A multi-source heterogeneous technology information resource fusion and management system

Abstract

This invention relates to the field of information resource fusion management technology, specifically disclosing a fusion and management system for multi-source heterogeneous scientific and technological information resources. It parses equipment fault chains from unstructured text in historical operation and maintenance logs, extracts rated parameter constraints from structured tables in equipment manuals, and collects operational feature vectors from real-time sensor data streams to generate a knowledge graph containing N entity relationships. Based on the entity attribute constraint rules of the knowledge graph, a bidirectional attention mapping network is designed to calculate the semantic similarity weights between multi-source data and knowledge nodes. A weighted graph embedding vector set is generated through Hadamard product operations. This embedding vector set is then input into a pre-trained graph neural network model to locate the root cause equipment set leading to feature shifts. This process not only achieves efficient fault diagnosis and location but also provides decision support for subsequent preventative maintenance strategies, thereby improving system reliability and operational efficiency.

Description

Technical Field

[0001] This invention belongs to the field of information resource integration and management technology, and relates to a system for the integration and management of multi-source heterogeneous scientific and technological information resources. Background Technology

[0002] In modern power systems, power grids face the challenge of accessing diverse and heterogeneous data resources from various sensors, monitoring equipment, and information systems. These data are characterized by diversity and complexity. Therefore, employing graph neural networks to detect contradictions in the feature space and tracing the root causes of data conflicts through path reasoning is particularly important and necessary for management. Firstly, the operation and management of the power grid rely on accurate data analysis; any data conflict can lead to erroneous decisions, thereby affecting the safety and stability of the power grid. Analyzing the feature space of multi-source data using graph neural networks can effectively identify contradictions and inconsistencies between data, and then trace the root causes of data conflicts through path reasoning techniques. This process not only improves data quality but also provides a reliable basis for optimized power grid scheduling and fault diagnosis, ensuring the efficient operation of the power system.

[0003] However, current technologies using graph neural networks to detect feature space conflicts in heterogeneous multi-source data resources of power grids still have some significant shortcomings and drawbacks. First, the performance of graph neural networks often depends on the quality and quantity of training data, but in real-world power grid applications, data often suffers from missing data, noise, and inconsistencies. These data quality issues directly affect training effectiveness, leading to a decrease in the model's accuracy in feature space conflict detection. Second, existing technologies often rely on fixed inference rules and graph structures during path reasoning, which is poorly adaptable to the dynamically changing power grid environment. During power grid operation, system states and data flows may change frequently, and fixed inference rules may fail to reflect these changes in a timely manner, resulting in inaccurate data conflict tracing. Furthermore, the lack of sufficient contextual information during path reasoning may lead to an incomplete understanding of the root causes of data conflicts, hindering effective resolution of practical problems. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention provides a fusion and management system for multi-source heterogeneous scientific and technological information resources to solve the above-mentioned technical problems.

[0005] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows: This invention provides a fusion and management system for multi-source heterogeneous scientific and technological information resources, the system comprising: The knowledge graph construction module parses equipment fault chains from unstructured text in historical operation and maintenance logs, extracts rated parameter constraints from structured tables in equipment manuals, and collects operational feature vectors from real-time sensor data streams. Simultaneously, it uses the Hypergraph storage engine to fuse equipment topology connection relationships and fault mode causal relationships to generate a knowledge graph containing N entity relationships. Multimodal semantic fusion module: Based on entity attribute constraint rules of knowledge graph, it performs sliding window standardization on temperature sensor time series data, generates spectrum graphs by wavelet time-frequency transformation of vibration signals, and extracts current harmonic feature vectors; it designs a bidirectional attention mapping network to calculate the semantic similarity weights between multi-source data and knowledge nodes, and generates a weighted graph embedding vector set through Hadamard product operation; Feature contradiction tracing module: Input the embedded vector set into the pre-trained graph neural network model to construct a multi-dimensional feature space distribution map of device status; calculate the cosine similarity matrix between nodes of the same type of device to identify abnormal node pairs below a set threshold; perform transmission path reasoning based on the topological connection rules of the knowledge graph to locate the root cause device set that causes feature offset.

[0006] For example, the operational logic of the knowledge graph construction module includes: Step S11: Perform unstructured text parsing on historical operation and maintenance logs, extract entity relationship triples {device ID, fault type, maintenance measures} through the fault event causal relationship chain, and generate a fault chain entity relationship set; Step S12: Parse the structured tables in the equipment manual, match the rated voltage and maximum load rate parameter constraint rules based on regular expressions, and generate a set of equipment parameter constraint rules; Step S13: Collect sensor data streams in real time, extract the transformer temperature average, vibration spectrum main frequency, and current harmonic distortion rate features, and construct a feature vector set of equipment operating status; Step S14: Establish a dynamic conflict resolution mechanism. When the deviation between the manual parameters and the statistical mean of the feature vector set of the equipment operating status exceeds the set threshold, the constraint rules are dynamically updated based on the sensor data in the sliding window. Step S15: By fusing the fault chain set, the updated constraint rule set, and the state feature set through the Hypergraph storage engine, the topological connection relationship of the transformer nodes and the historical fault modes are associated to generate a knowledge graph containing N entity relationships. Each transformer entity node includes a load rate threshold, environmental temperature and humidity coupling coefficient, and a set of attributes for the number of maintenance times in the past three years.

