Power grid fault intelligent diagnosis method and system based on multi-mode cooperation

Abstract

The invention discloses a power grid fault intelligent diagnosis method and system based on multi-mode cooperation, and belongs to the technical field of power system fault diagnosis. The method comprises the following steps: collecting and off-line importing multi-modal data of primary equipment and secondary equipment of a multi-space power system in a power grid fault range in real time and standardizing the multi-modal data, and sequentially inputting the standardized multi-modal data into a local edge professional model, the method comprises the following steps: preliminarily reasoning fault properties in combination with a local knowledge base, inputting a knowledge-embedded causal reasoning large model deployed at a cloud end, generating an edge-cloud fault collaborative diagnosis model, iteratively updating the edge-cloud fault collaborative diagnosis model in combination with a multi-modal fault case library, and converting newest collected and imported power grid fault data into a multi-modal fault case library; and inputting the updated edge-cloud fault collaborative diagnosis model, generating a latest fault diagnosis report, and realizing the intelligent diagnosis of the power grid fault based on multi-modal collaboration. According to the invention, the power grid fault reason analysis performance and the identification accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power system fault diagnosis, and more particularly relates to a power grid fault intelligent diagnosis method and system based on multi-modal collaboration. BACKGROUND

[0002] With the expansion of the scale of the power system and the improvement of the intelligent level, the accurate diagnosis of power grid faults faces multiple challenges. Traditional methods mostly rely on single modal data analysis, such as judging the protection action logic only through electrical quantity time series waveform, or relying on artificial experience to analyze fault recording data and parse fault information text. However, in actual fault scenarios, multi-source data generated by primary equipment and secondary equipment, such as recording time series data, protection action events, waveform images, and meteorological disaster information, are often scattered and independent, with weak spatio-temporal correlation, leading to fragmented fault feature extraction. Especially in complex fault scenarios, noise interference of a single data source easily covers up key evidence, and semantic fragmentation of cross-modal information further hinders the deep mining of fault root causes, making it difficult to meet the analysis needs of high reliability.

[0003] The existing technology has significant limitations in multi-modal collaboration and intelligent reasoning. On the one hand, edge devices are limited by computing resources and are difficult to process high sampling rate recording data or extract fine electrical quantity indicators such as harmonic distortion and wavelet time-frequency features in real time, and model updates lag behind the dynamic changes of the power grid; on the other hand, cloud analysis systems mostly focus on data-driven models, lacking deep integration with domain knowledge such as power grid topology and relay protection logic, resulting in incomplete fault reasoning chain. In addition, the time scales of multi-source data are not synchronized or the synchronization accuracy is insufficient, and the coupling mechanism of external environmental factors such as meteorology and geography with electrical characteristics is not clear, further restricting the accurate identification of complex faults. SUMMARY

[0004] To solve the problems in the prior art, the present application provides a power grid fault intelligent diagnosis method and system based on multi-modal collaboration, which constructs a knowledge-embedded causal reasoning large model and a lightweight edge professional small model collaboration system by fusing multi-modal data within the fault range, forms a data-model-business closed loop combining power grid fault domain knowledge and multi-modal fault case library, and realizes accurate extraction of electrical characteristics and multi-modal semantic fusion using a hierarchical reasoning mechanism, thereby improving fault cause analysis performance and identification accuracy.

[0005] The present application adopts the following technical solutions.

[0006] The first aspect of the present application provides a power grid fault intelligent diagnosis method based on multi-modal collaboration, characterized in that:

[0007] Real-time acquisition and offline import of multi-modal data of power system primary equipment and secondary equipment in the fault range of the power grid multi-space, wherein the multi-modal data includes environmental data and fault recording data;

[0008] Preprocessing and electrical quantity feature extraction are performed on the multi-modal data, and the multi-modal data after preprocessing and feature extraction is standardized to obtain standardized multi-modal data;

[0009] A local electrical quantity analysis engine is constructed in combination with the standardized multi-modal data, the local electrical quantity analysis engine is compressed according to channel pruning and mixed quantization compression technology, the fault feature extraction and pattern recognition capability of the benchmark model deployed in the cloud is migrated to the compressed local electrical quantity analysis engine by using knowledge distillation technology, an edge professional model deployed locally is constructed, the standardized multi-modal data is combined with power system operation and power grid fault field knowledge to generate rule constraints, a knowledge-embedded multi-level causal graph network is constructed, and a knowledge-embedded causal reasoning large model deployed in the cloud is generated.

[0010] The standardized multi-modal data is sequentially input into the edge professional model deployed locally, the electrical features are extracted in combination with the local knowledge base, and the fault properties are initially reasoned, the initially reasoned fault properties are input into the knowledge-embedded causal reasoning large model deployed in the cloud, the power grid fault reason is reasoned, the edge-cloud fault collaborative diagnosis model is generated, and the edge-cloud fault collaborative diagnosis model is iteratively updated in combination with the multi-modal fault case library.

[0011] The latest collected and imported power grid fault data is input into the updated edge-cloud fault collaborative diagnosis model, the power grid fault reason is hierarchically and collaboratively reasoned, and the latest fault diagnosis report is generated, so as to realize the intelligent diagnosis of the power grid fault based on multi-modal collaboration.

[0012] Preferably, the multi-modal data of the power system primary equipment and secondary equipment in the fault range of the power grid multi-space specifically includes:

[0013] Meteorological information, geographic information and disaster information obtained from the regulation and control cloud platform;

[0014] Topology data of the primary equipment, associated circuit breaker and switch state, SOE (Sequence of Events, event sequence record) position information, SCADA comprehensive intelligent alarm data, event information, alarm information, state position information of the secondary equipment, fault brief, fault report, configuration file, title file, data file, information file, dmf (Disturbance Message File, power system transient data exchange format file) file, recording report and equipment patrol report of the relay protection device and centralized recording device.

[0015] Preferably, the electrical quantity feature extraction specifically includes:

[0016] Filtering the collected fault recording data according to the abnormal operation state information to obtain operation state fault recording data;

[0017] According to the configuration file, title file, data file, information file and dmf file of the relay protection device and the centralized recording device in the collected fault recording data, combined with the recording channel name, recording channel current and voltage change in the recording data, the fault recording channel is identified and automatically labeled, combined with the fault brief, fault report, equipment inspection report and environmental information, the fault nature and fault environment are marked to obtain a fault recording marking file;

[0018] According to the operation state fault recording data and the fault recording marking file, electrical quantity characteristics are extracted and obtained.

[0019] Preferably, the local electrical quantity analysis engine is compressed according to the channel pruning and mixed quantization compression technology, specifically comprising:

[0020] An L1-norm pruning algorithm is used for the local electrical quantity analysis engine, the L1-norm of each channel weight in the convolution layer and the fully connected layer is calculated, the channel with L1-norm lower than the preset threshold is identified and removed, and the pruned local electrical quantity analysis engine is obtained;

[0021] The pruned local electrical quantity analysis engine is subjected to grouped INT8 quantization and key layer FP16 reservation strategy to obtain a quantized local electrical quantity analysis engine;

[0022] The volume of the quantized local electrical quantity analysis engine is compressed by entropy coding, and a compressed local electrical quantity analysis engine is obtained, which is deployed to the edge side device.

