation, combined with the real-time resource status of the edge gateway. The cleaned data stream refers to the data stream after noise suppression and outlier removal; the reconstructed data stream refers to the data stream after missing interval imputation; and the enhanced data stream refers to the final data stream after systematic bias compensation. The residual error tensor is a multidimensional array composed of the residual errors of each data frame in the enhanced data stream relative to the device health baseline or physical prediction value, used to quantify the inadequacy of the enhancement effect. The edge-side data value density distribution refers to the assessment result of the importance of each data frame to downstream control decisions or status monitoring, used to guide the priority allocation of subsequent transmission shaping.

[0060] Understandably, the resource state of edge gateways is time-varying. When CPU load suddenly increases at a certain moment, the complexity of the cleaning algorithm or the reconstruction window length must be reduced to avoid affecting the execution of real-time control tasks. Tiered enhancement execution, by directly linking resource state quantities to game model strategies, achieves dynamic adaptation between enhancement strength and resource availability, preventing edge node overload and crashes, while ensuring optimal enhancement effects within the allowable resource range.

[0061] In one feasible implementation, step S40 may include: extracting Pareto optimal strategy combinations from the multi-objective augmented game model, and parsing the cleaning decision parameters, reconstruction decision parameters, and compensation decision parameters; determining the computing resource quotas of the cleaning strategy participants based on the CPU utilization rate in the real-time resource state variables, and performing noise suppression and outlier removal on the original data block sequence based on the physical violation point set according to the cleaning decision parameters and the computing resource quotas to obtain the cleaned data stream; determining the reconstruction window length of the reconstruction strategy participants based on the memory remaining in the real-time resource state variables, and performing noise suppression and outlier removal on the original data block sequence based on the reconstruction decision parameters and the reconstruction window length. The missing intervals in the cleaned data stream are spatiotemporally interpolated based on a trusted benchmark cluster to obtain a reconstructed data stream. The compensation update frequency of the compensation strategy participants is determined based on the bandwidth margin in the real-time resource state quantity. Based on the compensation decision parameters and the compensation update frequency, residual compensation correction is performed on the systematic deviations in the reconstructed data stream using an industrial physical constraint map to obtain an enhanced data stream. Frame-by-frame residual error is calculated in the enhanced data stream to obtain a residual error tensor. The value density is assessed based on the impact of each data frame in the original data block sequence on downstream control decisions to obtain the edge-side data value density distribution.

[0062] It should be noted that the cleaning decision parameters include filter type, filter order, threshold coefficient, and effective range; the reconstruction decision parameters include interpolation method, neighborhood radius, spatiotemporal weight coefficient, and confidence threshold; and the compensation decision parameters include compensator gain, integral time constant, compensation direction, and update frequency. Computational resource quotas refer to the allocation of resources based on the current CPU utilization rate. With maximum allowable occupancy rate The difference ΔR cpu The maximum available CPU cycles are allocated proportionally to the cleaning strategy. The reconstruction window length refers to the maximum time series length allowed to be processed in a single reconstruction operation, which is affected by the available memory M. avail Constraints: A longer window allows for more historical data to be utilized, resulting in higher reconstruction accuracy, but also higher memory consumption. The compensation update frequency refers to the time interval between residual corrections executed by the compensation strategy, which is affected by the bandwidth margin B. avai The higher the update frequency, the faster the tracking of time-varying deviations, but the more bandwidth resources are consumed.

[0063] Understandably, the resource requirements of the three strategies differ significantly: cleaning operations are typically computationally point-by-point and CPU-intensive; reconstruction operations require maintaining historical data buffers and are memory-intensive; compensation operations may require downloading cloud calibration parameters or uploading residual statistics and are bandwidth-intensive. By precisely matching resource quotas with resource types, the enhancement benefits per unit of resource can be maximized.

[0064] In actual implementation, the Pareto optimal strategy combination is used. Select the solution that best matches the current resource state. If the Pareto solution with the smallest Euclidean distance from the current resource reserve is selected, the cleaning decision parameters, reconstruction decision parameters, and compensation decision parameters can be obtained analytically.

[0065] Determine the cleaning strategy and calculate resource quotas. Assume the total CPU processing capacity of the edge gateway is C. total The current utilization rate, measured in MIPS or cores, is R. cpu The available cleaning resource quota is:

[0066] in, This refers to the available cleaning resource quota, which is the maximum CPU processing capacity quota that the cleaning strategy can obtain. The maximum CPU utilization allowed for the edge gateway is set to 80%. The resource allocation coefficients for the cleaning strategy are output by the game theory model and satisfy... .

[0067] according to Choose the complexity of the cleaning algorithm. For example, when the quota is sufficient, use multi-scale decomposition denoising based on wavelet transform, and use moving average or median filtering when the quota is limited.

[0068] Cleaning is performed on the original data block sequence D. For noise and outlier defects marked in the physical violation point set, the cleaning is performed based on the filter strength coefficients. Determine the filter window length ,in The base window length is used for the data sequence within the window. Perform filtering. After filtering, data points that deviate from the device health baseline by more than the outlier detection threshold are removed and replaced with the filtered predicted values, resulting in the cleaned data stream D. clean .

[0069] Determine the reconstruction window length. Assume the memory usage of each historical data frame is m. frame The remaining memory is M avail The maximum reconstruction window length is... for:

[0070] in, To reference the number of neighboring nodes in the baseline domain, This is the resource allocation coefficient for the restructuring strategy.

[0071] Actual reconstruction window length ,in This is the minimum effective window length recommended by the game theory model based on the length of the missing interval. If... , If the minimum effective length is 5 sampling points, then the missing interval is marked as an unreconstructable interval and left for cloud processing to avoid false reconstructions caused by insufficient data on the edge side.

[0072] Data stream D after cleaning clean Spatiotemporal collaborative interpolation is performed on the missing intervals in the data. Let the missing interval be... , length is Select the topological distance d from the trusted baseline cluster T to the missing node. topo ≤R nb The reference node is extracted in the reconstruction window. The data sequence within is used as a reference sample. Based on the spatiotemporal weighting coefficient β... ST The weights of interpolation in the time dimension and the spatial dimension are balanced when β ST →1 emphasizes temporal continuity and is suitable for slowly varying processes; when β ST →0 emphasizes spatial consistency and is suitable for field distributions with small spatial gradients. A confidence test is performed on the reconstruction results; if the confidence level is lower than γ... conf If the reconstruction is not successful, it is marked as a low-confidence reconstruction and given a larger weight in the residual error tensor to obtain the reconstructed data stream D. recon .

[0073] Determine the compensation update frequency. Let D be the amount of reference data to be transmitted in a single compensation operation. ref The bandwidth margin is B. avail The maximum allowed update frequency is:

[0074] in, This is the resource allocation coefficient for the compensation strategy.

