Power distribution system fault diagnosis method and system based on multi-dimensional data coordination analysis

Abstract

The present application relates to the technical field of power distribution system monitoring, in particular to a power distribution system fault diagnosis method based on multi-dimensional data coordination analysis, the present application deploys multi-dimensional data collection of power quality, temperature, insulation, fire risk, lightning protection and grounding resistance, uses dynamic confidence interval and cross-module correlation entropy to realize early warning of faults, combines graph neural network and fault feature database to realize accurate positioning and type discrimination, and evaluates equipment life based on cumulative damage degree model, thereby solving the problems of data island, early warning lag, rough positioning and one-sided evaluation in the prior art, which can early warn potential faults, effectively improve discrimination accuracy, and significantly improve the intelligent operation and maintenance level and power supply reliability of the power distribution system.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power distribution systems, specifically to a fault diagnosis method and system for power distribution systems based on multi-dimensional data coordination analysis. Background Technology

[0002] Currently, with the continuous advancement of smart grid construction, the requirements for the safety and reliability of power distribution systems are increasing. Traditional power distribution system monitoring methods mainly suffer from the following problems: 1. Single monitoring dimension: Existing technologies usually only focus on monitoring a single parameter, such as simple temperature monitoring or power quality monitoring. They operate independently and lack the ability to coordinate and analyze multi-dimensional data, making it difficult to detect cross-module related fault symptoms. 2. Insufficient early warning capability: Traditional threshold alarm modes cannot identify gradual trends during the fault latency period, resulting in untimely warnings. For example, in the early stages of cable insulation deterioration, there are slight power quality distortions and localized temperature rises, which are difficult for single-system monitoring to capture. 3. Difficulty in fault location: After a fault occurs, manual inspection is required, and there is a lack of segment location capability based on multi-source information fusion. Especially in complex branch lines, it is difficult to quickly isolate the fault point, which affects the efficiency of emergency repair. 4. Lack of life assessment: Existing assessment methods are mostly based on a single parameter (such as temperature or insulation resistance), without comprehensively considering the coupled effects of multiple factors such as operating environment, load impact, and harmonic accumulation effect, resulting in low confidence of assessment results; 5. Lack of intelligent lightning protection and grounding monitoring: Traditional lightning protection systems are only passive protection and cannot monitor the dynamic changes of lightning current parameters and grounding resistance in real time, making it difficult to assess the effectiveness of lightning protection and the aging status of the grounding system.

[0003] 6. Data silos between systems: Each monitoring system operates independently, and the data cannot be effectively integrated, making it difficult to achieve synergistic effects.

[0004] To address the aforementioned issues, it is necessary to break down data barriers, achieve cross-module collaborative analysis, and construct a comprehensive diagnostic technology capable of collaborative analysis of multi-dimensional monitoring data to provide early warning and precise location capabilities, thereby improving the safety and reliability of the power distribution system. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method for power distribution systems based on multi-dimensional data coordination analysis, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, one aspect of the present invention provides the following technical solution: A fault diagnosis method for power distribution systems based on multi-dimensional data coordination analysis includes the following steps: S1. Multi-dimensional data acquisition: Multi-dimensional data of the power distribution system is acquired through multiple monitoring modules. Each module operates synchronously and acquires data independently, and data synchronization is ensured through a unified time base. S2. Data preprocessing and feature extraction: Cleaning, transforming, and dimensionality-reducing the raw data collected by multiple monitoring modules to extract effective features that can be used for collaborative analysis; S3. Spatiotemporal alignment and fusion of multi-source data: Unify the time base of the effective features extracted in step S2 and perform cross-module fusion to provide a consistent data foundation for subsequent collaborative analysis; S4. Dynamic Confidence Interval Modeling and Primary Early Warning: Dynamic confidence intervals are established using the isolated forest algorithm to achieve primary early warning of single-parameter anomalies; S5. Cross-module correlation entropy analysis and collaborative early warning: Capture cross-module coupling anomalies through correlation entropy analysis to achieve collaborative early warning of multi-parameter correlation; S6. Fault type identification and section location: Based on a multi-dimensional feature database and machine learning algorithms, accurate fault classification is achieved, and based on graph neural networks and power distribution topology analysis, accurate fault location is achieved. S7. Equipment life assessment and diagnostic output: Based on the cumulative damage model and Weibull distribution, predict the remaining life of the equipment and convert it into executable operation and maintenance decision release to realize human-computer interaction; S8. Closed-loop feedback and model self-learning optimization: Continuously optimize model parameters by comparing diagnostic results with actual operation and maintenance feedback.

[0007] Preferably, step 1 specifically includes: Power quality monitoring module: Collects data including three-phase voltage, three-phase current, frequency, power factor, active / reactive power, harmonic content, three-phase imbalance, and voltage sag / surge depth; Distribution cabinet temperature monitoring module: Collects data including the temperature of busbars, switches, and connectors inside the distribution cabinet; Cable insulation monitoring module: Collects data including cable insulation resistance and leakage current; Electrical fire monitoring module: Collects data including residual current, abnormal temperature rise, and arc fault; Intelligent lightning protection monitoring module: Collects data including surge arrester status, number of lightning strikes, and lightning current amplitude; Grounding resistance monitoring module: Collects data including the grounding resistance value of the power distribution system and its changing trend; All monitoring modules achieve sub-microsecond clock synchronization via IEEE 1588 PTP.

[0008] Preferably, step 2 specifically includes: S21. Data cleaning: Clean the original data by wavelet thresholding, 3σ outlier handling, and missing value completion. S22. Feature Extraction: Extract time-domain features, frequency-domain features, and time-frequency-domain features from the cleaned data; S23. Data Standardization: Use Z-score to standardize the extracted features, unifying feature data of different dimensions to the same numerical scale; S24. Feature dimensionality reduction: Principal component analysis is used to retain principal components with a cumulative contribution rate of ≥95%, and independent component analysis is used to separate non-Gaussian independent signal sources to obtain a standardized feature vector set. S25. Output: Construct a multi-dimensional fused feature vector based on the standardized feature vector set. X(t) = [X PQ (t), X Temp (t), X Ins (t), X Fire (t), X SPD (t), X GR (t) ] (1; In formula (1): X(t) represents the multidimensional standardized feature vector at time t; X PQ (t) represents the power quality characteristic subvector; X Temp (t) represents the temperature feature vector; X Ins (t) represents the characteristic sub-vector of the insulation state; X Fire (t) represents the fire risk feature subvector; X SPD (t) represents the lightning protection feature subvector; X GR (t) represents the characteristic sub-vector of grounding resistance.

[0009] Preferably, step 3 specifically includes: S31. Spatiotemporal alignment: Based on the IEEE 1588 PTP timestamp to unify the time t, the multi-dimensional standardized feature vector X(t) is time-aligned: high-frequency data is downsampled and low-frequency data is forward-padded to ensure that each feature sub-vector has a valid value at the same time t. S32. Bayesian network fusion: Using the 6-dimensional feature sub-vectors of the multi-dimensional standardized feature vector X(t) as input nodes, a 6-node Bayesian network is constructed. The conditional probability table is learned through historical fault cases, and the posterior probability of the fault is calculated. S33. Construction of the association strength matrix: Calculate the fused data vector and construct the normalized association strength matrix: Fusion data vector computation: X(t) fused = X(t) W fusion (2); In formula (2): X(t) fused W represents the fused data vector at time t; fusion The fusion weight matrix is ​​determined by the posterior probability of the fault in the Bayesian network. Correlation strength matrix calculation: A ij = MI(X i (t); X j (t)) / max(MI) (3; In formula (3): A ij X represents the (i,j)th element of the normalized correlation strength matrix, with a value range of [0,1]; i (t) and X j (t) represent the feature sub-vectors corresponding to the i-th and j-th monitoring modules in X(t), respectively; MI(X i (t);X j (t) represents X i (t) and X j The mutual information value of (t); max(MI) represents the normalization of the maximum mutual information value of all monitoring modules; MI(X i (t);X j The formula for calculating (t) is as follows: MI(X i (t);X j (t))=∫∫p(x i ,x j )·ln[p(x i ,x j ) / (p(x i )·p(x j ))]dx i dx j (4); In formula (4): p(x) i ,x j Let p(x) be the joint probability density. i ), p(x j ) represents the marginal probability density, x i and x j This represents the actual values ​​of the data collected by the i-th and j-th monitoring modules.