[0023] Preferably, the knowledge distillation technology is used to migrate the fault feature extraction and pattern recognition capability of the benchmark model deployed in the cloud to the compressed local electrical quantity analysis engine, specifically comprising:

[0024] A multi-modal benchmark model is trained in the cloud combined with standardized multi-modal data, and is set as a teacher model;

[0025] The same standardized multi-modal data is continuously input into the teacher model to obtain the intermediate layer feature map before the output layer of the teacher model and the final output soft label, and the intermediate layer feature map and the soft label of the teacher model are obtained;

[0026] The compressed local electrical quantity analysis engine is set as a student model, and the student model is trained according to the intermediate layer feature map and the soft label of the teacher model, and the knowledge transferred compressed model is obtained after the student model is trained, and the edge professional model deployed locally is generated.

[0027] Preferably, the standardized multi-modal data is combined with power system operation and power grid fault domain knowledge to generate node, edge and edge weight, hierarchy and knowledge embedding and rule constraint, build a knowledge-embedded multi-level causal graph network, and generate a knowledge-embedded causal reasoning large model deployed in the cloud, specifically including:

[0028] According to the standardized multi-modal data combined with power system operation and power grid fault domain knowledge, the nodes of the knowledge-embedded multi-level causal graph network are generated, including device-level nodes, fault-level nodes and causal-level nodes;

[0029] The edges of the knowledge-embedded multi-level causal graph network are set, and the edges connect the nodes to represent causal and influence relationships, and the weights of the edges are used to represent the strength and probability of the causal relationship;

[0030] The hierarchy of the knowledge-embedded multi-level causal graph network is set, including device layer, fault phenomenon layer and root cause layer;

[0031] The knowledge embedding and rule constraints of the knowledge-embedded multi-level causal graph network are constructed, including the prior rules and constraints of the graph network, the deterministic rules of the power grid and the probabilistic association rules;

[0032] The neural network is selected, and the obtained nodes, edges, hierarchy and knowledge embedding and rule constraints are used to train the knowledge-embedded multi-level causal graph network;

[0033] The input data is reduced and fused by the trained knowledge-embedded multi-level causal graph network, the prompt template and prompt are set for the natural language interaction task, the output of the knowledge-embedded multi-level causal graph network is guided to conform to the reasoning result of the business specification, and the knowledge-embedded causal reasoning large model deployed in the cloud is generated.

[0034] Preferably, the input data is reduced and fused by the trained knowledge-embedded multi-level causal graph network, the prompt template and prompt are set for the natural language interaction task, the output of the knowledge-embedded multi-level causal graph network is guided to conform to the reasoning result of the business specification, and the knowledge-embedded causal reasoning large model deployed in the cloud is generated, specifically including:

[0035] The multi-modal features input to the cloud causal reasoning large model are reduced and fused to generate context feature representation;

[0036] For the natural language interaction task, the context feature representation is combined, the prompt template and prompt are set, and the knowledge-embedded multi-level causal graph network combined with the causal graph network reasoning ability is guided to output a business report conforming to the business specification;

[0037] In filling the placeholders in the prompt template and the prompt, the activated fault level nodes and their attributes in the knowledge embedded multi-level causal graph network are utilized to integrate the required standard structure and elements of the business report, associate the prompt template and the prompt with the knowledge embedded multi-level causal graph network, generate the filled and business standard compliant fault report, and build the knowledge embedded causal reasoning large model deployed in the cloud.

[0038] Preferably, the standardized multi-modal data is sequentially input into the edge professional model deployed locally, the electrical characteristics are extracted and the fault property is preliminarily reasoned in combination with the local knowledge base, the preliminarily reasoned fault property is input into the knowledge embedded causal reasoning large model deployed in the cloud, the power grid fault cause is reasoned, and the edge-cloud fault collaborative diagnosis model is generated, specifically including:

[0039] The standardized multi-modal data is input into the edge professional model deployed locally, real-time analysis is performed in combination with the local knowledge base, the feature vector containing the electrical characteristics and the preliminary diagnosis result is extracted, the most possible preliminary fault property is output and the fault brief is generated, and the feature vector and the fault brief are uploaded to the cloud;

[0040] The feature vector and the fault brief are received in the cloud and set as text semantic features, the text semantic features, the environmental data in the standardized multi-modal data and the recording waveform image features are aligned and semantically fused across modalities through a multi-head self-attention mechanism to obtain cross-modal features;

[0041] The cross-modal features are input into the knowledge embedded causal graph network, the multi-modal information is fused, the causal reasoning is performed, the graph network reasoning result is obtained and the fault cause decision path is generated, the confidence of each type of fault root cause hypothesis in the path and the supporting evidence are calculated, the structured prompt template and the graph network reasoning result are combined to generate the edge-cloud collaborative fault diagnosis report, and the edge-cloud fault collaborative diagnosis model is generated.

[0042] Preferably, the edge-cloud fault collaborative diagnosis model is iteratively updated in combination with the multi-modal fault case library, specifically including:

[0043] The newly added multi-modal fault cases in the daily multi-modal fault case library are utilized to automatically trigger the training and adjustment of the parameters of the edge professional model deployed locally and the parameters of the graph neural network in the knowledge embedded causal reasoning large model deployed in the cloud in the edge-cloud fault collaborative diagnosis model;

[0044] The causal graph structure is dynamically updated by adding new fault modes, root cause nodes and connecting edges in the knowledge embedded causal reasoning large model deployed in the cloud, and the weights of the edges in the causal graph are dynamically adjusted by using a graph learning algorithm;

[0045] The new device parameters verified to be valid, the protection logic changes, the fault mode descriptions and the environmental influence experiences are updated to the local knowledge base and the cloud power system operation and power grid fault field knowledge base;

[0046] A case similarity retrieval algorithm is established, similarity with historical cases is retrieved based on feature vectors and graph node embedding, and the most possible new case diagnosis is updated according to the similarity, so that the edge-cloud fault collaborative diagnosis model is updated.

[0047] The second aspect of the present application provides a power grid fault intelligent diagnosis system based on multi-modal collaboration, which runs the power grid fault intelligent diagnosis method based on multi-modal collaboration.

[0048] The data acquisition module is used for real-time acquisition and offline import of multi-modal data of power system primary equipment and secondary equipment in the power grid fault range multi-space;

[0049] The standardization module is used for pre-processing and electrical quantity feature extraction of the multi-modal data, and standardization of the multi-modal data after pre-processing and feature extraction, to obtain standardized multi-modal data;

[0050] The edge and cloud model construction module is used for combining the standardized multi-modal data to construct a local electrical quantity analysis engine, compressing the local electrical quantity analysis engine according to the channel pruning and mixed quantization compression technology, migrating the fault feature extraction and pattern recognition capability of the benchmark model deployed in the cloud to the compressed local electrical quantity analysis engine by using the knowledge distillation technology, constructing the edge professional model deployed locally, combining the standardized multi-modal data with the power system operation and power grid fault field knowledge to generate rule constraints, constructing a multi-level causal graph network embedded with knowledge, and generating a knowledge-embedded causal reasoning large model deployed in the cloud;

[0051] The fault collaborative diagnosis model construction module is used for inputting the standardized multi-modal data into the edge professional model deployed locally in sequence, extracting electrical features and preliminarily reasoning fault properties in combination with the local knowledge base, inputting the preliminarily reasoned fault properties into the knowledge-embedded causal reasoning large model deployed in the cloud, reasoning the power grid fault cause, generating an edge-cloud fault collaborative diagnosis model, and iteratively updating the edge-cloud fault collaborative diagnosis model in combination with the multi-modal fault case library;

[0052] The diagnosis report generation module is used for inputting the latest collected and imported power grid fault data into the updated edge-cloud fault collaborative diagnosis model, hierarchically and collaboratively reasoning the power grid fault cause and generating the latest fault diagnosis report, to realize the power grid fault intelligent diagnosis based on multi-modal collaboration.