[0007] For example, the operation logic of step S14 is as follows: Step S141: Calculate the absolute deviation between the parameters in the equipment manual and the statistical mean of the feature vector set of equipment operating status, and generate the parameter deviation vector D; Step S142: When the deviation of any dimension in D exceeds the set threshold, a real-time data acquisition command for a sliding time window of T=30min is triggered. Step S143: Perform a time-series stationarity test on the sensor data stream within window T, and select a set Q of continuous sampling points that meet the requirements of normal distribution; Step S144: Based on the statistical distribution characteristics of set Q, a second-order polynomial fitting algorithm is used to generate the parameter correction coefficient matrix C; Step S145: Inject the correction coefficient C into the equipment parameter constraint rule set to dynamically update the rated voltage allowable range and maximum load rate threshold of the corresponding equipment; Step S146: Generate the updated constraint rule set P' and bind it to the knowledge graph node with version tags.

[0008] For example, the operation logic of the multimodal semantic fusion module is as follows: Step S21: Standardize the temperature sensor time series data in a 5-minute sliding window, calculate the z-score of the data in each window, and generate a standardized temperature matrix; Step S22: Perform wavelet packet decomposition on the vibration sensor signal, extract the energy proportion characteristics of the 0 to 500 Hz frequency band, and generate a vibration frequency spectrum. Step S23: Collect current sensor readings, extract the amplitudes of the 3rd, 5th, and 7th harmonic components through fast Fourier transform, and construct the current harmonic feature vector; Step S24: Encode the attribute of the device rated parameter node in the knowledge graph into a knowledge node vector, and design a bidirectional attention network to calculate the three sets of semantic similarity weights between the temperature matrix, vibration spectrum, and harmonic vector and the knowledge node, respectively. Step S25: Perform the Hadamard product operation on the weight matrix and the multimodal feature dataset to generate a set of weighted graph embedding vectors.

[0009] For example, the operation logic of step S24 includes: Step S241: Extract the rated parameter node attributes of the equipment from the knowledge graph, including the rated voltage value, maximum load rate and allowable temperature rise threshold, and encode them into numerical vectors of the same dimension; Step S242: Normalize the standardized temperature matrix generated by the temperature sensor and calculate the root mean square value of the temperature change rate within each time window. Step S243: Convert the vibration spectrum into a grayscale pixel matrix and extract the image texture features as the vibration representation vector; Step S244: Construct a bidirectional attention network model, setting temperature change rate, vibration texture features, and current harmonic vector as the query end, and knowledge node vector as the key end; Step S245: Calculate the cosine similarity between the query end and the key end through a multi-head attention mechanism, and generate three sets of semantic similarity weights between the temperature matrix, vibration spectrum, and harmonic vector and the knowledge node, respectively.

[0010] For example, step S25 further includes: Step S251: Perform softmax normalization on the three sets of semantic similarity weights to output a semantic similarity weight matrix with uniform dimensions; load the semantic similarity weight matrix and multimodal feature dataset generated by the bidirectional attention network, wherein the multimodal feature dataset includes a standardized temperature matrix, a reference vibration time spectrum, and a reference current harmonic vector. Step S252: Verify the consistency between the column dimension of the weight matrix and the row dimension of the multimodal features. If the dimensions do not match, trigger the feature vector interpolation completion algorithm. Step S253: Perform the Hadamard product operation in the order of feature channels. The calculation method is to multiply the elements of the weight matrix with the corresponding feature values ​​one by one. Step S254: Add a weight source identifier to each product result, including three types of tags: {temperature weight, vibration weight, harmonic weight}; Step S255: Aggregate all labeled product results, reorganize them by dimension according to device number and timestamp, and generate a graph embedding vector set.

[0011] For example, the feature contradiction tracing module includes: Step S31: Load the graph embedding vector set and the pre-trained graph neural network model, wherein the embedding vector set includes equipment temperature, vibration, harmonic features and weight labels; Step S32: Spatial mapping of the embedded vector input graph neural network to generate a multidimensional feature space distribution map, where each node represents the state vector of a single device; Step S33: Calculate the cosine similarity between nodes of the same type. When the similarity is detected to be lower than the set threshold, it is marked as an abnormal node pair. Step S34: Draw the electrical connection conduction path of the abnormal node based on the device topology connection relationship in the knowledge graph; Step S35: Perform reverse gradient propagation calculation along the propagation path to locate the initial abnormal device node that caused the feature offset; Step S36: Aggregate the initial abnormal nodes and their associated devices to generate a root cause device set and label the deviation type.

[0012] For example, the sub-steps of step S34 include: Step S341: Load the device topology connection relationship data and abnormal node set in the knowledge graph, and obtain the electrical connection direction information between devices, wherein the electrical connection direction information is divided into transmission / distribution direction; Step S342: Starting from the abnormal node, perform a bidirectional breadth-first search along the electrical connection direction, traversing the path with a hop count not exceeding 6 cascaded devices; Step S343: Filter valid conduction paths according to equipment type and eliminate reactive power compensation equipment path branches; equipment types include transformers / circuit breakers / capacitors; Step S344: Perform physical connection verification on the screened conduction paths, check the impedance matching of adjacent device ports, and remove mismatched path segments; Step S345: Mark the effective conduction paths as a topology link diagram with arrows, and generate conduction path data containing {path length, device type sequence, impedance value}.

[0013] For example, the sub-steps of step S35 include: Step S351: Load the transmission path data and device feature offset values, and establish a gradient calculation index table for path nodes; Step S352: Initialize the backpropagation gradient starting from the feature offset of the device node at the end of the path; Step S353: Calculate by backtracking hop by hop along the conduction path, and update the formula as Δi=Δj×(impedance matching degree×0.8+equipment health degree×0.2); where Δj represents the gradient value of the downstream node of the current node in the conduction path, and Δi represents the gradient value of the upstream node currently being calculated. Step S354: When the gradient change rate between adjacent nodes exceeds a set threshold, mark the node as a candidate anomaly source; Step S355: Select the candidate node with the largest cumulative gradient value and confirm it as the initial abnormal device node.