[0010] Preferably, step 4 specifically includes: S41. Modeling from historical normal data: For each monitoring module i, the Isolation Forest algorithm is used, with the historical normal time period data of that module as the training set, to construct a random binary tree: S42. Generate dynamic confidence intervals: For each monitoring module i, use EWMA to dynamically update the mean and standard deviation to generate dynamic confidence intervals; S43. Anomaly Score Calculation and Primary Early Warning Trigger: For each module i, based on the fused feature data X of that module... i (t) fused Calculate the anomaly score s for each monitoring module. i : s i = 2^{-E[h(X i (t) fused )] / c(n)} (5; In formula (5): s i X represents the isolated forest anomaly score of the i-th monitoring module; i (t) fused h(X) represents the fused feature data of the i-th monitoring module; n represents the total number of training samples for the i-th module; i (t) fused ) represents the path length of the fused feature data of the i-th module in the isolated tree; E[h(X i (t) fused )] represents the expected path length; c(n) represents the average path length given the number of samples n, c(n) = 2H(n-1) - 2(n-1) / n, where H(·) is the harmonic number; When s i When the value is greater than 0.6, the monitoring module will trigger a primary warning. S44. Anomaly Module Filtering: Based on the anomaly score s of each monitoring module. i Filter to meet s i >0.6 and X i (t) fused Monitoring modules that exceed the dynamic confidence interval and whose duration is greater than Δt are identified as an abnormal module index set I. abnormal = {i | Module i satisfies the abnormal condition}.

[0011] Preferably, step 5 specifically includes: S51, Cross-module association entropy calculation: S assoc =-Σ{i <j∈I abnormal}w ij ·A ij ·ln(A ij (6); In formula (6): S assoc Indicates the cross-module association entropy; w ij w represents the weighting coefficient between the i-th and j-th monitoring modules. ij =(s i +sj ) / 2; X(t) abnormal ={X i (t) fused |i∈I abnormal} (7); In formula (7): X(t) abnormal Indicates a subset of abnormal data; S52, Collaborative Early Warning Trigger: When the correlation entropy Sassoc < 0.5 and I abnorma A collaborative early warning is triggered when l≥2.

[0012] Preferably, step 6 specifically includes: S61. Fault Feature Database Construction: Based on historical operating data, establish a fault feature database containing six typical fault types, X(t). abnormal The feature vectors corresponding to each monitoring module are denoted as X. PQ X Temp X Ins X Fire X SPD X GR Each type of fault is characterized by the combined features of all monitoring modules: Insulation aging: Physical rules are manifested as X PQ The total harmonic distortion (THDi) of the current increases by more than 8%, X Temp Temperature gradient anomaly >10 K / m and X Ins insulation resistance R Ins The voltage drop is less than 100 MΩ. Poor contact: Physical rules are manifested as X PQ The imbalance mutation and X Temp The rate of temperature rise is >5 K / min; Harmonic resonance: Physical rules are manifested as X PQ The 5th / 7th harmonic amplification is greater than 3 times; Arc fault: Physical rules manifest as X Fire The residual current pulse is >300 mA and X PQ High-frequency noise is 3 times greater than background noise; Lightning damage: Physical rules are represented by X SPD The peak lightning current is >10 kA and X SPD _status Surge protector activated; Grounding failure: Physically manifested as X GR The grounding resistance increased by more than 50% compared to the initial value and X PQ Abnormal contact voltage; S62. Fault type identification: F type=argmax{k∈{1,...,6}}P k (X(t) abnormal (8); In formula (8): F type This represents the fault type determination result, selecting the category with the highest confidence level; X(t) abnormal P represents a subset of abnormal data output by S5. k P represents the confidence score for class k faults, output by the SVM classifier. k =SVMk(X(t) abnormal ), k=1~6 represent the following respectively: insulation aging, poor contact, harmonic resonance, arc fault, lightning damage, and grounding failure; SVMk represents the support vector machine classification model for the k-th type of fault. S63. Result Verification: When max(P) k When the value is less than 0.85, the multi-model voting mechanism is activated, and the physical rules in S61 are used for verification. The final fault type is then output. S64. Graph Structure Construction: The power distribution system is abstracted as an undirected graph structure G=(V,E), where nodes V = monitoring points, edges E = line segments, and edge features include line impedance Ze, length Le, and historical failure rate λe. The initial features of node V are given by X(t). abnormal The result was obtained after dimensionality reduction; S65. Construction of Graph Convolutional Network Localization Model: v =argmax{v∈V}P fault (v) (9); In formula (9): v Pfault(v) represents the location of the faulty node; Pfault(v) represents the probability of failure of node v. Pfault(v)=fGCN(Xabnormal(t),Ftype,G) (10); In formula (10): fGCN(·) represents a graph convolutional network, and the input is an anomalous data subset X(t). abnormal Fault type Ftype and undirected graph structure G; S66. Location result output: Output fault node v Fault distance d fault =Σ{e∈path(v ,root)}Le, where root represents the starting point of the power supply input terminal of the power distribution system; path(v ,root): From v The path to the root consists of all the line segments traversed; e represents a certain segment of the path; Le: the length of the line segment e.

[0013] Preferably, step 7 specifically includes: S71. Calculate the cumulative damage degree by integrating the damage factors of each monitoring module: D total =k1·H eq +k2·θ rise +k3·D PD +β env ·N lightning (11); In formula (11): D total H represents the cumulative damage level, with a value range of [0,1]. eq Represents the equivalent harmonic thermal effect; θ rise Indicates the cumulative amount of temperature exceeding the rated value; D PD Indicates cumulative damage from partial discharge; N lightning Indicates the number of lightning strikes; k1, k2, k3 represent damage weighting coefficients; β env Indicates the environmental impact correction factor; S72. Remaining life prediction: E[L rem ]=η eff ·Γ(1+1 / β w (12); In formula (12): E[L rem ] represents the expected remaining lifetime; η eff Indicates the corrected effective characteristic lifetime; β w Represents the Weibull shape parameter; Γ(·) represents the gamma function; η eff =η0·(1-D total ) γdmg (13); In formula (13): η0 represents the initial characteristic lifetime; γ dmg Indicates the damage sensitivity coefficient; S73. Lifespan Classification: When E[Lrem] < 0.3, the lifespan level is healthy; When 0.3 ≤ E[Lrem] < 0.5, the lifespan rating is "Caution". When 0.5 ≤ E[Lrem] < 0.7, the lifespan level is a warning. When 0.7 ≤ E[Lrem] < 0.9, the lifespan rating is hazardous; When E[Lrem] ≥ 0.9, the lifetime level is critical; S74. Warning Issuance: Warning information shall be issued through means including but not limited to audible and visual alarms, SMS messages, APP push notifications, telephone calls, and emergency platforms. S75. Generate diagnostic reports: including but not limited to: Basic information: time of failure, affected circuits, and list of anomaly monitoring modules; Anomaly description: Abnormal parameters and the degree of exceeding limits in the monitoring module; Correlation Analysis: Correlation Graph and Correlation Strength Matrix A of Anomaly Monitoring Module ij ; Diagnostic conclusion: Fault type F type Location result v and df ault Lifespan rating; Recommended course of action: Based on fault type F type Targeted operation and maintenance strategies; S76. Closed-loop feedback: Compare the diagnostic report with the actual inspection record, calculate the fault identification accuracy, location error, and life deviation, and update the fault feature database. S77, Operation and Maintenance Decisions: Generate an operation and maintenance strategy matrix, link it with the spare parts inventory system, and optimize power outage maintenance plans.