[0053] Compared with the prior art, the present application has at least the following beneficial effects:

[0054] The application significantly improves the accuracy and efficiency of power grid fault analysis through deep fusion and layered collaborative reasoning mechanism of multi-modal data, deploys edge professional models in the local based on edge-cloud collaborative architecture, realizes real-time feature extraction and preliminary diagnosis of high sampling rate recording wave data in a lightweight manner, effectively reduces data transmission burden and cloud computing delay; the knowledge embedded causal reasoning large model deployed in the cloud fuses the power grid topology structure, relay protection logic and external environmental factors through cross-modal semantic alignment and causal graph network reasoning, constructs multi-dimensional fault decision path, solves the misjudgment problem caused by data island in traditional methods, and improves the accuracy of power grid fault diagnosis.

[0055] Combined with the domain knowledge base and the dynamically updated multi-modal case base, the system can adaptively optimize model parameters and reasoning rules, and still maintain high generalization ability when facing new faults or small sample scenarios, thereby improving the generalization ability of power grid fault diagnosis.

[0056] Meanwhile, through millisecond-level time synchronization and accurate calculation of fine electrical quantity features such as transition resistance and harmonic distortion rate during preprocessing, the positioning accuracy of complex cascading faults of the power grid is significantly improved, the demand for manual intervention is reduced, reliable support is provided for intelligent diagnosis and rapid recovery of power grid faults, and the reliability of power grid fault diagnosis is improved.

[0057] Through the decision path of the causal graph network and the attribution of the key features of the edge professional model, the system has certain fault cause analysis and explanation ability, so that the operation and maintenance personnel can intuitively understand and trace the fault judgment basis, and the fault cause explanation, understandability and traceability of the power grid system are improved. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a structural schematic diagram of the power grid fault intelligent diagnosis method provided by the embodiment of the application. DETAILED DESCRIPTION

[0059] In order to make the object, technical scheme and advantages of the application clearer, the technical scheme of the application will be described clearly and completely below in combination with the drawings in the embodiments of the application. The described embodiments are only a part of the embodiments of the application, not all the embodiments. Based on the spirit of the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0060] As shown in Figure 1 Embodiment 1 of the application provides a power grid fault intelligent diagnosis method based on multi-modal collaboration, including the following steps:

[0061] Step 1, collecting and offline importing multi-modal data of power system primary equipment and secondary equipment in multi-space of power grid fault range in real time, wherein the multi-modal data comprises environmental data and fault recording data.

[0062] In a preferred but non-limiting embodiment of the present application, the multi-modal data comprises:

[0063] The environmental data comprises meteorological information, geographic information, disaster information and the like obtained from a regulation cloud platform;

[0064] The fault recording data comprises topology data of primary equipment, associated circuit breaker and switch state, SOE displacement information, SCADA comprehensive intelligent alarm data, event information of secondary equipment, alarm information, state displacement information, fault brief report, fault report, configuration file of relay protection device and centralized recording device, title file, data file, information file, dmf file, recording report and equipment patrol report and the like.

[0065] Step 2, pre-processing and electrical quantity feature extraction are performed on the multi-modal data of power system primary equipment and secondary equipment in multi-space of power grid fault range collected or offline imported, and the multi-modal data after pre-processing and feature extraction is standardized to obtain standardized multi-modal data.

[0066] In a preferred but non-limiting embodiment of the present application, step 2 comprises:

[0067] Step 2.1, abnormal operating state information filtering is performed on the fault recording data collected in step 1 according to the topology data of primary equipment, associated circuit breaker and switch state, SOE displacement information, SCADA comprehensive intelligent alarm data, event information of secondary equipment, alarm information, state displacement information, fault brief report, fault report and the like to obtain operating state fault recording data;

[0068] Step 2.2, according to the configuration file, title file, data file, information file and dmf file of relay protection device and centralized recording device in the fault recording data collected in step 1, recording channel name, recording channel current and voltage change in the recording data are analyzed, fault recording channel is identified and automatically labeled, manual correction is supported at the same time, and the fault recording mark file is obtained by marking fault nature and fault environment in combination with fault brief report, fault report, equipment inspection report and environmental information;

[0069] Step 2.3, electrical quantity feature extraction is performed on the operating state fault recording data obtained in step 2.1 and the fault recording mark file obtained in step 2.2 to obtain electrical quantity feature.

[0070] Further preferably, step 2.3 comprises:

[0071] The event information of the secondary equipment in the running state fault recording data, the recording report, the title file of the relay protection device and the centralized recording device, etc. are identified to recognize the fault phase, the reclosing condition, the fault voltage, the fault current, the distance measurement, etc.

[0072] The data file and the fault recording mark file in the running state fault recording data are calculated to obtain the fault phase, the reclosing condition, the fault current, the fault voltage, the impedance angle, the initial angle, the distance measurement, etc. The electrical quantity characteristic parameters which are not identified are calculated.

[0073] The data file and the fault recording mark file in the running state fault recording data are calculated and judged to obtain the broken line characteristics of the fault under the conditions that the three-phase voltage is normal before and after the fault, the current is less than the threshold value, and the zero sequence current is greater than the threshold value.

[0074] The first-order and second-order differences of the current and voltage channel waveform data are calculated by using the data file and the fault recording mark file in the running state fault recording data to obtain the fault waveform mutation and curvature characteristics.

[0075] The fundamental effective value and the phase angle offset in the set cycle range before and after the fault are calculated and extracted by using the data file and the fault recording mark file in the running state fault recording data.

[0076] The direct current component, the 2-13 harmonic amplitude ratio and the total harmonic distortion rate are extracted by using the discrete Fourier front cycle rectangular integral algorithm by using the data file and the fault recording mark file in the running state fault recording data.

[0077] The line parameters are automatically obtained according to the configuration file, the title file, the dmf file, etc. in the running state fault recording data, the distance measurement is calculated according to the recording data, and the transition resistance is further calculated.

[0078] More preferably, the transition resistance is calculated according to the recording data, specifically including:

[0079] The distance measurement is calculated by using the double-end distance measurement algorithm by using the data file and the fault recording mark file in the running state fault recording data to obtain the fault position.

[0080] The three-phase voltage and current at the protection installation of the line at both ends are used to calculate the positive, negative and zero sequence voltage and current.

