Data integration and multi-mode diagnosis method based on power transmission and distribution scene

Abstract

The invention relates to the technical field of power transmission and distribution production, and discloses a data integration and multi-modal diagnosis method based on a power transmission and distribution scene, and the method comprises the following steps: S1, enhanced integration of multi-source heterogeneous data: collecting time sequence monitoring data, text procedures and image data of power transmission and distribution equipment, constructing an equipment topological correlation graph through a graph attention neural network, and carrying out the enhanced integration of the multi-source heterogeneous data; node feature embedding is optimized through self-supervised comparative learning, an adversarial variational auto-encoder is designed for edge data, and an enhanced sample is generated in combination with physical constraints of equipment. According to the data integration and multi-modal diagnosis method based on the power transmission and distribution scene, the field adaptability and reliability of a diagnosis result are improved while the model fine tuning cost is reduced, and the knowledge migration problem of a general model in the power transmission and distribution scene is solved; the introduction of a dynamic knowledge graph and a multi-dimensional evaluation system realizes the real-time integration of new regulation knowledge and the comprehensive evaluation of model performance, and ensures the sustainable evolution ability and decision transparency of the diagnosis model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission and distribution production, in particular to a data integration and multi-modal diagnosis method based on a power transmission and distribution scene. BACKGROUND

[0002] In the power transmission and distribution production command scene, the existing technology has problems such as a large number of power equipment categories, large differences in data formats and structures, and complex and inefficient processing when dealing with multi-source heterogeneous data, resulting in low data cleaning and preprocessing efficiency and affecting the accuracy of equipment health diagnosis; multi-modal data (text, image, table, etc.) fusion is difficult, the feature alignment accuracy is insufficient, visual errors and text hallucinations are prone to occur, and it is difficult to support deep knowledge understanding; the large model has poor knowledge transferability in the power transmission and distribution field, high fine-tuning cost, insufficient professional knowledge question answering and equipment diagnosis capability, and the reward model is prone to manipulation, and it is difficult to integrate new knowledge (such as technical standards and maintenance procedures), making it difficult to realize continuous learning and evolution; at the same time, there is a lack of a unified evaluation system, which cannot scientifically quantify the multi-dimensional capabilities of the large model in the power transmission and distribution field, restricting its efficient application and continuous optimization in the production command scene.

[0003] Based on this, the present application provides a data integration and multi-modal diagnosis method based on a power transmission and distribution scene. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a data integration and multi-modal diagnosis method based on a power transmission and distribution scene, which has the advantages of enhanced integration of multi-source heterogeneous data, deep interactive fusion of multi-modal features, dynamic adaptation and optimization of diagnosis models, and knowledge enhancement and evaluation closed loop, solving the problems raised in the background art.

[0005] The present application provides the following technical scheme: a data integration and multi-modal diagnosis method based on a power transmission and distribution scene, comprising the following steps:

[0006] S1. Enhanced integration of multi-source heterogeneous data: collect time series monitoring data, text procedures and image data of power transmission and distribution equipment, construct a device topology graph using a graph attention neural network, optimize node feature embedding through self-supervised contrastive learning, design an adversarial variational autoencoder for edge data, generate enhanced samples combined with device physical constraints, and the node attention weight calculation formula in the graph attention neural network is:

[0007] e ij =a(Wh i ,Wh j )

[0008]

[0009] Self-supervised contrastive learning uses InfoNCE loss:

[0010]

[0011] The adversarial variational autoencoder loss function is:

[0012]

[0013] L KL =D KL (q θ (z|x)||p(z));

[0014] S2. Deep interaction fusion of multi-modal features: Construct a spatio-temporal interleaved attention network, the text path adopts a pre-trained power BERT combined with a time convolution network to extract dynamic semantic features, the image path introduces a device region mask mechanism through an improved visual Transformer (ViT), and a bidirectional modal conversion module (text → image generation + image → text description) is used to realize feature mutual enhancement;

[0015] S3. Dynamic adaptation and optimization of the diagnosis model: a dynamic-rank LoRA fine-tuning strategy is proposed, the rank value of the low-rank matrix is adaptively adjusted according to the layer importance, the reward model of reinforcement learning is added with a physical rule consistency score, and the diagnosis decision is optimized through deep deterministic policy gradient (DDPG);

[0016] S4. Knowledge enhancement and evaluation closed loop: an incremental knowledge graph updating mechanism is adopted, entity linking and relationship reasoning are used to realize real-time fusion of new procedures, a physical constraint satisfaction degree index is added to the evaluation system, and the logical necessity of the diagnosis result is verified through causal inference.