[0075] Actual update frequency ,in This is the minimum tracking frequency determined based on the time-varying characteristics of the systematic deviation, such as 2-5 times the Nyquist sampling frequency of the deviation.

[0076] For the reconstructed data stream D reconPerform residual compensation correction. Based on the conservation constraint equations in the industrial physical constraint map, calculate the physical residual of the current data frame. The compensation amount is generated based on the compensation decision parameters, and then superimposed onto the reconstructed data stream according to the compensation direction to obtain the enhanced data stream D. enhanced .

[0077] Frame-by-frame residual error calculation is performed on the enhanced data stream. Let the physical prediction value of the i-th node in the k-th frame after enhancement be... The residual error is predicted from the industrial physical constraint map based on the state and state evolution of neighboring nodes. Therefore, the residual error is the difference between the predicted physical value and the actual measured value after enhancement. The residual errors of all nodes are arranged according to frame time sequence and node number to obtain the residual error tensor.

[0078] Value density assessment is performed based on the impact of each data frame in the original data block sequence on downstream control decisions. Assuming downstream control decisions involve J key controlled variables, such as reactor temperature, pipeline pressure, and motor speed, a value density measure based on sensitivity analysis is used.

[0079] in, Let k be the data value density of the k-th frame. For the j-th key controlled variable, Let k be the enhanced data vector of the k-th frame. The sensitivity coefficient of the controlled variable to the data frame is determined by the parameter coupling relationship in the industrial physical constraint map or the mathematical model of the control loop. The control weight coefficient for the j-th controlled variable is determined by the importance level of this control loop to production safety or product quality.

[0080] The value density of each frame is normalized to obtain the value density distribution of the edge data.

[0081] In this implementation, the three strategies of cleaning, reconstruction, and compensation are dynamically adapted to the real-time resource status of the edge gateway through hierarchical resource quota allocation, ensuring the continuity and stability of the enhancement process when CPU, memory, and bandwidth resources fluctuate. Through targeted cleaning based on physical violation point sets, spatiotemporal collaborative interpolation based on trusted benchmark clusters, and residual compensation based on industrial physical constraint maps, layered and progressive governance of data quality defects is achieved. Through frame-by-frame quantization of residual error tensors and data value density assessment, refined quantitative inputs are provided for subsequent physical consistency closed-loop verification and dynamic transmission shaping, significantly improving the integrity, controllability, and resource utilization efficiency of edge-side data quality enhancement.

[0082] Step S50: Perform physical consistency closed-loop verification based on the trusted benchmark cluster and the residual error tensor to obtain the quality gain index and strategy regret value.

[0083] It should be noted that physical consistency closed-loop verification refers to the process of quantitatively evaluating the physical rationality and quality improvement of the enhanced data stream after the data quality enhancement is completed, using a trusted benchmark cluster as a high-confidence reference. The quality gain index is a comprehensive indicator characterizing the degree of improvement in data quality of the enhanced data stream relative to the original data stream. Its value range is typically a percentage ratio of 0-1 or greater than 1; a larger value indicates a more significant enhancement effect. The strategy regret value is an indicator used in game theory to measure the utility gap between the actual implemented strategy and the theoretically optimal strategy. In this implementation, it specifically refers to the normalized deviation between the expected marginal utility and the actual marginal utility of each participant in the enhancement strategy, used to guide the adaptive adjustment of the strategy in the next round of the game.

[0084] Understandably, without a closed-loop verification mechanism, the edge gateway cannot know whether the current enhancement operation has truly improved data quality, nor can it determine whether the resource consumption is worthwhile. In dynamically changing industrial environments, fixed enhancement strategies may become ineffective as operating conditions shift. It is necessary to continuously evaluate the effectiveness of the strategy through closed-loop feedback and trigger strategy re-optimization based on the strategy regret value, forming an adaptive closed loop of execution-verification-adjustment.

[0085] In one feasible implementation, step S50 may include: using the trusted benchmark cluster as a benchmark reference set for physical consistency verification, determining the physical consistency fit of the enhanced data stream relative to the benchmark reference set; determining the ratio of the data quality metric before and after enhancement based on the deviation of the residual error tensor relative to the industrial physical constraint map, combined with the physical consistency fit, to obtain the quality gain index; and determining the strategy utility loss of each participant based on the deviation between the expected marginal utility and the actual marginal utility of each participant in the Pareto optimal strategy combination, to obtain the strategy regret value.

[0086] It should be noted that physical consistency conformity refers to the proportion of data frames in the enhanced data stream that satisfy the conservation relationships of the industrial physical constraint map, or the average degree of conformity between the data frame state variables and the physical prediction values. The deviation of the residual error tensor relative to the industrial physical constraint map refers to the magnitude of the projection of the residual error onto the constraint boundary, used to quantify the degree of physical inconsistency that still exists in the enhanced data.

[0087] Understandably, the data in the trusted benchmark cluster comes from high-trust nodes selected by multi-node consensus voting. These nodes do not directly participate in the enhancement operation but are only used as a reconstruction reference. Therefore, they can be used as an independent third-party benchmark to verify the enhancement effect and avoid verification bias.

[0088] In actual implementation, the trusted baseline cluster T is used as the baseline reference set. Let the enhanced data stream be = For each frame The physical consistency with the baseline reference set is determined. First, based on the conservation constraints in the industrial physical constraint map, the frame-level physical consistency is determined:

[0089] in, Let C represent the physical consistency fit of the k-th frame, and C be the total number of conservation constraints in the industrial physical constraint map. Let σ be the residual function value of the k-th frame under the c-th conservation constraint. c The permissible standard deviation for the c-th constraint is determined by the historical residual statistics in the equipment health baseline. ∈(0,1], the closer the value is to 1, the more the frame conforms to physical laws.

[0090] Determine the global physical consistency. :

[0091] Where Φ represents the global physical consistency degree, and K represents the total number of data stream frames.

[0092] The deviation from the industrial physical constraint map is determined based on the residual error tensor. Since the residual error already includes the physical prediction bias, its root mean square value is directly calculated as a measure of deviation.

[0093] Determine the original data quality metrics before enhancement. Similarly, calculate the physical consistency ratio Φ for the original data block sequence D. (raw) Deviation from constraints This allows for the determination of the mass metric before and after enhancement. The mass gain index G Q To enhance the ratio of the quality measure before and after, when G Q When G > 1, it indicates that the enhancement operation effectively improved the data quality; when G Q When the value is ≤1, it indicates that the enhancement strategy has failed or that over-processing has led to quality degradation, and the strategy needs to be re-solved.

