A multi-modal power grid health assessment and analysis method

Abstract

The present invention proposes a multimodal-based power grid health assessment and analysis method, which includes deploying multiple sensors, collecting multimodal data of power grid equipment, fusing multimodal features, and forming a comprehensive assessment model; constructing a federated learning model; introducing a self-supervised learning mechanism in the operation data of power grid equipment, automatically generating pseudo labels for the operation data and training a deep learning model, while collecting the operation data, maintenance records, and fault history of power grid equipment to construct a device knowledge graph; identifying normal operation data and abnormal patterns, and identifying defects based on abnormal patterns to locate the fault points of power grid equipment; constructing a fault prediction model to evaluate the health status of power grid equipment and predict potential faults. The present invention achieves comprehensive monitoring, accurate diagnosis, and dynamic maintenance of power grid equipment by introducing multiple advanced artificial intelligence algorithms and technologies.

Description

Technical Field

[0001] The present invention belongs to the field of power grid status monitoring, and in particular relates to a multi-modal-based power grid health assessment and analysis method. Background Art

[0002] With the continuous development of the power system and the increasing scale of the power grid, the operation and maintenance of power grid equipment has become more complex and important. The existing monitoring and maintenance of power grid equipment mainly rely on regular inspections and manual maintenance. This method is not only inefficient, but also prone to missed detections and false detections, making it difficult to ensure the safe and stable operation of the power grid. In recent years, with the rapid development of big data and artificial intelligence technologies, data-driven power grid equipment monitoring and maintenance methods have gradually emerged. However, the current technology still has some obvious shortcomings:

[0003] Limitations of single-mode data: Most existing power grid monitoring systems rely on single-mode data (such as temperature, vibration, image, etc.) for analysis. This single-mode data often cannot fully reflect the health status of the equipment when faced with complex power grid equipment operating conditions, resulting in insufficient accuracy and reliability of monitoring results.

[0004] Data privacy and security issues: Traditional centralized data storage and processing methods are prone to data leakage and security risks, so data privacy protection is particularly important.

[0005] Insufficient accuracy and timeliness of fault diagnosis: Most existing fault diagnosis systems rely on predefined rules and models, lack intelligence and adaptive capabilities, and are unable to cope with the complex and changing power grid operating environment. Due to the lack of real-time intelligent diagnosis methods, fault identification and processing are often not timely, affecting the reliability and stability of the power grid.

[0006] Insufficient optimization of maintenance strategies: Existing power grid maintenance strategies mainly rely on experience and predetermined plans, lacking a dynamic adjustment mechanism based on real-time data, making it difficult to achieve accurate maintenance and optimization of equipment.

[0007] Faced with the above problems, how to effectively use big data and artificial intelligence technologies to achieve multimodal monitoring, intelligent diagnosis and dynamic maintenance of power grid equipment has become an important issue in current power system research and application. Summary of the invention

[0008] The purpose of this invention is to design a multimodal power grid health assessment and analysis method, which realizes comprehensive monitoring, accurate diagnosis and dynamic maintenance of power grid equipment by introducing a variety of advanced artificial intelligence algorithms and technologies, and at the same time introducing self-supervised learning and federated learning for power grid equipment health assessment, fault diagnosis and maintenance optimization.

[0009] In order to achieve the above object, the present invention provides a multi-modal power grid health assessment and analysis method, the method comprising:

[0010] Step S1, deploying a variety of sensors to collect multimodal data of power grid equipment in real time, preprocessing the multimodal data, extracting features from the preprocessed multimodal data to form multimodal features, and fusing the multimodal features to form a comprehensive evaluation model; wherein the multimodal data includes temperature data, image data and vibration data; and the multimodal features include temperature features, image features and vibration features;

[0011] Step S2: construct a federated learning model, perform local federated learning model training on multiple distributed nodes, aggregate model parameters using a federated averaging algorithm, and update and distribute the aggregated global model to each node;

[0012] Step S3: introduce a self-supervised learning mechanism into the operation data of the power grid equipment, automatically generate pseudo labels for the operation data and train the deep learning model, and at the same time collect the operation data, maintenance records and fault history of the power grid equipment to build a device knowledge graph;