[0081] The positive, negative and zero sequence voltage and current of the fault position are calculated by using the positive, negative and zero sequence at both ends of the line.

[0082] The positive, negative and zero sequence voltage is averaged, and the positive, negative and zero sequence current is summed.

[0083] The fault location voltage and current are calculated by positive, negative and zero sequence, and the transition resistance is obtained by dividing the voltage by the current.

[0084] The wavelet transform is performed on the recording wave signal to extract energy, variance and entropy value of wavelet coefficients in each frequency band.

[0085] Step 2.4, according to the primary topology and channel relationship, the spatial correlation of different spatial fault recording data sources in the fault recording data is established, and a group of multi-space fault recording marking files is obtained;

[0086] The millisecond-level time synchronization is performed on the recording wave data generated by different devices in the fault recording data, and the recording wave data synchronized in time by different devices is obtained.

[0087] For the fault recording data, the inter-phase current mutation, zero sequence current mutation and zero sequence auxiliary start criterion are used to identify the accurate fault time, and the fault time is obtained.

[0088] According to the fault time and the fault recording channel marked in the fault recording marking file of step 2.2, the recording wave form normalization image related to each fault is automatically formed, and the recording wave form image feature is obtained.

[0089] The recording data with a sampling rate higher than a set threshold 1 is resampled according to the set sampling rate, and the recording data with a sampling rate lower than the threshold 1 is interpolated and sampled, and a unified time sequence data sampling rate is obtained.

[0090] The group of multi-space fault recording marking files, the recording data synchronized in time by different devices, the fault time, the normalization image and the unified time sequence data sampling rate are combined to obtain the preprocessed data.

[0091] Step 2.5, the operating state fault recording data, the fault recording marking file, the electrical quantity feature, the preprocessed data obtained in steps 2.1-2.4 and the environment data in step 1 are combined to obtain the multi-modal data after preprocessing and feature extraction, and the multi-modal data is standardized to obtain the standardized multi-modal data.

[0092] Step 3, a local electrical quantity analysis engine is constructed, the local electrical quantity analysis engine is compressed according to the channel pruning and mixed quantization compression technology, the fault feature extraction and pattern recognition ability of the benchmark model deployed in the cloud is migrated to the compressed local electrical quantity analysis engine, and the edge professional model deployed locally is constructed in combination with the standardized multi-modal data in step 2.

[0093] In the preferred but non-limiting embodiments of the application, step 3 comprises:

[0094] Step 3.1, a lightweight convolutional neural network is set to construct the local electrical quantity analysis engine of the edge professional model deployed locally.

[0095] Further preferably, step 3.1 comprises:

[0096] A local electrical quantity analysis engine is constructed, which is a core component of the edge lightweight professional model. The engine receives the standardized multi-modal data obtained in step 2, including the running state fault recording data and its label file and the extracted electrical quantity features, and analyzes the electrical quantity features in real time, such as but not limited to fault phase, fault current / voltage effective value, harmonic content, mutation feature, ranging result, transition resistance estimation value, time-frequency domain feature, etc. These analysis results are used for local preliminary fault identification on the one hand, and form a high-dimensional feature vector, which is provided to the cloud-side knowledge-embedded causal reasoning large model.

[0097] Step 3.2, channel pruning and mixed quantization compression technology is used for the local electrical quantity analysis engine in step 3.1, the engine parameter quantity is compressed to meet the memory and computing power constraints of the edge device, and the compressed local electrical quantity analysis engine is obtained.

[0098] Further preferably, step 3.2 comprises:

[0099] Step 3.2.1, L1-norm pruning algorithm is used for the local electrical quantity analysis engine, the L1-norm of each channel weight in the convolution layer and the fully connected layer is calculated, the channels with L1-norm lower than the preset threshold are identified and removed, and the pruned local electrical quantity analysis engine is obtained;

[0100] Step 3.2.2, grouping INT8 quantization and key layer FP16 reservation strategy are used to group the weights and activation values of the pruned local electrical quantity analysis engine in step 3.2.1;

[0101] For most layers in the pruned local electrical quantity analysis engine, such as but not limited to intermediate feature extraction layers, INT8 symmetric quantization is applied, the scaling factor and zero point of each group of weights / activation values are calculated, the original FP32 values are mapped to the INT8 range of -128 to 127, and efficient INT8 operation cores are used during inference;

[0102] For key layers in the pruned local electrical quantity analysis engine that are extremely sensitive to accuracy, such as but not limited to output layers and initial feature fusion layers, FP16 precision is reserved to avoid significant accuracy loss caused by quantization, and the quantized local electrical quantity analysis engine is obtained.

[0103] Step 3.2.3, entropy encoding is performed on the quantized local electrical quantity analysis engine in step 3.2.2 to compress the volume of the local electrical quantity analysis engine, further reducing the volume of the local electrical quantity analysis engine, and the compressed local electrical quantity analysis engine is deployed to the edge side device.

[0104] Step 3.3, using the knowledge distillation technique, migrates the fault feature extraction and pattern recognition capabilities of the benchmark model deployed in the cloud to the compressed local electrical quantity analysis engine deployed to the edge side device in step 3.2, obtaining a compressed model after knowledge migration, that is, an edge professional model deployed locally.

[0105] Further preferably, step 3.3 comprises:

[0106] Step 3.3.1, in the cloud, using the standardized multi-modal data obtained in step 2, training a high-precision multi-modal benchmark model, set as a teacher model;

[0107] Step 3.3.2, continue to input the same standardized multi-modal data into the teacher model of step 3.3.1, obtain the intermediate layer feature map before the output layer of the teacher model and the final output soft label, obtain the intermediate layer feature map and the soft label of the teacher model, wherein the soft label contains a probability distribution vector;

[0108] Step 3.3.3, set the compressed local electrical quantity analysis engine of step 3.2 as a student model, train the student model according to the intermediate layer feature map and the soft label of the teacher model obtained in step 3.3.2, and obtain a compressed model after knowledge migration after training. The total loss function L total , is expressed as follows:

[0109] L total = a·L hard + b·L soft + g·L feat (1)

[0110] In the formula,

[0111] a, b, g are adjustable hyperparameters that balance the importance of different supervision signals, and by minimizing L_total, the student model inherits the feature extraction and discrimination ability of the teacher model while maintaining lightweight,

[0112] L hard represents the standard cross-entropy loss between the student model prediction result and the true hard label, wherein the true hard label contains the fault property, fault phase and fault reason in the standardized multi-modal data obtained in step 2;

[0113] L soft represents the KL divergence loss between the probability distribution S of the final output layer of the student model and the soft label T of the teacher model obtained in step 3.3.2, L soft = KL(T||S);

[0114] L feat represents the specific intermediate layer feature map F sThe intermediate layer feature map F of the teacher model obtained in step 3.3.2 t Mean squared error loss or cosine similarity loss between them, L feat =MSE(F s ||F t This makes the intermediate representation of the student model approximate the teacher model.

[0115] Step 4 involves combining the standardized multimodal data obtained in Step 2 with knowledge of power system operation and grid faults to generate node, edge, and edge weights, hierarchical structure, as well as knowledge embedding and rule constraints. This constructs a multi-level causal graph network for knowledge embedding, generating a large-scale causal reasoning model deployed in the cloud. The multi-level causal graph network for knowledge embedding serves as the core reasoning engine of the cloud-based causal reasoning model, used to integrate multimodal features and knowledge of grid faults for interpretable causal reasoning.