[0014] Preferably, step 8 specifically includes: S81. Result Verification: Maintenance personnel fill in the actual fault information and calculate the early warning lead time, location error, judgment accuracy, life deviation, false alarm rate, and missed alarm rate. S82, Model Update: Update the calculation parameters of dynamic confidence interval, cross-module correlation entropy, fault type discrimination, graph convolutional network localization model, and remaining lifetime prediction according to different preset periods; S83. Knowledge Graph Construction: Accumulate ≥100 cases in the first year, ≥500 cases after three years, forming a complete knowledge graph; S84. Performance Indicator Setting: Early warning lead time ≥7 days, positioning error <5m, discrimination accuracy ≥95%, lifetime error <20%, false alarm rate <5%, and false negative rate <1%; S85, Output: Feed the updated model calculation parameters back to S3 for data fusion optimization in the next diagnostic cycle.

[0015] Another aspect of the present invention provides a power distribution system fault diagnosis system based on multi-dimensional data coordination analysis, for implementing the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described above, including: The multi-dimensional data acquisition layer enables full-dimensional status perception of the power distribution system through various monitoring modules. Each module operates synchronously and collects data independently, and data synchronization is ensured through a unified time base. The data preprocessing and feature extraction layer cleans, standardizes, extracts features and reduces the dimensionality of the raw monitoring data output from the multi-dimensional data acquisition layer. It unifies the raw monitoring data with different units and sampling frequencies into a standardized feature vector and outputs a multi-dimensional fused feature vector. The multi-source data spatiotemporal alignment and fusion layer takes the multi-dimensional fusion feature vector output from the data preprocessing and feature extraction layer as input to achieve a unified time base and cross-module fusion for each monitoring module, and outputs a fusion data vector and a correlation strength matrix. The dynamic confidence interval modeling and primary early warning layer takes the fused data vector output by the spatiotemporal alignment and fusion layer of multi-source data as input, calculates the anomaly score independently for each sub-vector, filters out the anomaly modules, and forms an anomaly module index set. The cross-module correlation entropy analysis and collaborative early warning layer takes the abnormal module index set output by the dynamic confidence interval modeling and primary early warning layer, and the correlation strength matrix output by the multi-source data spatiotemporal alignment and fusion layer as input. It captures cross-module coupling anomalies through correlation entropy analysis, realizes collaborative early warning of multi-parameter correlation, and outputs cross-module correlation entropy and a subset of abnormal data. The fault type discrimination and section location layer takes the abnormal data subset output by the cross-module correlation entropy analysis and collaborative early warning layer as input. Based on a multi-dimensional feature database and machine learning algorithm, it realizes the accurate classification of 6 types of power distribution system faults and outputs the fault type. Based on graph neural network and power distribution topology analysis, it realizes the accurate location of fault points and outputs the fault node and fault distance. The equipment life assessment and diagnosis output layer takes the abnormal data subset output by the cross-module correlation entropy analysis and collaborative early warning layer as input, predicts the remaining life of the faulty equipment based on the cumulative damage degree model and Weibull distribution, integrates the diagnosis results and transforms them into executable operation and maintenance decision releases, and realizes human-computer interaction. The closed-loop feedback and model self-learning optimization layer takes the diagnostic results and operation and maintenance decisions from the equipment life assessment and diagnostic output layer as input to achieve continuous system evolution. By comparing the diagnostic results with actual operation and maintenance feedback, the calculation parameters of the model are continuously optimized.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the problems of data silos, delayed warnings, coarse positioning, and one-sided assessments in existing technologies by deploying multi-dimensional data collection on power quality, temperature, insulation, fire risk, lightning protection, and grounding resistance. It utilizes dynamic confidence intervals and cross-module correlation entropy to achieve early fault warnings, combines graph neural networks and fault feature databases to achieve precise location and type identification, and assesses equipment lifespan based on a cumulative damage degree model. This invention can provide early warnings of potential faults, effectively improve the accuracy of identification, and significantly enhance the intelligent operation and maintenance level and power supply reliability of power distribution systems. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis in an embodiment of the present invention; Figure 2This is a framework diagram of multi-dimensional data acquisition in the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis in an embodiment of the present invention; Figure 3 This is a flowchart of data preprocessing and feature extraction in the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis in an embodiment of the present invention; Figure 4 This is a flowchart of the primary early warning, collaborative early warning, fault identification and location, and life assessment in the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis in this embodiment of the invention. Figure 5 This is a flowchart of closed-loop feedback and model self-learning optimization in the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis in the embodiments of the invention. Figure 6 This is a structural block diagram of a power distribution system fault diagnosis system based on multi-dimensional data coordination analysis in an embodiment of the invention. Detailed Implementation

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

[0019] Figure 1 This is an overall flowchart of the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis provided by the present invention. Figure 1 As shown, embodiments of the present invention provide a method for fault diagnosis of power distribution systems based on multi-dimensional data coordination analysis, including the following steps: S1. Multi-dimensional data acquisition: Multi-dimensional data of the power distribution system is acquired through multiple monitoring modules. Each module operates synchronously and acquires data independently, and data synchronization is ensured through a unified time base. S2. Data preprocessing and feature extraction: Cleaning, transforming, and dimensionality-reducing the raw data collected by multiple monitoring modules to extract effective features that can be used for collaborative analysis; S3. Spatiotemporal alignment and fusion of multi-source data: Unify the time base of the effective features extracted in step S2 and perform cross-module fusion to provide a consistent data foundation for subsequent collaborative analysis; S4. Dynamic Confidence Interval Modeling and Primary Early Warning: Dynamic confidence intervals are established using the isolated forest algorithm to achieve primary early warning of single-parameter anomalies; S5. Cross-module correlation entropy analysis and collaborative early warning: Capture cross-module coupling anomalies through correlation entropy analysis to achieve collaborative early warning of multi-parameter correlation; S6. Fault type identification and section location: Based on a multi-dimensional feature database and machine learning algorithms, accurate fault classification is achieved, and based on graph neural networks and power distribution topology analysis, accurate fault location is achieved. S7. Equipment life assessment and diagnostic output: Based on the cumulative damage model and Weibull distribution, predict the remaining life of the equipment and convert it into executable operation and maintenance decision release to realize human-computer interaction; S8. Closed-loop feedback and model self-learning optimization: Continuously optimize model parameters by comparing diagnostic results with actual operation and maintenance feedback.

[0020] This invention addresses the problems of data silos, delayed warnings, coarse positioning, and one-sided assessments in existing technologies by deploying multi-dimensional data collection on power quality, temperature, insulation, fire risk, lightning protection, and grounding resistance. It utilizes dynamic confidence intervals and cross-module correlation entropy to achieve early fault warnings, combines graph neural networks and fault feature databases to achieve precise location and type identification, and assesses equipment lifespan based on a cumulative damage degree model. This invention can provide early warnings of potential faults, effectively improve the accuracy of identification, and significantly enhance the intelligent operation and maintenance level and power supply reliability of power distribution systems.

[0021] Figure 2 This is a framework diagram of multi-dimensional data acquisition in a power distribution system fault diagnosis method based on multi-dimensional data coordination analysis, as described in one embodiment of the present invention. Figure 2 As shown, step 1 specifically includes: Power quality monitoring module: Collects data including three-phase voltage, three-phase current, frequency, power factor, active / reactive power, harmonic content, three-phase unbalance, and voltage sag / rise depth; sampling frequency is 256 points per cycle (12.8 kHz @ 50 Hz); deployment locations include the 10 kV incoming line side, the low-voltage outgoing line side of the transformer, and the outgoing line side of the UPS power supply; Distribution cabinet temperature monitoring module: Collects data including the temperature of busbars, switches, and connectors inside the distribution cabinet; collected by fluorescent fiber optic temperature sensors (-40~200℃, ±1℃), with 4~6 collection points per cabinet and a sampling period of 30 seconds; Cable insulation monitoring module: Collects data including cable insulation resistance and leakage current; monitors insulation resistance (10 kΩ~10 GΩ) online using ultra-low frequency (10 Hz) current injection method, and continuously monitors partial discharge (sensitivity 1 pC); Electrical fire monitoring module: Collects data including residual current, abnormal temperature rise, and arc fault; monitors jointly through a residual current transformer (sensitivity 100mA), an arc fault detector, and a smoke sensor, with a sampling frequency of 1 kHz; Intelligent lightning protection monitoring module: Collects data including surge arrester status, number of lightning strikes, and lightning current amplitude; detects lightning current peak value of 0~1100 kA and waveform resolution of 1μs through direct lightning strike monitoring device (Rogowski coil) and intelligent SPD; Grounding resistance monitoring module: Collects data including the grounding resistance value of the power distribution system and its changing trend; online measurement is performed using the heterodyne method (47 Hz / 53 Hz), with an accuracy of ±2% and a sampling period of 4 hours; All monitoring modules achieve sub-microsecond clock synchronization via IEEE 1588 PTP; The power distribution system achieves full-dimensional status perception through six monitoring modules. Each module operates synchronously and collects data independently, and data synchronization is ensured through a unified time base.