[0013] Step S4: Identify normal operation data and abnormal patterns, identify defects based on the abnormal patterns, and locate the fault point of the power grid equipment;

[0014] Step S5: construct a fault prediction model to evaluate the health status of power grid equipment and predict potential faults;

[0015] In step S2, a federated learning model is constructed, local federated learning model training is performed on multiple distributed nodes, model parameters are aggregated using a federated averaging algorithm, and the aggregated global model is updated and distributed to each node, specifically including:

[0016] S201. Initialize a global model on the central server and distribute its parameters to each edge node, as shown below:

[0017] ,

[0018] ,

[0019] in, represents the initial parameters of the global model, Indicates that the initial model parameters Distribute to each edge node;

[0020] S202, after receiving the global model parameters, each edge node uses local data to perform global model training, as shown below:

[0021] ,

[0022] in, represents the model parameters of the i-th node at the t-th iteration, is the learning rate, is the loss function Parameters The gradient of Represents a local dataset;

[0023] S203, after each edge node completes the local global model training, it sends the updated model parameters back to the central server, as shown below:

[0024] ,

[0025] in, Indicates sending updated model parameters back to the central server;

[0026] S204: After receiving the model parameters updated by all edge nodes, the central server performs weighted average aggregation to generate new global model parameters, which are expressed as follows:

[0027] ,

[0028] in, represents the global model parameters after the t+1th iteration, K is the number of edge nodes, N is the total number of data samples of all nodes, is the number of data samples of the i-th node;

[0029] S205: Distribute the updated global model parameters to each edge node, as shown below:

[0030] ,

[0031] in, Indicates distribution to each edge node;

[0032] S206, perform federated learning iterations until a predetermined number of iterations is reached, as shown below:

[0033] :

[0034] :

[0035] ,

[0036] ,

[0037] ,

[0038] ;

[0039] S207. After the training is completed, the global model is verified and tested on the central server.

[0040] Optionally, the method further comprises:

[0041] Step S6: Construct a power grid equipment maintenance strategy and perform dynamic optimization and closed-loop control:

[0042] Collect multimodal data in real time and upload it to the central server, build a dynamic maintenance optimization model and train the dynamic maintenance optimization model to learn the optimal maintenance strategy, define a multi-objective optimization function to optimize the optimal maintenance strategy of power grid equipment, and provide real-time feedback and adjustment of the maintenance strategy based on the real-time operation data of the equipment;

[0043] Step S7: perform system integration and application:

[0044] Integrate steps S1-S6 into the system framework, implement data interaction and function calls of each module through the API interface, and deploy them to the actual power grid environment.

[0045] Optionally, in step S1, the temperature feature includes statistical features and temperature frequency domain features; the image feature includes edge features and thermal image features; the vibration feature includes time domain features and vibration frequency domain features;

[0046] The comprehensive assessment model includes:

[0047] The temperature features, image features and vibration features are concatenated into a high-dimensional feature vector F, which is expressed as follows:

[0048] ,

[0049] in, represents the temperature eigenvector, represents the vibration eigenvector, represents the image feature vector;

[0050] Use principal component analysis PCA to reduce the dimension of high-dimensional feature vectors, as shown below:

[0051] ,

[0052] in, , and are eigenvectors, eigenvalues ​​and eigenvector matrices respectively, Represents the feature vector after dimensionality reduction;

[0053] For the feature vector Min-Max normalization is performed to scale the eigenvalues ​​to the [0,1] interval, as shown below:

[0054] ,

[0055] in, represents the normalized feature vector;

[0056] For the feature vector Standardization is performed so that the mean of the eigenvector is 0 and the standard deviation is 1, as shown below:

[0057] ,

[0058] in, and They represent the mean and standard deviation of the normalized eigenvectors respectively.

[0059] Optionally, the verifying and testing the global model on the central server includes:

[0060] The verification is expressed as follows:

[0061] ,

[0062] in, Indicates that the global model is evaluated on the validation dataset performance;

[0063] The test is expressed as follows:

[0064] ,

[0065] in, Indicates testing the model on the test data set performance and obtain the final result.

