A measurement data intelligent management method and system for a fusion architecture

Abstract

The present application relates to a kind of measurement data intelligent management methods and systems for fusion architecture, comprising the following steps: S1: four unified management target and range boundary are built, to unify data model, unified interface standard, unified security policy, unified operation and management as top constraint, determine measurement data global range, management object and application target, synchronously define key quality indicators and SLA;S2: unified access layer is built to obtain data, after protocol analysis, do field normalization and unit conversion, and StreamID, source, sampling rate, original timestamp are forced to be marked, then time series is windowed, and feature vector is generated;S3: intelligent detection model is built based on physical constraint residual error auto-encoder, based on feature vector, obtain detection result QualityFinding, and repair task RepairTask to be repaired;S4: according to repair task RepairTask to be repaired, repair strategy is selected and AI is introduced to generate repair report RepairTag.The present application guarantees the reliable supply of measurement data in multiple business scenarios and SLA is achieved.

Description

Technical Field

[0001] This invention relates to the field of data governance, and in particular to a method and system for intelligent governance of measurement data oriented towards a converged architecture. Background Technology

[0002] With the advancement of new power system construction, dispatch control, condition assessment, fault diagnosis, predictive maintenance, and line loss analysis are placing higher demands on the coverage, timeliness, and reliability of measurement data. Grid-side measurement data comes from a wide range of sources, including telemetry and telecontrol from SCADA / EMS, synchronization phasors from PMUs, and electricity consumption information from AMIs, as well as various data formats such as integrated online monitoring, protection and fault recording, and edge intelligent terminals. This data exhibits characteristics such as multi-source heterogeneity, strong time-series characteristics, high frequency and high volume, and cross-professional coupling. All of this data needs to be uniformly aggregated, interpreted, and serviced across the entire network to support cross-scenario integrated applications.

[0003] In engineering practice, measurement data commonly suffers from the following problems: different systems use different point tables and naming rules, making it difficult to consistently identify the same device / measurement point across platforms; inconsistent protocols and field definitions lead to semantic ambiguity regarding units, ranges, and phases; differences in station-end time synchronization conditions and network jitter cause timestamp offsets, drifts, out-of-order sequences, and duplicates, thus affecting event correlation, phasor analysis, and state estimation; communication link fluctuations and equipment failures result in quality defects such as missing data, abrupt changes, and drifts, exhibiting high frequency, distributed characteristics, and difficulty in troubleshooting at large-scale locations. Traditional governance methods, primarily based on manual inspections, offline scripts, or single rules, often fail to cover complex anomaly patterns and lack effective location of anomaly causes and traceable evidence chains, resulting in unstable governance effects and high maintenance costs.

[0004] On the other hand, the construction of data platforms and converged architectures has promoted the "four unifications" requirements of "unified data model, unified interface standard, unified security policy, and unified operation and maintenance management," emphasizing the standardization, service-orientation, and compliant controllability of data throughout its entire lifecycle. However, in measurement data scenarios, without unified access, time governance, quality assessment, intelligent repair, lineage versioning, and closed-loop operation and maintenance mechanisms tailored to the characteristics of time-series data, the problem of "data being able to be aggregated but not usable, usable but not verifiable, and verifiable but not operable" will still arise, making it difficult to form sustainable data asset capabilities. Summary of the Invention

[0005] To address the aforementioned issues, the present invention aims to provide an intelligent governance method and system for measurement data in a converged architecture. By combining time reliability modeling, intelligent anomaly detection, and traceable repair, an end-to-end quality closed loop is formed from data collection to application, thereby ensuring reliable supply and SLA achievement of measurement data in multiple business scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent governance of measurement data for converged architectures includes the following steps: S1: Construct four unified governance goals and scope boundaries, with unified data model, unified interface standard, unified security policy, and unified operation and maintenance management as top-level constraints, to determine the full scope of measurement data, governance objects and application goals, and simultaneously define key quality indicators and SLAs; S2: Build a unified access layer to acquire data, perform field normalization and unit conversion after protocol parsing, and forcibly add StreamID, source, sampling rate, and original timestamp. Then, window the time series data to generate feature vectors. S3: Construct an intelligent detection model based on a physical constraint residual autoencoder, and obtain the detection result QualityFinding and the repair task RepairTask based on the feature vector; S4: Select a repair strategy based on the repair task (RepairTask) and introduce AI to generate a repair report (RepairTag).

[0007] Furthermore, the four unified governance goals and scope boundaries are defined as follows: First, clarify the governance boundaries and top-level constraints, unify the data model level, and establish a standardized data dictionary, point-to-table mapping specifications, and metadata management system covering the entire power system. This ensures that heterogeneous sources of data, including SCADA telemetry and telecontrol, PMU phasor data, AMI power consumption information, online monitoring status variables, protection waveform and fault waveform recording, and edge intelligent terminal sensing data, undergo field normalization, unit standardization, and semantic alignment before entering the system. Second, unify the interface standards level by formulating standardized access adaptation specifications based on IEC104, IEC61850, PMU protocols, and AMI communication protocols, and unify API gateways, message buses, and data service external interfaces to avoid point-to-point integration and interface fragmentation. Third, unify the security strategy level by establishing an end-to-end security control mechanism covering identity authentication, access control, data anonymization, and access auditing. Fourth, unify the operation and maintenance management level by constructing a centralized monitoring and alarm, fault handling, performance optimization, and capacity management system to achieve standardization, automation, and visualization of the governance process. Under the constraints of the four unifications, the system sorts out the entire scope, core objects and target scenarios of measurement data governance; and establishes a quantifiable and monitorable key quality indicators and service level agreement (SLA) system based on the governance objects and application goals.

[0008] Furthermore, a unified access layer is constructed, specifically as follows: streaming data is written to the message bus through a real-time acquisition channel, and data from different sites, device types, or business domains are carried by topics and partitions; batch data enters the ODS landing area, retaining the original file and table structure and landing time; the access layer performs unified access orchestration and flow control governance for all sources, including connection management, back pressure and rate limiting, breakpoint resumption, retry and deduplication strategies, and incorporates the basic operating indicators of the acquisition link into unified operation and maintenance monitoring.