[0014] As described above, the fusion and management system for multi-source heterogeneous scientific and technological information resources provided by the present invention has at least the following beneficial effects: This invention constructs a hypergraph-structured knowledge graph that integrates equipment topology connections and fault mode causal relationships by parsing equipment fault chains from historical operation and maintenance logs, extracting rated parameter constraints from equipment manuals, and collecting operational feature vectors from real-time sensor data streams. This not only enables semantic modeling of complex industrial systems but also provides a highly interpretable and logically clear framework for subsequent analysis. Compared to traditional purely data-driven methods, this knowledge-driven modeling approach significantly improves the system's robustness and reasoning capabilities when facing sparse or anomalous data. Secondly, in the multimodal semantic fusion module, heterogeneous sensor data such as temperature, vibration, and current are standardized, subjected to wavelet transform, and feature extraction. A bidirectional attention mechanism is then used to semantically map these data to entity nodes in the knowledge graph, further enhancing the depth and breadth of the system's perception of device status. This module utilizes Hadamard product operations to generate weighted graph embedding vector sets, enabling the fusion of data from different sources in a unified semantic space. This effectively addresses the semantic gap between multi-source heterogeneous data, improving the overall perception accuracy and generalization capability of the system. Finally, the feature contradiction tracing module uses a graph neural network to construct a multi-dimensional feature space distribution map of device status. It identifies abnormal device nodes by calculating cosine similarity and, combined with the topological structure of the knowledge graph, performs propagation path reasoning to accurately locate the root cause device set leading to feature shifts. This process not only achieves efficient fault diagnosis and location but also provides decision support for subsequent preventative maintenance strategies, thereby improving system reliability and operational efficiency. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0017] The following description, in conjunction with the implementation of the present invention, is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0018] Example 1 Please see Figure 1 As shown, a multi-source heterogeneous scientific and technological information resource fusion and management system is provided. The system includes a knowledge graph construction module, a multimodal semantic fusion module, and a feature contradiction tracing module. The above modules are connected by wired and / or wireless connections to realize data transmission between the modules. Knowledge graph construction module: It parses equipment fault chains from unstructured text in historical operation and maintenance logs, extracts rated parameter constraints from structured tables in equipment manuals, and collects operational feature vectors from real-time sensor data streams. Simultaneously, it uses the Hypergraph storage engine to fuse equipment topology connection relationships and fault mode causal relationships to generate a knowledge graph containing N entity relationships.

[0019] The operational logic of the knowledge graph construction module includes: Step S11: Perform unstructured text parsing on historical operation and maintenance logs, extract entity relationship triples {device ID, fault type, maintenance measures} through the fault event causal relationship chain, and generate a fault chain entity relationship set; Step S12: Parse the structured tables in the equipment manual, match the rated voltage and maximum load rate parameter constraint rules based on regular expressions, and generate a set of equipment parameter constraint rules; Step S13: Collect sensor data streams in real time, extract the transformer temperature average, vibration spectrum main frequency, and current harmonic distortion rate features, and construct a feature vector set of equipment operating status; Step S14: Establish a dynamic conflict resolution mechanism. When the deviation between the manual parameters and the statistical mean of the feature vector set of the equipment operating status exceeds the set threshold, the constraint rules are dynamically updated based on the sensor data in the sliding window. Step S15: By fusing the fault chain set, the updated constraint rule set, and the state feature set through the Hypergraph storage engine, the topological connection relationship of the transformer nodes and the historical fault modes are associated to generate a knowledge graph containing N entity relationships. Each transformer entity node includes a load rate threshold, environmental temperature and humidity coupling coefficient, and a set of attributes for the number of maintenance times in the past three years.

[0020] This invention first performs deep text mining on historical maintenance logs, using natural language processing to identify equipment identifiers, abnormal phenomena, and handling measures in fault events. Through event time correlation analysis, a fault propagation chain is constructed, automatically generating a triplet set containing equipment, fault, and maintenance actions. Each relationship is weighted by fault occurrence frequency and repair time. For structured data from equipment manuals, rated parameters are extracted using optical character recognition and table parsing techniques. Regular matching rules are designed to capture voltage ranges and load rate thresholds. An aging compensation coefficient is introduced to dynamically adjust thresholds based on the equipment's operational years, forming a parameter constraint rule library with version identifiers. In the real-time data processing stage, a streaming computing framework is deployed. Sliding window statistics are applied to raw sensor data. Temperature features are calculated using a weighted average within the window (with higher weight for recent data). The dominant vibration frequency is determined using the spectral energy peak detection method, and the harmonic distortion rate is calculated by the root ratio of the square sum of the fundamental and third harmonic amplitudes, constructing a standard feature vector containing spatiotemporal stamps. Dynamic conflict resolution is achieved through a dual-threshold mechanism. When the monitored feature mean continuously exceeds the parameter setting range in the manual and the fluctuation variance is higher than the warning line, a parameter self-correction program is initiated. Based on recent equipment health status data, the allowable fluctuation range is recalculated using the data quantile method, and the constraint rule version in the knowledge graph is updated synchronously. The knowledge graph storage adopts a hypergraph relationship model, which maps the rated parameters, real-time features, and historical fault records of equipment entity nodes in three dimensions. The load rate threshold is calculated by superimposing the initial parameters of the equipment with the average annual aging rate. The environmental temperature and humidity coupling coefficient is derived from the regression analysis of the equipment's full life cycle data. The number of maintenance times in the past three years is based on the actual fault maintenance records statistically analyzed by the work order system and weighted by the time decay factor, ultimately forming a knowledge graph that integrates the equipment's physical characteristics, operating status, and historical knowledge.

