A new energy power distribution fault diagnosis system and method based on multi-source data fusion

Abstract

The application provides a new energy power distribution fault diagnosis system and method based on multi-source data fusion, and relates to the technical field of power system fault diagnosis. The system comprises a data processing module, a fusion reasoning module, a graph structure diagnosis module and a control response module. The data processing module uniformly and structurally processes multi-source operation data to generate a feature tensor; the fusion reasoning module constructs a dynamic fusion weight based on the information entropy of each data source and outputs a fusion confidence discriminant result; the graph structure diagnosis module combines the power distribution network topology to construct a graph structure model and realizes node-level fault positioning through a graph neural network; and the control response module generates a control instruction based on the diagnosis result, collects and executes feedback to update the graph structure attributes, thereby realizing a diagnosis closed loop. The application has the advantages of high diagnosis precision, strong adaptability, timely response and the like, and is suitable for intelligent identification and dynamic processing of faults in the new energy power distribution scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system fault diagnosis, and in particular to a new energy distribution fault diagnosis system and method based on multi-source data fusion. BACKGROUND

[0002] With the continuous increase of new energy penetration, a large number of distributed photovoltaic, wind power, energy storage and other new power sources are connected to the distribution network, and the structure and operation state of the distribution system are increasingly complex. The traditional fault diagnosis method relying on single source measurement data and fixed model identification rules cannot meet the requirements of high real-time, high accuracy and high robustness for fault location in the new energy scenario.

[0003] In the prior art, a classification model based on voltage, current and other single channel data is often used to identify fault states, and some schemes introduce neural networks, decision trees and other algorithms to improve classification accuracy. However, due to the lack of cross-channel information fusion mechanism, there are still the following technical limitations: multi-source information cannot be fully fused: the data dimensions collected by various edge devices are different, and the sampling frequencies are inconsistent. The traditional feature-level splicing method cannot realize information alignment, which easily leads to fusion failure or redundant interference; the fusion mechanism lacks self-adaptive ability: most fusion algorithms have fixed weights or are set based on prior experience, which cannot dynamically reflect the credibility changes of each data source, especially in the data abnormal or local failure scenario, which leads to unstable fusion results; the graph structure cannot be included in the diagnosis closed loop: most of the current distribution diagnosis methods based on graph neural networks are based on static graph structure, and the one-time graph modeling is only used in the reasoning stage. Control feedback information is not included in the graph state update process, and lacks system evolution and strategy self-adjustment ability; fault response and diagnosis logic are disconnected: in most existing schemes, the diagnosis result is difficult to act on the subsequent diagnosis mechanism. The system behavior is a linear "perception-judgment-execution" chain, and a learnable and evolvable closed-loop control system is not formed.

[0004] Therefore, the current new energy distribution system urgently needs an intelligent fault diagnosis method and system structure that can fuse multi-source heterogeneous data, has information confidence adjustment ability, and can evolve graph structure and control feedback cooperatively, to improve the diagnosis intelligence level, fault tolerance ability and response efficiency in the new energy scenario. SUMMARY

[0005] In view of the problems of multi-source heterogeneous data fusion difficulty, lack of self-adaptive adjustment ability of fusion mechanism, disconnection of graph structure diagnosis and control feedback, and non-closed loop evolution of system in the existing new energy distribution system, the present application provides a new energy distribution fault diagnosis system and method based on multi-source data fusion.

[0006] To solve the above problems, the present application realizes the following technical solutions:

[0007] A new energy power distribution fault diagnosis system based on multi-source data fusion, comprising:

[0008] A data processing module is configured to receive multi-source operation data from a new energy power distribution system, and perform unified structured processing on the multi-source operation data, including normalization, standardization and multi-dimensional feature extraction, to generate a corresponding feature tensor;

[0009] A fusion inference module is configured to receive the feature tensor, calculate an information entropy value based on a fault classification probability distribution of each data source within a preset time window, and construct a dynamic fusion weight vector based on the information entropy value, the dynamic fusion weight vector being used to guide weighted fusion processing of multi-source confidence information, and output a fusion confidence discriminant result of each type of fault;

[0010] A graph structure diagnosis module is configured to receive the fusion confidence discriminant result, and construct a graph structure model in combination with a topology structure of a current power distribution network, the graph structure model including nodes representing power distribution equipment, edges or structural attribute fields of edges representing electrical connection relationships, the fusion result being mapped as an embedded feature of a node in the graph, and a node-level fault recognition and positioning being performed through a graph neural network model containing an attention mechanism, and a fault diagnosis result being output;

[0011] A control response module is configured to receive the fault diagnosis result, generate a corresponding isolation control instruction, and receive control execution feedback, and update the feedback information to a structural attribute field of a node or an edge in the graph structure model, for dynamic modification of the graph structure in a subsequent diagnosis cycle.