[0009] Furthermore, after protocol parsing, field normalization and unit conversion are performed, and StreamID, source, sampling rate, and original timestamp are forcibly added. Then, the time series is windowed to generate feature vectors, as follows: In the data parsing and normalization stage, the unified access layer performs protocol parsing through the protocol adapter, extracting the point identifier, measurement value, quality bit, timestamp, and device and channel identifier fields; then, based on the unified data model and metadata dictionary, field normalization and unit conversion are performed, and basic format verification is conducted; the access layer forcibly completes each record and writes it into the unified metadata tag; after normalization, the access layer performs windowing processing on the time series data, organizing continuously arriving samples into fixed windows or sliding windows, and tolerating and merging out-of-order and late data according to the watermark strategy, outputting a window sequence with a consistent structure, and generating feature vectors based on the window sequence for subsequent quality detection and root cause analysis.

[0010] Furthermore, an intelligent detection model is constructed based on a physically constrained residual autoencoder. Based on the feature vector, the detection result QualityFinding and the repair task RepairTask are obtained, as detailed below: S3-1: Based on the feature vector obtained in S2, construct a window feature vector and standardize it, and construct a missing mask vector for missing, invalid or unavailable feature dimensions; S3-2: Construct the physically constrained residual autoencoder model PC-AE; S3-3: Train the model with reconstruction consistency, physical constraint consistency, topological neighborhood consistency and spectral consistency as joint objectives to obtain the intelligent detection model; S3-4: Based on the intelligent detection model, online inference calculation of fused anomaly scores, output anomaly windows and anomaly intervals, and generate evidence based on residual contribution and physical violation items, outputting QualityFinding; S3-5: Generate RepairTask based on QualityFinding.

[0011] Furthermore, a physically constrained residual autoencoder model, PC-AE, is constructed as follows: For each measurement data stream identified as StreamID, the features output by S2 are windowed according to a preset window length W and step size H, forming a window feature vector x at window time t. s,t Where s is StreamID; the window feature vector is standardized to obtain the standardized window feature vector: ; Wherein, μ and σ are obtained from the mean and variance of the training reference set statistics. To prevent division by zero of constants; Build encoder f θ (·) and decoder g φ (·), reconstruct the standardized window features: ; Among them, z s,t For latent variable representation, θ and φ are respectively the encoder f θ (·) and decoder g φ (·) model parameters; Calculate the reconstructed residual vector: ; Residual branches are introduced to characterize anomalous pattern summaries, and a residual encoder h is constructed. ψ (·): q s,t =h ψ (r s,t ); Where, q s,t The output of the residual branch is used for subsequent type discrimination evidence and RepairTask generation.

[0012] Furthermore, based on the intelligent detection model, online inference calculations are performed to fuse anomaly scores, and anomaly windows and anomaly intervals are output, as detailed below: (1) During the online reasoning stage, input is processed for each window. Output Calculate the reconstruction error term : ; Where s is the data stream identifier; t is time; d is the number of feature dimensions; j is the feature dimension index; m s,t,j This is a missing dimension mask; 1 indicates that the j-th dimension is valid, and 0 indicates that it is missing; w j Let the j-th dimension feature weight be denoted as 'j'. To prevent extremely small constants with a denominator of 0; (2) Calculate physical violations : ; Where β1, β2, β3 are the weighting coefficients of the physical sub-constraints; κ X,s,t The index for the imbalance of the ordinal components of the original data is denoted by X, where X represents the type of physical quantity. The imbalance index corresponding to the reconstructed data; P and Q represent active power and reactive power, respectively. The active power and reactive power are calculated from the reconstructed data; The rate of change of the input data; To reconstruct the frequency change rate of the data; (3) Calculate topology violations : ; in, The mean vector of the input features over the topological neighborhood; To reconstruct the mean vector of the features in the topological neighborhood; It is the square norm 2; (4) Calculate spectrum violations: ; Among them, e s,t The spectral feature vector of the input data; To reconstruct the spectral feature vector of the data; (5) Define the fusion anomaly score A s,t : ; Where, γ phy ,γ topo ,γ spec These are the weighting coefficients that map physical, topological, and spectral violations to the fusion score; (6) To accommodate the statistical differences among different StreamIDs, a threshold τ is determined for each StreamID based on the quantile of the reference set. s : ; Among them, Q p (·) is the quantile function; it returns the p-th quantile value of the sample set; p is the quantile level; ref is the high confidence data window; when A s,t >τ s This window has been identified as abnormal. (7) Merge adjacent abnormal windows according to the maximum interval to form an abnormal interval: ; Among them, I k This represents the k-th abnormal interval; T start T is the start time of the interval; end This is the end time of the interval; And define the interval intensity S k : .

[0013] Furthermore, evidence is generated based on the residual contribution and physical violations, and QualityFinding is output as follows: (1) Calculate the contribution of the one-dimensional residual c s,t,j : ; (2) According to feature group G g Aggregate to obtain group contribution: ; Where g is the feature group index; G g C is the set of feature dimension indices contained in the g-th group; s,t,g P represents the total contribution value of the g-th feature group. s,t,g Let g be the probability of the g-th group contributing to the anomaly. And in the abnormal interval I k The above calculation group's evidence statistics: ; Among them, |I k | is the interval I k Number of time points included; Time interval I k The average contribution probability of the g-th feature group; (3) Calculate the physical violation statistics within the interval: ; in, This represents the average physical constraint violation rate within the interval. This represents the average violation rate of the topological neighborhood within the interval. (4) For type sets Constructing evidence points: ; in For the event Weight coefficients of the g-th feature For the event The coefficient of the u-th additional feature; For interval I k The u-th eigenvalue Weighting coefficients for physical and topological violations, respectively. An abnormal event (5) Confidence level: ; ; ; ; Where, p score σ is the anomaly intensity confidence level; σ(·) is the Sigmoid activation function; α is the intensity weight coefficient; τ s p is the anomaly detection threshold for measurement point s; dur γ is the duration confidence level; γ is the time decay coefficient; 1-exp(-γ∣I k |) represents a nonlinear mapping with higher confidence levels the longer the duration; p type η is the confidence level for the anomaly type; E is the type discrimination coefficient; (1) E (2) Feature scores for two candidate anomalous events; σ(η(E) (1) -E (2) The larger the score difference, the higher the confidence level of the type determination; Conf represents the overall confidence level. (6) Scope of influence: Calculate the co-occurrence rate R for the inflow s′ within the candidate set N(s):

[0014] When R≥r0, it is included in the scope of influence.