[0066] Optionally, the step S3 specifically includes:

[0067] S301, construct a deep learning model, introduce a self-supervised learning mechanism, automatically generate pseudo labels for deep learning model training to form a multimodal deep learning model;

[0068] S302, constructing a knowledge graph;

[0069] S303: Integrate the deep learning model and the knowledge graph to diagnose and predict the faults of power grid equipment.

[0070] Optionally, the loss function of the deep learning model is Defined as:

[0071] ,

[0072] in, Representation vector and The similarity of Representation sample The representation vector of Representation sample The representation vector of is the temperature parameter, Representation vector and The similarity of Representation sample The representation vector of

[0073] The loss function of the multimodal deep learning model Defined as:

[0074] ,

[0075] in, and The feature extraction networks representing temperature data and vibration data respectively, and They are temperature and vibration data respectively;

[0076] The triples of the knowledge graph are expressed as follows:

[0077] ,

[0078] in, A set of triples representing a knowledge graph, is a triple, representing an entity With entity Through the relationship connected;

[0079] Using graph neural network for knowledge graph reasoning is expressed as follows:

[0080] ,

[0081] in, represents the representation vector of node v at layer l+1, represents the representation vector of node v at layer l, represents the set of neighbor nodes of node v, represents the normalization coefficient, represents the weight matrix of the lth layer, represents the activation function;

[0082] The fusion of the deep learning model and the knowledge graph includes: concatenating the representation vector obtained by self-supervised learning with the node representation vector of the knowledge graph, as shown below:

[0083] ,

[0084] in, represents the fused representation vector, represents the representation vector obtained by self-supervised learning, Represents the representation vector obtained by knowledge graph reasoning;

[0085] Design a fault diagnosis loss function and train the fault diagnosis model on the fused representation vector; wherein the fault diagnosis loss function , which is expressed as follows:

[0086] ,

[0087] in, represents the cross entropy loss function, represents the fault diagnosis model, Indicates a fault tag.

[0088] Optionally, the step S4 specifically includes:

[0089] S401, using a generative adversarial network to generate normal operation data samples; wherein the generator Used to generate simulated normal operation data samples, discriminator Used to distinguish between real data and generated data;

[0090] S402: Use adversarial training to detect abnormal patterns in device operation data and set thresholds , if the anomaly score exceeds the threshold, the current sample is marked as an anomaly point;

[0091] S403, combine convolutional neural network to identify defects and locate equipment failure points:

[0092] Extract and enhance the features of the abnormal points, build a defect classification model and use the multimodal features of step S1 to train the defect classification model, input the feature-extracted and enhanced abnormal point data into the defect classification model to identify the fault type, and analyze the identified fault type in combination with the knowledge graph to generate maintenance suggestions and repair plans; wherein the loss function of the defect classification model is It is expressed as follows:

[0093] ,

[0094] in, represents the defect classification model, Indicates the fault tag, represents the cross entropy loss function.

[0095] Optionally, the step S5 specifically includes:

[0096] S501. Constructing comprehensive health assessment indicators based on temperature features, image features and vibration features , and the exponential smoothing method is used to calculate the health index for the comprehensive health assessment index , according to the health index Classify health levels;

[0097] S502. Use long short-term memory network to build a fault prediction model to predict the future health index of the equipment ;

[0098] S503. Based on the predicted health index , estimate the probability of failure , and according to the probability Predict when equipment will fail ;

[0099] S504, according to the health index and probability Combined, calculate the comprehensive health score ;

[0100] S505, based on comprehensive health score Develop equipment operation and maintenance strategies;

[0101] S506: Verify and optimize the fault prediction model.

[0102] Optionally, the dynamic maintenance optimization model includes:

[0103] The equipment operation state is defined as the state vector, and the executable maintenance action is defined as the action space. The reward function is designed as:

[0104] ,

[0105] in, represents the reward function, Represents the reward value, Indicates maintenance action The cost, Indicates the change in health index;

[0106] The dynamic maintenance optimization model is trained using the deep Q network algorithm to learn the optimal maintenance strategy, which is expressed as follows:

[0107] ,

[0108] in, represents the Q-value function, Represents the discount factor.

[0109] Optionally, the multi-objective optimization function , which is expressed as follows:

[0110] ,

[0111] in, is the total maintenance cost, is the average health index, is the average failure rate.