[0017] Preferably, the construction method of the graph neural network in S1 comprises: taking power equipment as nodes and electrical connection relationship between devices as edges to construct an initial topological graph, using a node embedding algorithm to map device attribute data into a high-dimensional feature vector, capturing the dependency relationship between nodes through a graph convolution layer, and outputting a structured device correlation feature matrix.

[0018] Preferably, the improved variational autoencoder in S1 comprises: introducing an attention mechanism in the encoder to assign higher reconstruction weight to edge data samples, using a residual connection structure in the decoder to improve the robustness to noise data, optimizing model parameters through weighted sum of KL divergence and reconstruction loss, and realizing adaptive enhancement of edge data.

[0019] Preferably, the dual-path network of the spatio-temporal attention mechanism in S2 includes: the text path uses a bidirectional LSTM to extract time sequence features, and an attention weight matrix constructed by combining a domain dictionary to strengthen the representation of professional terms; the image path extracts spatial features through a convolutional neural network, introduces a timestamp encoding layer to fuse the time sequence information of the device state, and uses a mutual attention mechanism in the cross-modal interaction layer to calculate a similarity matrix of the text-image features, and dynamically adjusts the modal fusion weight through a gating unit.

[0020] Preferably, the mixed fine-tuning strategy in S3 includes: using the LoRA technology to inject a low-rank matrix to the bottom layer of the pre-trained model, using an adapter module to adapt the domain features to the top layer, using gradient clipping to suppress overfitting, and using a learning rate decay strategy to balance the learning weights of global knowledge and domain knowledge.

[0021] Preferably, the reinforcement learning algorithm of the dynamic reward mechanism in S3 includes: the reward model is composed of a fact accuracy score, a time sequence consistency score, and a domain compliance score, wherein the fact accuracy score is calculated by comparing the diagnosis result with the actual state of the device, the time sequence consistency score is obtained by analyzing the change trend of the continuous time diagnosis result, and the domain compliance score is generated based on the rule matching of the latest maintenance regulations.

[0022] Preferably, the construction method of the dynamic knowledge graph in S3 includes: using a time sequence knowledge graph to store the version iteration information of the device regulations, using a time-aware graph attention mechanism to retrieve the most relevant knowledge for the current diagnosis task, and using an incremental learning strategy in the knowledge update module to automatically update the attribute information of the graph nodes and edges after entity recognition and relationship extraction of newly entered regulation documents.

[0023] Preferably, the multi-dimensional diagnosis evaluation system in S4 includes: a robustness index determined by calculating the diagnosis accuracy rate reduction rate under an adversarial sample (generated by an FGSM algorithm), an explainability index obtained by calculating the similarity of feature importance sorting and expert decision path, and a time sequence consistency index evaluated by analyzing the Markov chain transition probability matrix of continuous diagnosis results.

[0024] Preferably, the output of the time sequence convolution network is:

[0025]

[0026] The visual Transformer region mask probability is:

[0027]

[0028] The cross-modal conversion loss is:

[0029]

[0030] The dynamic rank LoRA parameter update is:

[0031] W = W0 + B l A l , rank(B l ) = rank(A l ) = r(l)

[0032] The reinforcement learning reward function is:

[0033] R = alpha * R fact + beta * R temp + gamma * R reg + delta * R phys

[0034] The entity link similarity calculation is:

[0035]

[0036] The robustness index calculation formula is:

[0037] Compared with the prior art, the present application has the following beneficial effects:

[0038] The data integration and multi-modal diagnosis method based on power transmission and distribution scenarios, by adopting a graph neural network to construct a device topology correlation graph, and combining an improved variational autoencoder to enhance edge data, improves the integration efficiency and integrity of multi-source heterogeneous data, solves the problems of large data format difference and low utilization rate of edge data; a dual-path network with a spatio-temporal attention mechanism, strengthens the temporal and spatial correlation of text-image data, and through a dynamic gating unit, adapts the modal weight of different diagnosis tasks, improves the fusion accuracy of multi-modal features, and reduces the cross-modal semantic deviation; a reinforcement learning algorithm combining a hybrid fine-tuning strategy and a dynamic reward mechanism reduces the model fine-tuning cost while improving the field adaptability and reliability of the diagnosis result, solving the knowledge transfer problem of general models in the power transmission and distribution scenario; the introduction of a dynamic knowledge graph and a multi-dimensional evaluation system realizes the real-time integration of new regulation knowledge and the comprehensive evaluation of model performance, ensuring the continuous evolution ability and decision transparency of the diagnosis model. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The method step diagram of the present application. DETAILED DESCRIPTION

[0040] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0041] Please refer to Figure 1 , the data integration and multi-modal diagnosis method based on power transmission and distribution scenarios includes the following steps:

[0042] S1. Enhanced integration of multi-source heterogeneous data: Collecting time series monitoring data, text regulations and image data of power transmission and distribution equipment, constructing a device topology graph using a graph attention neural network, optimizing node feature embedding through self-supervised contrastive learning, designing an adversarial variational autoencoder for edge data, and generating enhanced samples combined with device physical constraints;

[0043] The node attention weight calculation formula in the graph attention neural network is:

[0044] e ij =a(Wh i ,Wh j )

[0045]

[0046] Self-supervised contrastive learning uses InfoNCE loss:

[0047]

[0048] The loss function of the adversarial variational autoencoder is:

[0049]

[0050] L KL =D KL (q θ (z|x)||p(z))

[0051] S2. Deep interactive fusion of multi-modal features: Constructing a spatio-temporal interwoven attention network, using a pre-trained power BERT combined with a time series convolutional network to extract dynamic semantic features in the text path, and introducing a device region mask mechanism through an improved visual Transformer (ViT) in the image path, and using a bidirectional modal conversion module (text→image generation + image→text description) to realize feature mutual enhancement between cross-modal layers;

[0052] S3. Dynamic adaptation optimization of diagnostic model: a dynamic-rank LoRA fine-tuning strategy is proposed to adaptively adjust the rank of low-rank matrix according to layer importance, and a physical rule consistency score is added to the reinforcement learning reward model. Deep deterministic policy gradient (DDPG) is used to optimize the diagnostic decision;

[0053] S4. Knowledge enhancement and evaluation closed loop: an incremental knowledge graph updating mechanism is adopted to realize real-time integration of new procedures through entity linking and relation reasoning. A physical constraint satisfaction degree index is added to the evaluation system to verify the logical necessity of the diagnostic results through causal inference.

[0054] By using graph neural network to construct device topology correlation graph, and combining improved variational autoencoder to enhance edge data, the integration efficiency and integrity of multi-source heterogeneous data are improved, solving the problem of large data format difference and low utilization rate of edge data. The dual-path network with spatio-temporal attention mechanism strengthens the temporal and spatial correlation of text-image data, and through the dynamic gating unit, the modal weight of different diagnostic tasks is adapted to improve the fusion accuracy of multi-modal features and reduce the cross-modal semantic deviation. The reinforcement learning algorithm combining mixed fine-tuning strategy and dynamic reward mechanism reduces the model fine-tuning cost while improving the domain adaptability and reliability of the diagnostic results, solving the knowledge transfer problem of general models in power transmission and distribution scenarios. The introduction of dynamic knowledge graph and multi-dimensional evaluation system realizes the real-time integration of new procedure knowledge and the comprehensive evaluation of model performance, ensuring the continuous evolution ability and decision transparency of the diagnostic model

[0055] In S1, the construction method of graph neural network includes: taking power equipment as node and electrical connection relationship between devices as edge to construct initial topology graph, using node embedding algorithm to map device attribute data into high-dimensional feature vector, and capturing the dependency relationship between nodes through graph convolution layer to output structured device correlation feature matrix.

[0056] In S1, the improved variational autoencoder includes: introducing attention mechanism in the encoder to assign higher reconstruction weight to edge data samples, using residual connection structure in the decoder to improve the robustness to noisy data, and optimizing model parameters through weighted sum of KL divergence and reconstruction loss to realize adaptive enhancement of edge data.