[0021] The operation logic for step S14 is as follows: Step S141: Calculate the absolute deviation between the parameters in the equipment manual and the statistical mean of the feature vector set of equipment operating status, and generate the parameter deviation vector D; Step S142: When the deviation of any dimension in D exceeds the set threshold, a real-time data acquisition command for a sliding time window of T=30min is triggered. Step S143: Perform a time-series stationarity test on the sensor data stream within window T, and select a set Q of continuous sampling points that meet the requirements of normal distribution; Step S144: Based on the statistical distribution characteristics of set Q, a second-order polynomial fitting algorithm is used to generate the parameter correction coefficient matrix C; Step S145: Inject the correction coefficient C into the equipment parameter constraint rule set to dynamically update the rated voltage allowable range and maximum load rate threshold of the corresponding equipment; Step S146: Generate the updated constraint rule set P' and bind it to the knowledge graph node with version tags.

[0022] The embodiment in the instruction manual first calculates the deviation by comparing the rated parameters in the equipment manual with the statistical mean of the real-time operating feature vector, using a sliding window mechanism to ensure data timeliness. Specifically, for key parameters such as rated voltage and maximum load rate, the average sensor data within the most recent 5-minute window is statistically analyzed in real time, and the absolute value difference is calculated with the parameters in the manual. If the deviation of any parameter exceeds a set threshold for three consecutive cycles (e.g., voltage deviation exceeds 15%, load rate deviation exceeds 20%), a 30-minute high-density data acquisition window is triggered. The raw data collected during this window period needs to be verified for time-series stationarity, outliers caused by sudden interference are removed, and a continuous set of sampling points that conforms to the characteristics of a normal distribution is retained. Based on the filtered effective data, a parameter correction model is established using second-order polynomial regression analysis. By calculating the gradient change relationship between actual operating parameters and theoretical values, a correction coefficient matrix is ​​dynamically generated. The voltage correction value is taken as the weighted average of the 95th percentile of the window data, and the load rate threshold is dynamically adjusted by combining the equipment aging coefficient and 10% of the recent peak load. When the updated constraint rule set is associated with knowledge graph nodes, the load rate threshold is compensated in segments based on the equipment's years of operation. The environmental temperature and humidity coupling coefficient is determined through regression analysis of historical fault data. When counting maintenance frequency over the past three years, planned maintenance and fault maintenance are distinguished (only the latter is counted), and weighted by time decay. The updated parameter set is injected into the knowledge graph incrementally.

[0023] Multimodal semantic fusion module: Based on entity attribute constraint rules of knowledge graph, it performs sliding window standardization on temperature sensor time series data, generates spectrum graphs by wavelet time-frequency transformation of vibration signals, and extracts current harmonic feature vectors; it designs a bidirectional attention mapping network to calculate the semantic similarity weights between multi-source data and knowledge nodes, and generates a weighted graph embedding vector set through Hadamard product operation.

[0024] The operation logic of the multimodal semantic fusion module is as follows: Step S21: Standardize the temperature sensor time series data in a 5-minute sliding window, calculate the z-score of the data in each window, and generate a standardized temperature matrix; Step S22: Perform wavelet packet decomposition on the vibration sensor signal, extract the energy proportion characteristics of the 0 to 500 Hz frequency band, and generate a vibration frequency spectrum. Step S23: Collect current sensor readings, extract the amplitudes of the 3rd, 5th, and 7th harmonic components through fast Fourier transform, and construct the current harmonic feature vector; Step S24: Encode the attribute of the device rated parameter node in the knowledge graph into a knowledge node vector, and design a bidirectional attention network to calculate the three sets of semantic similarity weights between the temperature matrix, vibration spectrum, and harmonic vector and the knowledge node, respectively. Step S25: Perform the Hadamard product operation on the weight matrix and the multimodal feature dataset to generate a set of weighted graph embedding vectors.

[0025] In this embodiment of the invention, during multimodal data fusion processing, the time-series data from temperature sensors is first dynamically standardized. The mean and standard deviation of the temperature values ​​within each 5-minute time window are calculated, converting the original temperature values ​​into z-scores of a standard normal distribution. This eliminates baseline drift caused by environmental changes, forming a standardized temperature matrix arranged in time sequence. For vibration signal processing, a six-layer wavelet packet decomposition technique is used. After decomposition, the effective frequency band signal from 0 to 500 Hz is reconstructed, and the percentage of energy in each sub-band to the total energy is calculated, converting it into a 128×128 pixel grayscale image representing the spatiotemporal distribution characteristics of vibration energy. For current sensor data, based on the 50 Hz fundamental frequency, the 3rd, 5th, and 7th harmonic components are accurately separated using a Fast Fourier Transform with a Hanning window. The ratio of each harmonic amplitude to the fundamental amplitude is used to construct a three-dimensional harmonic feature vector. Equipment parameter nodes in the knowledge graph are converted into numerical matrices through vectorization encoding. Attributes such as rated voltage, load rate threshold, and environmental coefficient are mapped to different dimensions of the vector, forming a semantically labeled knowledge node matrix. The bidirectional attention network employs a three-channel parallel structure to calculate the semantic similarity weights of the temperature matrix, vibration spectrum, harmonic vector, and knowledge nodes. The temperature channel calculates the association strength using a sliding window mean comparison; the vibration channel evaluates the matching degree using the intersection-union ratio of the spectrum texture feature histograms; and the harmonic channel determines similarity based on the correlation coefficient of frequency band energy distribution. Finally, the spatial alignment of multimodal features and weights is achieved through Hadamard product operations. The weight matrix is ​​element-wise multiplied with the normalized feature matrix of the corresponding modality to generate a weighted 3D embedding vector for each time point. Temperature is weighted in the red channel, vibration in the green channel, and harmonics in the blue channel, forming a weighted spectral embedding vector set.