[0112] The beneficial technical effects of the present invention are at least as follows:

[0113] (1) The present invention fuses multiple sensor data (such as temperature, vibration, image, etc.), uses a deep learning model to extract and fuse features of multimodal data, and constructs a comprehensive evaluation model to achieve a comprehensive evaluation of the health status of power grid equipment. This overcomes the limitations of single-modal data and improves the accuracy and reliability of monitoring results.

[0114] (2) The present invention adopts a federated learning algorithm to perform model training on multiple distributed nodes, and realizes joint modeling of distributed data through methods such as FedAvg, avoiding the security risks brought by centralized data storage and processing, while making full use of distributed data resources.

[0115] (3) The present invention combines self-supervised learning and knowledge graph technology to achieve intelligent fault diagnosis of power grid equipment. Self-supervised learning improves the recognition ability of unlabeled data by automatically generating pseudo-labels for model training; knowledge graphs provide fault cause analysis and repair suggestions by structured representation of equipment operation data and fault history, significantly improving the accuracy and timeliness of fault diagnosis.

[0116] (4) Generative adversarial networks (GANs) are used to generate normal operation data samples, and adversarial training is used to detect abnormal data. Convolutional neural networks (CNNs) are used to identify defects and locate equipment failure points, effectively improving the accuracy of anomaly detection and defect identification.

[0117] (5) The present invention uses deep reinforcement learning algorithms (such as DeepQ-Learning, DQN) to dynamically optimize the maintenance strategy of power grid equipment. By building a simulation model of the power grid equipment operating environment, simulating the equipment operating status, and adjusting the maintenance strategy based on real-time data, closed-loop control of the equipment operating status is achieved, overcoming the problem that traditional maintenance strategies lack a dynamic adjustment mechanism.

[0118] (6) The present invention solves the shortcomings of existing technologies in terms of data limitations, privacy security, fault diagnosis accuracy and maintenance strategy optimization through the comprehensive application of advanced technologies such as multimodal data fusion, federated learning, self-supervised learning, knowledge graphs, generative adversarial networks and reinforcement learning, and provides an innovative and practical solution for health assessment, fault diagnosis and maintenance optimization of power grid equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0120] Figure 1 This is a flow chart of a multi-modal power grid health assessment and analysis method of the present invention. DETAILED DESCRIPTION

[0121] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0122] In one or more embodiments, Figure 1 As shown, a multi-modal power grid health assessment and analysis method is disclosed, and the method includes steps S1-S7:

[0123] Step S1, deploying a variety of sensors to collect multimodal data of power grid equipment in real time, preprocessing the multimodal data, extracting features from the preprocessed multimodal data to form multimodal features, and fusing the multimodal features to form a comprehensive evaluation model; wherein the multimodal data includes temperature data, image data and vibration data; and the multimodal features include temperature features, image features and vibration features;

[0124] Specifically, temperature data collection involves installing temperature sensors at key locations of power grid equipment (such as transformers, switchgear, etc.). The sensors are placed in areas where hot spots may occur to ensure that the temperature changes of the equipment can be monitored in real time.

[0125] Vibration data collection is to install accelerometers on the mechanical parts of the equipment (such as motors, generators, etc.) to monitor vibration. The accelerometers should be installed along the three orthogonal directions (x, y, z) of the equipment to ensure that the vibration information is fully captured;

[0126] Image data collection is to use thermal imagers and high-definition cameras to regularly shoot the equipment to capture thermal and visible light images of the equipment. The camera should cover the main parts of the equipment and regularly shoot the status of the equipment during operation;

[0127] Furthermore, the preprocessing includes:

[0128] A. Data cleaning: Remove abnormal data points caused by sensor failure or environmental interference during the acquisition process. Use statistical methods such as the triple standard deviation method (3σ method) to detect and remove outliers;

[0129] ,

[0130] in, is the cleaned data. is the original data point, and are the mean and standard deviation of the data respectively;

[0131] B. Data normalization: In order to eliminate the dimensional differences of different modal data, the data needs to be normalized. Use the Min-Max normalization method to scale the data to the [0,1] interval;

[0132] ,

[0133] in, is the normalized data, is the original data, and are the minimum and maximum values ​​of the data respectively;

[0134] C. Data alignment: Since the acquisition frequencies of different modal data may be different, the data needs to be time aligned. The linear interpolation method is used to synchronize the data;

[0135] ,

[0136] in, It's at the time Interpolation data on and The time points and The data on.