[0057] In S2, the dual-path network with spatio-temporal attention mechanism includes: text path uses bidirectional LSTM to extract temporal features, and combines attention weight matrix constructed by domain dictionary to strengthen professional term representation; image path extracts spatial features through convolutional neural network and introduces timestamp encoding layer to fuse temporal information of device state; cross-modal interaction layer uses mutual attention mechanism to calculate similarity matrix of text-image features, and dynamically adjusts modal fusion weight through gating unit.

[0058] Wherein, the mixed fine-tuning strategy in S3 includes: using LoRA technology to inject low-rank matrices to the bottom layer of the pre-trained model, and using an adapter module to adapt domain features to the top layer; suppressing overfitting by gradient clipping, and balancing the learning weights of global knowledge and domain knowledge by using a learning rate decay strategy.

[0059] Wherein, the reinforcement learning algorithm of the dynamic reward mechanism in S3 includes: the reward model is composed of fact accuracy score, temporal consistency score and domain compliance score, wherein the fact accuracy score is calculated by comparing the diagnosis result with the actual state of the equipment, the temporal consistency score is obtained by analyzing the change trend of continuous diagnosis results, and the domain compliance score is generated based on the latest maintenance regulations.

[0060] Wherein, the construction method of the dynamic knowledge graph in S3 includes: using a time-series knowledge graph to store version iteration information of equipment regulations, and retrieving the most relevant knowledge for the current diagnosis task through a time-aware graph attention mechanism; the knowledge update module uses an incremental learning strategy to automatically update the attribute information of graph nodes and edges after performing entity recognition and relationship extraction on newly entered regulation documents.

[0061] Wherein, the multi-dimensional diagnosis evaluation system in S4 includes: the robustness index is determined by calculating the decrease rate of diagnosis accuracy under adversarial samples (generated by FGSM algorithm); the explainability index is obtained by calculating the similarity between feature importance ranking and expert decision path; and the temporal consistency index is evaluated by analyzing the Markov chain transition probability matrix of continuous diagnosis results.

[0062] Wherein, the output of the temporal convolution network is:

[0063]

[0064] The visual Transformer region mask probability is:

[0065]

[0066] The cross-modal conversion loss is:

[0067]

[0068] The dynamic rank LoRA parameter update is:

[0069] W = W0 + B l A l , rank(B l ) = rank(A l ) = r(l)

[0070] The reinforcement learning reward function is:

[0071] R = aR fact + bR temp + gR reg + dR phys

[0072] The entity link similarity calculation is:

[0073]

[0074] The robustness index calculation formula is:

[0075] Example one

[0076] Taking the transformer state diagnosis of a 110 kV substation as an example

[0077] Step 1: Data integration

[0078] Collection stage: For the core state parameters of the transformer, multiple types of sensors are combined to collect data - the oil temperature is collected every 5 minutes by a platinum resistance temperature sensor (precision ±0.5℃), the oil level is monitored in real time by an ultrasonic sensor (range 0-100%, error ±1%), and the winding temperature is continuously recorded by a distributed optical fiber sensor (spatial resolution 1m); the text regulations cover 12 types of current standard documents such as "DL / T 572-2010 Power Transformer Operation Regulations" and "Q / GDW 1168-2013 Power Transmission and Transformation Equipment Condition-based Maintenance Test Regulations", including key clauses such as "winding temperature limit" and "insulating oil quality requirements"; image data includes 3 infrared thermal images per day (resolution 640x512, temperature measurement range -20~150℃), and 1 appearance inspection image per hour (high-definition camera, focusing on key parts such as bushings, oil tanks, and tap changers).

[0079] Topology correlation graph construction: taking the transformer ontology (node ID: T001) as the center node, the associated nodes are refined to 12 auxiliary devices including cooling system (including submersible pump P001, cooling fan F001-F004), high-voltage bushing (B001-B003), low-voltage bushing (B004-B006), tap changer (S001), and gas relay (R001); the edge attribute is labeled with electrical connection type (such as "hard copper bar connection" and "cable connection") and rated current (such as 400A and 630A). Node embedding uses the GraphSAGE algorithm to map 15-dimensional attributes such as device rated capacity (such as 50MVA), date of manufacture (2020-06-15), and historical fault frequency (3 times) to 128-dimensional feature vectors; through 2 layers of graph convolution layers (32 neurons per layer), the node dependency relationship is captured, and finally a 12x128 correlation feature matrix is output (row represents node, column represents feature).