[0012] As a preferred scheme of the present application, the data processing module comprises:

[0013] A data receiving unit is configured to receive operation data from a plurality of edge collection terminals, including electrical quantities, state quantities and environmental parameters, and align and encapsulate the operation data according to a time stamp to construct a structured data input format;

[0014] A processing unit is configured to perform normalization processing based on Box-Cox transformation and Z-score standardization on the structured data;

[0015] A feature extraction unit is configured to extract multi-dimensional features including RMS value, kurtosis, wavelet energy density, spectral barycenter and frequency shift from each data channel within a sliding time window, and generate a feature tensor for subsequent processing.

[0016] As a preferred scheme of the present application, the fusion inference module comprises:

[0017] A feature archiving unit is configured to receive the feature tensor, and classify and archive the feature tensor according to a data source identifier to construct a fusion input set;

[0018] an uncertainty quantification unit configured to calculate an information entropy value based on a probability distribution P i ={p i1 ,p i2 ,…,p iK} of each data source in discriminating faults in a target time window, wherein the information entropy value is calculated according to the following formula:

[0019]

[0020] wherein H i is the information entropy value, i represents a data source number, k represents a fault type number, K is a total number of fault types, and p ik represents a probability of the i-th data source in discriminating the k-th fault type;

[0021] an entropy weight generation unit configured to construct a dynamic fusion weight w i according to the information entropy value, wherein the dynamic fusion weight w i is calculated according to the following formula:

[0022]

[0023] wherein N is a total number of data sources participating in fusion, j is a data source number in fusion, and H j is an information entropy value calculated for the j-th data source;

[0024] a confidence fusion unit configured to perform weighted synthesis on a confidence evidence of each data source and the dynamic fusion weight w i to output a fusion confidence result corresponding to a fault type;

[0025] a fusion feedback adjustment unit configured to adaptively update a weight generation strategy or a confidence threshold parameter in a subsequent time window based on the fusion confidence result.

[0026] As a preferred scheme of the present application, the fusion reasoning module further comprises:

[0027] a feature compression unit configured to perform dimension compression processing on a received feature tensor before performing uncertainty quantification, wherein the compression processing is based on a principal component analysis (PCA) algorithm, and after constructing a sample covariance matrix, the first several principal components are extracted, and low-order feature components with a cumulative contribution rate lower than a preset threshold are removed;

[0028] The compressed tensor output by the feature compression unit serves as an input basis for subsequent information entropy value calculation and weight generation, so as to improve feature alignment and fusion consistency among multiple data sources, and the compressed dimension can be dynamically adjusted according to a fusion feedback result.

[0029] As a preferred scheme of the present application, the graph structure diagnosis module comprises:

[0030] A graph model construction unit is configured to construct a graph structure model based on the topology structure and device connection information of the current new energy power distribution system

[0031] G=(V,E), wherein the node set V represents the electrical equipment unit, and the edge set E represents the physical connection relationship, each node and edge in the graph structure model is configured with a structure attribute field including a state label, a device type and a control attribute;

[0032] A feature mapping unit is configured to map the fusion confidence discrimination result to the corresponding node of the graph structure model according to the node device number, and splice the historical state features in the graph structure model to form an input feature vector;

[0033] A graph neural inference unit is configured to execute an attention graph neural network model based on the graph structure model and the node feature input, identify the fault and fault type label of each node;

[0034] A graph embedding output unit is configured to jointly output the intermediate state embedding result of the graph neural network and the fault discrimination result to form a structured diagnosis tensor.

[0035] As a preferred scheme of the present application, the graph neural inference unit adopts a structure attribute perception type attention mechanism, and the attention weight coefficient between nodes is calculated according to the following formula:

[0036]

[0037] Wherein: α ij represents the attention weight coefficient allocated by the i-th node to its adjacent node j; h i , h j are the input feature vectors of nodes i and j respectively; W1, W2 are trainable feature mapping matrices for embedding node features into a unified representation space; s i , s j are the structure state attribute vectors of nodes i and j respectively, including device type, current fault state and historical control record; φ(s i , s j ) is a joint embedding function between node attributes, used to reflect the correlation degree between structure states; is a trainable attention parameter vector; σ(·) is an activation mapping function; ‖ represents vector splicing operation; h k represents the input feature vector of the adjacent node , is the adjacent node set of node i.

[0038] As a preferred scheme of the present application, the attention graph neural network model further includes the following technical improvement modules:

[0039] An attribute-driven fine-tuning module is configured to dynamically fine-tune the attention weight according to the difference of the node state attribute vector in the graph structure during the attention coefficient calculation process.

[0040] A policy regulation module is configured to adjust the scale factor in the normalization process according to the degree distribution characteristics of the nodes in the current graph structure.

[0041] A connection inhibition module is configured to perform a pruning operation on the edge connection corresponding to the attention weight below the set threshold.

[0042] As a preferred scheme of the present application, the control response module comprises:

[0043] A response instruction generation unit is configured to extract the control attribute of the target node and the associated edge according to the fault diagnosis result, and generate an isolation control instruction containing a control target identifier, an operation type, a priority and an execution strategy parameter.