[0026] The operation logic of step S24 includes: Step S241: Extract the rated parameter node attributes of the equipment from the knowledge graph, including the rated voltage value, maximum load rate and allowable temperature rise threshold, and encode them into numerical vectors of the same dimension; Step S242: Normalize the standardized temperature matrix generated by the temperature sensor and calculate the root mean square value of the temperature change rate within each time window. Step S243: Convert the vibration spectrum into a grayscale pixel matrix and extract the image texture features as the vibration representation vector; Step S244: Construct a bidirectional attention network model, setting temperature change rate, vibration texture features, and current harmonic vector as the query end, and knowledge node vector as the key end; Step S245: Calculate the cosine similarity between the query end and the key end through a multi-head attention mechanism, and generate three sets of semantic similarity weights between the temperature matrix, vibration spectrum, and harmonic vector and the knowledge node, respectively.

[0027] In the semantic weight calculation process of this invention, the rated parameter nodes of the equipment are first extracted from the knowledge graph, including key attributes such as rated voltage, load rate threshold, and allowable temperature rise range. These parameters are scaled proportionally to the [0,1] interval through linear normalization and then concatenated into a fixed-dimensional numerical vector. For temperature sensor data, based on a pre-standardized temperature matrix, the temperature change gradient between adjacent sampling points is calculated with a 5-minute time window. The square root of the average of the gradient values ​​within the sliding window is used to quantify the severity of temperature changes, forming a time-stamped dynamic temperature feature vector. After grayscale conversion of the vibration spectrum, the image texture features are extracted using a local binary mode algorithm. The edge direction histogram within each 3×3 pixel region is statistically analyzed to generate a vibration feature vector representing the mechanical state of the equipment. The bidirectional attention network adopts a three-way parallel structure, using the temperature change rate, vibration texture features, and harmonic vector as independent query input streams, and the knowledge node vector as a shared key-value comparison end. Each attention head obtains the association weights of the three dimensions of temperature-knowledge, vibration-knowledge, and harmonic-knowledge by calculating the cosine similarity between the query vector and the key-value vector in different projection spaces. In practice, the temperature channel focuses on comparing the matching degree between the current dynamic changes and the rated temperature rise threshold, the vibration channel focuses on analyzing the similarity between the spectral texture and historical fault modes, and the harmonic channel evaluates the correlation between distortion characteristics and load rate constraints. Finally, three sets of semantic similarity weight values ​​are output.

[0028] Step S25 also includes: Step S251: Perform softmax normalization on the three sets of semantic similarity weights to output a semantic similarity weight matrix with uniform dimensions; load the semantic similarity weight matrix and multimodal feature dataset generated by the bidirectional attention network, wherein the multimodal feature dataset includes a standardized temperature matrix, a reference vibration time spectrum, and a reference current harmonic vector. Step S252: Verify the consistency between the column dimension of the weight matrix and the row dimension of the multimodal features. If the dimensions do not match, trigger the feature vector interpolation completion algorithm. Step S253: Perform the Hadamard product operation in the order of feature channels. The calculation method is to multiply the elements of the weight matrix with the corresponding feature values ​​one by one. Step S254: Add a weight source identifier to each product result, including three types of tags: {temperature weight, vibration weight, harmonic weight}; Step S255: Aggregate all labeled product results, reorganize them by dimension according to device number and timestamp, and generate a graph embedding vector set.

[0029] In the feature fusion stage of this invention, the three sets of semantic similarity weights output by the bidirectional attention network are first subjected to softmax normalization, compressing the weight values ​​of the three dimensions of temperature, vibration, and harmonics to the [0,1] interval while keeping the sum of 1, forming a semantic weight probability matrix. Simultaneously, a pre-processed standardized temperature matrix (obtained through z-score calculation within a window), a reference vibration time spectrum (an effective frequency band energy grayscale image reconstructed by wavelet packet decomposition), and a reference current harmonic vector (a three-dimensional vector constructed from the harmonic component amplitude ratio) are loaded. When performing spatial alignment of the weight matrix with the multimodal features, a dynamic interpolation strategy is used to address the dimensionality inconsistency problem: when a mismatch is detected between the number of rows in the temperature matrix (time series length) and the number of columns in the weights, linear interpolation is used to generate supplementary feature points at equal intervals on the time axis; when the resolution of the vibration spectrum does not match the weight dimensions, bicubic interpolation is used to adjust the image size. During the Hadamard product operation, the normalized value of each temperature data point is multiplied by its corresponding weight. Vibration spectrum pixels are weighted point-by-point, and the three components of the harmonic vector are multiplied by the three weight channels respectively, ensuring that features of different physical dimensions produce comparable fusion effects under the weighting. The product results are categorized by source type: temperature features are labeled with red spatial channel weights, vibration with green spectral channel weights, and harmonics with blue frequency domain channel weights. Finally, through spatiotemporal dimension recombination, the three-channel weight product results from the same device number and time stamp are superimposed and combined to form a graph embedding vector set with multidimensional semantic labels.