[0080] Edge data augmentation: focusing on 3 typical edge data (oil temperature rise ≥5℃ / 10min, oil level drop ≥3% / 5min, winding temperature jump ≥8℃ / min), improved adversarial variational autoencoder parameter settings: encoder contains 3 layers of LSTM (hidden layer dimension 128), attention mechanism assigns 0.8 weight to edge samples (0.2 to normal samples); decoder uses 3 layers of residual convolution layer (kernel_size=3), when reconstructing edge samples, constraints such as "oil temperature ≤105℃" and "oil level ≥15%" are applied; combined optimization by KL divergence (weight 0.2) and MSE reconstruction loss (weight 0.8), 800 augmented samples are generated for each type of edge data, increasing the proportion of edge data from 5% to 20%;

[0081] Step 2: Multimodal fusion

[0082] Text path feature extraction: pre-trained power BERT based on Transformer (12-layer encoder, hidden layer 768 dimensions), fine-tuned on "Electric Power Equipment Overhaul Procedures Compilation" (500,000 corpus), input "transformer winding overheating treatment procedures" text segment (length 512 tokens), output static semantic vector; Time convolution network (TCN) contains 3 layers of 1D convolution (kernel_size=3, step 1, padding=1), extracts dynamic features of "temperature threshold changes with load" in the text (such as "winding temperature ≤105℃ at full load, ≤90℃ at light load"), and finally fused into a 128-dimensional text feature vector.

[0083] Image path feature extraction: improved visual Transformer (ViT) input infrared thermal image (224×224 pixels, 3 channels), divided into 16×16 patches (each 14×14 pixels), generating 196 patch vectors (768 dimensions); region mask mechanism uses 20% mask rate for "winding region" and "sleeve joint" key areas (30%), and 50% mask rate for "background area" (70%), to strengthen key area feature learning; Introduce timestamp encoding (convert collection time 2024-07-1514:30 to sine cosine encoding vector), and after fusion with patch features, output 128-dimensional image feature vector.

[0084] Cross-modal interaction enhancement: In the bidirectional modal conversion module, the text→image generation adopts the Stable Diffusion model, inputs the "thermal image of winding temperature 95℃" text description, and generates simulated thermal image features (cosine similarity with real image features ≥0.85); the image→text description adopts the BLIP model, generates the "sleeve A phase temperature 82℃, 15℃ higher than B phase" text features for the infrared thermal image; calculate the cosine similarity matrix (128×128) of text-image features, dynamically adjust the weight through the gating unit (sigmoid function) - in the overheat diagnosis task, the image modality (temperature visualization information) weight is 0.6, and the text modality (regulation threshold) weight is 0.4, output 256-dimensional fusion feature vector.

[0085] Step 3: Model optimization

[0086] Dynamic rank LoRA fine-tuning: Based on the pre-trained BERT model, calculate the layer importance through gradient norms (top layer > middle layer > bottom layer), and adaptively set the low-rank matrix rank value: bottom layer (1-4 layers) rank = 4 (focus on general features), middle layer (5-8 layers) rank = 8 (balance general and domain features), top layer (9-12 layers) rank = 12 (strengthen domain features); additionally add an Adapter module (2-layer MLP, hidden layer 256 dimensions) to the top layer to adapt to "transformer-specific fault patterns" (such as "poor contact of tap changer leading to overheating"). The gradient clipping threshold is set to 1.0 (to suppress gradient explosion) during training, the initial learning rate is 5e-5, and the cosine annealing strategy is used (decay by 50% every 100 epochs) to balance the learning weights of global knowledge (general language model) and domain knowledge (power regulations).

[0087] Reinforcement learning optimization of diagnostic decision: The reward model is composed of 4 scores weighted (total score is 10): factual accuracy score (4 points) - compare the diagnostic results with the actual measured values of the optical fiber sensor (error ≤2℃ gets 4 points, 2-5℃ gets 2 points, >5℃ gets 0 points); time sequence consistency score (2 points) - analyze the Markov chain transition probability of continuous 3-hour diagnostic results (such as "normal → slight abnormal → severe abnormal" reasonable transition gets 2 points, jump transition gets 0 points); domain compliance score (2 points) - match the "overheating handling process" in the "Transformer Maintenance Guidelines" (full match gets 2 points, partial match gets 1 point); physical rule consistency score (2 points) - verify that the "oil temperature and winding temperature difference ≤10℃" (satisfies gets 2 points, otherwise deducts 1 point). Use deep deterministic policy gradient (DDPG) optimization: the Actor network (3-layer MLP, hidden layer 256 dimensions) outputs the diagnostic decision (such as "need to shut down for maintenance" "strengthen monitoring"), the Critic network (double network structure) evaluates the decision value, and the experience replay pool stores 10^6 samples, and updates the network parameters every 500 steps.