[0094] Calculate the strategy regret value. Let the expected marginal utility of the participant in the cleaning strategy within the Pareto optimal strategy portfolio be... This value is predicted by the game theory model based on the historical resource-quality elasticity coefficient and the current number of physical violations. After actual execution, the actual marginal utility is determined based on the actual number of physical violations removed and the actual increase in CPU resources consumed. The instantaneous regret value of the cleaning strategy is:

[0095] in, To prevent small positive numbers from being divided by zero.

[0096] Similarly, determine the instantaneous regret value of the reconstruction strategy. Instantaneous regret value with compensation strategy The historical instantaneous regret values ​​of each participant are exponentially weighted and accumulated to obtain the cumulative regret value. When the cumulative regret value exceeds a preset adaptive threshold θ... regret When this occurs, it indicates that the strategy's long-term performance is below expectations, and it is necessary to shrink its strategy space boundary or adjust the weights of the resource-quality elasticity coefficient in the next round of the game. θ regret Take 0.2-0.3.

[0097] By using a trusted benchmark cluster as an independent benchmark reference set for physical consistency assessment, the objectivity and credibility of the verification results are ensured. The quality gain index enables a quantitative comparison of data quality before and after enhancement, providing a clear criterion for judging the enhancement effect. The policy regret value quantifies the utility deviation between the game model prediction and the actual execution, enabling the edge gateway to detect policy failures in a timely manner and trigger adaptive correction, significantly improving the long-term stability and operational adaptability of the data quality enhancement system.

[0098] Step S60: Dynamically shape the uplink data stream of the edge gateway according to the quality gain index, policy regret value, edge-side data value density distribution and real-time resource status quantity to achieve adaptive enhancement of data quality at the edge of the industrial IoT.

[0099] It should be noted that dynamic transmission shaping refers to a comprehensive process of prioritizing, redundancy coding, integrity verification, and adjusting the transmission rate of the enhanced data stream based on data quality status, resource status, and network conditions. Its goal is to maximize the value density and reliability of transmitted data within limited bandwidth resources. Uplink data stream, or outbound data stream, refers to the data stream transmitted from the edge gateway to the cloud data platform or upper-layer application system. The target data quality adaptive enhancement result refers to the high-quality data stream and its accompanying quality metadata ultimately output to the cloud after transmission shaping.

[0100] Understandably, the uplink bandwidth of industrial edge gateways is often limited, such as the bandwidth of 4G / 5G IoT cards and the bandwidth quota of the factory's intranet. If all enhanced data is uploaded indiscriminately, it will not only waste bandwidth but may also lead to the loss of critical data due to congestion. By mapping data value density, quality gain index, and policy regret value to transmission attributes, fine-grained transmission management is achieved, ensuring that high-value, high-quality, and high-confidence data arrives at the cloud reliably and with priority.

[0101] In one feasible implementation, step S60 may include: obtaining the current bandwidth utilization and network link quality index of each uplink channel of the edge gateway; determining the transmission priority of each data frame based on the edge-side data value density distribution, wherein data frames with high value density levels are assigned high transmission priority; determining the transmission reliability level of each data frame based on the quality gain index, wherein data frames with a quality gain index lower than a preset threshold trigger increased transmission redundancy; determining the verification strength level of each data frame based on the policy regret value, wherein data frames with a policy regret value higher than a preset threshold trigger enhanced integrity verification; and determining the transmission priority and transmission reliability based on the data frame's transmission reliability. The enhanced data stream is dynamically shaped for transmission based on the level, verification strength level, and bandwidth margin in the real-time resource status quantity, resulting in a shaped transmission sequence. The transmission rate of the shaped transmission sequence is adaptively adjusted according to the real-time resource status quantity, driving each uplink channel to send data and monitoring the transmission quality response in real time. When the deviation between the transmission quality response and the expected transmission quality exceeds a preset tolerance threshold, the strategy re-solution of the multi-objective enhanced game model is triggered, dynamically updating the Pareto optimal strategy combination until the transmission quality of the enhanced data stream reaches the target set range, thereby achieving adaptive enhancement of data quality at the edge of the industrial IoT.

[0102] It should be noted that the network link quality index is a comprehensive indicator characterizing the current uplink transmission quality, typically including parameters such as packet loss rate, latency, jitter, and signal-to-noise ratio, and is obtained in real-time by the network interface layer of the edge gateway. Transmission priority refers to the order in which data frames enter the transmission queue; frames with higher value density are sent first. Transmission reliability level refers to the redundancy coding strength determined based on the quality status of the data frames; frames with low quality gain index require higher redundancy to ensure cloud recovery. Verification strength level refers to the integrity verification algorithm complexity determined based on the policy regret value; frames with high regret values ​​require stronger verification to detect potential enhancement errors. Adaptive transmission rate adjustment refers to dynamically adjusting the data transmission rate based on current bandwidth margin and network link quality to avoid network congestion or bandwidth waste. Transmission redundancy refers to the amount of redundant data added through forward error correction coding (FEC) or retransmission, used to recover the original data in packet-loss links. Enhanced integrity verification refers to verification algorithms with stronger error detection capabilities compared to standard cyclic redundancy check (CRC), such as CRC-32C, Adler-32, or lightweight hash check.

[0103] Understandably, industrial control data has strong real-time requirements, but not all data has the same urgency for control. For example, periodic inspection data from equipment in standby mode has low value density and can tolerate a certain delay; while fault alarm or safety interlock trigger data has extremely high value density and must be transmitted immediately with high reliability. Through dynamic transmission shaping, the edge gateway can automatically yield high-priority data when bandwidth is bottlenecked and automatically "harden" transmission guarantees when quality is low, achieving dynamic matching of network resources and data value.

[0104] In actual execution, the uplink channels of each edge gateway are acquired. Current bandwidth utilization With network link quality index The overall network link quality index is calculated as follows:

[0105] in, The packet loss rate of channel ch. For average round-trip time, For the maximum allowable delay, For average jitter, To the maximum permissible jitter, , , Let be the weighting coefficient, satisfying .

[0106] Based on edge-side data value density distribution Determine transmission priority. Normalize the value density of each frame to the [0,1] interval, and divide them into three priority levels: high, medium, and low, according to a preset quantile. The priority division threshold is set at the 75th quantile and the 25th quantile.

[0107] Based on the mass gain index G Q Determine the transmission reliability level. If G Q <G th G th A preset threshold, such as 1.2, indicates that the data frame enhancement effect is poor, with a large residual error, requiring increased transmission redundancy. Reliability level: Rel k With redundancy of redundancy coding γ red The mapping relationship is as follows:

[0108] in, This represents the target redundancy of the k-th frame. A larger value indicates higher redundancy and stronger packet loss recovery capability. Based on the basic redundancy, such as 10%, For maximum additional redundancy, such as 50%, when G Q ≥Gth Time to take = Reliability levels can also be divided into three levels: high, medium, and low, based on the degree of redundancy. The higher the redundancy, the higher the reliability level.