[0030] Feature contradiction tracing module: Input the embedded vector set into the pre-trained graph neural network model to construct a multi-dimensional feature space distribution map of device status; calculate the cosine similarity matrix between nodes of the same type of device to identify abnormal node pairs below a set threshold; perform transmission path reasoning based on the topological connection rules of the knowledge graph to locate the root cause device set that causes feature offset.

[0031] The feature contradiction tracing module includes: Step S31: Load the graph embedding vector set and the pre-trained graph neural network model, wherein the embedding vector set includes equipment temperature, vibration, harmonic features and weight labels; Step S32: Spatial mapping of the embedded vector input graph neural network to generate a multidimensional feature space distribution map, where each node represents the state vector of a single device; Step S33: Calculate the cosine similarity between nodes of the same type. When the similarity is detected to be lower than the set threshold, it is marked as an abnormal node pair. Step S34: Draw the electrical connection conduction path of the abnormal node based on the device topology connection relationship in the knowledge graph; Step S35: Perform reverse gradient propagation calculation along the propagation path to locate the initial abnormal device node that caused the feature offset; Step S36: Aggregate the initial abnormal nodes and their associated devices to generate a root cause device set and label the deviation type.

[0032] In the process of tracing the source of feature contradictions, this invention first loads a graph embedding vector set containing equipment temperature, vibration, harmonic features, and their weight labels, and then calls a pre-trained graph neural network model. Through hierarchical spatial mapping of the graph neural network, multi-dimensional feature vectors are projected onto a low-dimensional space to generate a feature distribution map representing the equipment's operating state. Each node corresponds to a unique device, and the edges between nodes represent electrical connections between devices. When evaluating the state similarity of nodes of the same type, the cosine similarity between node vectors is calculated (based on the cosine value of the vector angle to measure similarity). When the similarity is lower than a set threshold, it is marked as an abnormal node pair, and its spatial coordinates and equipment attributes are extracted. Based on the topological connection information in the knowledge graph, a bidirectional breadth-first search is performed starting from the abnormal node to draw a complete conduction path including related equipment such as switchgear, lines, and transformers. The path search depth is limited to 6 cascaded devices. When performing reverse gradient propagation analysis along the conduction path, starting from the terminal abnormal node, the gradient decay value is calculated based on impedance matching degree and equipment health status, and the process is backtracked to the initial abnormal node level by level. Finally, the initial abnormal node and its associated equipment on the same power supply bus are aggregated, and the deviation type of the root cause equipment set is marked according to the frequency of occurrence of the fault type, based on the fault modes in the historical maintenance records (such as overload, insulation aging, and harmonic exceedance).

[0033] The sub-steps of step S34 include: Step S341: Load the device topology connection relationship data and abnormal node set in the knowledge graph, and obtain the electrical connection direction information between devices, wherein the electrical connection direction information is divided into transmission / distribution direction; Step S342: Starting from the abnormal node, perform a bidirectional breadth-first search along the electrical connection direction, traversing the path with a hop count not exceeding 6 cascaded devices; Step S343: Filter valid conduction paths according to equipment type and eliminate reactive power compensation equipment path branches; equipment types include transformers / circuit breakers / capacitors; Step S344: Perform physical connection verification on the screened conduction paths, check the impedance matching of adjacent device ports, and remove mismatched path segments; Step S345: Mark the effective conduction paths as a topology link diagram with arrows, and generate conduction path data containing {path length, device type sequence, impedance value}.

[0034] In the transmission path drawing process of this invention, the device topology connection relationships and the set of marked abnormal nodes in the knowledge graph are first loaded. The electrical connection direction between devices is analyzed (transmission direction refers to the direction from the upper-level substation to the current-level busbar, and distribution direction refers to the direction from the current-level busbar to the user end), and an adjacency matrix with directional attributes is established. Taking the abnormal node as the search starting point, a bidirectional breadth-first search is performed simultaneously along both the transmission and distribution directions. The traversal path depth is limited to 6 cascaded devices, such as abnormal transformer → switchgear → line → lower-level transformer. During this process, the device type sequence of the path nodes is recorded in real time. Based on the power system transmission characteristics, path branches containing reactive power compensation devices such as capacitors and reactors are eliminated, and effective transmission paths of key devices such as transformers, circuit breakers, and lines are retained. When verifying the physical connectivity of the initially screened paths, the impedance ratio of adjacent device ports is calculated. The calculation logic is the percentage matching degree between the output impedance of the upper-level device and the input impedance of the lower-level device. When the impedance mismatch between three consecutive nodes exceeds a set threshold, it is determined to be an invalid path segment and removed. The final effective conduction path is presented as a topology diagram with directional arrows. The total number of hops in the path is marked as the length identifier. The equipment types are arranged in the conduction order (e.g., transformer-circuit breaker-line). The total impedance of the path is taken as the geometric mean of the port impedance of each node, generating a standardized conduction path dataset.