[0137] Furthermore, it is necessary to perform wavelet decomposition on each data sequence to extract details and approximation parts at different scales;

[0138] ,

[0139] in, is the signal, wavelet coefficient Corresponding scale and displacement , is the wavelet basis function;

[0140] Apply soft thresholding to the details to remove high-frequency noise;

[0141] ,

[0142] in, is the processed wavelet coefficient, is the threshold value;

[0143] Perform wavelet reconstruction to obtain the denoised signal.

[0144] .

[0145] Furthermore, the preprocessed data is effectively stored and managed for subsequent analysis and processing;

[0146] Use distributed databases (such as Hadoop HDFS) to store large-scale power grid equipment data to ensure efficient data reading and writing and scalability. Establish data index and metadata management systems to facilitate fast retrieval and query. Back up data regularly to ensure data security and reliability.

[0147] Furthermore, it provides a data service interface for subsequent feature extraction and model training to ensure efficient access and call of data;

[0148] Design RESTful API interface to achieve efficient access and call of data. Provide batch data reading and real-time data stream interface to meet different processing requirements;

[0149] Specifically, the preprocessed multimodal data is subjected to feature extraction to form multimodal features, and multimodal feature fusion is performed to form a comprehensive evaluation model, including:

[0150] Extracting features from the temperature data includes:

[0151] (1) Statistical feature extraction:

[0152] Calculate the mean, standard deviation, maximum value, minimum value, kurtosis and skewness of the temperature data;

[0153] ,

[0154] ,

[0155] ,

[0156] ,

[0157] ,

[0158] ,

[0159] in, Indicates The temperature value at a time point, is the total number of data points;

[0160] (2) Frequency domain feature extraction:

[0161] Fast Fourier transform (FFT) is used to convert temperature data from time domain to frequency domain and extract the main frequency components;

[0162] ,

[0163] in, Indicates frequency The Fourier transform result on is, Indicates The temperature value at a time point.

[0164] The vibration data feature extraction, vibration data is also time series data, and time domain and frequency domain features need to be extracted, including:

[0165] (1) Time domain feature extraction:

[0166] Calculate the mean, standard deviation, maximum, minimum, root mean square (RMS) and crest factor (CrestFactor) of vibration data;

[0167] ,

[0168] ,

[0169] in, Indicates The vibration value at a time point;

[0170] (2) Frequency domain feature extraction:

[0171] Use Fast Fourier Transform (FFT) to convert vibration data from the time domain to the frequency domain and extract frequency components and energy spectral density;

[0172] ,

[0173] in, Indicates frequency The power spectral density on .

[0174] The image data feature extraction uses a physics-based method to extract features from image data, including:

[0175] (1) Edge detection: using the Sobel operator to perform edge detection and extract edge features of the image;

[0176] ,

[0177] in, and are the gradients in the horizontal and vertical directions respectively, is the input image;

[0178] (2) Thermal image features: Calculate the average temperature of the thermal image and the standard deviation of the temperature distribution;

[0179] ,

[0180] ,

[0181] in, Indicates the image Line The temperature value of the column, and are the number of rows and columns of the image, respectively.

[0182] Furthermore, the extracted temperature, vibration and image features are fused, including:

[0183] Concatenate all extracted features into a high-dimensional feature vector;

[0184] ,

[0185] in, represents the temperature eigenvector, represents the vibration eigenvector, represents the image feature vector;

[0186] Use principal component analysis PCA to reduce the dimension of high-dimensional feature vectors, as shown below:

[0187] ,

[0188] in, , and are eigenvectors, eigenvalues, and eigenvector matrices, respectively;

[0189] Normalize and standardize the fused feature vectors to ensure that the dimensions of the features are consistent;

[0190] Perform Min-Max normalization on the eigenvector and scale the eigenvalue to the interval [0,1], as shown below:

[0191] ,

[0192] The eigenvector is standardized so that its mean is 0 and its standard deviation is 1, including:

[0193] ,

[0194] in, and They represent the mean and standard deviation of the normalized eigenvectors respectively.