[0109] The verification strength level is determined based on the policy regret value. The average cumulative regret value of the nodes involved in each frame is determined. If the average cumulative regret value is greater than θ... regret When the condition is met, an enhanced integrity check is triggered, using CRC-32C or a 32-bit lightweight hash; otherwise, standard CRC-16 is used.

[0110] Based on transmission priority, reliability level, parity strength level, and bandwidth margin B avail Perform dynamic transmission shaping. Construct a multi-objective scheduling optimization problem and solve it using a greedy algorithm based on priority queues or a dynamic programming algorithm to obtain the shaped transmission sequence.

[0111] The transmission rate is adaptively adjusted based on real-time resource status. Let the current bandwidth margin be B. avail The network link quality is Q. link Then the adaptive transmission rate is:

[0112] in, This represents the maximum transmission rate of the physical interface. ∈(0,1] is the safety margin coefficient, which is set to 0.8 to prevent congestion caused by rate estimation errors.

[0113] The system drives each uplink channel to send data according to the shaped transmission sequence and monitors the transmission quality response in real time, including cloud acknowledgment rate, end-to-end latency, and retransmission rate. Let the expected transmission quality be Q. exp The actual transmission quality is Q. act .like , The preset tolerance threshold indicates that the current enhancement and transmission strategies are not adapted to changes in network state. This triggers a strategy re-solution in the multi-objective enhancement game model, dynamically updates the Pareto optimal strategy combination, and re-executes hierarchical enhancement and transmission shaping until the transmission quality reaches the target set range.

[0114] In one feasible implementation, to improve transmission reliability and bandwidth utilization efficiency, the enhanced data stream is dynamically shaped based on the transmission priority, transmission reliability level, parity strength level, and bandwidth margin in the real-time resource status quantity to obtain a shaped transmission sequence. This includes: fountain code encoding of data frames in the enhanced data stream based on the transmission reliability level to obtain a fountain code encoding sequence, wherein high-reliability-level data frames use high-redundancy fountain code encoding, and low-reliability-level data frames use low-redundancy fountain code encoding; differential parity coding of data frames in the fountain code encoding sequence based on the parity strength level to obtain a differential redundancy coding sequence, wherein high-parity-level data frames use enhanced cyclic redundancy check (CRC), and low-parity data frames use standard cyclic redundancy check (CCRC); and transmission queue scheduling of the differential redundancy coding sequence based on the transmission priority and bandwidth margin in the real-time resource status quantity to obtain the shaped transmission sequence.

[0115] It should be noted that fountain codes are a type of rateless erasure coding, whose encoded symbols can be generated infinitely. The receiver only needs to receive slightly more encoded symbols than the original data to successfully decode the data, making them particularly suitable for industrial wireless links with uncertain packet loss rates. Differential check coding refers to using CRC checksums with different bit widths or polynomials based on the importance and reliability of the data frame, achieving a flexible trade-off between checksum strength and computational overhead.

[0116] Understandably, traditional fixed-redundancy coding, such as repeating it three times, wastes bandwidth when the link quality is good and lacks recovery capability when the link quality is poor. Fountain codes, with their rateless nature, allow the sender to continuously transmit coded symbols until it receives confirmation from the receiver, achieving redundancy on demand. By dynamically adjusting the coding density of the fountain code (i.e., the number of coded symbols generated for each original data packet) in conjunction with the reliability level, redundancy overhead can be minimized while ensuring the probability of recovery.

[0117] In actual implementation, for data frames with high reliability levels, systematic fountain codes such as RaptorQ or LT codes are used for encoding, with a high coding density ρ set. fountain =2.0~3.0, meaning generating 2~3 times the amount of encoded symbols as the original data, ensuring recovery even under severe link conditions with a 50% packet loss rate. For data frames with low reliability levels, low-density fountain code encoding is used, ρ fountain =1.2~1.5, saving bandwidth.

[0118] Differential check coding selects the CRC polynomial based on the check strength level. Standard cyclic redundancy check uses CRC-16-CCITT, suitable for low check strength levels; enhanced cyclic redundancy check uses CRC-32C or CRC-32K, which have a larger Hamming distance and can detect more burst errors, suitable for high check strength levels.

[0119] Transmission queue scheduling is performed based on transmission priority and bandwidth availability. A multi-level priority queue structure is adopted, with high-priority frames entering queue Q. High Medium priority to enter Q Med Low priority enters Q Low The scheduler uses a weighted round-robin algorithm, with weights dynamically adjusted based on bandwidth availability.

[0120] in, > > For priority weight coefficients, satisfying + + , , , These are the scheduling weights for high, medium, and low priority queues, respectively. The larger the value, the higher the proportion of transmission bandwidth allocated to them.

[0121] When bandwidth is sufficient =0.5, =0.3, =0.2; when bandwidth is tight, increase Reduced to 0.7 The priority is reduced to 0.1, and even low-priority queue scheduling is suspended. Through the combination of fountain code encoding, differential verification, and priority queue scheduling, the shaped transmission sequence is obtained.

[0122] This embodiment provides an adaptive data quality enhancement method for industrial edge gateways. By introducing an industrial physical constraint map, the physical conservation relationships, parameter coupling boundaries, and state evolution laws of equipment operation are embedded into the data quality analysis process, enabling root cause localization of data quality defects based on physical consistency. By constructing a multi-objective enhancement game model, cleaning, reconstruction, and compensation strategies are modeled as game participants, and a Pareto collaborative optimization mechanism based on marginal utility evaluation is used to achieve resource balance and conflict resolution among strategies. Through physical consistency closed-loop verification and strategy regret value calculation, adaptive correction of enhancement strategies is achieved. By driving dynamic transmission shaping through the data value density distribution on the edge side, differentiated and precise allocation of edge computing resources and network bandwidth is achieved, significantly improving the accuracy, resource utilization efficiency, and transmission reliability of data quality enhancement on the industrial edge side.

[0123] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, step S20 includes steps S201 to S204: Step S201: Perform physical consistency verification on the original data block sequence based on the industrial physical constraint map to obtain a set of physical violation points. The physical consistency verification includes at least energy conservation verification, parameter coupling boundary verification, and state evolution continuity verification.

[0124] It should be noted that physical consistency verification refers to the process of using the pre-set equipment mechanism conservation equations, parameter coupling boundaries, and state evolution laws in the industrial physical constraint map to perform frame-by-frame compliance checks on each data frame in the original data block sequence, identifying and marking data points that violate physical laws. The physical violation point set refers to the set of all data frames judged to violate physical consistency, and each violation point contains at least the following attribute information: node number, timestamp, type of violation constraint, and degree of violation.