[0116] More preferably, step 4 includes:

[0117] Step 4.1: Based on standardized multimodal data and knowledge of power system operation and grid faults, generate device-level nodes, fault-level nodes, and causal-level nodes of a knowledge-embedded multi-level causal graph network.

[0118] More preferably, step 4.1 includes:

[0119] Define device-level nodes, which represent key equipment in the power grid, such as but not limited to lines, transformers, circuit breakers, and protection devices. Attributes include device type, ratings and impedance parameters, and real-time status from SCADA / SOE. Specific implementations include:

[0120] Key equipment such as lines, transformers, circuit breakers, and protection devices in the power grid are set as equipment-level nodes, and the equipment type, rating, impedance parameters, and real-time status from SCADA / SOE are set as attributes of the equipment-level nodes.

[0121] Define fault-level nodes, which represent possible fault events and phenomena, such as, but not limited to, phase A ground fault, phase B open circuit, protection malfunction, and lightning strike. Attributes include fault characteristics, fault voltage, fault current, transition resistance, harmonic characteristics, time-frequency characteristics, and environmental characteristics from the standardized multimodal data obtained in step 2. Specific implementation includes:

[0122] Set fault events and phenomena as fault-level nodes, and set the fault nature, fault voltage, fault current, transition resistance, harmonic characteristics, time-frequency characteristics and environmental characteristics in the standardized multimodal data obtained in step 2 as attributes of the fault-level nodes.

[0123] define a causal level node, the causal level node represents a fault cause and root cause, for example but not limited to lightning strike, bushfire, pollution flashover, tree barrier, bird damage, equipment defect and misoperation, attributes include confidence, evidence chain, and specific implementation includes:

[0124] set the fault cause and root cause as the causal level node, and set the confidence and evidence chain as the attributes of the causal level node.

[0125] Step 4.2, set the edge of the knowledge-embedded multi-level causal graph network, the edge connects the nodes, and is used to represent the causal relationship and influence relationship. For example but not limited to lightning strike -> causes -> line A phase overvoltage -> leads to -> A phase insulation breakdown -> shows as -> A phase ground fault & triggers -> protection device action, and the weight of the edge is used to represent the strength and probability of the causal relationship.

[0126] Step 4.3, according to the standardized multi-modal data combined with the power system operation and power grid fault field knowledge, set the hierarchical structure of the device layer, fault phenomenon layer and root cause layer of the knowledge-embedded multi-level causal graph network.

[0127] More preferably, step 4.3 includes:

[0128] set the first layer of the knowledge-embedded multi-level causal graph network as the device layer, which is used to represent the devices in the power grid topology and their states;

[0129] set the second layer of the knowledge-embedded multi-level causal graph network as the fault phenomenon layer, which is used to represent the fault events and features that can be directly observed / analyzed caused by abnormal device states and external factors, and receives the feature vectors uploaded from the edge model in step 3.1 and the environmental features in step 2.

[0130] set the third layer of the knowledge-embedded multi-level causal graph network as the root cause layer, which is used to represent the root causes leading to the occurrence of fault phenomena.

[0131] Step 4.4, build the prior rules and constraints, power grid deterministic rules and probabilistic association rules of the graph network of the knowledge-embedded multi-level causal graph network.

[0132] More preferably, step 4.4 includes:

[0133] encode the power grid fault field knowledge, for example but not limited to protection configuration logic, equipment fault mode, meteorological influence mechanism and electrical quantity change law, etc., as the prior rules and constraints of the graph network:

[0134] inject power grid deterministic rules, for example but not limited to when a single-phase ground fault occurs, the zero sequence current significantly increases; when a fault occurs in a lightning strike area, the fault current waveform contains high-frequency components; when the circuit breaker is in the tripping state, the current of the associated line is zero;

[0135] Based on the historical case library statistical learning to obtain probabilistic association rules, such as but not limited to the specific harmonic feature combination + high humidity environment and the existence of high probability association of insulator pollution flashover;

[0136] A rule updating mechanism is set, after the deployment of the knowledge-embedded causal reasoning large model, combined with the daily new multi-modal fault case library in step 5, the weights of the causal edges are dynamically adjusted by using a graph learning algorithm, such as but not limited to gradient-based optimization or Bayesian update, to learn new fault mode associations and realize the continuous evolution of the rule base.

[0137] Step 4.5, select a neural network, combine the nodes, edges, hierarchical structure and knowledge embedding obtained in steps 4.1-4.4 with the rule constraints, and train the knowledge-embedded multi-level causal graph network.

[0138] More preferably, step 4.5 includes:

[0139] Using a graph neural network, such as but not limited to a graph convolution network or a GAT (Graph Attention Tnetwork), as the basic architecture, learning node representation and edge weight, using the idea of Bayesian network or structural equation model for probabilistic reasoning, when training the knowledge-embedded multi-level causal graph network, using the samples containing fault properties, fault voltage, fault current and other real labels in the standardized multi-modal data obtained in step 2, through supervised learning, such as but not limited to node classification loss and edge prediction loss, optimize the network parameters, so that it can accurately predict the fault root cause node and its confidence.

[0140] Step 4.6, through the trained knowledge-embedded multi-level causal graph network, the input data is reduced and fused, the effectiveness and accuracy of the input of the knowledge-embedded multi-level causal graph network are improved, for natural language interaction scenarios, such as but not limited to generating power grid fault briefs, etc., combined with structured prompt templates and prompt sentences, guide the output of the knowledge-embedded multi-level causal graph network to conform to the reasoning results of the business specification, and generate a knowledge-embedded causal reasoning large model deployed in the cloud.

[0141] Further preferably, step 4.6 includes:

[0142] Step 4.6.1, reduce and fuse the multi-modal features input to the cloud causal reasoning large model to generate context feature representation, specifically including:

[0143] Through the multi-head self-attention mechanism, the knowledge-embedded multi-level causal graph network learns the internal relationship and relative importance between different features, such as but not limited to the current mutation feature and the specific harmonic feature, and the specific protection action information and the weather information, automatically focuses on the key information, generates a fused and more discriminative context feature representation, and solves the problem of large difference and high dimension of multi-modal features, wherein the multi-modal features input into the cloud causal reasoning large model come from the electrical quantity feature vector, environmental feature text, protection action information text, and wave recording image features parsed by the edge model in step 3.1.

[0144] In step 4.6.2, for natural language interaction tasks such as but not limited to generating fault analysis reports and answering fault reason queries, a specific prompt template and prompt are set based on the context feature representation, guiding the knowledge-embedded multi-level causal graph network with reasoning capability to output structured and business specification-compliant results. In the template structure example, the following fault information is analyzed:

[0145] 1. Fault nature: Fill in the fault time, fault nature, fault type, and key protection action information marked in the step 2 standardized multi-modal data;

[0146] 2. Key electrical features: Fill in the key feature summary uploaded by the edge model in step 3.1, such as maximum fault current, harmonic distortion rate, and transition resistance estimate;

[0147] 3. Environmental features: Fill in the external environmental feature labels.