[0195] Step S2: construct a federated learning model, perform local federated learning model training on multiple distributed nodes, aggregate model parameters using a federated averaging algorithm, and update and distribute the aggregated global model to each node, including:

[0196] S201. Initialize a global model on the central server and distribute its parameters to each edge node, as shown below:

[0197] ,

[0198] ,

[0199] in, represents the initial parameters of the global model, Indicates that the initial model parameters Distribute to each edge node;

[0200] S202, after receiving the global model parameters, each edge node uses local data to perform global model training, as shown below:

[0201] ,

[0202] in, represents the model parameters of the i-th node at the t-th iteration, is the learning rate, is the loss function Parameters The gradient of Represents a local dataset;

[0203] S203, after each edge node completes the local global model training, it sends the updated model parameters back to the central server, as shown below:

[0204] ,

[0205] in, Indicates sending updated model parameters back to the central server;

[0206] S204: After receiving the model parameters updated by all edge nodes, the central server performs weighted average aggregation to generate new global model parameters, which are expressed as follows:

[0207] ,

[0208] in, represents the global model parameters after the t+1th iteration, K is the number of edge nodes, N is the total number of data samples of all nodes, is the number of data samples of the i-th node;

[0209] S205: Distribute the updated global model parameters to each edge node, as shown below:

[0210] ,

[0211] in, Indicates distribution to each edge node;

[0212] S206, perform federated learning iterations until a predetermined number of iterations is reached, as shown below:

[0213] :

[0214] :

[0215] ,

[0216] ,

[0217] ,

[0218] ;

[0219] S207. After the training is completed, the global model is verified and tested on the central server, including:

[0220] Model Validation:

[0221] ,

[0222] Evaluate the global model on the validation dataset performance, are the model parameters after the final iteration;

[0223] Model testing:

[0224] ,

[0225] Test the model on the test dataset performance and obtain the final result.

[0226] Step S3: introduce a self-supervised learning mechanism into the operation data of the power grid equipment, automatically generate pseudo labels for the operation data and train the deep learning model, and collect the operation data, maintenance records and fault history of the power grid equipment to build a device knowledge graph, including:

[0227] S301, construct a deep learning model, introduce a self-supervised learning mechanism, automatically generate pseudo labels for deep learning model training to form a multimodal deep learning model, including:

[0228] Apply data enhancement techniques such as rotation, scaling, noise addition, etc. to each sample to generate multiple enhanced samples;

[0229] ,

[0230] in, represents a set of data augmentation operations, To enhance operations;

[0231] For each original sample Generate augmented samples , :

[0232] ,

[0233] in, For the Enhanced samples, For the corresponding enhancement operation;

[0234] Furthermore, a contrast loss function is defined to maximize the similarity between enhanced samples and minimize the similarity between different samples. The loss function of the deep learning model is Defined as:

[0235] ,

[0236] in, Representation vector and The similarity of Representation sample The representation vector of Representation sample The representation vector of is the temperature parameter, Representation vector and The similarity of Representation sample The representation vector of

[0237] Furthermore, using the loss function of contrastive learning, a feature extraction network is trained to map the input data into a feature space:

[0238] ,

[0239] in, It is a sample The characteristic representation of It is the feature extraction network;

[0240] Joint contrastive learning is performed on multimodal data, and the loss function of the multimodal deep learning model is Defined as:

[0241] ,

[0242] in, and The feature extraction networks representing temperature data and vibration data respectively, and They are temperature and vibration data respectively.

[0243] S302, constructing a knowledge graph;

[0244] Furthermore, the knowledge graph represents equipment operation data and fault information in a structured way, providing knowledge support for fault diagnosis;

[0245] Knowledge graph initialization. Extract entities such as equipment, components, fault types, fault causes, and treatment methods from historical data and expert knowledge:

[0246] ,

[0247] in, Represents a collection of entities, For the Entity;

[0248] Define the relationship between entities, such as "the component is contained in the equipment" and "the fault occurs in the component".