[0044] An instruction issuing and executing unit is configured to issue the isolation control instruction to the execution terminal of the corresponding power distribution equipment, and trigger the action of the corresponding switch device, including the opening and closing, load switching or feeder isolation control operation.

[0045] An execution state acquisition unit is configured to acquire the operation feedback information of the execution terminal in real time, and the feedback information includes the actual state of the switch, the action completion flag, the response time and the abnormal execution identifier.

[0046] A graph attribute updating unit is configured to map and update the operation feedback information to the state structure attribute field of the corresponding node or edge in the graph structure model, including the line state flag, the execution success flag and the response time delay value, for the dynamic adjustment and structure correction of the fusion strategy in the subsequent graph neural network diagnosis process.

[0047] As a preferred scheme of the present application, the control response module further comprises:

[0048] A policy adjustment unit is configured to dynamically adjust the information entropy value threshold parameter, the fusion weight generation strategy or the node state confidence threshold in the fusion reasoning module according to the state parameter, the response time delay or the execution abnormality flag contained in the control execution feedback.

[0049] The policy adjustment unit constructs a feedback information driven fusion mechanism adaptive path, so that the system can strategically optimize the fusion accuracy according to the historical control behavior in the next diagnosis cycle.

[0050] A new energy power distribution fault diagnosis method based on multi-source data fusion, which applies a new energy power distribution fault diagnosis system based on multi-source data fusion as described above, the method comprises the following steps:

[0051] Step S1: receiving multi-source operation data from a new energy power distribution system, including voltage, current, load power and environmental state parameters; structurally sampling the multi-source operation data according to time series, completing uniform format packaging, and generating a multi-source feature tensor through normalization, standardization and feature extraction;

[0052] Step S2: performing principal component analysis processing on the feature tensor, removing redundant feature components with a cumulative contribution rate lower than a preset threshold, and constructing a compressed feature tensor;

[0053] Step S3: constructing a fault classification probability distribution in a time window corresponding to each data source, calculating an information entropy value, constructing a dynamic fusion weight based on the information entropy value, and performing weighted synthesis processing of confidence information by the inference module according to the weight, and outputting a fusion confidence discriminant result;

[0054] Step S4: combining the topological structure relationship and device connection information of the new energy power distribution network to construct a graph structure model, mapping the fusion confidence result to a feature vector of a node in the graph structure, and performing node-level aggregation and reasoning through a structure attribute perception attention graph neural network model to generate a node embedding vector and a corresponding fault type label;

[0055] Step S5: generating an isolation control instruction containing a control target identifier, an operation type, a priority and a strategy parameter according to the diagnosis label and the node control attribute, and issuing the isolation control instruction to the corresponding power distribution device terminal for execution;

[0056] Step S6: collecting the terminal execution feedback information, including action completion state, response time and abnormal flag, and writing the feedback result into the structure attribute field of the corresponding node or edge in the graph structure model;

[0057] Step S7: dynamically adjusting the information entropy value threshold parameter, the fusion weight generation strategy or the confidence discriminant threshold according to the updated graph structure state, realizing the adaptive evolution of the diagnosis strategy, and forming a closed-loop intelligent fault diagnosis and control mechanism.

[0058] The application has the advantages of strong data fusion capability, high fault identification precision, good response control closed loop and the like. The system performs unified structured processing on multi-source operation data from the new energy power distribution system through a data processing module, completes normalization, standardization and multi-dimensional feature extraction, generates a feature tensor with good consistency, and effectively solves the problems of multi-source data heterogeneity and time sequence asynchronization. The fusion reasoning module uses information entropy to measure the discriminant uncertainty of each data source, dynamically constructs a fusion weight, guides the weighted fusion of multi-source confidence information, and improves the accuracy and robustness of the fusion result. The graph structure diagnosis module constructs a graph structure model based on the current power distribution network topology, maps the fusion result to node embedding features, and realizes node-level fault identification and positioning by combining an attention mechanism graph neural network, thereby enhancing the perception and diagnosis capability of the system for complex electrical connection relationships. The control response module generates isolation control instructions according to the fault diagnosis result and executes them, and simultaneously collects execution feedback information to dynamically update the structure attribute fields of the nodes or edges of the graph structure model, so that the graph structure evolves in real time with the control behavior, and a closed loop mechanism of diagnosis-control-feedback-re-diagnosis is formed, thereby significantly improving the intelligent level and adaptability of the system, and having good engineering implementability and application promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0060] Among them:

[0061] Figure 1 It is a schematic diagram of the overall structure of the system of the present application.

[0062] Figure 2 It is a flow chart of the method of the present application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the following will combine the drawings of the embodiments of the present application to clearly and completely describe the technical scheme of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those skilled in the art belong to the scope of protection of the present application.