[0148] Based on the power grid fault knowledge and the causal graph network, infer the most likely fault cause and list it in priority, provide key evidence and support confidence, and generate a brief fault analysis report.

[0149] In step 4.6.3, the placeholder content in the prompt template and prompt, such as but not limited to "key electrical features", is filled in by using the fault level node and its attributes activated in the knowledge-embedded multi-level causal graph network in step 4.5, which integrates the required specification structure and key elements of the business report. The prompt template and prompt are associated with the knowledge-embedded multi-level causal graph network to generate a filled and business specification-compliant fault report.

[0150] The feature optimization in step 4.6 provides higher quality and more integrated input features for the causal graph network in step 4.5, improving the accuracy of node representation. The structured prompt template in step 4.6 is a bridge and standardizer for the root cause node and evidence chain of the causal graph network reasoning results to output in natural language to users. The key information reasoned by the causal graph network is extracted and filled into the corresponding position of the template.

[0151] Step 5, input the step 2 standardized multi-modal data into the edge professional model deployed in step 3 locally, combine with the local knowledge base to analyze and extract electrical features and preliminarily infer the fault nature, adopt hierarchical collaborative reasoning, input the preliminary inference fault nature into the knowledge embedded causal reasoning large model deployed in the cloud in step 4, fuse the electrical feature extraction and multi-modal semantics, infer the power grid fault cause, obtain the edge-cloud fault collaborative fault diagnosis report, generate the edge-cloud fault collaborative diagnosis model, and combine the multi-modal fault case library to iteratively update the edge-cloud fault collaborative diagnosis model.

[0152] In the preferred but non-limiting embodiments of the present application, step 4 comprises:

[0153] Step 5.1, input the step 2 standardized multi-modal data into the edge professional model deployed in step 3 locally, combine with the local knowledge base for real-time analysis, extract the feature vector containing electrical features and preliminary diagnosis results, preliminarily identify the fault cause according to the feature vector, output 1-3 most possible preliminary fault natures, generate a fault brief according to the preliminary fault nature, and upload the feature vector and the fault brief to the cloud, wherein the fault brief contains: fault accurate time, fault equipment / line, preliminary fault nature, key electrical quantity summary such as maximum fault current / voltage, associated protection action and brief environmental marker.

[0154] Step 5.2, receive the feature vector and the fault brief in the cloud and set them as text semantic features, perform cross-modal alignment and semantic fusion on the text semantic features, environmental data in the standardized multi-modal data and recording waveform image features through a multi-head self-attention mechanism to obtain cross-modal features;

[0155] Input the cross-modal features into the knowledge embedded causal graph network, fuse multi-modal information, perform causal reasoning, obtain graph network reasoning results and generate a fault cause decision path, calculate the confidence and supporting evidence of each type of fault root cause hypothesis in the path, combine the structured prompt template and the graph network reasoning results to generate an edge-cloud collaborative fault diagnosis report, and generate an edge-cloud fault collaborative diagnosis model.

[0156] Further preferably, step 5.2 comprises:

[0157] Receive the feature vector and the fault brief uploaded by the edge professional model deployed in step 5.1 locally and set them as text semantic features, combine the environmental data and recording waveform image features of the step 2 standardized multi-modal data, perform cross-modal alignment and deep semantic fusion on the text semantic features, environmental data and recording waveform image features through a multi-head self-attention mechanism to obtain cross-modal features;

[0158] The cross-modal feature input knowledge embedding causal graph network is input, and through message passing of a GNN (Graph Neural Network), a graph neural network), the semantic relationship of the device node connection in the power grid topology structure, the protection action node association in the protection logic, and the attribute definition of the environment / cause node in the external environment factor is deeply fused, an interpretable causal reasoning is performed, and a graph network reasoning result is obtained.

[0159] A fault cause decision path is generated according to the graph network reasoning result, and specifically includes:

[0160] An activation path of the device state, the fault feature, the environment, and the protection action from the high-confidence cause node to the initial evidence node is traced back, and a reasoning process including a multi-dimensional evidence chain of time, space, electricity, logic, environment, and the like is formed;

[0161] The confidence and supporting evidence of each type of fault cause hypothesis in the fault cause decision path are calculated, a structured prompt template is used, the graph network reasoning result is combined, an edge-cloud collaborative fault diagnosis report is generated, and an edge-cloud fault collaborative diagnosis model is generated.

[0162] Step 5.3, the hybrid model of the fused electrical feature extraction and the multi-modal semantic is iteratively optimized in combination with the multi-modal fault case library, an updated edge-cloud fault collaborative diagnosis model is obtained, and the model performance and recognition accuracy are improved.

[0163] Further preferably, step 5.3 includes:

[0164] The multi-modal fault cases added to the daily multi-modal fault case library, including preprocessed data, features, and verification labels / causes, are used to automatically trigger the training and adjustment of the edge professional model parameters deployed locally and the graph neural network parameters in the knowledge-embedded causal reasoning large model deployed in the cloud in the edge-cloud fault collaborative diagnosis model;

[0165] The multi-modal fault cases added to the daily multi-modal fault case library are used to dynamically update the causal graph structure by adding new fault modes and cause nodes and connection edges of the knowledge-embedded causal reasoning large model deployed in the cloud in the hybrid model of the fused electrical feature extraction and the multi-modal semantic, dynamically adjust the weights of the edges in the causal graph using a graph learning algorithm such as gradient-based optimization or Bayesian update, reflect the changes in the strength of the causal relationship or newly discovered statistical associations, dynamically update / add deterministic rules and probabilistic rules based on case diagnosis and expert input.

[0166] Valid new device parameters, protection logic changes, fault mode descriptions, and environmental impact experiences are updated to the local knowledge base and the cloud power system operation and power grid fault domain knowledge base;

[0167] Set up performance degradation early warning indicators to monitor the accuracy, recall rate and other indicators of the model in the validation set or online diagnosis, and trigger an alarm.

[0168] Compare the decision path / evidence chain generated by the model with the analysis of the domain expert or the actual fault report to evaluate and improve the model's explainability.

[0169] Establish a case similarity retrieval algorithm based on feature vectors or graph node embedding to quickly retrieve similar historical cases, whose experience can be used to assist in new case diagnosis or model updating, and update the most likely new case diagnosis according to the similarity with historical cases to update the edge-cloud fault collaborative diagnosis model, improving the accuracy in small sample scenarios. This closed loop ensures continuous adaptive optimization of the system.

[0170] Step 6: Input the latest collected and imported power grid fault data into the edge-cloud fault collaborative diagnosis model of step 5, perform hierarchical collaborative reasoning on the power grid fault cause, and generate the latest fault diagnosis report, realizing intelligent diagnosis of power grid faults based on multi-modal collaboration.

[0171] Embodiment 2 of the present application provides a power grid fault intelligent diagnosis system based on multi-modal collaboration, which runs the power grid fault intelligent diagnosis method based on multi-modal collaboration described in embodiment 1, comprising:

[0172] A data acquisition module is used to acquire and import multi-modal data of power system primary equipment and secondary equipment in multiple spaces of power grid fault range in real time;

[0173] A standardization module is used to preprocess and extract electrical quantity features from the multi-modal data, and standardize the preprocessed and feature-extracted multi-modal data to obtain standardized multi-modal data.