[0022] Figure 3 This is a flowchart illustrating the data preprocessing and feature extraction process in a power distribution system fault diagnosis method based on multi-dimensional data coordination analysis, as described in one embodiment of the present invention. Figure 3 As shown, step 2 specifically includes: S21. Data cleaning: Clean the original data by wavelet thresholding, 3σ outlier handling, and missing value completion. S22. Feature extraction: Extract time-domain features (mean / variance / peak value), frequency-domain features (harmonics / THD / spectral entropy), and time-frequency-domain features (STFT / wavelet packet) from the cleaned data. S23. Data Standardization: Use Z-score to standardize the extracted features, unifying feature data of different dimensions to the same numerical scale; S24. Feature dimensionality reduction: Principal component analysis is used to retain principal components with a cumulative contribution rate of ≥95%, and independent component analysis is used to separate non-Gaussian independent signal sources to obtain a standardized feature vector set. S25. Output: Construct a multi-dimensional fused feature vector based on the standardized feature vector set. X(t) = [X PQ (t), X Temp (t), X Ins (t), X Fire (t), X SPD (t), X GR (t) ] (1; In formula (1): X(t) represents the multidimensional standardized feature vector at time t; X PQ (t) represents the power quality characteristic subvector; X Temp (t) represents the temperature feature vector; X Ins (t) represents the characteristic sub-vector of the insulation state; X Fire(t) represents the fire risk feature subvector; X SPD (t) represents the lightning protection feature subvector; X GR (t) represents the characteristic sub-vector of grounding resistance.

[0023] By cleaning, standardizing, extracting features, and reducing the dimensionality of the raw monitoring data output by S1, the feature data of the six monitoring modules with different dimensions and sampling frequencies are unified into a standardized feature tensor X(t), providing input for subsequent data fusion.

[0024] Figure 4 This is a flowchart illustrating the primary early warning, coordinated early warning, fault identification and location, and lifespan assessment processes in a power distribution system fault diagnosis method based on multi-dimensional data coordination analysis, as described in one embodiment of the present invention. Figure 4 As shown, step 3 specifically includes: S31. Spatiotemporal alignment: Based on the IEEE 1588 PTP timestamp to unify the time t, the multi-dimensional standardized feature vector X(t) is time-aligned: high-frequency data is downsampled and low-frequency data is forward-padded to ensure that each feature sub-vector has a valid value at the same time t. S32. Bayesian network fusion: Using the 6-dimensional feature sub-vectors of the multi-dimensional standardized feature vector X(t) as input nodes, a 6-node Bayesian network is constructed. The conditional probability table (CPT) is learned through historical failure cases to calculate the posterior probability of the failure. S33. Construction of the association strength matrix: Calculate the fused data vector and construct the normalized association strength matrix: Fusion data vector computation: X(t) fused = X(t) W fusion (2); In formula (2): X(t) fused W represents the fused data vector at time t; fusion The fusion weight matrix is ​​determined by the posterior probability of the fault in the Bayesian network. Correlation strength matrix calculation: A ij = MI(X i (t); X j (t)) / max(MI) (3; In formula (3): A ij X represents the (i,j)th element of the normalized correlation strength matrix, with a value range of [0,1]; i (t) and X j (t) represent the feature sub-vectors corresponding to the i-th and j-th monitoring modules in X(t), respectively; MI(X i(t);X j (t) represents X i (t) and X j The mutual information value of (t); max(MI) represents the normalization of the maximum mutual information value of all monitoring modules; MI(X i (t);X j The formula for calculating (t) is as follows: MI(X i (t);X j (t))=∫∫p(x i ,x j )·ln[p(x i ,x j ) / (p(x i )·p(x j ))]dx i dx j (4); In formula (4): p(x) i ,x j Let p(x) be the joint probability density. i ), p(x j ) represents the marginal probability density, x i and x j This represents the actual values ​​of the data collected by the i-th and j-th monitoring modules.

[0025] Using the multi-dimensional standardized feature vector X(t) output by S2 as input, a unified time base and cross-module fusion of data from six monitoring modules are achieved, and the fused data vector X(t) is output. fused and correlation strength matrix A ij This provides a data foundation for subsequent anomaly detection and correlation analysis.

[0026] In one embodiment of the present invention, such as Figure 4 As shown, step 4 specifically includes: S41. Modeling based on historical normal data: For each monitoring module i, the isolated forest algorithm is used to construct a random binary tree using the historical normal time period data of the module as the training set; where i∈{PQ, Temp, Ins, Fire, SPD, GR}, which correspond to the 6 monitoring modules of the present invention respectively; S42. Generate dynamic confidence intervals: For each monitoring module i, use EWMA (exponentially weighted moving average, α=0.05) to dynamically update the mean and standard deviation, and generate dynamic confidence intervals. S43. Anomaly Score Calculation and Primary Early Warning Trigger: For each module i, based on the fused feature data X of that module... i (t) fusedCalculate the anomaly score s for each monitoring module. i : s i = 2^{-E[h(X i (t) fused )] / c(n)} (5; In formula (5): s i X represents the isolated forest anomaly score of the i-th monitoring module; i (t) fused h(X) represents the fused feature data of the i-th monitoring module; n represents the total number of training samples for the i-th module; i (t) fused ) represents the path length of the fused feature data of the i-th module in the isolated tree; E[h(X i (t) fused )] represents the expected path length; c(n) represents the average path length given the number of samples n, c(n) = 2H(n-1) - 2(n-1) / n, where H(·) is the harmonic number; When s i When the value is greater than 0.6, the monitoring module will trigger a primary warning. S44. Anomaly Module Filtering: Based on the anomaly score s of each monitoring module. i Filter to meet s i >0.6 and X i (t) fused Monitoring modules that exceed the dynamic confidence interval and whose duration is greater than Δt are identified as an abnormal module index set I. abnormal = {i | Module i satisfies the abnormal condition}.

[0027] The fused data tensor X(t) output by S3 fused As input, X(t) fused The six monitoring modules are decomposed into six sub-vectors. Anomaly scores are calculated independently for each sub-vector to identify anomalous modules, forming an anomalous module index set I. abnormal This provides information on abnormal modules for subsequent correlation analysis.

[0028] In one embodiment of the present invention, such as Figure 4 As shown, step 5 specifically includes: S51, Cross-module association entropy calculation: S assoc =-Σ{i <j∈I abnormal}w ij ·A ij ·ln(A ij (6); In formula (6): S assoc Indicates the cross-module association entropy; w ijw represents the weighting coefficient between the i-th and j-th monitoring modules. ij =(s i +s j ) / 2; X(t) abnormal ={X i (t) fused |i∈I abnormal} (7); In formula (7): X(t) abnormal Indicates a subset of abnormal data; S52, Collaborative Early Warning Trigger: When the correlation entropy Sassoc < 0.5 and I abnorma A collaborative early warning is triggered when l≥2.

[0029] The set of exception module indexes I output by S4 abnormal The correlation strength matrix A output by S3 ij Using the input parameter as input, cross-module coupling anomalies are captured through correlation entropy analysis, enabling collaborative early warning of multi-parameter correlations, and outputting the correlation entropy S. assoc and the anomalous data subset X(t) abnormal This provides evidence of related anomalies for subsequent fault type identification.