[0015] Finally, the structured output provides a QualityFinding for each anomalous interval.

[0016] Furthermore, a repair strategy is selected based on the repair task (RepairTask), and AI is used to generate a repair report (RepairTag), as detailed below: Based on the RepairTask output by S3, the optimal repair strategy is determined by combining a multidimensional decision tree and a rule engine. First, the key fields in the RepairTask are analyzed, including TaskType, AnomalyType, Confidence, ImpactScope, Evidence, and Priority, to construct the repair decision vector d. repair : d repair =[TaskType,AnomalyType,Conf,∣ImpactScope∣,Evidence,Priority] T ; Based on the classification system of power data quality anomalies, a repair strategy library is established. Strategy matching adopts a similarity-based retrieval algorithm to calculate the similarity between RepairTask and historical cases in the strategy library. After determining the candidate remediation strategies, a multi-strategy fusion decision-making mechanism is used to select the optimal remediation scheme and optimize the remediation parameters, establishing a strategy combination scoring function: ; Where, α s The strategy weights are defined as follows: Effectiveness(s) is the strategy effectiveness score, and Compatibility(s,S) is the strategy weight. combo The score represents the compatibility rating between strategies, and Cost(s) represents the cost of implementing the strategies. Repair parameter optimization is achieved through a Bayesian optimization algorithm, establishing a parameter space Θ={θ1,θ2,…,θ} for key parameters of different repair strategies. k Define the objective function: ; Among them Quality after and Quality before These are the data quality scores before and after the repair, respectively, and Stability(θ) is the repair stability index. A Gaussian process is used to establish a mapping relationship between parameters and the objective function. The parameter search is guided by the acquisition function to achieve automatic optimization of the repair parameters. RepairTag, an AI-generated repair report, employs a hybrid architecture that combines a large language model with a domain knowledge base to automate document generation and knowledge summarization of the repair process.

[0017] A measurement data intelligent governance system for converged architecture includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the measurement data intelligent governance method for converged architecture as described above.

[0018] The present invention has the following beneficial effects: 1. This invention uses four unifications as the top-level constraint, forming a closed loop from governance objectives, scope boundaries, objects and applications to quality indicators and SLAs, covering all domain measurement data and all elements, significantly reducing the integration costs and management risks caused by inconsistent data standards across systems, fragmented interfaces and decentralized security operations and maintenance; 2. This invention achieves protocol parsing, field normalization, unit conversion, and mandatory metadata annotation through a unified access layer. It also generates statistical multidimensional features by windowing time series data and realizes anomaly detection by combining autoencoder reconstruction error. This produces QualityFinding and RepairTask that can be structured and implemented, making anomaly location more accurate, alarms more interpretable, and governance actions automatically triggered. Attached Figure Description

[0019] Figure 1This is a flowchart of the method of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a method for intelligent governance of measurement data oriented towards a converged architecture is provided, including the following steps: S1: Construct four unified governance goals and scope boundaries, with unified data models, unified interface standards, unified security policies, and unified operation and maintenance management as top-level constraints. Define the full scope of measurement data (SCADA / PMU / AMI / online monitoring / protection waveform recording / edge intelligent terminals, etc.), governance objects (point tables, time series, events, alarms, equipment ledgers, topology and communication links), and application goals (scheduling and monitoring, status evaluation, fault diagnosis, predictive maintenance, line loss analysis, etc.). Simultaneously define key quality indicators (completeness rate, accuracy rate, consistency, timeliness, traceability) and SLA (end-to-end latency, availability, throughput). S2: Construct a unified access layer to acquire data, stream data enters the message bus, and batch data enters the ODS landing area. After protocol parsing, field normalization and unit conversion are performed, and StreamID, source, sampling rate, and original timestamp are forcibly added. Then, the time series is windowed to generate feature vectors: statistical features (mean / variance / kurtosis / skewness), time series features (autocorrelation, spectral energy, rate of change), phasor / three-correlation features, topological neighborhood features (correlation quantities of the same bus / same feeder / same transformer), and communication features (delay, jitter, packet loss). S3: Construct an intelligent detection model based on a physical constraint residual autoencoder. Based on feature vectors, obtain the detection result QualityFinding, including anomaly intervals, types (communication / clock / sensor / model mapping / operational status changes), confidence level and impact range, as well as the repair task to be repaired. S4: Select a repair strategy based on the repair task (RepairTask) and introduce AI to generate a repair report (RepairTag).