[0064] As shown in Figure 1 An embodiment of the present application provides a new energy power distribution fault diagnosis system based on multi-source data fusion, which comprises:

[0065] (1) Data processing module

[0066] For receiving multi-source operation data from a new energy power distribution system, and performing unified structured processing on the multi-source operation data, including normalization, standardization and multi-dimensional feature extraction, to generate corresponding feature tensors;

[0067] The data processing module comprises:

[0068] A data receiving unit configured to receive operation data from a plurality of edge collection terminals, including electrical quantities, state quantities and environmental parameters, and align and encapsulate the operation data according to timestamps to construct a structured data input format;

[0069] A processing unit configured to perform normalization processing on the structured data based on Box-Cox transformation and Z-score standardization;

[0070] A feature extraction unit configured to extract multi-dimensional features including RMS value, kurtosis, wavelet energy density, spectral barycenter and frequency shift from each data channel within a sliding time window, and generate feature tensors for subsequent processing.

[0071] In a specific implementation, the feature extraction unit can further integrate an edge recognition strategy module to perform rejection or weight weakening processing on data quality abnormal channels, to avoid interference of low confidence input on subsequent fusion models. At the same time, the sliding time window length and step size parameters can be dynamically set according to actual power distribution operation characteristics, for example, adjusting the time window width according to the fault response sensitive interval, to ensure real-time performance while improving diagnostic accuracy. The generation of feature tensors supports tensor parallel encapsulation mechanism, which facilitates direct connection with the data interface of the downstream fusion module, reduces system memory load and delay pressure.

[0072] (2) Fusion inference module

[0073] For receiving the feature tensors, and calculating information entropy values based on the fault classification probability distribution of each data source within a preset time window, and constructing a dynamic fusion weight vector based on the information entropy values, the dynamic fusion weight vector is used to guide the weighted fusion processing of multi-source confidence information, and output the fusion confidence discrimination result of each type of fault;

[0074] In a preferred embodiment of the present application, the fusion inference module is used to complete the whole process from information credibility measurement to fusion discrimination output according to data feature tensors of different sources. The module mainly includes four key sub-units in a specific implementation: uncertainty quantification, weight generation, confidence fusion and feedback adjustment, each unit has a clear technical responsibility and interactive data structure, ensuring that the fusion decision has high adaptability and high robustness.

[0075] The fusion inference module comprises:

[0076] a feature archiving unit configured to receive the feature tensor and classify and archive the feature tensor according to a data source identifier, and construct a fusion input set;

[0077] an uncertainty quantification unit configured to calculate an information entropy value based on a fault discrimination probability distribution P i ={p i1 ,p i2 ,…,p iK} of each data source in a target time window, and the calculation formula is:

[0078]

[0079] wherein H i is the information entropy value, i represents a data source number, k represents a fault type number, K is the total number of fault categories, and p ik represents a discrimination probability of the i-th data source for the k-th fault type;

[0080] It should be noted that, unlike the fixed fusion weight or data confidence prior in the prior art, the information entropy is first introduced as a weight generation basis in this step, which has real-time and self-feedback capabilities within a source.

[0081] an entropy weight generation unit configured to construct a dynamic fusion weight w i based on the information entropy value, and the calculation formula is:

[0082]

[0083] wherein N is the total number of data sources participating in fusion, j is a data source number in fusion, and H j is the information entropy value calculated for the j-th data source;

[0084] It should be noted that, compared with the traditional linear average and hard rule screening strategy, the entropy-driven normalized weight has the characteristics of dynamic, interpretable, adjustable, and is suitable for various data quality states.

[0085] a confidence fusion unit configured to weight and synthesize the confidence evidence of each data source and the dynamic fusion weight w i , and output a fusion confidence result corresponding to a fault type;

[0086] a fusion feedback adjustment unit configured to adaptively update a weight generation strategy or a confidence threshold parameter in a subsequent time window based on the fusion confidence result.

[0087] In this embodiment, the fusion feedback adjustment unit is responsible for listening to the fusion output trend of multiple consecutive cycles. If it is found that the entropy value of a certain data source is long-term higher than the average value or its weight fluctuates greatly, the participation weight upper and lower limits of the corresponding channel are automatically adjusted or the "soft removal" mechanism is enabled, the data source is temporarily frozen and the historical average value is replaced.

[0088] At the same time, the entropy fluctuation trend is also used to drive the subsequent window step adjustment. For example, when the fusion entropy jitter increases, the system automatically shortens the time window, improves the time resolution, and improves the diagnostic sensitivity.

[0089] The system first realizes the closed-loop control path of information uncertainty, weight adjustment, confidence discrimination and feedback callback. It not only has static fusion capability, but also has self-evolution adaptability.

[0090] The fusion reasoning module further comprises:

[0091] The feature compression unit is configured to perform dimension compression processing on the received feature tensor before performing uncertainty quantization. The compression processing is based on a principal component analysis (PCA) algorithm. After constructing a sample covariance matrix, the first several principal components are extracted, and low-order feature components with a cumulative contribution rate lower than a preset threshold are removed.