[0030] In one embodiment of the present invention, such as Figure 4 As shown, step 6 specifically includes: S61. Fault Feature Database Construction: Based on historical operating data, establish a fault feature database containing six typical fault types, X(t). abnormal The feature vectors corresponding to each monitoring module are denoted as X. PQ X Temp X Ins X Fire X SPD X GR Each type of fault is characterized by the combined features of all monitoring modules: Insulation aging: Physical rules are manifested as X PQ The total harmonic distortion (THDi) of the current increases by more than 8%, X Temp Temperature gradient anomaly >10 K / m and X Ins insulation resistance R Ins The voltage drop is less than 100 MΩ. Poor contact: Physical rules are manifested as X PQ The imbalance mutation and X Temp The rate of temperature rise is >5 K / min; Harmonic resonance: Physical rules are manifested as X PQ The 5th / 7th harmonic amplification is greater than 3 times; Arc fault: Physical rules manifest as XFire The residual current pulse is >300 mA and X PQ High-frequency noise is 3 times greater than background noise; Lightning damage: Physical rules are represented by X SPD The peak lightning current is >10 kA and X SPD _status Surge protector activated; Grounding failure: Physically manifested as X GR The grounding resistance increased by more than 50% compared to the initial value and X PQ Abnormal contact voltage; S62. Fault type identification: F type =argmax{k∈{1,...,6}}P k (X(t) abnormal (8); In formula (8): F type This represents the fault type determination result, selecting the category with the highest confidence level; X(t) abnormal P represents a subset of abnormal data output by S5. k P represents the confidence score for class k faults, output by the SVM classifier. k =SVMk(X(t) abnormal ), k=1~6 represent the following respectively: insulation aging, poor contact, harmonic resonance, arc fault, lightning damage, and grounding failure; SVMk represents the support vector machine classification model for the k-th type of fault. S63. Result Verification: When max(P) k When the confidence level is less than 0.85 (i.e., the highest confidence level is below 0.85), a multi-model voting mechanism is activated, and the physical rules in S61 are used for verification. The final fault type is then output. For example, insulation aging must meet X... PQ The total harmonic distortion (THDi) of the current increases by more than 8%, X Temp Temperature gradient anomaly >10 K / m and X Ins insulation resistance R Ins The voltage drop is less than 100 MΩ. S64. Graph Structure Construction: The power distribution system is abstracted as an undirected graph structure G=(V,E), where nodes V = monitoring points, edges E = line segments, and edge features include line impedance Ze, length Le, and historical failure rate λe. The initial features of node V are given by X(t). abnormal The result was obtained after dimensionality reduction; S65. Construction of Graph Convolutional Network Localization Model: v =argmax{v∈V}P fault (v) (9); In formula (9): v Pfault(v) represents the location of the faulty node; Pfault(v) represents the probability of failure of node v. Pfault(v)=fGCN(Xabnormal(t),Ftype,G) (10); In formula (10): fGCN(·) represents a graph convolutional network, and the input is an anomalous data subset X(t). abnormal Fault type Ftype and undirected graph structure G; S66. Location result output: Output fault node v Fault distance d fault =Σ{e∈path(v ,root)}Le, where root represents the starting point of the power supply input terminal of the power distribution system; path(v ,root): From v The path to the root consists of all the line segments traversed; e represents a certain segment of the path; Le: the length of the line segment e.

[0031] The anomalous data subset X(t) output by S5. abnormal Using a fault feature database and machine learning algorithms as input, this algorithm accurately classifies six types of power distribution system faults and outputs the fault type F. type Furthermore, based on graph neural networks and power distribution topology analysis, the system achieves precise fault location and outputs the fault node v. and fault distance d fault .

[0032] In one embodiment of the present invention, such as Figure 4 As shown, step 7 specifically includes: S71. Calculate the cumulative damage degree by integrating the damage factors of each monitoring module: D total =k1·H eq +k2·θ rise +k3·D PD +β env ·N lightning (11); In formula (11): D total H represents the cumulative damage level, with a value range of [0,1]. eq The equivalent harmonic thermal effect is represented by X(t). abnormal ZhongX PQ Sub-vector (i.e., the fusion of feature data from the power quality monitoring module) calculation; θ rise This represents the cumulative amount of temperature exceeding the rated value, expressed as X(t). abnormal ZhongX Temp Sub-vector (i.e., the fusion feature data of the distribution cabinet temperature monitoring module) calculation; DPD This represents the cumulative damage from partial discharge, expressed by X(t). abnormal ZhongX Ins Sub-vector (i.e., the fusion of feature data from the cable insulation monitoring module) calculation; N lightning The number of lightning strikes is represented by X(t). abnormal ZhongX SPD Sub-vector extraction (i.e., fusion feature data from the intelligent lightning protection monitoring module); k1, k2, k3 represent damage weight coefficients; β env This represents the environmental impact correction factor, derived from X(t). abnormal ZhongX GR The sub-vector (i.e., the fused feature data of the grounding resistance monitoring module) is determined; it should be noted that electrical fire is a fault result and early warning indicator, not a direct cause of equipment life loss, therefore formula (11) does not include the electrical fire monitoring module X. Fire Fusion of feature data; S72. Remaining life prediction: E[L rem ]=η eff ·Γ(1+1 / β w (12); In formula (12): E[L rem ] represents the expected remaining lifetime; η eff Indicates the corrected effective characteristic lifetime; β w Γ(·) represents the Weibull shape parameter, determined by statistical analysis of historical fault data; Γ(·) represents the gamma function. η eff =η0·(1-D total ) γdmg (13); In formula (13): η0 represents the initial characteristic life, which is determined by the equipment's factory parameters; γ dmg The damage sensitivity coefficient is determined by historical data. S73. Lifespan Classification: When E[Lrem] < 0.3, the lifespan level is healthy; When 0.3 ≤ E[Lrem] < 0.5, the lifespan rating is "Caution". When 0.5 ≤ E[Lrem] < 0.7, the lifespan level is a warning. When 0.7 ≤ E[Lrem] < 0.9, the lifespan rating is hazardous; When E[Lrem] ≥ 0.9, the lifetime level is critical; S74. Warning Issuance: Warning information shall be issued through means including but not limited to audible and visual alarms, SMS messages, APP push notifications, telephone calls, and emergency platforms; specifically as follows: Yellow alert (basic alert, single monitoring module malfunction): The alert is issued via audible and visual alarms; Orange Alert (Collaborative Alert, Multiple Monitoring Modules Linked to Abnormalities): Alerts are issued via SMS and APP push notifications; Red Alert (Emergency Warning): The warning is issued via telephone and emergency platform; S75. Generate diagnostic reports: including but not limited to: Basic information: time of failure, affected circuits, and list of anomaly monitoring modules; Anomaly description: Abnormal parameters and the degree of exceeding limits in the monitoring module; Correlation Analysis: Correlation Graph and Correlation Strength Matrix A of Anomaly Monitoring Module ij ; Diagnostic conclusion: Fault type F type Location result v and df ault Lifespan rating; Recommended course of action: Based on fault type F type Targeted operation and maintenance strategies; S76. Closed-loop feedback: Compare the diagnostic report with the actual inspection record, calculate the fault identification accuracy, location error, and life deviation, and update the fault feature database. S77, Operation and Maintenance Decisions: Generate an operation and maintenance strategy matrix, link it with the spare parts inventory system, and optimize power outage maintenance plans.

[0033] The anomalous data subset X(t) output by S5. abnormal Using the cumulative damage model and Weibull distribution as input, the remaining lifespan of the faulty equipment is predicted. The diagnostic results of S4 to S7 are integrated, and the multi-dimensional analysis conclusions are transformed into actionable operation and maintenance information to achieve human-computer interaction.