[0021] In this embodiment, the four unified governance objectives and scope boundaries are constructed as follows: First, clarify the governance boundaries and top-level constraints, unify the data model level, and establish a standardized data dictionary, point-to-table mapping specifications, and metadata management system covering the entire power system. This ensures that heterogeneous sources of data, including SCADA telemetry and telecontrol, PMU phasor data, AMI power consumption information, online monitoring status variables, protection waveform and fault waveform recording, and edge intelligent terminal sensing data, undergo field normalization, unit standardization, and semantic alignment before entering the system. Second, unify the interface standards level by formulating standardized access adaptation specifications based on IEC104, IEC61850, PMU protocols, and AMI communication protocols, and unify API gateways, message buses, and data service external interfaces to avoid point-to-point integration and interface fragmentation. Third, unify the security strategy level by establishing an end-to-end security control mechanism covering identity authentication, access control, data anonymization, and access auditing. Fourth, unify the operation and maintenance management level by constructing a centralized monitoring and alarm, fault handling, performance optimization, and capacity management system to achieve standardization, automation, and visualization of the governance process. Under the constraints of the four unifications, the system systematically outlines the overall scope, core objects, and target scenarios of measurement data governance. The data sources encompass telemetry, telesignaling, and remote control data from Supervisory Control and Data Acquisition (SCADA) systems; voltage / current phasor and frequency data from synchronous phasor measurement units (PMUs); electricity consumption data from smart meters in Advanced Measurement Infrastructure (AMIs); temperature, vibration, and partial discharge data from online monitoring of equipment such as transformers, switchgear, and lines; waveform recording and action information from relay protection devices; and environmental perception and load forecasting data from intelligent terminals at the distribution network edge. The governance objects include equipment point tables (static attributes such as measurement point code, name, type, range, and accuracy), time-series data (real-time and historical measurement value sequences), event data (discrete events such as operations, alarms, and faults), and other related data. The system includes alarm data (alarm information such as exceeding limits, communication interruption, and equipment abnormalities), equipment ledgers (basic equipment information, technical parameters, maintenance records, etc.), topology relationships (electrical connections, geographical locations, logical groups, etc.), and communication links (network quality indicators such as communication status, latency, and packet loss rate). Application objectives focus on core business scenarios such as scheduling and monitoring (real-time status display, trend analysis, and operation assistance), status evaluation (equipment health and system security margin assessment), fault diagnosis (anomaly location, cause analysis, and impact assessment), predictive maintenance (prediction of equipment degradation trends and optimization of maintenance plans), and line loss analysis (theoretical calculation, actual measurement comparison, and anomaly investigation). Furthermore, based on the governance targets and application objectives, a quantifiable and monitorable key quality indicators and service level agreement (SLA) system is established. Regarding quality indicators, completeness is defined as the proportion of valid data to the total data to be collected (e.g., telemetry point completeness ≥ 99.5%, PMU data completeness ≥ 99.9%), accuracy is defined as the degree of deviation between data values ​​and true values ​​(e.g., analog quantity measurement error ≤ 0.2%, switch quantity status accuracy ≥ 99.99%), consistency is defined as the degree of logical conformity of multi-source data of the same type (e.g., correlation verification of multiple measurement points of the same device, three-phase data balance verification, etc.), and timeliness is defined as the end-to-end delay from data acquisition to usability (e.g., SCADA data ≤ 3 seconds, PMU data ≤ 100 milliseconds). For AMI data (≤15 minutes), traceability is defined as the ability to fully record and audit data lineage, processing, and quality changes; in terms of SLA indicators, end-to-end latency covers the entire link time requirements from terminal collection, communication transmission, system processing to application availability; availability is defined as the proportion of governance service uptime to total time (e.g., core service availability ≥99.95%); throughput is defined as the data volume processing capacity per unit time (e.g., supporting second-level concurrent processing of millions of test points); and a hierarchical alarm mechanism and automatic recovery strategy are established to ensure that the governance effect can be quantitatively evaluated and continuously improved.

[0022] In this embodiment, a unified access layer is constructed as follows: Streaming data (such as SCADA real-time telemetry, PMU high-frequency phasors, and online monitoring second-level data) is written to a message bus (e.g., Kafka / similar message middleware) using a real-time acquisition channel, with topics and partitions carrying data from different sites, device types, or business domains; Batch data (such as AMI meter reading batches, historical waveform files, and offline maintenance record-related data) enters the ODS landing area, preserving the original file and table structure and landing time to meet traceability and replay requirements; The access layer uniformly performs access orchestration and flow control management for all sources, including connection management, backpressure and rate limiting, breakpoint resumption, retry and deduplication strategies, and incorporates the basic operating indicators of the acquisition link (inbound rate, backlog depth, failure rate, etc.) into unified operation and maintenance monitoring.

[0023] In this embodiment, after protocol parsing, field normalization and unit conversion are performed, and StreamID, source, sampling rate, and original timestamp are forcibly added. Then, the time series is windowed to generate a feature vector, as follows: In the data parsing and normalization stage, the unified access layer uses a protocol adapter to parse protocols such as IEC104, IEC61850, PMU, AMI, and vendor-specific protocols, extracting point identifiers, measurement values, quality bits, timestamps, and device and channel identifier fields. Subsequently, based on the unified data model and metadata dictionary, field normalization (synonymous field mapping, type alignment, enumeration standardization, quality bit semantic alignment) and unit conversion (such as kV / V, A / mA, MW / kW, ℃ / K, etc.) are performed, along with basic format verification (numerical range, precision, missing values, etc.). The access layer forcibly completes and writes unified metadata tags for each record: StreamID (uniquely identifies a data stream / measurement point stream, usually generated by combining site ID, device ID, measurement point ID, protocol channel, etc.), source (system / plant / vendor / channel / collector), sampling rate (or upload cycle / trigger method), original timestamp (original device upload time and access time can be retained simultaneously), and if necessary, data version, parser version, mapping rule version, etc., to locate problems caused by "different calibers for the same point" or "rule changes"; after completing the normalization, the access layer performs windowing processing on the time series data, organizing continuously arriving samples according to fixed windows or sliding windows (e.g., according to 1 second / 5 seconds / 1 minute windows, or according to PMU). The system uses 20ms / 40ms windows and tolerates and merges out-of-order and late data according to a watermark strategy, outputting a consistent window sequence. Based on this window sequence, feature vectors are generated for subsequent quality inspection and root cause analysis. These feature vectors include statistical features (mean, variance, extreme values, kurtosis, skewness, quantiles, etc.), time-series features (autocorrelation, spectral energy, rate of change / slope, volatility, etc.), phasor / three-phase correlation features (such as three-phase imbalance, phase angle difference, sequence component / negative sequence index, etc., selected according to data type), topological neighborhood features (coordinated changes and mutual constraint relationship indicators of related measurement points under the same bus / feeder / transformer), and communication features (delay, jitter, packet loss, retransmission, backlog, etc.). Finally, the unified access layer outputs a normalized standard data stream / ODS standard table, complete metadata tags, and windowed feature vectors downstream, providing consistent input for subsequent time reliability modeling, anomaly detection, and repair decisions.

[0024] In this embodiment, an intelligent detection model is constructed based on a physically constrained residual autoencoder. Based on the feature vector, the detection result QualityFinding and the repair task RepairTask are obtained, as detailed below: S3-1: Based on the feature vector obtained in S2, construct a window feature vector and standardize it, and construct a missing mask vector for missing, invalid or unavailable feature dimensions; S3-2: Construct the physically constrained residual autoencoder model PC-AE; S3-3: Train the model with reconstruction consistency, physical constraint consistency, topological neighborhood consistency and spectral consistency as joint objectives to obtain the intelligent detection model; S3-4: Based on the intelligent detection model, online inference calculation of fused anomaly scores, output anomaly windows and anomaly intervals, and generate evidence based on residual contribution and physical violation items, outputting QualityFinding; S3-5: Generate RepairTask based on QualityFinding. RepairTask includes: task type, target object, time range, priority, set of suggested actions, evidence reference, and reference to the original data segment.