[0125] It is understandable that data anomalies in industrial settings are not entirely random, but often exhibit specific patterns inherent to the physical processes. For example, motor shaft power output cannot be negative, the outlet fluid temperature of a heat exchanger cannot consistently exceed the inlet temperature without external heating, and the algebraic sum of the mass flow rates of each branch at a pipeline junction must satisfy a conservation relationship. By using the explicitly coded physical laws in the industrial physical constraint map as verification benchmarks, data points that violate physical impossibilities can be quickly filtered from massive amounts of heterogeneous data, significantly improving the accuracy of anomaly detection and the interpretability of results, and avoiding false alarms and missed alarms caused by purely statistical methods neglecting physical mechanisms.

[0126] In actual implementation, based on the industrial physical constraint map The set of node conservation constraints in For each sensing node v i Establish a physical constraint residual calculation model corresponding to its device type for ∈V. Let the state measurement value of the i-th sensing node in the k-th time window of the original data block sequence be ∈V. For parameters such as temperature, pressure, and flow rate, the set of state measurements of their neighboring nodes is: .

[0127] For the energy conservation constraint Determine the energy balance residual. For the mass conservation constraint... Determine the quality balance residual. For parameter-coupled boundary constraints... Determine the boundary deviation residuals. Based on the above constraint residuals, determine the comprehensive physical violation index for the i-th node in the k-th frame. Set a physical violation threshold θ. vio Take a value of 0.05 to 0.15, which means the overall residual exceeds the nominal value by 5% to 15%. If >θ vio Then, the data frame (i,k) is marked as a physical violation point, and its violation type and degree are recorded. After traversing all nodes and all time windows, a set of physical violation points P is generated.

[0128] Step S202: Based on the timing drift, perform cross-node clock synchronization deviation analysis to identify sampling misalignment caused by clock out-of-sync, and obtain a set of timing misalignment nodes.

[0129] It should be noted that cross-node clock synchronization deviation analysis refers to the process of detecting and identifying sensing nodes with significant clock synchronization problems by using the timing drift of each sensing node relative to a unified time base, through statistical trend testing and stability analysis. The timing misalignment node set refers to the set of sensing nodes determined to have unacceptable clock synchronization deviations, whose timing drift exhibits characteristics of exceeding thresholds, non-stationarity, or abnormal jumps.

[0130] Understandably, the clock source quality of sensing nodes in industrial fields varies significantly. High-end PLCs are typically equipped with temperature-compensated crystal oscillators, while low-end sensors may only use ordinary ceramic crystal oscillators. Under long-term operation or drastic changes in ambient temperature, crystal oscillator frequency drift can lead to cumulative deviations in sampling timescales. Furthermore, in wireless transmission scenarios, media access control layer contention backoff, routing jumps, and network congestion can introduce random delay jitter, further exacerbating timing misalignments. If these timing misalignments are not identified and compensated for in a timely manner, false phase differences and pseudo-correlations will occur during multi-source data fusion, leading to misjudgments in physical consistency checks and severely impacting the accuracy of subsequent root cause localization.

[0131] In actual execution, for each sensing node i within the sliding time window W sync The time series drift sequence obtained by internal estimation Where n is the time window number, Multidimensional stability analysis was performed on the number of samples within the window.

[0132] The sample mean, sample standard deviation, and maximum absolute drift of the time series drift are determined, and the sample variance of the first-order difference sequence is determined to characterize the severity of the drift fluctuation.

[0133] Define a clock synchronization deviation criterion: if sensing node i satisfies any of the following conditions, it is included in the time-misaligned node set T. syncCondition 1 is that the cumulative drift exceeds the limit, that is, the sample mean of the time series drift is greater than the average drift threshold, and the average drift threshold is taken as 5%~10% of the sampling period; Condition 2 is that the random jitter is severe, that is, the sample standard deviation is greater than the standard deviation threshold, and the standard deviation threshold is taken as 2%~5% of the sampling period; Condition 3 is that the maximum instantaneous jump exceeds the limit, that is, the maximum absolute drift is greater than the jump threshold, and the jump threshold is taken as 20% of the sampling period; Condition 4 is that the difference variance is abnormal, that is, the sample variance is greater than the variance threshold, indicating that there are unpredictable random jumps or trend reversals in the drift.

[0134] For the identified timing misalignment nodes, their misalignment types are further classified: cumulative drift type, which satisfies condition 1 but not condition 3, and is characterized by monotonically increasing or decreasing timescale deviation, usually caused by crystal oscillator frequency deviation; random jitter type, which satisfies condition 2 but not condition 1, and is characterized by violent fluctuations around zero mean, usually caused by network transmission jitter; jump type, which satisfies condition 3, and is characterized by sudden timescale jumps, usually caused by clock source switching, route reconstruction or node restart; and composite type, which satisfies multiple conditions at the same time, and is characterized by complex timing misalignment characteristics.

[0135] The misalignment type and severity of each temporal misalignment node are recorded to provide temporal anomaly source information for subsequent pollution propagation tracing. The final result is a set T of temporal misalignment nodes. sync .

[0136] Step S203: Based on the device health baseline, perform multi-node consistency voting on the data of each sensing node in the original data block sequence, and include the sensing node data that deviates from the device health baseline by less than a preset threshold into the trusted benchmark cluster.

[0137] It should be noted that multi-node consensus voting is a method that utilizes the data correlation between similar or physically related sensing nodes deployed redundantly in an industrial site, and uses a statistical voting mechanism to select high-reliability sensing nodes. Its theoretical basis lies in the fact that in an industrial site, a single sensor may generate abnormal data due to localized faults, electromagnetic interference, or aging. However, the probability of multiple physically related or spatially adjacent sensors simultaneously experiencing the same-direction fault is extremely low. Therefore, the consistent state of the majority of nodes is more likely to reflect the true physical process state. A reliable benchmark cluster refers to the set of sensing nodes selected through consensus voting whose measurement data distribution highly matches the equipment health baseline and receives support from a majority of nodes in the topological neighborhood. This cluster provides a reliable reference for subsequent data reconstruction and physical consistency verification.

[0138] Understandably, key measurement points in industrial settings often exhibit hardware redundancy, such as dual thermocouples, redundant transmitters, or physical redundancy, like temperature, pressure, and flow measurement points upstream and downstream of the same thermal process. Even without explicit hardware redundancy, the energy, mass, and signal coupling relationships between devices revealed by the industrial physical constraint map provide logical redundancy. Through multi-node consensus voting, a credible majority and a suspect minority can be automatically distinguished from numerous sensing nodes without prior fault diagnosis information, avoiding the contamination of global data quality by a single node failure and providing a high-confidence benchmark data source for subsequent reconstruction operations.