[0092] The compressed tensor output by the feature compression unit is used as the input basis for subsequent information entropy value calculation and weight generation, so as to improve the feature alignment and fusion consistency between multiple sources of data. The compressed dimension can be dynamically adjusted according to the fusion feedback result.

[0093] In a specific implementation, in order to improve the adaptability of the fusion reasoning module to multi-scene, multi-quality heterogeneous data, the system introduces a dynamic weight adjustment window mechanism driven by information entropy, which allows temporary adjustment of the fusion contribution of the current data source according to the fault burst intensity change or data loss. Specifically, if a certain data channel appears obvious drift, distortion or mutation, its corresponding entropy value will be explicitly suppressed in weight calculation, ensuring that the fusion result is not affected by local distortion.

[0094] In addition, the entropy value normalization weight calculation unit supports an interpolation backtracking correction mechanism. Even if a certain type of data is missing or not reliable in the current time window, it can be estimated and calculated through the historical entropy weight trend of adjacent time windows, improving the system's continuous fault tolerance capability.

[0095] The fusion input interface part adopts a tensor formatting cache mode, so that the compressed principal component tensor can be directly input into the subsequent graph structure diagnosis module, avoiding redundant intermediate processing steps and reducing system response delay. This structure also facilitates the rapid migration of the system to different new energy power distribution edge node platforms, enhancing the practicality of the present application in actual industrial environments.

[0096] (3) Graph structure diagnosis module

[0097] for receiving the fusion confidence discrimination result, and constructing a graph structure model in combination with a topology structure of a current power distribution network, the graph structure model including nodes representing power distribution devices, edges or structural attribute fields of the edges representing electrical connection relationships, the fusion result being mapped as an embedded feature of a node in the graph, and node-level fault identification and positioning being performed through a graph neural network model including an attention mechanism, and a fault diagnosis result being output;

[0098] The graph structure diagnosis module includes:

[0099] a graph model construction unit configured to construct a graph structure model based on a topology structure and device connection information of a current new energy power distribution system

[0100] G=(V, E), where a node set V represents electrical device units, and an edge set E represents physical connection relationships, each node and edge in the graph structure model being configured with a structural attribute field including a state marker, a device type, and a control attribute;

[0101] a feature mapping unit configured to map the fusion confidence discrimination result to a corresponding node of the graph structure model according to a node device number, and splice the fusion confidence discrimination result with a historical state feature in the graph structure model to form an input feature vector;

[0102] a graph neural inference unit configured to execute an attention graph neural network model based on the graph structure model and the node feature input, and identify a fault and a fault type label of each node;

[0103] a graph embedding output unit configured to jointly output an intermediate state embedding result of the graph neural network and a fault discrimination result, and form a structured diagnosis tensor.

[0104] The graph neural inference unit adopts a structural attribute perception type attention mechanism, and an attention weight coefficient between nodes is calculated according to the following formula:

[0105]

[0106] wherein: α ij represents an attention weight coefficient allocated by an i-th node to its adjacent node j; h i , h j are input feature vectors of the nodes i and j, respectively; W1 and W2 are trainable feature mapping matrices for embedding the node features into a unified representation space; s i , s j are structural state attribute vectors of the nodes i and j, respectively, including a device type, a current fault state, and a historical control record; and φ(s i , s j ) is a joint embedding function between node attributes, for reflecting a correlation degree between structural states. is a trainable attention parameter vector; σ(·) is an activation mapping function; || represents a vector concatenation operation; h k denotes adjacent nodes is an input feature vector of node i, is a set of adjacent nodes of node i.

[0107] The attention graph neural network model further comprises the following technical improvement modules:

[0108] An attribute-driven fine-tuning module is configured to dynamically fine-tune the attention weight according to the difference of the node state attribute vector in the graph structure during the attention coefficient calculation process.

[0109] A strategy regulation module is configured to adjust the scale factor in the normalization process according to the degree distribution characteristics of the nodes in the current graph structure.

[0110] A connection inhibition module is configured to perform a pruning operation on the edge connection corresponding to the attention weight below the set threshold.

[0111] In a preferred embodiment, to further improve the upstream and downstream awareness linkage of fault location, the graph structure model supports dynamic evolution of the structure attribute, that is, after each diagnosis cycle, the system updates the node attribute state according to the fusion confidence result, and automatically triggers the local diffusion propagation mechanism of the corresponding graph adjacent edge attribute. The mechanism adopts a propagation coefficient adjustment method based on edge weight control, and only allows confidence information to propagate in the fusion region with strong consistency, thereby avoiding error amplification caused by invalid edges.

[0112] In addition, to support distributed deployment scenarios, the internal neural network parameters of the module support heterogeneous graph partition training, and the embedding features and weight parameters in each local subgraph can be locally converged and synchronized to the main graph model to form an efficient edge deployment architecture. The system also has a graph cache mechanism, which does not need to reconstruct the graph model every round in the slow topology change scenario, and only updates the difference of the attribute field in the graph that is dynamically updated, thereby significantly reducing the overall graph modeling overhead of the diagnosis cycle.