[0025] In this embodiment, a physically constrained residual autoencoder model PC-AE is constructed, as follows: For each measurement data stream identified as StreamID, the features output by S2 are windowed according to a preset window length W and step size H, forming a window feature vector x at window time t. s,t Where s is StreamID; the window feature vector is standardized to obtain the standardized window feature vector: ; Wherein, μ and σ are obtained from the mean and variance of the training reference set statistics. To prevent division by zero of constants; Build encoder f θ (·) and decoder g φ (·), reconstruct the standardized window features: ; Among them, z s,t For latent variable representation, θ and φ are respectively the encoder f θ (·) and decoder g φ (·) model parameters; Calculate the reconstructed residual vector: ; Residual branches are introduced to characterize anomalous pattern summaries, and a residual encoder h is constructed. ψ (·): q s,t =h ψ (r s,t ); Where, qs,t The output of the residual branch is used for subsequent type discrimination evidence and RepairTask generation.

[0026] The reconstruction consistency objective is a core component of the physically constrained residual autoencoder (PC-AE). Its basic principle is to learn the intrinsic representation of normal data by minimizing the difference between the input feature vector and the reconstructed output. Specifically, for the standardized window feature vector... encoder f θ (·) Map it to a low-dimensional latent space to obtain z s,t decoder g φ (·) Then reconstruct the latent variables back into the original feature space to obtain The consistency loss is reconstructed using a mask-weighted approach to handle missing or invalid feature dimensions. The physical constraint consistency target is based on the basic physical laws of the power system. It distinguishes between real operational changes and data quality anomalies by embedding power physical relationships during the autoencoder training process. The target includes three core physical constraints: (1) Three-phase sequence component balance constraint, based on the symmetrical component theory, requires that the ratio of the negative sequence component to the zero sequence component relative to the positive sequence component of the three-phase system be kept within a reasonable range. Mathematically, this is expressed as the sequence component transformation matrix acting on the three-phase phasors to obtain the sequence components; (2) Power relationship consistency constraint, based on the power triangle relationship, ensures that the mathematical relationship between active power, reactive power and apparent power remains unchanged before and after reconstruction; (3) Frequency change rate continuity constraint, requires that the frequency and its change rate satisfy physical continuity to avoid unreasonable frequency jumps. The topological neighborhood consistency objective is based on the network topology of the power system. It detects localized anomalies and identifies fault propagation patterns by constraining data consistency between topologically adjacent measuring points. This objective first requires establishing topological neighborhood relationships between measuring points, constructing an adjacency matrix by analyzing factors such as electrical connections, geographical location, and functional affiliation. The definition of topological associations includes multiple levels: first-level neighborhoods are direct electrical connections (same bus, same feeder, same transformer winding); second-level neighborhoods are primary electrical connections (adjacent buses, high and low voltage sides of transformers, series lines); and third-level neighborhoods are logical connections (same site, same voltage level, same control area). Neighborhood sets are defined based on the adjacency matrix, and connection weights are introduced to reflect the strength of neighborhood relationships. These weights can be determined based on factors such as electrical distance, impedance, and capacity ratio. The construction of the topology graph structure lays the foundation for subsequent neighborhood consistency constraints, enabling the model to utilize the inherent network characteristics of the power system for anomaly detection. The spectral consistency target constrains the frequency domain characteristics of power signals, detecting data quality anomalies during sampling, transmission, and storage by maintaining the consistency of the signal spectral characteristics before and after reconstruction. This target first requires extracting frequency domain features from the time-domain window data, using Fast Fourier Transform (FFT) to convert the time-series signal within the window to the frequency domain, and obtaining spectral coefficients. Based on the spectral coefficients, various frequency domain features are calculated: (1) band energy features, dividing the frequency domain into several bands and calculating the energy of each band; (2) characteristic harmonic amplitude, extracting the amplitude and phase of key frequency components (fundamental, odd harmonics, and even harmonics) of the power system; (3) spectral envelope and centroid features, reflecting the overall frequency domain distribution characteristics of the signal; (4) power spectral density, describing the distribution density of signal energy in the frequency domain. These frequency domain features can reflect the essential characteristics of power signals. When problems such as sampling rate errors, aliasing, and quantization noise occur during data transmission, the frequency domain features will change significantly, providing an effective basis for anomaly detection.

[0027] In this embodiment, based on the intelligent detection model, the fusion anomaly score is calculated online, and the anomaly window and anomaly interval are output, as follows: (1) During the online reasoning stage, input is processed for each window. Output Calculate the reconstruction error term : ; Where s is the stream identifier (StreamID); t is the time; d is the number of feature dimensions; j is the feature dimension index; m s,t,j This is a missing dimension mask; 1 indicates that the j-th dimension is valid, and 0 indicates that it is missing; w j Let the j-th dimension feature weight be denoted as 'j'. To prevent extremely small constants with a denominator of 0; (2) Calculate physical violations (selected by enabled physical constraints) : ; Where β1, β2, β3 are the weighting coefficients of the physical sub-constraints; κ X,s,t X represents the order component imbalance index of the original data (e.g., negative order ratio), and X represents the type of physical quantity (e.g., voltage V or current I). The imbalance index corresponding to the reconstructed data; P and Q represent active power and reactive power, respectively. The active power and reactive power are calculated from the reconstructed data; The rate of change of the input data; To reconstruct the frequency change rate of the data; (3) Calculate topology violations : ; in, The mean vector of the input features over the topological neighborhood; To reconstruct the mean vector of the features in the topological neighborhood; It is the square norm 2; (4) Calculate spectrum violations: ; Among them, e s,t The spectral feature vector of the input data (such as band energy, harmonic energy, etc.); To reconstruct the spectral feature vector of the data; (5) Define the fusion anomaly score A s,t : ; Where, γ phy ,γ topo ,γ spec These are the weighting coefficients that map physical, topological, and spectral violations to the fusion score; (6) To accommodate the statistical differences among different StreamIDs, a threshold τ is determined for each StreamID based on the quantile of the reference set. s : ; Among them, Q p (·) is the quantile function; it returns the p-th quantile value of the sample set; p is the quantile level; ref is the high confidence data window; when A s,t >τ s This window has been identified as abnormal. (7) Merge adjacent abnormal windows according to the maximum interval to form an abnormal interval: ; Among them, I k This represents the k-th abnormal interval; T start T is the start time of the interval; end This is the end time of the interval; And define the interval intensity S k : .