[0088] Dynamic knowledge graph update: Adopt time knowledge graph (TKG) storage procedure version iteration, nodes contain "procedure ID, release time, core clauses", edges contain "updated from, replaced" relationships (such as 2024 version of "Transformer Maintenance Guide" replaces 2019 version); Through time-aware graph attention mechanism (assigning 0.7 attention weight to procedures in the past 3 years), retrieve the most relevant clauses for the current task (overheating diagnosis in July 2024); When new 2024 version procedures are entered, use BERT-CRF model to identify entities such as "insulating oil breakdown voltage ≥ 30kV", and TPLinker model to extract relationships such as "insulating oil → detection period → 1 year", automatically update the attribute information of 23 nodes and 15 edges in the graph.

[0089] Step 4: Evaluation optimization

[0090] Incremental knowledge fusion: Entity linking uses entity similarity calculation (Jaccard coefficient ≥ 0.8) to link "smart patrol robot detection standard" in new procedures to "patrol method" node in the graph; Relationship reasoning uses TransE algorithm to infer new relationships such as "winding temperature too high → need to detect insulating oil degradation" from "winding temperature too high → need to detect insulating oil" and "insulating oil degradation → winding poor heat dissipation", and supplement to the knowledge graph.

[0091] Multi-dimensional evaluation system:

[0092] Robustness index: Use FGSM algorithm to generate adversarial samples (perturbation ε = 0.01), calculate the diagnosis accuracy rate from 92% to 88%, the decrease rate is 4.3% (<5%, meet the requirements);

[0093] Explainability index: Sort feature importance by SHAP value (infrared temperature > winding sensor data > procedure threshold), and the edit distance similarity of expert decision path ("first look at infrared thermal image, then check sensor data, finally compare with procedure") is 0.85;

[0094] Temporal consistency index: In the Markov chain transition matrix of continuous diagnosis results, the probability of "normal → slight abnormal" is 0.2, and the probability of "slight abnormal → serious abnormal" is 0.3, which conforms to the transformer degradation law (score 0.91);

[0095] Physical constraint satisfaction degree: In 1000 diagnoses, 985 meet the constraints such as "oil temperature ≤ 105℃" and "oil level ≥ 15%" (satisfaction degree 98.5%).

[0096] Based on the evaluation results, the gating unit weights are dynamically adjusted through reinforcement learning. Since the infrared image is more critical in the overheat positioning, its weight is increased from 0.6 to 0.65. The final diagnosis accuracy is increased to 94%, and the fault positioning accuracy (such as distinguishing "winding overheat" from "sleeve overheat") reaches 92%.

[0097] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary and that changes in form and detail can be made without departing from the principles and spirit of the application. The scope of the application is therefore defined by the appended claims and their equivalents.

Claims

1. A data integration and multi-modal diagnosis method based on power transmission and distribution scenarios, characterized in that, Comprising the following steps: S1. Enhanced integration of multi-source heterogeneous data: Collect time-series monitoring data, text regulations and image data of power transmission and distribution equipment, construct a device topology graph using a graph attention neural network, optimize node feature embedding through self-supervised contrastive learning, design an adversarial variational autoencoder for edge data, and generate enhanced samples combined with device physical constraints. The node attention weight calculation formula in the graph attention neural network is: e ij = a(Wh i , Wh j ) Self-supervised contrastive learning uses InfoNCE loss: The adversarial variational autoencoder loss function is: L KL = D KL (q θ (z|x)||p(z)) S2. Deep interactive fusion of multi-modal features: Construct a spatio-temporal interwoven attention network, use pre-trained power BERT combined with time-series convolutional network to extract dynamic semantic features for text path, introduce device region mask mechanism through improved visual Transformer (ViT) for image path, and implement feature mutual enhancement through bidirectional modal conversion module (text->image generation + image->text description) across modal layers; S3. Dynamic adaptation and optimization of diagnosis model: Propose a dynamic-rank LoRA fine-tuning strategy to adaptively adjust the rank value of the low-rank matrix according to the layer importance, add physical rule consistency score to the reinforcement learning reward model, and optimize the diagnosis decision through deep deterministic policy gradient (DDPG); S4. Knowledge enhancement and evaluation closed loop: Use incremental knowledge graph update mechanism to realize real-time fusion of new regulations through entity linking and relationship reasoning, add physical constraint satisfaction degree index to the evaluation system, and verify the logical necessity of the diagnosis result through causal inference.