[0139] In actual execution, suppose the original data block sequence contains N sensing nodes, and within the k-th time window, the measurement value of node i is... It can be a univariate or multivariate vector. Based on the device health baseline. Determine the normalized deviation of node i in the k-th frame:

[0140] in, In the univariate case, the Mahalanobis distance between the measurement value of node i in the k-th frame and its historical normal operating condition mean is approximately equal to that value. This represents the sample mean in the equipment health baseline. Sample standard deviation Set a preset threshold θ trust Taking values ​​of 2.0 to 3.0 corresponds to a confidence interval of 95.4% to 99.7%. If... <θ trust Then node i casts a credible vote v in the kth frame. i,k =1; otherwise, cast a suspicious vote. i,k =0.

[0141] To further utilize spatial correlation information, the set of neighboring nodes of node i is determined based on the equipment topology correlation network in the industrial physical constraint map. ,in Let i be a directed coupled edge in the device topology graph. Determine the neighborhood consistency index of node i:

[0142] in, The coupling weight is the coupling edge between nodes i and j in the industrial physical constraint map, representing the degree of physical association between the two nodes. It is determined by energy transfer efficiency, mass flow rate ratio, or signal gain. This represents the weighted percentage of trusted votes cast within the neighborhood, with values ​​ranging from [0,1]. A value closer to 1 indicates a higher percentage of trusted nodes in the neighborhood of node i. Let represent the voting result of node j in frame k.

[0143] Set a consistency threshold θ consist The threshold is set to 0.5~0.7, which follows the majority rule. The criterion for determining whether node i is included in the trusted baseline cluster T is:

[0144] This means that the node's own deviation is below a threshold, and more than half of its associated nodes in its physical neighborhood are also in a trustworthy state. Traverse all nodes and all time windows to generate a trustworthy baseline cluster T.

[0145] For nodes not included in the trusted baseline cluster, the reasons for their exclusion are recorded, such as excessive deviation or insufficient neighborhood consistency, to provide auxiliary information for subsequent root cause localization. Data frames in the trusted baseline cluster will be marked with high-confidence tags, serving as reference benchmarks for spatiotemporal co-interpolation during subsequent reconstruction and as a reference set for physical consistency during closed-loop verification.

[0146] Step S204: Based on the device topology identifier and the set of time-series misaligned nodes, perform spatial correlation analysis of data anomalies to determine the propagation direction and impact depth of quality defects from the abnormal nodes along the device topology network to the associated nodes, and obtain the pollution propagation path.

[0147] It should be noted that spatial correlation analysis of data anomalies refers to the process by which identified anomaly sources, including anomalous nodes in the physical violation point set and anomalous source nodes in the temporal misalignment node set, are analyzed to propagate to topologically related nodes through energy transfer, mass flow, signal transmission, or data-dependent paths, utilizing the geometric connection structure and physical coupling relationships of the equipment topology network. A contamination propagation path is a quality defect propagation topology represented by a directed weighted graph or matrix. Its elements include the influence weight of the defect propagating from the source node to the target node, the propagation delay, and the attenuation coefficient. This is used to quantify the scope, direction, and depth of influence of the quality defect, providing precise boundaries for defining the strategy influence domain in subsequent multi-objective augmented game models.

[0148] Understandably, data anomalies in industrial systems often exhibit significant propagation and cascading effects. For example, a faulty upstream temperature sensor can cause erroneous heating control commands to be transmitted to downstream actuators, leading to abnormal downstream temperature data; a clock out-of-synchronization at a critical node can cause false timing deviations in associated nodes involved in data fusion, differential calculation, or closed-loop control; and flow measurement deviations upstream in a pipeline can propagate along mass conservation constraints to all downstream nodes. Spatial correlation analysis can connect isolated anomalies into propagation chains with clear physical meaning, answering key questions such as where the anomaly originated, its path, its destination, and the extent of its impact. This prevents subsequent enhancement operations from merely addressing the surface symptoms while ignoring the root cause.

[0149] In actual implementation, a device topology association network is constructed based on device topology identifiers. , where the set of nodes For a sensing node, there is a set of directed edges. It represents the physical connection or data dependency between nodes, such as material flow direction, energy transfer direction, control signal transmission path, etc.

[0150] The time-displaced node set T sync The nodes in the set are labeled as temporal anomaly source nodes, and the nodes in the physical violation point set P are labeled as physical anomaly manifestation nodes. Anomaly source nodes are the starting points of quality defects, and anomaly manifestation nodes are the points where the effects of defects become apparent. Together, they constitute the input for anomaly propagation analysis.

[0151] For each directed edge e ij ∈ Based on the coupling weight w in the industrial physical constraint map ij With transmission delay τ ij Determine the basic propagation impact coefficient:

[0152] in, This represents the basic impact percentage of a defect propagating directly from node i to node j. Let i be the set of outgoing neighbor nodes of node i.

[0153] Considering the exponential decay effect of propagation delay, a delay decay factor β is defined. ij For ∈(0,1], the longer the time delay, the more significant the attenuation.

[0154] By combining coupling weights and time delay attenuation, the direct propagation influence weights are determined.

[0155] For multi-hop propagation, i.e., indirect propagation, a propagation accumulation model based on topology paths is adopted. Let the propagation originate from the anomaly source node v. s to target node v t There exists a directed path of length L. The propagation impact weight of this path is:

[0156] in, ∈(0,1] is the multi-hop attenuation coefficient, which characterizes the proportion of energy, quality, or signal loss after each hop of propagation, and is taken as 0.5~0.9. For path p from source node v s to target node v t The total propagation impact weight, The total influence weight for direct propagation between adjacent nodes.

[0157] If there are multiple lines from v s to v t For a directed path, the total propagation influence weight is the sum of the weights of each path. To control computational complexity, the maximum path length L is usually limited. max ≤3~5, meaning only local propagation effects within 3 to 5 hops are considered.

[0158] The total propagation impact weights between all source nodes and target nodes are arranged into a propagation path weight matrix. Further, the comprehensive pollution exposure index of each node is determined based on the propagation path weight matrix. The propagation path weight matrix and the pollution exposure index of each node are combined as the pollution propagation path output, providing input for the subsequent delineation of the strategy influence domain and the constraint of the enhanced operation boundary in the multi-objective enhanced game model.

[0159] In this embodiment, high-precision identification and physical interpretability of data anomalies are achieved through physical consistency verification based on industrial physical constraint maps. Cross-node clock synchronization deviation analysis reveals the root causes of anomalies in the time series dimension, distinguishing various out-of-synchronization modes such as cumulative drift, random jitter, and sudden jumps. A highly reliable benchmark dataset is established through multi-node consensus voting, providing a reliable reference benchmark for subsequent reconstruction and verification. Spatial correlation analysis based on equipment topology association networks and multi-hop propagation accumulation models enables quantitative tracing of pollution propagation paths, clarifying the propagation direction, impact depth, and affected area of ​​defects. These steps provide accurate defect distribution, reliable benchmarks, and propagation boundary inputs for subsequent multi-objective augmented game theory models, significantly improving the accuracy of root cause localization, the targeting of augmentation strategies, and the effectiveness of resource allocation.