[0028] In this embodiment, evidence is generated based on the residual contribution and physical violation items, and QualityFinding is output as follows: (1) Calculate the contribution of the one-dimensional residual c s,t,j : ; (2) According to feature group G gAggregate to obtain group contribution: ; Where g is the feature group index (such as statistical group, time series group, phasor group, etc.); G g C is the set of feature dimension indices contained in the g-th group; s,t,g P represents the total contribution value of the g-th feature group. s,t,g Let g be the probability of the g-th group contributing to the anomaly. And in the abnormal interval I k The above calculation group's evidence statistics: ; Among them, |I k | is the interval I k Number of time points included; Time interval I k The average contribution probability of the g-th feature group; (3) Calculate the physical violation statistics within the interval: ; in, This represents the average physical constraint violation rate within the interval. This represents the average violation rate of the topological neighborhood within the interval. (4) For type sets Constructing evidence points: ; in For the event Weight coefficients of the g-th feature For the event The coefficient of the u-th additional feature; For interval I k The u-th feature value (such as duration, peak intensity, etc.) Weighting coefficients for physical and topological violations, respectively. An abnormal event The comprehensive feature score; f eatIk,u Includes: communication latency / jitter / packet loss statistics, time synchronization alarm status, differences between multiple sources at the same point, consistency of multiple measurement points on the same device, etc.; argmax is calculated. It is an exception type; (5) Confidence level: ; ; ; ; Where, p scoreσ is the anomaly intensity confidence level; σ(·) is the Sigmoid activation function; α is the intensity weight coefficient; τ s p is the anomaly detection threshold for measurement point s; dur γ is the duration confidence level; γ is the time decay coefficient; 1-exp(-γ∣I k |) represents a nonlinear mapping with higher confidence levels the longer the duration; p type η is the confidence level for the anomaly type; E is the type discrimination coefficient; (1) E (2) Feature scores for two candidate anomalies (typically the highest and second-highest scores); σ(η(E) (1 )-E (2) The larger the score difference, the higher the confidence level of the type determination; Conf represents the overall confidence level. (6) Scope of influence: Calculate the co-occurrence rate R for the inflow s′ within the candidate set N(s):

[0029] When R≥r0, it is included in the scope of influence.

[0030] Finally, the structured output provides a QualityFinding for each anomalous interval, including: StreamID=s, Interval=I k ; ScoreSummary: maxA, mean, intensity Sk, percentage exceeding threshold;

[0031] Confidence = Conf; Evidence: Top group , Key deviation; ImpactScope: List of affected Streams and co-occurrence rate R.

[0032] In this embodiment, a repair strategy is selected based on the repair task (RepairTask), and AI is used to generate a repair report (RepairTag), as detailed below: Based on the RepairTask output by S3, the optimal repair strategy is determined by combining a multidimensional decision tree and a rule engine. First, the key fields in the RepairTask are analyzed, including TaskType, AnomalyType, Confidence, ImpactScope, Evidence, and Priority, to construct the repair decision vector d.repair : d repair =[TaskType,AnomalyType,Conf,∣ImpactScope∣,Evidence,Priority] T ; Based on the classification system of power data quality anomalies, a repair strategy library is established: (1) the interpolation repair strategy library Sinterp, which is suitable for anomalies such as communication interruption and data loss; (2) the correction repair strategy library Scorrect, which is suitable for anomalies such as sensor drift and range error; (3) the replacement repair strategy library Sreplace, which is suitable for anomalies such as equipment failure and mapping error; (4) the filtering repair strategy library Sfilter, which is suitable for anomalies such as noise interference and signal distortion; and (5) the labeling repair strategy library Slabel, which is suitable for anomalies such as changes in operating status and event confirmation. The strategy matching adopts a similarity-based retrieval algorithm to calculate the similarity between RepairTask and historical cases in the strategy library; After determining the candidate remediation strategies, a multi-strategy fusion decision-making mechanism is used to select the optimal remediation scheme and optimize the remediation parameters. For complex anomalies, a single remediation strategy may not be able to completely solve the problem; therefore, a strategy combination scoring function is established. ; Where, α s The strategy weights are defined as follows: Effectiveness(s) is the strategy effectiveness score, and Compatibility(s,S) is the strategy weight. combo The score represents the compatibility rating between strategies, and Cost(s) represents the cost of implementing the strategies. Repair parameter optimization is achieved through a Bayesian optimization algorithm. A parameter space Θ={θ1,θ2,…,θ} is established for key parameters of different repair strategies (such as interpolation window length, filter cutoff frequency, correction coefficients, etc.). k Define the objective function: ; Among them Quality after and Quality before These are the data quality scores before and after the repair, respectively, and Stability(θ) is the repair stability index. A Gaussian process is used to establish a mapping relationship between parameters and the objective function. The parameter search is guided by the acquisition function (such as EI, UCB) to achieve automatic optimization of the repair parameters. Risk assessment and safety assurance mechanisms are incorporated into the remediation strategy selection process to ensure that remediation operations will not negatively impact power system operation. The risk assessment model establishes a multi-dimensional risk matrix based on factors such as the scope of impact of the remediation operation, remediation confidence level, and data importance. R=[Rsafety,Raccuracy,Rlatency,Rrollback]T; Among them, Rsafety is the safety risk (the impact of repair errors on system operation), Raccuracy is the accuracy risk (the possibility of data quality deterioration after repair), Rlatency is the latency risk (the impact of excessive repair time on real-time performance), and Rrollback is the rollback risk (the risk that the repair operation is irreversible). A risk threshold matrix Rthreshold is established. When any risk component exceeds the threshold, a safety protection mechanism is triggered: (1) In high safety risk scenarios, a read-only labeling strategy is adopted, and no data repair is performed. Only anomalies are marked and manually confirmed; (2) In high accuracy risk scenarios, a conservative repair strategy is adopted, and a repair method with a small impact range and high success rate is selected; (3) In high latency risk scenarios, a fast repair strategy is adopted, and real-time requirements are given priority; (4) In high rollback risk scenarios, multi-version data saving is forcibly enabled to ensure that the repair operation is completely reversible. The AI-generated repair report RepairTag uses a hybrid architecture that combines a large language model (LLM) with a domain knowledge base to automate the generation of documentation and knowledge summarization of the repair process.