[0113] (4) Control response module

[0114] configured to receive the fault diagnosis result, generate corresponding isolation control instructions, receive control execution feedback, and update the feedback information to the structure attribute field of the node or edge in the graph structure model for dynamic modification of the graph structure in the subsequent diagnosis cycle.

[0115] The control response module comprises:

[0116] A response instruction generation unit is configured to extract the control attribute of the target node and the associated edge according to the fault diagnosis result, and generate isolation control instructions containing the control target identifier, operation type, priority, and execution strategy parameters.

[0117] The instruction issuing execution unit is configured to issue the isolation control instruction to an execution terminal of a corresponding power distribution device, and trigger corresponding switch device actions, including switching on or off, load switching, or feeder isolation control operation.

[0118] The execution state acquisition unit is configured to acquire operation feedback information of the execution terminal in real time, and the feedback information includes switch actual state, action completion flag, response time, and abnormal execution identifier.

[0119] The graph attribute updating unit is configured to map and update the operation feedback information to a state structure attribute field of a corresponding node or edge in the graph structure model, including line state flag, execution success flag, and response time delay value, for dynamic adjustment and structure correction of a fusion strategy in a subsequent graph neural network diagnosis process.

[0120] The control response module further includes:

[0121] The strategy adjustment unit is configured to dynamically adjust an information entropy value threshold parameter, a fusion weight generation strategy, or a node state confidence threshold value in the fusion reasoning module according to state parameters, response time delay, or execution abnormality flag contained in the control execution feedback.

[0122] The strategy adjustment unit constructs a feedback information driven fusion mechanism adaptive path, so that the system can strategically optimize the fusion accuracy according to historical control behaviors in the next diagnosis cycle.

[0123] In the embodiment of the application, the control response module is the end decision execution layer of the diagnosis chain, and its function is not only to responsively execute isolation operation, but also to continuously optimize the system fusion strategy through feedback information. In actual deployment, there is a clear data transmission and state linkage mechanism between response instruction generation, execution instruction issuing, and feedback acquisition. Each unit realizes quasi-real-time closed-loop operation through a message queue or a control bus.

[0124] After the feedback information is mapped to the graph structure attribute field, it will be embedded as part of the input features of the graph neural network in the next diagnosis cycle, and then affect the node representation learning process. The strategy adjustment unit indirectly affects the weight generation and confidence calculation path in the fusion reasoning module by modifying the fusion parameters after the feedback is triggered, so that the system fusion result can gradually adapt to environmental disturbance and control behavior deviation.

[0125] Through the dynamic cooperation of the above structure and data flow, the application realizes the multi-stage linkage logic of fusion judgment-graph structure mapping-control response-strategy adjustment, builds a stable and high-response self-adjusting mechanism for the diagnosis system, and ensures that fault identification and control execution form a logical closed loop.

[0126] AsFigure 2 As shown, it is another embodiment of the application, which provides a new energy power distribution fault diagnosis method based on multi-source data fusion, applying the above-mentioned new energy power distribution fault diagnosis system based on multi-source data fusion, the method comprises the following steps:

[0127] Step S1: receiving multi-source operation data from the new energy power distribution system, including voltage, current, load power and environmental state parameters; structuring sampling the multi-source operation data according to time sequence, completing unified format packaging, and generating multi-source feature tensor through normalization, standardization and feature extraction;

[0128] Step S2: performing principal component analysis processing on the feature tensor, removing redundant feature components with cumulative contribution rate lower than a preset threshold, and constructing a compressed feature tensor;

[0129] Step S3: constructing the fault classification probability distribution in the time window corresponding to each data source, calculating the information entropy value, and constructing the dynamic fusion weight based on the information entropy value; the inference module performs weighted synthesis processing of confidence information according to the weight, and outputs the fusion confidence discriminant result;

[0130] Step S4: combining the topological structure relationship and device connection information of the new energy power distribution network to construct a graph structure model, mapping the fusion confidence result to a feature vector of a node in the graph structure, and performing node-level aggregation and reasoning through a structure attribute perception attention graph neural network model to generate a node embedding vector and a corresponding fault type label;

[0131] Step S5: generating an isolation control instruction containing a control target identifier, an operation type, a priority and a strategy parameter according to the diagnosis label and the node control attribute, and issuing the isolation control instruction to the corresponding power distribution equipment terminal for execution;

[0132] Step S6: collecting the terminal execution feedback information, including action completion state, response time and abnormal flag, and writing the feedback result into the structure attribute field of the corresponding node or edge in the graph structure model;

[0133] Step S7: according to the updated graph structure state, dynamically adjusting the information entropy value threshold parameter, the fusion weight generation strategy or the confidence discriminant threshold value, realizing the adaptive evolution of the diagnosis strategy, and forming a closed-loop intelligent fault diagnosis and control mechanism.