[0035] The sub-steps of step S35 include: Step S351: Load the transmission path data and device feature offset values, and establish a gradient calculation index table for path nodes; Step S352: Initialize the backpropagation gradient starting from the feature offset of the device node at the end of the path; Step S353: Calculate by backtracking hop by hop along the conduction path, and update the formula as Δi=Δj×(impedance matching degree×0.8+equipment health degree×0.2); where Δj represents the gradient value of the downstream node of the current node in the conduction path, and Δi represents the gradient value of the upstream node currently being calculated. An example of the calculation process is as follows: In the conduction path A→B→C→D: The calculation starts from the end D and proceeds in reverse (the initial gradient of D is Δ=1.0). Calculate the gradient at node C: ΔC = ΔD × (impedance matching degree from C to D × 0.8 + health degree of C × 0.2); Calculate the gradient of node B: ΔB = ΔC × (impedance matching degree from B to C × 0.8 + health degree of B × 0.2); At this point: when calculating node B, Δj = ΔC (gradient of downstream node C); when calculating node B, Δi = ΔB (gradient of node B currently being calculated).

[0036] Step S354: When the gradient change rate between adjacent nodes exceeds a set threshold, mark the node as a candidate anomaly source; Step S355: Select the candidate node with the largest cumulative gradient value and confirm it as the initial abnormal device node.

[0037] In the anomaly source localization process of this invention, firstly, conduction path data containing electrical connection relationships and feature offset values ​​of equipment is loaded, and a gradient calculation table is constructed with equipment nodes as indices and impedance matching degree and health degree as attributes. Starting from the equipment node at the end of the conduction path, its feature offset value is normalized to an initial gradient value, set to 1.0, and the calculation is performed step by step back along the reverse direction of the electrical connection. The gradient value of each upstream node is determined by multiplying the gradient value of its direct downstream node by a comprehensive weighting coefficient, where the weighting coefficient consists of the impedance matching degree between adjacent devices and the health degree score of the device itself. The impedance matching degree is calculated by the ratio of the output impedance of the upstream device to the input impedance of the downstream device, and the health degree score is derived from a comprehensive evaluation of the equipment's years of operation, number of maintenance operations, and real-time monitoring data. When the gradient value increase between adjacent nodes exceeds a set threshold, the node is marked as a candidate anomaly source. After traversing the complete conduction path, the cumulative gradient values ​​of each candidate node are sorted. The cumulative value is the weighted sum of the gradient values ​​of all downstream paths of that node. The node with the largest cumulative gradient value is selected as the initial anomaly equipment node, and abrupt changes exceeding the threshold are recorded as auxiliary localization criteria.

[0038] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0039] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0040] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0041] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fusion and management system for multi-source heterogeneous scientific and technological information resources, characterized in that, include: The knowledge graph construction module parses equipment fault chains from unstructured text in historical operation and maintenance logs, extracts rated parameter constraints from structured tables in equipment manuals, and collects operational feature vectors from real-time sensor data streams. Simultaneously, it uses the Hypergraph storage engine to fuse equipment topology connection relationships and fault mode causal relationships to generate a knowledge graph containing N entity relationships. Multimodal semantic fusion module: Based on entity attribute constraint rules of knowledge graph, it performs sliding window standardization on temperature sensor time series data, generates spectrum graphs by wavelet time-frequency transformation of vibration signals, and extracts current harmonic feature vectors; it designs a bidirectional attention mapping network to calculate the semantic similarity weights between multi-source data and knowledge nodes, and generates a weighted graph embedding vector set through Hadamard product operation; Feature contradiction tracing module: Input the embedded vector set into the pre-trained graph neural network model to construct a multi-dimensional feature space distribution map of the device status; Calculate the cosine similarity matrix between nodes of the same type to identify abnormal node pairs below a set threshold; Based on the topological connection rules of knowledge graphs, the propagation path reasoning is performed to locate the set of root cause devices that lead to feature shift; The feature contradiction tracing module includes: Step S31: Load the graph embedding vector set and the pre-trained graph neural network model, wherein the embedding vector set includes equipment temperature, vibration, harmonic features and weight labels; Step S32: Spatial mapping of the embedded vector input graph neural network to generate a multidimensional feature space distribution map, where each node represents the state vector of a single device; Step S33: Calculate the cosine similarity between nodes of the same type. When the similarity is detected to be lower than the set threshold, it is marked as an abnormal node pair. Step S34: Draw the electrical connection conduction path of the abnormal node based on the device topology connection relationship in the knowledge graph; Step S35: Perform reverse gradient propagation calculation along the propagation path to locate the initial abnormal device node that caused the feature offset; Step S36: Aggregate the initial abnormal nodes and their associated devices to generate a root cause device set and label the deviation type; The sub-steps of step S34 include: Step S341: Load the device topology connection relationship data and abnormal node set in the knowledge graph, and obtain the electrical connection direction information between devices, wherein the electrical connection direction information is divided into transmission / distribution direction; Step S342: Starting from the abnormal node, perform a bidirectional breadth-first search along the electrical connection direction, traversing the path with a hop count not exceeding 6 cascaded devices; Step S343: Filter valid conduction paths according to equipment type and eliminate reactive power compensation equipment path branches; equipment types include transformers / circuit breakers / capacitors; Step S344: Perform physical connection verification on the screened conduction paths, check the impedance matching of adjacent device ports, and remove mismatched path segments; Step S345: Mark the effective conduction paths as a topology link diagram with arrows, and generate conduction path data containing {path length, device type sequence, impedance value}; The sub-steps of step S35 include: Step S351: Load the transmission path data and device feature offset values, and establish a gradient calculation index table for path nodes; Step S352: Initialize the backpropagation gradient starting from the feature offset of the device node at the end of the path; Step S353: Calculate by backtracking hop by hop along the conduction path, and update the formula as Δi=Δj×(impedance matching degree×0.8+equipment health degree×0.2); where Δj represents the gradient value of the downstream node of the current node in the conduction path, and Δi represents the gradient value of the upstream node currently being calculated. Step S354: When the gradient change rate between adjacent nodes exceeds a set threshold, mark the node as a candidate anomaly source; Step S355: Select the candidate node with the largest cumulative gradient value and confirm it as the initial abnormal device node.