[0033] The hybrid architecture combining a large language model and a domain knowledge base includes four core modules: (1) a repair context encoder, which encodes information such as RepairTask, selected repair strategy, repair parameters, and execution results into a structured vector representation ccontext=Encoder([RepairTask,Strategy,Parameters,Results]); (2) a power domain knowledge injection module, which provides domain knowledge support for report generation through retrieval-enhanced generation (RAG) based on a pre-trained power professional knowledge graph and terminology dictionary; (3) a multimodal report generator, which combines text generation, chart drawing, and data visualization functions to generate a comprehensive report containing text descriptions, data charts, and repair effect comparisons; and (4) a report quality assessment and optimization module, which continuously optimizes report quality through automatic assessment indicators (such as BLEU, ROUGE, and professional terminology coverage) and human feedback mechanisms. The RepairTag adopts a hierarchical standard data structure, which includes five main parts: repair overview, detailed analysis, execution record, effect assessment, and recommended measures. The repair overview section includes basic information such as repair task ID, execution time, repair type, and scope of impact, stored in structured JSON format.

[0034] A measurement data intelligent governance system for converged architecture includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the measurement data intelligent governance method for converged architecture as described above.

[0035] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0036] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0037] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0038] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent governance of measurement data for converged architectures, characterized in that, Includes the following steps: S1: Construct four unified governance goals and scope boundaries, with unified data model, unified interface standard, unified security policy, and unified operation and maintenance management as top-level constraints, to determine the full scope of measurement data, governance objects and application goals, and simultaneously define key quality indicators and SLAs; S2: Build a unified access layer to acquire data, perform field normalization and unit conversion after protocol parsing, and forcibly add StreamID, source, sampling rate, and original timestamp. Then, window the time series data to generate feature vectors. S3: Construct an intelligent detection model based on a physical constraint residual autoencoder, and obtain the detection result QualityFinding and the repair task RepairTask based on the feature vector; S4: Select a repair strategy based on the repair task (RepairTask) and introduce AI to generate a repair report (RepairTag).

2. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 1, characterized in that, The specific objectives and scope boundaries for constructing the four unified governance frameworks are as follows: First, clarify the governance boundaries and top-level constraints, unify the data model level, and establish a standardized data dictionary, point-to-table mapping specifications, and metadata management system covering the entire power system. This ensures that heterogeneous sources of SCADA telemetry and telecontrol data, PMU phasor data, AMI electricity consumption information, online monitoring status data, protection waveform and fault waveform recording, and edge intelligent terminal sensing data are normalized, standardized in units, and semantically aligned before entering the lake. Second, unify the interface standards level, formulate standardized access adaptation specifications based on IEC104, IEC61850, PMU protocol, and AMI communication protocol, and unify API gateways, message buses, and data service external interfaces to avoid point-to-point integration and interface fragmentation. At the unified security strategy level, establish an end-to-end security control mechanism covering identity authentication, access control, data anonymization, and access auditing; At the unified operation and maintenance management level, a centralized monitoring and alarm, fault handling, performance optimization and capacity management system is built to achieve standardization, automation and visualization of the governance process; Under the constraints of the four unifications, the system systematically sorts out the overall scope, core objects and target scenarios of measurement data governance; Based on the governance targets and application objectives, establish a quantifiable and monitorable key quality indicators and service level agreement (SLA) system.

3. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 1, characterized in that, The construction of the unified access layer is as follows: streaming data is written to the message bus through a real-time acquisition channel, and data from different sites, device types or business domains are carried by topics and partitions; batch data enters the ODS landing area, retaining the original file and table structure and landing time. The access layer performs unified access orchestration and flow control management for all sources, including connection management, back pressure and rate limiting, breakpoint resumption, retry and deduplication strategies, and incorporates the basic operating indicators of the acquisition link into unified operation and maintenance monitoring.

4. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 3, characterized in that, After protocol parsing, field normalization and unit conversion are performed, and StreamID, source, sampling rate, and original timestamp are forcibly added. Then, the time series is windowed to generate feature vectors, as follows: In the data parsing and normalization stage, the unified access layer performs protocol parsing through the protocol adapter, extracting the point identifier, measurement value, quality bit, timestamp, device and channel identifier fields; then, based on the unified data model and metadata dictionary, field normalization and unit conversion are performed, and basic format verification is carried out. The access layer forces each record to be completed and writes it into a unified metadata tag; After normalization, the access layer performs windowing processing on the time-series data, organizing continuously arriving samples into fixed or sliding windows, and tolerating and merging out-of-order and late-arriving data according to the waterline strategy, outputting a window sequence with a consistent structure, and generating feature vectors based on the window sequence for subsequent quality detection and root cause analysis.

5. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 1, characterized in that, The intelligent detection model, constructed based on a physically constrained residual autoencoder, obtains the detection result QualityFinding and the repair task RepairTask based on feature vectors, as detailed below: S3-1: Based on the feature vector obtained in S2, construct a window feature vector and standardize it, and construct a missing mask vector for missing, invalid or unavailable feature dimensions; S3-2: Construct the physically constrained residual autoencoder model PC-AE; S3-3: Train the model with reconstruction consistency, physical constraint consistency, topological neighborhood consistency and spectral consistency as joint objectives to obtain the intelligent detection model; S3-4: Based on the intelligent detection model, online inference calculation of fused anomaly scores, output anomaly windows and anomaly intervals, and generate evidence based on residual contribution and physical violation items, outputting QualityFinding; S3-5: Generate RepairTask based on QualityFinding.

6. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 5, characterized in that, The construction of the physically constrained residual autoencoder model PC-AE is as follows: For each measurement data stream identified as StreamID, the features output by S2 are windowed according to a preset window length W and step size H, forming a window feature vector x at window time t. s,t Where s is StreamID; the window feature vector is standardized to obtain the standardized window feature vector: ； Wherein, μ and σ are obtained from the mean and variance of the training reference set statistics. To prevent division by zero of constants; Build encoder f θ (·) and decoder g φ (·), reconstruct the standardized window features: ； Among them, z s,t For latent variable representation, θ and φ are respectively the encoder f θ (·) and decoder g φ (·) model parameters; Calculate the reconstructed residual vector: ； Residual branches are introduced to characterize anomalous pattern summaries, and a residual encoder h is constructed. ψ (·): q s,t =h ψ (r s,t ); Where, q s,t The output of the residual branch is used for subsequent type discrimination evidence and RepairTask generation.

7. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 6, characterized in that, The intelligent detection model is used to perform online inference calculation and fusion of anomaly scores, and output anomaly windows and anomaly intervals, as detailed below: (1) During the online reasoning stage, input is processed for each window. Output Calculate the reconstruction error term : ； Where s is the data stream identifier; t is time; d is the number of feature dimensions; j is the feature dimension index; m s,t,j This is a missing dimension mask; 1 indicates that the j-th dimension is valid, and 0 indicates that it is missing; w j Let the j-th dimension feature weight be denoted as 'j'. To prevent extremely small constants with a denominator of 0; (2) Calculate physical violations : ； Where β1, β2, β3 are the weighting coefficients of the physical sub-constraints; κ X,s,t The index for the imbalance of the ordinal components of the original data is denoted by X, where X represents the type of physical quantity. The imbalance index corresponding to the reconstructed data; P and Q represent active power and reactive power, respectively. The active power and reactive power are calculated from the reconstructed data; The rate of change of the input data; To reconstruct the frequency change rate of the data; (3) Calculate topology violations : ； in, The mean vector of the input features over the topological neighborhood; To reconstruct the mean vector of the features in the topological neighborhood; It is the square norm 2; (4) Calculate spectrum violations: ； Among them, e s,t The spectral feature vector of the input data; To reconstruct the spectral feature vector of the data; (5) Define the fusion anomaly score A s,t : ； Where, γ phy ,γ topo ,γ spec These are the weighting coefficients that map physical, topological, and spectral violations to the fusion score; (6) To accommodate the statistical differences among different StreamIDs, a threshold τ is determined for each StreamID based on the quantile of the reference set. s : ； Among them, Q p (·) is the quantile function; it returns the p-th quantile value of the sample set; p is the quantile level; ref is the high confidence data window; when A s,t >τ s This window has been identified as abnormal. (7) Merge adjacent abnormal windows according to the maximum interval to form an abnormal interval: ； Among them, I k This represents the k-th abnormal interval; T start T is the start time of the interval; end This is the end time of the interval; And define the interval intensity S k : 。 8. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 7, characterized in that, The evidence is generated based on the residual contribution and physical violation items, and the output QualityFinding is as follows: (1) Calculate the contribution of the one-dimensional residual c s,t,j : ； (2) According to feature group G g Aggregate to obtain group contribution: ； Where g is the feature group index; G g C is the set of feature dimension indices contained in the g-th group; s,t,g P represents the total contribution value of the g-th feature group. s,t,g Let g be the probability of the g-th group contributing to the anomaly. And in the abnormal interval I k The above calculation group's evidence statistics: ； Among them, |I k | is the interval I k Number of time points included; Time interval I k The average contribution probability of the g-th feature group; (3) Calculate the physical violation statistics within the interval: ； in, This represents the average physical constraint violation rate within the interval. This represents the average violation rate of the topological neighborhood within the interval. (4) For type sets Constructing evidence points: ； in For the event Weight coefficients of the g-th feature For the event The coefficient of the u-th additional feature; For interval I k The u-th eigenvalue Weighting coefficients for physical and topological violations, respectively. An abnormal event (5) Confidence level: ； ； ； ； Where, p score σ is the anomaly intensity confidence level; σ(·) is the Sigmoid activation function; α is the intensity weight coefficient; τ s p is the anomaly detection threshold for measurement point s; dur γ is the duration confidence level; γ is the time decay coefficient; 1-exp(-γ∣I k |) represents a nonlinear mapping with higher confidence levels the longer the duration; p type η is the confidence level for the anomaly type; E is the type discrimination coefficient; (1) E (2) Feature scores for two candidate anomalous events; σ(η(E) (1) -E (2) The larger the score difference, the higher the confidence level of the type determination; Conf represents the overall confidence level. (6) Scope of influence: Calculate the co-occurrence rate R for the inflow s′ within the candidate set N(s): ； When R≥r0, it is included in the scope of influence; Finally, the structured output provides a QualityFinding for each anomalous interval.

9. The intelligent governance method for measurement data oriented towards a converged architecture according to claim 1, characterized in that, The process of selecting a repair strategy based on the repair task (RepairTask) and using AI to generate a repair report (RepairTag) is as follows: Based on the RepairTask output by S3, the optimal repair strategy is determined by combining a multidimensional decision tree and a rule engine. First, the key fields in the RepairTask are analyzed, including TaskType, AnomalyType, Confidence, ImpactScope, Evidence, and Priority, to construct the repair decision vector d. repair : d repair =[TaskType,AnomalyType,Conf,∣ImpactScope∣,Evidence,Priority] T ; Based on the classification system of power data quality anomalies, a repair strategy library is established. Strategy matching adopts a similarity-based retrieval algorithm to calculate the similarity between RepairTask and historical cases in the strategy library. After determining the candidate remediation strategies, a multi-strategy fusion decision-making mechanism is used to select the optimal remediation scheme and optimize the remediation parameters, establishing a strategy combination scoring function: ； Where, α s The strategy weights are defined as follows: Effectiveness(s) is the strategy effectiveness score, and Compatibility(s,S) is the strategy weight. combo The score represents the compatibility rating between strategies, and Cost(s) represents the cost of implementing the strategies. Repair parameter optimization is achieved through a Bayesian optimization algorithm, establishing a parameter space Θ={θ1,θ2,…,θ} for key parameters of different repair strategies. k Define the objective function: ； Among them Quality after and Quality before These are the data quality scores before and after the repair, respectively, and Stability(θ) is the repair stability index. A Gaussian process is used to establish a mapping relationship between parameters and the objective function. The parameter search is guided by the acquisition function to achieve automatic optimization of the repair parameters. RepairTag, an AI-generated repair report, employs a hybrid architecture that combines a large language model with a domain knowledge base to automate document generation and knowledge summarization of the repair process.

10. A measurement data intelligent governance system for a converged architecture, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the measurement data intelligent governance method for a converged architecture as described in any one of claims 1-9.

Images