[0174] An edge and cloud model construction module is used to construct a local electrical quantity analysis engine based on the standardized multi-modal data, compress the local electrical quantity analysis engine according to channel pruning and mixed quantization compression technology, migrate the fault feature extraction and pattern recognition capability of the benchmark model deployed in the cloud to the compressed local electrical quantity analysis engine using knowledge distillation technology, construct an edge professional model deployed locally, combine the standardized multi-modal data with power system operation and power grid fault domain knowledge to generate rule constraints, construct a knowledge-embedded multi-level causal graph network, and generate a knowledge-embedded causal reasoning large model deployed in the cloud.

[0175] The fault collaborative diagnosis model construction module is used for sequentially inputting the standardized multi-modal data into the edge professional model deployed locally, extracting electrical characteristics and preliminarily reasoning the fault property in combination with the local knowledge base, inputting the preliminarily reasoned fault property into the knowledge-embedded causal reasoning large model deployed in the cloud, reasoning the power grid fault cause, generating an edge-cloud fault collaborative diagnosis model, and iteratively updating the edge-cloud fault collaborative diagnosis model in combination with the multi-modal fault case library.

[0176] The diagnosis report generation module is used for inputting the latest collected and imported power grid fault data into the updated edge-cloud fault collaborative diagnosis model, hierarchically and collaboratively reasoning the power grid fault cause and generating a latest fault diagnosis report, and realizing the power grid fault intelligent diagnosis based on multi-modal collaboration.

[0177] Compared with the prior art, the present application has at least the following beneficial effects:

[0178] The present application significantly improves the accuracy and efficiency of power grid fault analysis through the deep fusion of multi-modal data and the hierarchical collaborative reasoning mechanism, and based on the edge-cloud collaborative architecture, the edge professional model deployed locally realizes the real-time feature extraction and preliminary diagnosis of high sampling rate recording wave data in a lightweight manner, effectively reducing the data transmission burden and cloud computing delay; the knowledge-embedded causal reasoning large model deployed in the cloud deeply fuses the power grid topology structure, relay protection logic and external environmental factors through cross-modal semantic alignment and causal graph network reasoning, constructs a multi-dimensional fault decision path, solves the misjudgment problem caused by data island in the traditional method, and improves the accuracy of power grid fault diagnosis.

[0179] In combination with the domain knowledge base and the dynamically updated multi-modal case library, the system can adaptively optimize the model parameters and reasoning rules, and still maintain high generalization ability when facing new faults or small sample scenarios, thereby improving the generalization ability of power grid fault diagnosis.

[0180] Meanwhile, through millisecond-level time synchronization and accurate calculation of fine electrical quantity characteristics such as transition resistance and harmonic distortion rate during preprocessing, the positioning accuracy of complex cascading faults of the power grid is significantly improved, the demand for manual intervention is reduced, reliable support is provided for intelligent diagnosis and rapid recovery of power grid faults, and the reliability of power grid fault diagnosis is improved.

[0181] Through the decision path of the causal graph network and the key feature attribution of the edge professional model, the system has certain fault cause analysis and explanation ability, so that the operation and maintenance personnel can intuitively understand and trace the fault judgment basis, and the explanation, understandability and traceability of the fault cause of the power grid system are improved.

[0182] The present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present disclosure.

[0183] It should be noted that the above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit it. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or equivalent replaced without departing from the spirit and scope of the present application. Any modification or equivalent replacement should be covered within the protection scope of the claims of the present application.

Claims

1. A multi-modal collaborative-based power grid fault intelligent diagnosis method, characterized in that: real-time collection and offline import of multi-modal data of power system primary equipment and secondary equipment in multiple spaces of power grid fault range; pre-processing and electrical quantity feature extraction of multi-modal data, and standardization of multi-modal data after pre-processing and feature extraction to obtain standardized multi-modal data; construction of a local electrical quantity analysis engine combined with standardized multi-modal data, compression of the local electrical quantity analysis engine according to channel pruning and mixed quantization compression technology, migration of fault feature extraction and pattern recognition capability of a benchmark model deployed in the cloud to the compressed local electrical quantity analysis engine using knowledge distillation technology, construction of an edge professional model deployed locally, generation of rule constraints combined with power system operation and power grid fault domain knowledge, construction of a knowledge-embedded multi-level causal graph network, and generation of a knowledge-embedded causal reasoning large model deployed in the cloud; input of the standardized multi-modal data into the edge professional model deployed locally in sequence, extraction of electrical features and preliminary reasoning of fault properties combined with the local knowledge base, input of the preliminary reasoning of fault properties into the knowledge-embedded causal reasoning large model deployed in the cloud, reasoning of power grid fault causes, generation of an edge-cloud fault collaborative diagnosis model, and iterative updating of the edge-cloud fault collaborative diagnosis model combined with a multi-modal fault case library; input of the latest collected and imported power grid fault data into the updated edge-cloud fault collaborative diagnosis model, hierarchical collaborative reasoning of power grid fault causes, and generation of the latest fault diagnosis report.

2. The multi-modal collaborative-based power grid fault intelligent diagnosis method according to claim 1, characterized in that: the multi-modal data of the power system primary equipment and secondary equipment in multiple spaces of the power grid fault range comprises: meteorological information, geographic information, and disaster information obtained from a control cloud platform; topology data of primary equipment, associated circuit breaker and switch status, SOE displacement information, SCADA comprehensive intelligent alarm data, event information, alarm information, state displacement information of secondary equipment, fault brief, fault report, configuration files, title files, data files, information files, dmf files, wave recording report, and equipment patrol report of relay protection devices and centralized wave recording devices.

3. The multi-modal collaborative-based power grid fault intelligent diagnosis method according to claim 1, characterized in that: the electrical quantity feature extraction specifically comprises: filtering of collected fault recording data according to abnormal operating state information to obtain operating state fault recording data; identification of fault recording channels and automatic labeling combined with fault recording channel names, fault recording channel current and voltage changes in the configuration files, title files, data files, information files, and dmf files of relay protection devices and centralized wave recording devices in the collected fault recording data, labeling of fault properties and fault environment combined with fault briefs, fault reports, equipment inspection reports, and environmental information to obtain fault recording label files; and extraction of electrical quantity features according to the operating state fault recording data and the fault recording label files. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 4. The method of claim 1, wherein the local electrical quantity analysis engine is compressed based on channel pruning and mixed-quantization compression technology. The L1-norm pruning algorithm is used for the local electrical quantity analysis engine to calculate the L1-norm of each channel weight in the convolution layer and the fully connected layer, identify and remove channels with L1-norm lower than a preset threshold, and obtain a pruned local electrical quantity analysis engine. The pruned local electrical quantity analysis engine is quantized using the grouped INT8 quantization and key layer FP16 reservation strategy to obtain a quantized local electrical quantity analysis engine. The volume of the quantized local electrical quantity analysis engine is compressed through entropy coding to obtain a compressed local electrical quantity analysis engine, which is deployed to an edge device.