2. The fusion and management system for multi-source heterogeneous scientific and technological information resources according to claim 1, characterized in that, The operational logic of the knowledge graph construction module includes: Step S11: Perform unstructured text parsing on historical operation and maintenance logs, extract entity relationship triples {device ID, fault type, maintenance measures} through the fault event causal relationship chain, and generate a fault chain entity relationship set; Step S12: Parse the structured tables in the equipment manual, match the rated voltage and maximum load rate parameter constraint rules based on regular expressions, and generate a set of equipment parameter constraint rules; Step S13: Collect sensor data streams in real time, extract the transformer temperature average, vibration spectrum main frequency, and current harmonic distortion rate features, and construct a feature vector set of equipment operating status; Step S14: Establish a dynamic conflict resolution mechanism. When the deviation between the manual parameters and the statistical mean of the feature vector set of the equipment operating status exceeds the set threshold, the constraint rules are dynamically updated based on the sensor data in the sliding window. Step S15: By fusing the fault chain set, the updated constraint rule set, and the state feature set through the Hypergraph storage engine, the topological connection relationship of the transformer nodes and the historical fault modes are associated to generate a knowledge graph containing N entity relationships. Each transformer entity node includes a load rate threshold, environmental temperature and humidity coupling coefficient, and a set of attributes for the number of maintenance times in the past three years.

3. The fusion and management system for multi-source heterogeneous scientific and technological information resources according to claim 2, characterized in that, The operation logic for step S14 is as follows: Step S141: Calculate the absolute deviation between the parameters in the equipment manual and the statistical mean of the feature vector set of equipment operating status, and generate the parameter deviation vector D; Step S142: When the deviation in any dimension of D exceeds the set threshold, a real-time data acquisition command for a sliding time window of T=30min is triggered. Step S143: Perform a time-series stationarity test on the sensor data stream within window T, and select a set Q of continuous sampling points that meet the requirements of normal distribution; Step S144: Based on the statistical distribution characteristics of set Q, a second-order polynomial fitting algorithm is used to generate the parameter correction coefficient matrix C; Step S145: Inject the correction coefficient C into the equipment parameter constraint rule set to dynamically update the rated voltage allowable range and maximum load rate threshold of the corresponding equipment; Step S146: Generate the updated constraint rule set P' and bind it to the knowledge graph node with version tags.

4. The fusion and management system for multi-source heterogeneous scientific and technological information resources according to claim 1, characterized in that, The operation logic of the multimodal semantic fusion module is as follows: Step S21: Standardize the temperature sensor time series data in a 5-minute sliding window, calculate the z-score of the data in each window, and generate a standardized temperature matrix; Step S22: Perform wavelet packet decomposition on the vibration sensor signal, extract the energy proportion characteristics of the 0 to 500 Hz frequency band, and generate a vibration frequency spectrum. Step S23: Collect current sensor readings, extract the amplitudes of the 3rd, 5th, and 7th harmonic components through fast Fourier transform, and construct the current harmonic feature vector; Step S24: Encode the attribute of the device rated parameter node in the knowledge graph into a knowledge node vector, and design a bidirectional attention network to calculate the three sets of semantic similarity weights between the temperature matrix, vibration spectrum, and harmonic vector and the knowledge node, respectively. Step S25: Perform the Hadamard product operation on the weight matrix and the multimodal feature dataset to generate a set of weighted graph embedding vectors.

5. The fusion and management system for multi-source heterogeneous scientific and technological information resources according to claim 4, characterized in that, The operation logic of step S24 includes: Step S241: Extract the rated parameter node attributes of the equipment from the knowledge graph, including the rated voltage value, maximum load rate and allowable temperature rise threshold, and encode them into numerical vectors of the same dimension; Step S242: Normalize the standardized temperature matrix generated by the temperature sensor and calculate the root mean square value of the temperature change rate within each time window. Step S243: Convert the vibration spectrum into a grayscale pixel matrix and extract the image texture features as the vibration representation vector; Step S244: Construct a bidirectional attention network model, setting temperature change rate, vibration texture features, and current harmonic vector as the query end, and knowledge node vector as the key end; Step S245: Calculate the cosine similarity between the query end and the key end through a multi-head attention mechanism, and generate three sets of semantic similarity weights between the temperature matrix, vibration spectrum, and harmonic vector and the knowledge node, respectively.

6. The fusion and management system for multi-source heterogeneous scientific and technological information resources according to claim 4, characterized in that, Step S25 also includes: Step S251: Perform softmax normalization on the three sets of semantic similarity weights to output a semantic similarity weight matrix with uniform dimensions; load the semantic similarity weight matrix and multimodal feature dataset generated by the bidirectional attention network, wherein the multimodal feature dataset includes a standardized temperature matrix, a reference vibration time spectrum, and a reference current harmonic vector. Step S252: Verify the consistency between the column dimension of the weight matrix and the row dimension of the multimodal features. If the dimensions do not match, trigger the feature vector interpolation completion algorithm. Step S253: Perform the Hadamard product operation in the order of feature channels. The calculation method is to multiply the elements of the weight matrix with the corresponding feature values ​​one by one. Step S254: Add a weight source identifier to each product result, including three types of tags: {temperature weight, vibration weight, harmonic weight}; Step S255: Aggregate all labeled product results, reorganize them by dimension according to device number and timestamp, and generate a graph embedding vector set.

Images