[0160] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the data quality adaptive enhancement method for industrial edge gateways in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0161] This application also provides a data quality adaptive enhancement device for industrial edge gateways; please refer to [reference needed]. Figure 2 The data quality adaptive enhancement device for industrial edge gateways includes: The acquisition module 10 is used to acquire the raw data stream of the heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway, and to perform protocol stripping and time coarse alignment on the raw data stream and the real-time resource status to obtain the raw data block sequence, time drift, equipment health baseline and industrial physical constraint map. The traceability module 20 is used to locate the root cause of data quality defects and trace the spread of pollution based on the original data block sequence, time drift amount, equipment health baseline and industrial physical constraint map, and to obtain the physical violation point set, credible benchmark cluster and pollution propagation path; The construction module 30 is used to construct a multi-objective augmented game model based on the physical violation point set, the credible benchmark cluster and the pollution propagation path. The multi-objective augmented game model includes a cleansing strategy participant, a reconstruction strategy participant and a compensation strategy participant, and the participants achieve strategy equilibrium through a Pareto collaborative optimization mechanism based on marginal utility evaluation. Execution module 40 is used to perform edge-side hierarchical enhancement execution based on the multi-objective enhanced game model and the real-time resource state quantity, to obtain the enhanced data stream, residual error tensor and edge-side data value density distribution; Verification module 50 is used to perform physical consistency closed-loop verification based on the trusted benchmark cluster and the residual error tensor to obtain the quality gain index and the strategy regret value. The transmission module 60 is used to dynamically shape the uplink data stream of the edge gateway according to the quality gain index, the policy regret value, the edge-side data value density distribution and the real-time resource status quantity, so as to obtain the target data quality adaptive enhancement result of the edge side of the industrial Internet of Things.

[0162] The data quality adaptive enhancement device for industrial edge gateways provided in this application, employing the data quality adaptive enhancement method for industrial edge gateways in the above embodiments, can solve the technical problems of poor data identification accuracy, unreasonable resource allocation, and poor transmission reliability in the process of data quality enhancement on the industrial edge side. Compared with the prior art, the beneficial effects of the data quality adaptive enhancement device for industrial edge gateways provided in this application are the same as those of the data quality adaptive enhancement method for industrial edge gateways provided in the above embodiments, and other technical features in the data quality adaptive enhancement device for industrial edge gateways are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0163] This application also proposes a data quality adaptive enhancement device for an industrial edge gateway, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the data quality adaptive enhancement method for the industrial edge gateway as described above.

[0164] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data quality adaptive enhancement method for industrial edge gateways as described above.

[0165] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for adaptively enhancing data quality in an industrial edge gateway, characterized in that, The method includes: The raw data streams of heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway are obtained. The raw data streams are then stripped of protocols and coarsely aligned in time to obtain the raw data block sequence, time drift, equipment health baseline and industrial physical constraint map. Based on the original data block sequence, time-series drift, equipment health baseline, and industrial physical constraint map, the root cause of data quality defects and the traceability of pollution propagation are performed to obtain the set of physical violation points, the trusted benchmark cluster, and the pollution propagation path. A multi-objective augmented game model is constructed based on the physical violation point set, credible benchmark cluster, and pollution propagation path. The multi-objective augmented game model includes cleanup strategy participants, reconstruction strategy participants, and compensation strategy participants. The participants achieve strategy equilibrium through a Pareto collaborative optimization mechanism based on marginal utility evaluation. Based on the multi-objective enhanced game model and the real-time resource state quantity, the edge-side hierarchical enhanced execution is performed to obtain the enhanced data stream, residual error tensor and edge-side data value density distribution. Physical consistency closed-loop verification is performed based on the trusted benchmark cluster and the residual error tensor to obtain the quality gain index and the strategy regret value. The uplink data stream of the edge gateway is dynamically shaped based on the quality gain index, policy regret value, edge-side data value density distribution, and real-time resource status quantity to achieve adaptive enhancement of data quality at the edge of the industrial IoT.

2. The method as described in claim 1, characterized in that, The process of acquiring the raw data streams of heterogeneous sensing nodes in the industrial field and the real-time resource status quantities of the edge gateway, and performing protocol stripping and coarse timing alignment on the raw data streams to obtain the raw data block sequence, timing drift, device health baseline, and industrial physical constraint map includes: The raw data streams of heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway are obtained. The raw data streams include at least the raw sampled values, sampling timestamps, communication protocol types and device topology identifiers of each sensing node. The real-time resource status includes at least the CPU utilization rate, memory availability, bandwidth availability and edge cache queue depth. The original data stream is subjected to protocol stripping and frame structure standardization to extract payload data and obtain a standardized frame sequence. The standardized frame sequence is coarsely aligned based on the sampling time scale to estimate the temporal drift of each sensing node relative to the unified time base, and a set of time-aligned data frames is obtained. Historical statistical features are extracted from the time-aligned data frame set to establish the data distribution characteristics of each sensing node under normal operating conditions, and the equipment health baseline is obtained. Based on the equipment topology identifier and the pre-set industrial equipment mechanism model library, an industrial physical constraint map is established to characterize the physical conservation relationship, parameter coupling boundary and state evolution law of equipment operation; The time-aligned data frame set is divided and aggregated according to a preset time window to obtain the original data block sequence.

3. The method as described in claim 2, characterized in that, The process of locating the root cause of data quality defects and tracing the propagation of contamination based on the original data block sequence, time-series drift, equipment health baseline, and industrial physical constraint map yields a set of physical violation points, a cluster of reliable benchmarks, and a contamination propagation path, including: The physical consistency of the original data block sequence is verified based on the industrial physical constraint map to obtain a set of physical violation points. The physical consistency verification includes at least energy conservation verification, parameter coupling boundary verification, and state evolution continuity verification. Based on the time drift, cross-node clock synchronization deviation analysis is performed to identify sampling misalignment caused by clock out-of-sync, and a set of time-misaligned nodes is obtained. Based on the device health baseline, multi-node consistency voting is performed on the data of each sensing node in the original data block sequence, and the data of sensing nodes that deviate from the device health baseline by less than a preset threshold are included in the trusted benchmark cluster. Based on the device topology identifier and the set of time-series misaligned nodes, spatial correlation analysis of data anomalies is performed to determine the propagation direction and impact depth of quality defects from the abnormal nodes along the device topology network to the associated nodes, thereby obtaining the pollution propagation path.