2. The method of claim 1, wherein the method is based on power distribution scenarios. The construction method of the graph neural network in S1 includes: taking power equipment as nodes and electrical connection relationship between devices as edges to construct an initial topology graph, using node embedding algorithm to map device attribute data into high-dimensional feature vector, capturing the dependency relationship between nodes through graph convolution layer, and outputting structured device correlation feature matrix. 3.The data integration and multi-modal diagnosis method based on power distribution scenario according to claim 1, characterized in that: The improved variational autoencoder in S1 includes: introducing attention mechanism in the encoder to assign higher reconstruction weight to edge data samples, using residual connection structure in the decoder to improve the robustness to noisy data, optimizing model parameters through weighted sum of KL divergence and reconstruction loss, and realizing adaptive enhancement of edge data. 4.The data integration and multi-modal diagnosis method based on power distribution scenario according to claim 1, characterized in that: The dual-path network with spatio-temporal attention mechanism in S2 includes: using bidirectional LSTM to extract time-series features for text path, combining attention weight matrix constructed based on domain dictionary to strengthen professional term representation, extracting spatial features through convolutional neural network for image path, introducing timestamp encoding layer to fuse time-series information of device state, using mutual attention mechanism to calculate similarity matrix of text-image features in cross-modal interaction layer, and dynamically adjusting modal fusion weight through gate unit. 5.The data integration and multi-modal diagnosis method based on power distribution scenario according to claim 1, characterized in that: The mixed fine-tuning strategy in S3 includes: using LoRA technology to inject low-rank matrix to the bottom layer of pre-trained model, using adapter module for domain feature adaptation to the top layer, suppressing overfitting through gradient clipping, and balancing learning weights of global knowledge and domain knowledge through learning rate decay strategy. 6.The data integration and multi-modal diagnosis method based on power distribution scenario according to claim 1, characterized in that: The reinforcement learning algorithm of the dynamic reward mechanism in S3 includes that the reward model is composed of fact accuracy score, time sequence consistency score and domain compliance score, the fact accuracy score is calculated by comparing the diagnosis result with the actual state of the equipment, the time sequence consistency score is obtained by analyzing the change trend of the diagnosis result at continuous time, and the domain compliance score is generated based on the rule matching of the latest maintenance procedure. 7.The data integration and multi-modal diagnosis method based on power distribution scenario according to claim 1, characterized in that: The construction method of the dynamic knowledge graph in S3 includes that the version iteration information of the equipment procedure is stored by using the time sequence knowledge graph, the most relevant knowledge for the current diagnosis task is retrieved through the time-aware graph attention mechanism, and the knowledge update module adopts the incremental learning strategy to automatically update the attribute information of the graph nodes and edges after entity recognition and relation extraction are performed on the newly entered procedure document. 8.The data integration and multi-modal diagnosis method based on power distribution scenario according to claim 1, characterized in that: The multi-dimensional diagnosis evaluation system in S4 includes that the robustness index is determined by calculating the diagnosis accuracy reduction rate under the adversarial sample (generated by using the FGSM algorithm), the explainability index is obtained by calculating the similarity of the feature importance sorting and the expert decision path, and the time sequence consistency index is evaluated by analyzing the Markov chain transition probability matrix of the continuous diagnosis result. 9.The data integration and multi-modal diagnosis method based on power distribution scenario according to claim 1, characterized in that: The output of the time sequence convolution network is: The visual Transformer region mask probability is: The cross-modal conversion loss is: The dynamic rank LoRA parameter update is: W = W0+ B l A l , rank(B l ) = rank(A l ) = r(l) The reinforcement learning reward function is: R = aR + βR fact + γR temp + δR reg + εR phys The entity link similarity calculation is: The robustness index calculation formula is:

Images