5. The method of claim 1, wherein the knowledge distillation technology is used to migrate the fault feature extraction and pattern recognition capability of the benchmark model deployed in the cloud to the compressed local electrical quantity analysis engine. The standardized multi-modal data is combined to train a multi-modal benchmark model in the cloud as a teacher model. The same standardized multi-modal data is input into the teacher model to obtain the intermediate layer feature map before the output layer of the teacher model and the final output soft label, and the intermediate layer feature map and the soft label of the teacher model are obtained. The compressed local electrical quantity analysis engine is set as a student model, and the student model is trained based on the intermediate layer feature map and the soft label of the teacher model. After the student model is trained, a knowledge-migrated compressed model is obtained, and an edge professional model is generated for deployment on the local.

6. The method of claim 1, wherein the knowledge-embedded causal reasoning large model deployed in the cloud is generated. The standardized multi-modal data is combined with the knowledge of power system operation and power grid fault to generate nodes of the knowledge-embedded multi-level causal graph network, including device-level nodes, fault-level nodes, and causal-level nodes. The edges of the knowledge-embedded multi-level causal graph network are set to connect the nodes and represent the causal relationship and influence relationship, and the weights of the edges represent the strength and probability of the causal relationship. The hierarchical structure of the knowledge-embedded multi-level causal graph network is set to include device layers, fault phenomenon layers, and root cause layers. The knowledge embedding and rule constraints of the knowledge-embedded multi-level causal graph network are constructed to include the prior rules and constraints of the graph network, the deterministic rules of the power grid, and the probabilistic association rules. A neural network is selected to train the knowledge-embedded multi-level causal graph network based on the obtained nodes, edges, hierarchical structure, and knowledge embedding and rule constraints. The trained knowledge-embedded multi-level causal graph network is used to reduce the dimensionality and fuse the input data. For natural language interaction tasks, a prompt template and a prompt are set to guide the output of the knowledge-embedded multi-level causal graph network to comply with the reasoning results of the business specification, and a knowledge-embedded causal reasoning large model deployed in the cloud is generated. ​ ​ ​ 7. The multi-modal collaborative power grid fault intelligent diagnosis method according to claim 6, characterized in that: the knowledge-embedded causal reasoning large model deployed in the cloud comprises: dimensionality reduction and fusion are performed on the multi-modal features input into the cloud causal reasoning large model to generate context feature representations; for natural language interaction tasks, a prompt template and a prompt are set in combination with the context feature representations to guide the knowledge-embedded multi-level causal graph network with reasoning capability of the causal graph network to output a business report conforming to business specifications; in filling in the placeholders in the prompt template and the prompt, the activated fault level nodes and their attributes in the knowledge-embedded multi-level causal graph network are used to incorporate the required specification structure and elements of the business report, the prompt template and the prompt are associated with the knowledge-embedded multi-level causal graph network, and a filled-in fault report conforming to the business specifications is generated, thereby constructing the knowledge-embedded causal reasoning large model deployed in the cloud.

8. The multi-modal collaborative power grid fault intelligent diagnosis method according to claim 1, characterized in that: the edge-cloud fault collaborative diagnosis model comprises: standardized multi-modal data is input into the edge professional model deployed locally, real-time analysis is performed in combination with the local knowledge base, a feature vector containing electrical characteristics and preliminary diagnosis results is extracted, the most likely preliminary fault property is output and a fault brief is generated, and the feature vector and the fault brief are uploaded to the cloud; the feature vector and the fault brief are received in the cloud and set as text semantic features, the text semantic features, environmental data in the standardized multi-modal data, and recording waveform image features are aligned and semantically fused across modalities through a multi-head self-attention mechanism to obtain cross-modal features; the cross-modal features are input into the knowledge-embedded causal graph network, multi-modal information is fused, causal reasoning is performed, a graph network reasoning result is obtained and a fault cause decision path is generated, the confidence of each type of fault root cause hypothesis in the path and the supporting evidence are calculated, a structured prompt template and the graph network reasoning result are combined to generate an edge-cloud collaborative fault diagnosis report, and the edge-cloud fault collaborative diagnosis model is generated.

9. The multi-modal collaborative power grid fault intelligent diagnosis method according to claim 1, characterized in that: the iterative updating of the edge-cloud fault collaborative diagnosis model in combination with the multi-modal fault case library comprises: the newly added multi-modal fault cases in the daily multi-modal fault case library are used to automatically trigger the training and adjustment of the parameters of the edge professional model deployed locally and the graph neural network parameters in the knowledge-embedded causal reasoning large model deployed in the cloud in the edge-cloud fault collaborative diagnosis model; the graph structure is dynamically updated by adding new fault modes, root cause nodes, and connection edges in the knowledge-embedded causal reasoning large model deployed in the cloud, and the weights of the edges in the causal graph are dynamically adjusted using a graph learning algorithm; valid new device parameters, protection logic changes, fault mode descriptions, and environmental impact experiences are updated to the local knowledge base and the cloud power system operation and power grid fault domain knowledge base. A case similarity retrieval algorithm is established to retrieve the similarity with historical cases based on feature vectors and graph node embedding, and the edge-cloud fault collaborative diagnosis model is updated according to the similarity of the most possible new case diagnosis.

10. A power grid fault intelligent diagnosis system based on multi-modal collaboration, which runs an intelligent diagnosis method of power grid fault based on multi-modal collaboration according to any one of claims 1 to 9, characterized in that: a data acquisition module for real-time acquisition and offline import of multi-modal data of power system primary equipment and secondary equipment in multiple spaces of power grid fault range; a standardization module for pre-processing and electrical quantity feature extraction of multi-modal data, and standardizing the pre-processed and feature-extracted multi-modal data to obtain standardized multi-modal data; an edge and cloud model construction module for constructing a local electrical quantity analysis engine combined with standardized multi-modal data, compressing the local electrical quantity analysis engine according to channel pruning and mixed quantization compression technology, migrating the fault feature extraction and pattern recognition capability of the benchmark model deployed in the cloud to the compressed local electrical quantity analysis engine using knowledge distillation technology, constructing an edge professional model deployed locally, combining standardized multi-modal data with power system operation and power grid fault field knowledge to generate rule constraints, constructing a knowledge-embedded multi-level causal graph network, and generating a knowledge-embedded causal reasoning large model deployed in the cloud; a fault collaborative diagnosis model construction module for inputting standardized multi-modal data into the edge professional model deployed locally in sequence, extracting electrical features and preliminarily reasoning fault properties combined with the local knowledge base, inputting the preliminarily reasoned fault properties into the knowledge-embedded causal reasoning large model deployed in the cloud, reasoning the cause of the power grid fault, generating an edge-cloud fault collaborative diagnosis model, and iteratively updating the edge-cloud fault collaborative diagnosis model combined with a multi-modal fault case library; a diagnosis report generation module for inputting the latest collected and imported power grid fault data into the updated edge-cloud fault collaborative diagnosis model, hierarchically and collaboratively reasoning the cause of the power grid fault and generating the latest fault diagnosis report, and realizing intelligent diagnosis of power grid fault based on multi-modal collaboration.

Images