[0134] In summary, the application constructs a new energy power distribution fault diagnosis system based on multi-source data fusion, integrates and optimizes multiple key technical links such as data acquisition standardization, fusion reasoning self-adaption, graph structure perception diagnosis and response control closed loop, and forms a dynamic closed loop system from data processing to diagnosis execution to strategy feedback. The system not only improves the fault identification accuracy and response efficiency in the multi-source data environment, but also enhances the intelligent adjustment capability and operation adaptability of the system, and is suitable for various practical application scenarios such as distributed energy stations, intelligent power distribution networks and power edge nodes, and has good engineering implementation and broad industrialization prospect.

[0135] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0136] Any process or method descriptions in flow charts or described elsewhere herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for performing specific logic functions or steps in the process. And the various embodiments of the application can include additional or fewer steps or processes in addition to or other than those of the specific embodiments described.

[0137] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A new energy power distribution fault diagnosis system based on multi-source data fusion, characterized in that, include: The data processing module is used to receive multi-source operation data from the new energy power distribution system and perform unified structured processing on the multi-source operation data, including normalization, standardization and multi-dimensional feature extraction, to generate corresponding feature tensors. The fusion inference module is used to receive the feature tensor, calculate the information entropy value based on the fault classification probability distribution of each data source within a preset time window, and construct a dynamic fusion weight vector based on the information entropy value. The dynamic fusion weight vector is used to guide the weighted fusion processing of multi-source confidence information and output the fusion confidence discrimination result of various faults. The graph structure diagnosis module is used to receive the fusion confidence judgment result and construct a graph structure model in combination with the topology of the current power distribution network. The graph structure model includes nodes representing power distribution equipment, edges representing electrical connection relationships, or edge structure attribute fields. The fusion result is mapped to the embedding features of nodes in the graph, and node-level fault identification and localization are performed through a graph neural network model containing an attention mechanism, and fault diagnosis results are output. The control response module is used to receive the fault diagnosis results, generate corresponding isolation control commands, receive control execution feedback, and update the structural attribute fields of nodes or edges in the graph structure model with the feedback information for dynamic correction of the graph structure in subsequent diagnosis cycles. The control response module includes: The response instruction generation unit is used to extract the control attributes of the target node and associated edges based on the fault diagnosis results, and generate an isolation control instruction containing the control target identifier, operation type, priority and execution strategy parameters. The instruction issuing and execution unit is used to issue the isolation control instruction to the execution terminal of the corresponding power distribution equipment and trigger the corresponding switching equipment to perform actions, including opening and closing, load switching or feeder isolation control operations. The execution status acquisition unit is used to collect operation feedback information from the execution terminal in real time. The feedback information includes the actual status of the switch, the action completion flag, the response time, and the abnormal execution flag. The graph attribute update unit is used to map and update the operation feedback information to the state structure attribute fields of the corresponding nodes or edges in the graph structure model, including row state markers, execution success flags and response delay values, for dynamic adjustment and structural correction of the fusion strategy in the subsequent graph neural network diagnosis process.

2. The new energy power distribution fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The data processing module includes: The data receiving unit is used to receive operational data from multiple edge acquisition terminals, including electrical quantities, status quantities and environmental parameters, and to align and encapsulate the operational data according to timestamps to construct a structured data input format; The processing unit is used to perform normalization processing on the structured data based on Box-Cox transformation and Z-score normalization; The feature extraction unit is used to extract multidimensional features, including RMS value, kurtosis, wavelet energy density, spectral centroid and frequency shift, from each data channel within a sliding time window, and generate feature tensors for subsequent processing.

3. The new energy power distribution fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The fusion reasoning module includes: The feature archiving unit is used to receive the feature tensors, classify and archive the feature tensors according to the data source identifier, and construct a fusion input set; Uncertainty quantization unit, used to determine the probability distribution of fault identification based on each data source within the target time window. The information entropy value is calculated using the following formula: ； in: The information entropy value. Indicates the data source number. Indicates the fault type number. This represents the total number of fault categories. Indicates the first The data source for the first The probability of classifying faults; Entropy weight generation unit, used to construct dynamic fusion weights based on the information entropy value. The calculation formula is: ； in: The total number of data sources participating in the integration. Number the data sources in the merging process. For the first Information entropy values ​​calculated from each data source; The confidence fusion unit is used to combine confidence evidence from each data source with dynamic fusion weights. Perform weighted synthesis and output the fusion confidence result corresponding to the fault type; The fusion feedback adjustment unit is used to adaptively update the weight generation strategy or confidence threshold parameter in subsequent time windows based on the fusion confidence result.