[0034] Figure 5 This is a flowchart illustrating closed-loop feedback and model self-learning optimization in a power distribution system fault diagnosis method based on multi-dimensional data coordination analysis, as described in one embodiment of the invention. Figure 5 As shown, step 8 specifically includes: S81. Result Verification: Maintenance personnel fill in the actual fault information and calculate the early warning lead time, location error, judgment accuracy, life deviation, false alarm rate, and missed alarm rate. S82, Model Update: Update the calculation parameters of dynamic confidence interval, cross-module correlation entropy, fault type discrimination, graph convolutional network localization model, and remaining lifetime prediction according to different preset periods; S83. Knowledge Graph Construction: Accumulate ≥100 cases in the first year, ≥500 cases after three years, forming a complete knowledge graph; S84. Performance Indicator Setting: Early warning lead time ≥7 days, positioning error <5m, discrimination accuracy ≥95%, lifetime error <20%, false alarm rate <5%, and false negative rate <1%; S85, Output: Feed the updated model calculation parameters back to S3 for data fusion optimization in the next diagnostic cycle.

[0035] Using S7's diagnostic results and operational decisions as input, the diagnostic methods are continuously evolved. By comparing the diagnostic results with actual operational feedback, the calculation parameters of the model are continuously optimized.

[0036] Figure 6 This is a structural block diagram of a power distribution system fault diagnosis system based on multi-dimensional data coordination analysis, as described in an embodiment of the invention. The invention also provides a power distribution system fault diagnosis system based on multi-dimensional data coordination analysis, used to implement the power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described above, such as... Figure 6 As shown, it includes: The multi-dimensional data acquisition layer enables full-dimensional status perception of the power distribution system through various monitoring modules. Each module operates synchronously and collects data independently, and data synchronization is ensured through a unified time base. The data preprocessing and feature extraction layer cleans, standardizes, extracts features and reduces the dimensionality of the raw monitoring data output from the multi-dimensional data acquisition layer. It unifies the raw monitoring data with different units and sampling frequencies into a standardized feature vector and outputs a multi-dimensional fused feature vector. The multi-source data spatiotemporal alignment and fusion layer takes the multi-dimensional fusion feature vector output from the data preprocessing and feature extraction layer as input to achieve a unified time base and cross-module fusion for each monitoring module, and outputs a fusion data vector and a correlation strength matrix. The dynamic confidence interval modeling and primary early warning layer takes the fused data vector output by the spatiotemporal alignment and fusion layer of multi-source data as input, calculates the anomaly score independently for each sub-vector, filters out the anomaly modules, and forms an anomaly module index set. The cross-module correlation entropy analysis and collaborative early warning layer takes the abnormal module index set output by the dynamic confidence interval modeling and primary early warning layer, and the correlation strength matrix output by the multi-source data spatiotemporal alignment and fusion layer as input. It captures cross-module coupling anomalies through correlation entropy analysis, realizes collaborative early warning of multi-parameter correlation, and outputs cross-module correlation entropy and a subset of abnormal data. The fault type discrimination and section location layer takes the abnormal data subset output by the cross-module correlation entropy analysis and collaborative early warning layer as input. Based on a multi-dimensional feature database and machine learning algorithm, it realizes the accurate classification of 6 types of power distribution system faults and outputs the fault type. Based on graph neural network and power distribution topology analysis, it realizes the accurate location of fault points and outputs the fault node and fault distance. The equipment life assessment and diagnosis output layer takes the abnormal data subset output by the cross-module correlation entropy analysis and collaborative early warning layer as input, predicts the remaining life of the faulty equipment based on the cumulative damage degree model and Weibull distribution, integrates the diagnosis results and transforms them into executable operation and maintenance decision releases, and realizes human-computer interaction. The closed-loop feedback and model self-learning optimization layer takes the diagnostic results and operation and maintenance decisions from the equipment life assessment and diagnostic output layer as input to achieve continuous system evolution. By comparing the diagnostic results with actual operation and maintenance feedback, the calculation parameters of the model are continuously optimized.

[0037] The power distribution system fault diagnosis system provided by this invention, based on multi-dimensional data coordination and analysis, achieves measurable lightning protection grounding through online recording of lightning current waveforms and dynamic monitoring of grounding resistance, filling the gap in lightning protection system status perception. It achieves earlier early warning through correlation entropy and dynamic confidence interval mechanisms, detecting latent anomalies 7-30 days before a fault occurs, 3-5 times earlier than traditional methods. It improves positioning accuracy by fusing topological constraints and multi-source features using graph neural networks, achieving a fault section positioning error of less than 5 meters and reducing isolation time by 70%. Based on a fault feature database, the fault type identification accuracy reaches over 95%, avoiding ineffective maintenance due to misjudgment. It comprehensively considers electrical stress, thermal stress, discharge damage, and lightning impact, achieving a life prediction error of <20%, providing a more comprehensive life assessment and supporting precise operation and maintenance decisions.

[0038] The overall technical concept of this invention is further illustrated below with specific implementation examples: Implementation Case: Application of Power Distribution System in a Project at xxx Base System Deployment: The fault diagnosis system for power distribution systems based on multi-dimensional data coordination analysis provided by this invention will be deployed in the 0.4kV power distribution system of a certain project at the xxx base. The power distribution system includes two 10kV incoming lines, seven 10kV high-voltage switchgear cabinets, two 1250kVA dry-type transformers, 23 low-voltage switchgear cabinets, one 300kVA UPS unit and its matching 15min battery pack.

[0039] Sensor configuration: Power quality monitoring modules are installed on the two 10kV incoming lines, the low-voltage outgoing lines of the two transformers, and the outgoing lines of the UPS power supply. These modules can monitor power quality in real time, record fault waveforms, and perform data analysis. The sampling frequency is 256 (per cycle). 36 fluorescent fiber optic temperature sensors (measuring range -40-200℃, accuracy ±1℃) are installed on the low-voltage distribution cabinet busbars and cable splices. Inject 10Hz ultra-low frequency current into 5 key power supply cables and monitor the insulation resistance online (range 10kΩ-10GΩ). Each circuit is equipped with an electrical fire monitoring transformer to monitor fault arcs in the electrical circuit (sensitivity 100mA). Install direct lightning strike monitoring devices on outdoor independent lightning rods, and install intelligent SPD monitoring modules (lightning current peak measurement range 1100kA, waveform resolution 1μs) in the main distribution cabinet. Install an online grounding resistance monitoring instrument (47Hz / 53Hz, measurement accuracy ±2%) on the down conductor at point 3 of the grounding grid.

[0040] Fault diagnosis example: One day, the diagnostic system detected the following abnormal characteristics in a certain circuit: The total harmonic distortion (THDi) of the current gradually increased from 5.2% to 8.7%. The temperature of the cable termination head rose by 12°C in 48 hours, reaching 78°C (ambient temperature 25°C). The insulation resistance decreased from 500MΩ to 120MΩ, and the dielectric loss factor tanδ increased from 0.003 to 0.018. The residual current exhibits intermittent 200mA pulses; Other loop parameters are normal, and the correlation entropy has decreased from the normal value of 0.85 to 0.32.

[0041] Diagnostic Process: Fault information was uploaded to the feature vector server. A graph convolutional network localization model, combined with topology analysis, determined the fault to be insulation degradation at a cable joint with partial discharge. The fault location was determined to be a cable tray section 47 meters from the distribution cabinet. The lifespan assessment model calculated a cumulative damage degree L=0.68, predicting a remaining lifespan of approximately 42 days. The system issued an orange alert, recommending a power outage for maintenance within 48 hours.