4. The method as described in claim 1, characterized in that, The construction of a multi-objective enhanced game model based on the set of physical violation points, the credible benchmark cluster, and the contamination propagation path includes: Based on the defect type distribution and severity level of the physical violation point set, the initial strategy spaces of the cleaning strategy participants, reconstruction strategy participants, and compensation strategy participants are determined respectively. The strategy influence domain of each participant is determined based on the pollution propagation path, wherein the strategy influence domain is used to limit the boundary of action of each participant when performing enhancement operations; The reference benchmark domains of the reconstruction strategy participants are determined based on the trusted benchmark cluster, wherein the reference benchmark domains are used to provide neighborhood trusted reference values ​​for missing or anomalous data segments. Construct the marginal utility function for each participant, whereby the marginal utility function characterizes the amount of data quality improvement brought about by a unit investment of edge computing resources; Based on the initial policy space, policy influence domain, reference domain, and marginal utility function, the Pareto optimal policy combination is obtained by solving the non-dominated policy set through the Pareto collaborative optimization mechanism. Based on the real-time resource state quantities, the resource feasibility of the Pareto optimal strategy combination is verified to obtain a multi-objective augmented game model.

5. The method as described in claim 1, characterized in that, The step of performing hierarchical edge-side enhancement based on the multi-objective enhanced game model and the real-time resource state variables to obtain the enhanced data stream, residual error tensor, and edge-side data value density distribution includes: The Pareto optimal strategy combination is extracted from the multi-objective augmentation game model, and the cleaning decision parameters, reconstruction decision parameters, and compensation decision parameters are obtained analytically. The computing resource quotas of the participants in the cleaning strategy are determined based on the CPU utilization rate in the real-time resource status. The original data block sequence is subjected to noise suppression and outlier removal based on the physical violation point set according to the cleaning decision parameters and the computing resource quotas to obtain the cleaned data stream. Based on the memory remaining amount in the real-time resource status quantity, the reconstruction window length of the reconstruction strategy participants is determined. Based on the reconstruction decision parameters and the reconstruction window length, the missing intervals in the cleaned data stream are spatiotemporally co-impregnated according to a trusted benchmark cluster to obtain the reconstructed data stream. Based on the bandwidth margin in the real-time resource status quantity, the compensation update frequency of the compensation strategy participants is determined. Based on the compensation decision parameters and the compensation update frequency, the systematic deviation in the reconstructed data stream is corrected by residual compensation based on the industrial physical constraint map to obtain the enhanced data stream. The residual error is calculated frame by frame for the enhanced data stream to obtain the residual error tensor; The value density distribution of edge-side data is obtained by evaluating the impact of each data frame in the original data block sequence on downstream control decisions.

6. The method as described in claim 4, characterized in that, The physical consistency closed-loop verification based on the trusted benchmark cluster and the residual error tensor, to obtain the quality gain index and policy regret value, includes: The trusted benchmark cluster is used as the benchmark reference set for physical consistency verification to determine the physical consistency degree of the enhanced data stream relative to the benchmark reference set. Based on the deviation of the residual error tensor from the industrial physical constraint map, and combined with the physical consistency fit, the ratio of the data quality metric before and after enhancement is determined, and the quality gain index is obtained. Based on the deviation between the expected marginal utility and the actual marginal utility of each participant in the Pareto optimal strategy combination, the strategy utility loss of each participant is determined, and the strategy regret value is obtained.

7. The method as described in claim 1, characterized in that, The step of dynamically shaping the uplink data stream of the edge gateway based on the quality gain index, policy regret value, edge-side data value density distribution, and real-time resource status parameters to achieve adaptive enhancement of data quality at the edge of the Industrial Internet of Things includes: Obtain the current bandwidth utilization and network link quality index of each uplink channel of the edge gateway; The transmission priority of each data frame is determined based on the data value density distribution at the edge, wherein data frames with higher value density levels are assigned higher transmission priority. The transmission reliability level of each data frame is determined based on the quality gain index, wherein data frames with a quality gain index lower than a preset threshold trigger increased transmission redundancy. The verification strength level of each data frame is determined based on the policy regret value, wherein data frames with a policy regret value higher than a preset threshold trigger enhanced integrity verification. Based on the transmission priority, transmission reliability level, verification strength level, and bandwidth margin in the real-time resource status, the enhanced data stream is dynamically shaped to obtain a shaped transmission sequence. The transmission rate of the shaped transmission sequence is adaptively adjusted according to the real-time resource status, driving each uplink channel to send data and monitoring the transmission quality response in real time. When the deviation between the transmission quality response and the expected transmission quality exceeds a preset tolerance threshold, the strategy re-solution of the multi-objective enhanced game model is triggered, and the Pareto optimal strategy combination is dynamically updated until the transmission quality of the enhanced data stream reaches the target set range, so as to realize the adaptive enhancement of data quality at the edge of the industrial Internet of Things.

8. A data quality adaptive enhancement device for an industrial edge gateway, characterized in that, The device includes: The acquisition module is used to acquire the raw data stream of heterogeneous sensing nodes in the industrial field and the real-time resource status of the edge gateway, and to perform protocol stripping and time-series coarse alignment on the raw data stream and the real-time resource status to obtain the raw data block sequence, time-series drift, equipment health baseline and industrial physical constraint map. The traceability module is used to locate the root cause of data quality defects and trace the spread of contamination based on the original data block sequence, time drift, equipment health baseline and industrial physical constraint map, so as to obtain the physical violation point set, credible benchmark cluster and contamination propagation path; The construction module is used to construct a multi-objective augmented game model based on the set of physical violation points, the cluster of credible benchmarks and the pollution propagation path. The multi-objective augmented game model includes cleansing strategy participants, reconstruction strategy participants and compensation strategy participants, and the participants achieve strategy equilibrium through a Pareto collaborative optimization mechanism based on marginal utility evaluation. The execution module is used to perform edge-side hierarchical enhancement execution based on the multi-objective enhanced game model and the real-time resource state quantity, so as to obtain the enhanced data stream, residual error tensor and edge-side data value density distribution. The verification module is used to perform physical consistency closed-loop verification based on the trusted benchmark cluster and the residual error tensor to obtain the quality gain index and the strategy regret value. The transmission module is used to dynamically shape the uplink data stream of the edge gateway according to the quality gain index, the policy regret value, the edge-side data value density distribution, and the real-time resource status quantity, so as to achieve adaptive enhancement of data quality at the edge side of the Industrial Internet of Things.

9. A data quality adaptive enhancement device for an industrial edge gateway, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the data quality adaptive enhancement method for an industrial edge gateway as claimed in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data quality adaptive enhancement method for industrial edge gateways as described in any one of claims 1 to 7.

Images