4. The new energy power distribution fault diagnosis system based on multi-source data fusion according to claim 3, characterized in that, The fusion reasoning module further includes: The feature compression unit is used to perform dimensionality compression processing on the received feature tensor before performing uncertainty quantization. The compression processing is based on the principal component analysis (PCA) algorithm. After constructing the sample covariance matrix, the first few principal components are extracted, and low-order feature components with a cumulative contribution rate lower than a preset threshold are removed. The compressed tensor output by the feature compression unit serves as the input basis for subsequent information entropy value calculation and weight generation, which is used to improve the feature alignment and fusion consistency among multi-source data. The compression dimension can be dynamically adjusted according to the fusion feedback results.

5. A new energy power distribution fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The graph structure diagnostic module includes: The graph model construction unit is used to construct a graph structure model based on the current topology and equipment connection information of the new energy power distribution system. , where the node set Represents electrical equipment unit, edge set To represent physical connection relationships, each node and edge in the graph structure model is configured with structural attribute fields containing state flags, device types, and control attributes. The feature mapping unit is used to map the fused confidence discrimination result to the corresponding node of the graph structure model according to the node device number, and to concatenate it with the historical state features in the graph structure model to form an input feature vector. The graph neural reasoning unit is used to execute an attention-based graph neural network model based on the graph structure model and node feature input to identify the fault and fault type label of each node. The graph embedding output unit is used to jointly output the intermediate state embedding results of the graph neural network with the fault discrimination results to form a structured diagnostic tensor.

6. A new energy power distribution fault diagnosis system based on multi-source data fusion according to claim 5, characterized in that, The graph neural reasoning unit employs a structural attribute-aware attention mechanism, and the attention weight coefficients between nodes are calculated according to the following formula: ； in: Indicates the first Each node relates its neighboring nodes. The assigned attention weight coefficients; , They are nodes , The input feature vector; , It is a trainable feature mapping matrix used to embed node features into a unified representation space; , They are nodes , The structural state attribute vector includes equipment type, current fault status, and historical control records; This is a joint embedding function among node attributes, used to reflect the degree of association between structural states; This is a trainable attention parameter vector; To activate the mapping function; This represents a vector concatenation operation; Indicates adjacent nodes The input feature vector, For nodes The set of adjacent nodes.

7. A new energy power distribution fault diagnosis system based on multi-source data fusion according to claim 6, characterized in that, The attention map neural network model further includes the following technical improvement modules: The attribute-driven fine-tuning module is used to dynamically fine-tune the attention weights based on the differences in the node state attribute vectors in the graph structure during the attention coefficient calculation process. The strategy control module is used to adjust the scale factor in the normalization process based on the degree distribution characteristics of the nodes in the current graph structure. The connection suppression module is used to prune edge connections corresponding to attention weights that are below a set threshold.

8. A new energy power distribution fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The control response module further includes: The strategy adjustment unit is used to dynamically adjust the information entropy threshold parameter, fusion weight generation strategy, or node state confidence threshold in the fusion inference module based on the state parameters, response delay, or execution anomaly flag contained in the control execution feedback. The strategy adjustment unit constructs an adaptive path for the fusion mechanism driven by feedback information, enabling the system to strategically optimize the fusion accuracy based on historical control behavior in the next diagnostic cycle.

9. A method for diagnosing faults in new energy power distribution based on multi-source data fusion, using a new energy power distribution fault diagnosis system based on multi-source data fusion as described in any one of claims 1-8, characterized in that, The method includes the following steps: Step S1: Receive multi-source operation data from the new energy power distribution system, including voltage, current, load power and environmental state parameters; perform structured sampling on the multi-source operation data according to the time series, complete unified format encapsulation, and generate multi-source feature tensors through normalization, standardization and feature extraction; Step S2: Perform principal component analysis on the feature tensor to remove redundant feature components with a cumulative contribution rate lower than a preset threshold, thus forming a compressed feature tensor; Step S3: Construct the fault classification probability distribution within the time window corresponding to each data source, calculate the information entropy value, construct dynamic fusion weights based on the information entropy value, and perform weighted synthesis processing of confidence information according to the weights in the fusion inference module to output the fusion confidence discrimination result; Step S4: Construct a graph structure model by combining the topological relationship of the new energy power distribution network with the equipment connection information, map the fusion confidence judgment result into the feature vector of the node in the graph structure, and perform node-level aggregation and reasoning through the attention graph neural network model with structural attribute awareness to generate node embedding vectors and corresponding fault type labels; Step S5: Based on the diagnostic tags and node control attributes, generate an isolation control command containing the control target identifier, operation type, priority, and strategy parameters, and send the isolation control command to the corresponding power distribution equipment terminal for execution; Step S6: Collect the terminal execution feedback information, including action completion status, response time and abnormal flag, and write the feedback results into the structural attribute field of the corresponding node or edge in the graph structure model; Step S7: Based on the updated graph structure state, dynamically adjust the information entropy threshold parameter, fusion weight generation strategy, or confidence discrimination threshold to achieve adaptive evolution of the diagnostic strategy and form a closed-loop intelligent fault diagnosis and control mechanism.

Images