[0042] Verification results: Maintenance personnel inspected the cable connectors based on the location information and found carbonization of the insulation and obvious signs of partial discharge. After replacement, the system returned to normal. The actual fault location deviated from the location by only 2.3 meters, verifying the effectiveness of the invention.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis method for power distribution systems based on multi-dimensional data coordination analysis, characterized in that, Includes the following steps: S1. Multi-dimensional data acquisition: Multi-dimensional data of the power distribution system is acquired through multiple monitoring modules. Each module operates synchronously and acquires data independently, and data synchronization is ensured through a unified time base. S2. Data preprocessing and feature extraction: Cleaning, transforming, and dimensionality-reducing the raw data collected by multiple monitoring modules to extract effective features that can be used for collaborative analysis; S3. Spatiotemporal alignment and fusion of multi-source data: Unify the time base of the effective features extracted in step S2 and perform cross-module fusion to provide a consistent data foundation for subsequent collaborative analysis; S4. Dynamic Confidence Interval Modeling and Primary Early Warning: Dynamic confidence intervals are established using the isolated forest algorithm to achieve primary early warning of single-parameter anomalies; S5. Cross-module correlation entropy analysis and collaborative early warning: Capture cross-module coupling anomalies through correlation entropy analysis to achieve collaborative early warning of multi-parameter correlation; S6. Fault type identification and section location: Based on a multi-dimensional feature database and machine learning algorithms, accurate fault classification is achieved, and based on graph neural networks and power distribution topology analysis, accurate fault location is achieved. S7. Equipment life assessment and diagnostic output: Based on the cumulative damage model and Weibull distribution, predict the remaining life of the equipment and convert it into executable operation and maintenance decision release to realize human-computer interaction; S8. Closed-loop feedback and model self-learning optimization: Continuously optimize model parameters by comparing diagnostic results with actual operation and maintenance feedback.

2. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 1, characterized in that, Step 1 specifically includes: Power quality monitoring module: Collects data including three-phase voltage, three-phase current, frequency, power factor, active / reactive power, harmonic content, three-phase imbalance, and voltage sag / surge depth; Distribution cabinet temperature monitoring module: Collects data including the temperature of busbars, switches, and connectors inside the distribution cabinet; Cable insulation monitoring module: Collects data including cable insulation resistance and leakage current; Electrical fire monitoring module: Collects data including residual current, abnormal temperature rise, and arc fault; Intelligent lightning protection monitoring module: Collects data including surge arrester status, number of lightning strikes, and lightning current amplitude; Grounding resistance monitoring module: Collects data including the grounding resistance value of the power distribution system and its changing trend; All monitoring modules achieve sub-microsecond clock synchronization via IEEE 1588 PTP.

3. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 2, characterized in that, Step 2 specifically includes: S21. Data cleaning: Clean the original data by wavelet thresholding, 3σ outlier handling, and missing value completion. S22. Feature Extraction: Extract time-domain features, frequency-domain features, and time-frequency-domain features from the cleaned data; S23. Data Standardization: Use Z-score to standardize the extracted features, unifying feature data of different dimensions to the same numerical scale; S24. Feature dimensionality reduction: Principal component analysis is used to retain principal components with a cumulative contribution rate of ≥95%, and independent component analysis is used to separate non-Gaussian independent signal sources to obtain a standardized feature vector set. S25. Output: Construct a multi-dimensional fused feature vector based on the standardized feature vector set. X(t)=[X PQ (t), X Temp (t), X Ins (t), X Fire (t), X SPD (t), X GR (t) ] (1); In formula (1): X(t) represents the multidimensional standardized feature vector at time t; X PQ (t) represents the power quality characteristic subvector; X Temp (t) represents the temperature feature vector; X Ins (t) represents the characteristic sub-vector of the insulation state; X Fire (t) represents the fire risk feature subvector; X SPD (t) represents the lightning protection feature subvector; X GR (t) represents the characteristic sub-vector of grounding resistance.

4. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 3, characterized in that, Step 3 specifically includes: S31. Spatiotemporal alignment: Based on the IEEE 1588 PTP timestamp to unify the time t, the multi-dimensional standardized feature vector X(t) is time-aligned: high-frequency data is downsampled and low-frequency data is forward-padded to ensure that each feature sub-vector has a valid value at the same time t. S32. Bayesian network fusion: Using the 6-dimensional feature sub-vectors of the multi-dimensional standardized feature vector X(t) as input nodes, a 6-node Bayesian network is constructed. The conditional probability table is learned through historical fault cases, and the posterior probability of the fault is calculated. S33. Construction of the association strength matrix: Calculate the fused data vector and construct the normalized association strength matrix: Fusion data vector computation: X(t) fused = X(t) W fusion (2); In formula (2): X(t) fused W represents the fused data vector at time t; fusion The fusion weight matrix is ​​determined by the posterior probability of the fault in the Bayesian network. Correlation strength matrix calculation: A ij = MI(X i (t); X j (t)) / max(MI) (3); In formula (3): A ij X represents the (i,j)th element of the normalized correlation strength matrix, with a value range of [0,1]; i (t) and X j (t) represent the feature sub-vectors corresponding to the i-th and j-th monitoring modules in X(t), respectively; MI(X i (t);X j (t) represents X i (t) and X j The mutual information value of (t); max(MI) represents the normalization of the maximum mutual information value of all monitoring modules; MI(X i (t);X j The formula for calculating (t) is as follows: MI(X i (t);X j (t))=∫∫p(x i ,x j )·ln[p(x i ,x j ) / (p(x i )·p(x j ))]dx i dx j (4); In formula (4): p(x) i ,x j Let p(x) be the joint probability density. i ), p(x j ) represents the marginal probability density, x i and x j This represents the actual values ​​of the data collected by the i-th and j-th monitoring modules.

5. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 4, characterized in that, Step 4 specifically includes: S41. Modeling from historical normal data: For each monitoring module i, the Isolation Forest algorithm is used, with the historical normal time period data of that module as the training set, to construct a random binary tree: S42. Generate dynamic confidence intervals: For each monitoring module i, use EWMA to dynamically update the mean and standard deviation to generate dynamic confidence intervals; S43. Anomaly Score Calculation and Primary Early Warning Trigger: For each module i, based on the fused feature data X of that module... i (t) fused Calculate the anomaly score s for each monitoring module. i : s i = 2^{-E[h(X i (t) fused )] / c(n)} (5); In formula (5): s i X represents the isolated forest anomaly score of the i-th monitoring module; i (t) fused h(X) represents the fused feature data of the i-th monitoring module; n represents the total number of training samples for the i-th module; i (t) fused ) represents the path length of the fused feature data of the i-th module in the isolated tree; E[h(X i (t) fused )] represents the expected path length; c(n) represents the average path length given the number of samples n, c(n) = 2H(n-1) - 2(n-1) / n, where H(·) is the harmonic number; When s i When the value is > 0.6, the monitoring module triggers a primary warning. S44. Anomaly Module Filtering: Based on the anomaly score s of each monitoring module. i Filter to meet s i > 0.6 and X i (t) fused Monitoring modules that exceed the dynamic confidence interval and whose duration is greater than Δt are identified as an abnormal module index set I. abnormal = {i | Module i satisfies the abnormal condition}.

6. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 5, characterized in that, Step 5 specifically includes: S51, Cross-module association entropy calculation: SH assoc =-Σ{i <j∈I abnormal }w ij ·AM ij ·ln(A ij ) (6) In formula (6): S assoc Indicates the cross-module association entropy; w ij w represents the weighting coefficient between the i-th and j-th monitoring modules. ij =(s i +s j ) / 2; X(t) abnormal ={X i (t) fused |i∈I abnormal } (7); In formula (7): X(t) abnormal Indicates a subset of abnormal data; S52, Collaborative Early Warning Trigger: When the correlation entropy Sassoc < 0.5 and I abnorma A collaborative early warning is triggered when l≥2.

7. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 6, characterized in that, Step 6 specifically includes: S61. Fault Feature Database Construction: Based on historical operating data, establish a fault feature database containing six typical fault types, X(t). abnormal The feature sub-vectors corresponding to each monitoring module are denoted as X. PQ X Temp X Ins X Fire X SPD X GR Each type of fault is characterized by the combined features of all monitoring modules: Insulation aging: Physical rules are manifested as X PQ The total harmonic distortion (THDi) of the current increases by more than 8%, X Temp Temperature gradient anomaly >10 K / m and X Ins insulation resistance R Ins The voltage drop is less than 100 MΩ. Poor contact: Physical rules are manifested as X PQ The imbalance mutation and X Temp The rate of temperature rise is >5 K / min; Harmonic resonance: Physical rules are manifested as X PQ The 5th / 7th harmonic amplification is greater than 3 times; Arc fault: Physical rules manifest as X Fire The residual current pulse is >300 mA and X PQ High-frequency noise is 3 times greater than background noise; Lightning damage: Physical rules are represented by X SPD The peak lightning current is >10 kA and X SPD _status Surge protector activated; Grounding failure: Physically manifested as X GR The grounding resistance increased by more than 50% compared to the initial value and X PQ Abnormal contact voltage; S62. Fault type identification: F type =argmax{k∈{1,...,6}}P k (X(t) abnormal ) (8); In formula (8): F type This represents the fault type determination result, selecting the category with the highest confidence level; X(t) abnormal P represents a subset of abnormal data output by S5. k P represents the confidence score for class k faults, output by the SVM classifier. k =SVMk(X(t) abnormal ), k=1~6 represent the following respectively: insulation aging, poor contact, harmonic resonance, arc fault, lightning damage, and grounding failure; SVMk represents the support vector machine classification model for the k-th type of fault. S63. Result Verification: When max(P) k When the value is less than 0.85, the multi-model voting mechanism is activated, and the physical rules in S61 are used for performance verification. The final fault type is then output. S64. Graph Structure Construction: The power distribution system is abstracted as an undirected graph structure G=(V,E), where nodes V = monitoring points, edges E = line segments, and edge features include line impedance Ze, length Le, and historical failure rate λe. The initial features of node V are given by X(t). abnormal The result was obtained after dimensionality reduction; S65. Construction of Graph Convolutional Network Localization Model: v =argmax{v∈V}P fault (v) (9); In formula (9): v Pfault(v) represents the location of the faulty node; Pfault(v) represents the probability of failure of node v. Pfault(v)=fGCN(Xabnormal(t),Ftype,G) (10); In formula (10): fGCN(·) represents a graph convolutional network, and the input is an anomalous data subset X(t). abnormal Fault type Ftype and undirected graph structure G; S66. Location result output: Output fault node v Fault distance d fault =Σ{e∈path(v ,root)}Le, where root represents the starting point of the power supply input terminal of the power distribution system; path(v ,root): From v The path to the root consists of all the line segments traversed; e represents a certain segment of the path; Le: the length of the line segment e.

8. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 7, characterized in that, Step 7 specifically includes: S71. Calculate the cumulative damage degree by integrating the damage factors of each monitoring module: D total =k1·H eq +k2·θ rise +k3·D PD +b env ·N lightning (11); In formula (11): D total H represents the cumulative damage level, with a value range of [0,1]. eq Represents the equivalent harmonic thermal effect; θ rise Indicates the cumulative amount of temperature exceeding the rated value; D PD Indicates cumulative damage from partial discharge; N lightning Indicates the number of lightning strikes; k1, k2, k3 represent damage weighting coefficients; β env Indicates the environmental impact correction factor; S72. Remaining life prediction: E[L rem ]=n eff ·C(1+1 / b w ) (12); In formula (12): E[L rem ] represents the expected remaining lifetime; η eff Indicates the corrected effective characteristic lifetime; β w Represents the Weibull shape parameter; Γ(·) represents the gamma function; or eff =η0·(1-D total ) γdmg (13); In formula (13): η0 represents the initial characteristic lifetime; γ dmg Indicates the damage sensitivity coefficient; S73. Lifespan Classification: When E[Lrem] < 0.3, the lifespan level is healthy; When 0.3 ≤ E[Lrem] < 0.5, the lifespan rating is "Caution". When 0.5 ≤ E[Lrem] < 0.7, the lifespan level is a warning. When 0.7 ≤ E[Lrem] < 0.9, the lifespan rating is hazardous; When E[Lrem] ≥ 0.9, the lifetime level is critical; S74. Warning Issuance: Warning information shall be issued through means including but not limited to audible and visual alarms, SMS messages, APP push notifications, telephone calls, and emergency platforms. S75. Generate diagnostic reports: including but not limited to: Basic information: time of failure, affected circuits, and list of anomaly monitoring modules; Anomaly description: Abnormal parameters and the degree of exceeding limits in the monitoring module; Correlation Analysis: Correlation Graph and Correlation Strength Matrix A of Anomaly Monitoring Module ij ; Diagnostic conclusion: Fault type F type Location result v and df ault Lifespan rating; Recommended course of action: Based on fault type F type Targeted operation and maintenance strategies; S76. Closed-loop feedback: Compare the diagnostic report with the actual inspection record, calculate the fault identification accuracy, location error, and life deviation, and update the fault feature database. S77, Operation and Maintenance Decisions: Generate an operation and maintenance strategy matrix, link it with the spare parts inventory system, and optimize power outage maintenance plans.

9. The power distribution system fault diagnosis method based on multi-dimensional data coordination analysis as described in claim 8, characterized in that, Step 8 specifically includes: S81. Result Verification: Maintenance personnel fill in the actual fault information and calculate the early warning lead time, location error, judgment accuracy, life deviation, false alarm rate, and missed alarm rate. S82, Model Update: Update the calculation parameters of dynamic confidence interval, cross-module correlation entropy, fault type discrimination, graph convolutional network localization model, and remaining lifetime prediction according to different preset periods; S83. Knowledge Graph Construction: Accumulate ≥100 cases in the first year, ≥500 cases after three years, forming a complete knowledge graph; S84. Performance Indicator Settings: Early warning lead time ≥7 days, positioning error <5m, discrimination accuracy ≥95%, lifetime error <20%, false alarm rate <5%, and false negative rate <1%; S85, Output: Feed the updated model calculation parameters back to S3 for data fusion optimization in the next diagnostic cycle.

10. A power distribution system fault diagnosis system based on multi-dimensional data coordination analysis, characterized in that, The method for fault diagnosis of power distribution systems based on multi-dimensional data coordination analysis as described in any one of claims 1-9 includes: The multi-dimensional data acquisition layer enables full-dimensional status perception of the power distribution system through various monitoring modules. Each module operates synchronously and collects data independently, and data synchronization is ensured through a unified time base. The data preprocessing and feature extraction layer cleans, standardizes, extracts features and reduces the dimensionality of the raw monitoring data output from the multi-dimensional data acquisition layer. It unifies the raw monitoring data with different units and sampling frequencies into a standardized feature vector and outputs a multi-dimensional fused feature vector. The multi-source data spatiotemporal alignment and fusion layer takes the multi-dimensional fusion feature vector output from the data preprocessing and feature extraction layer as input to achieve a unified time base and cross-module fusion for each monitoring module, and outputs a fusion data vector and a correlation strength matrix. The dynamic confidence interval modeling and primary early warning layer takes the fused data vector output by the spatiotemporal alignment and fusion layer of multi-source data as input, calculates the anomaly score independently for each sub-vector, filters out the anomaly modules, and forms an anomaly module index set. The cross-module correlation entropy analysis and collaborative early warning layer takes the abnormal module index set output by the dynamic confidence interval modeling and primary early warning layer, and the correlation strength matrix output by the multi-source data spatiotemporal alignment and fusion layer as input. It captures cross-module coupling anomalies through correlation entropy analysis, realizes collaborative early warning of multi-parameter correlation, and outputs cross-module correlation entropy and a subset of abnormal data. The fault type discrimination and section location layer takes the abnormal data subset output by the cross-module correlation entropy analysis and collaborative early warning layer as input. Based on a multi-dimensional feature database and machine learning algorithm, it realizes the accurate classification of 6 types of power distribution system faults and outputs the fault type. Based on graph neural network and power distribution topology analysis, it realizes the accurate location of fault points and outputs the fault node and fault distance. The equipment life assessment and diagnosis output layer takes the abnormal data subset output by the cross-module correlation entropy analysis and collaborative early warning layer as input, predicts the remaining life of the faulty equipment based on the cumulative damage degree model and Weibull distribution, integrates the diagnosis results and transforms them into executable operation and maintenance decision releases, and realizes human-computer interaction. The closed-loop feedback and model self-learning optimization layer takes the diagnostic results and operation and maintenance decisions from the equipment life assessment and diagnostic output layer as input to achieve continuous system evolution. By comparing the diagnostic results with actual operation and maintenance feedback, the calculation parameters of the model are continuously optimized.

Images