[0249] ,

[0250] in, Represents a set of relations, For the a relationship;

[0251] Build triples of knowledge graphs based on entities and relations;

[0252] ,

[0253] in, A set of triples representing a knowledge graph, is a triple, representing an entity With entity Through the relationship connected;

[0254] Then, a graph neural network (GNN) is used to perform knowledge graph reasoning to obtain new fault diagnosis rules and knowledge;

[0255] ,

[0256] in, represents the representation vector of node v at layer l+1, represents the representation vector of node v at layer l, represents the set of neighbor nodes of node v, represents the normalization coefficient, represents the weight matrix of the lth layer, Represents the activation function.

[0257] S303, integrating the deep learning model and the knowledge graph to diagnose and predict faults of power grid equipment;

[0258] Specifically, the fusion of the deep learning model and the knowledge graph includes: concatenating the representation vector obtained by self-supervised learning with the node representation vector of the knowledge graph, as shown below:

[0259] ,

[0260] in, represents the fused representation vector, represents the representation vector obtained by self-supervised learning, Represents the representation vector obtained by knowledge graph reasoning;

[0261] Design a fault diagnosis loss function and train the fault diagnosis model on the fused representation vector; wherein the fault diagnosis loss function , which is expressed as follows:

[0262] ,

[0263] in, represents the cross entropy loss function, represents the fault diagnosis model, Indicates the fault label;

[0264] Furthermore, after the training is completed, the model is verified and put into practical use;

[0265] Model validation: Evaluate model performance on the validation dataset and adjust model parameters;

[0266] ;

[0267] Model application: Deploy the model on actual power grid equipment to perform online fault diagnosis and prediction;

[0268] .

[0269] Step S4, identifying normal operation data and abnormal patterns, and identifying defects according to the abnormal patterns to locate the fault point of the power grid equipment, including:

[0270] S401, using a generative adversarial network to generate normal operation data samples; wherein the generator Used to generate simulated normal operation data samples, discriminator Used to distinguish between real data and generated data;

[0271] Specifically, the goal of anomaly detection is to identify abnormal points in the equipment operation data that deviate from normal behavior. The present invention uses generative adversarial networks (GANs) to generate normal operation data samples, and detects abnormal patterns in the equipment operation data through adversarial training;

[0272] Among them, GANs are trained using normal operation data, where the generator Generate simulated normal operation data samples, discriminator Distinguish between real data and generated data:

[0273] ,

[0274] in, For real data, is random noise, is the generated data sample.

[0275] S402: Use adversarial training to detect abnormal patterns in device operation data and set thresholds , if the anomaly score exceeds the threshold, the current sample is marked as an anomaly point;

[0276] Specifically, we generate normal data samples through trained GANs, calculate the difference between actual data and generated data, and identify abnormal patterns:

[0277] ,

[0278] in, represents the Euclidean distance, is the actual data sample, is the generated data sample;

[0279] Furthermore, setting the threshold ,When the anomaly score exceeds the threshold, the sample is marked as an outlier;

[0280] .

[0281] S403, combine convolutional neural network to identify defects and locate equipment fault points, including:

[0282] Perform feature extraction and enhancement on the multimodal data of outliers to extract more information that is helpful for classification:

[0283] ,

[0284] in, is the enhanced feature vector, is the feature vector of the outlier point;

[0285] Using the multimodal features fused in the previous step, train a CNN for defect classification:

[0286] ,

[0287] in, represents the defect classification model, Indicates the fault tag, represents the cross entropy loss function;

[0288] Apply the trained defect classification model to the outliers to identify the specific defect type:

[0289] ;

[0290] Combined with the knowledge graph, the causes of the identified defect types are analyzed:

[0291] ;

[0292] Generate maintenance suggestions and repair plans based on fault diagnosis results and knowledge graphs:

[0293] ;

[0294] Monitor the equipment operation status in real time and perform online anomaly detection and defect identification on new data:

[0295] ,

[0296] ;

[0297] Furthermore, to ensure the accuracy and robustness of anomaly detection and defect identification, the performance of anomaly detection and defect identification models was evaluated on the validation dataset, and indicators such as accuracy, recall, and F1 score were calculated;

[0298] ,

[0299] ,

[0300] ,

[0301] ;

[0302] According to the verification results, the model parameters are tuned to improve the model performance:

[0303] ;

[0304] Through the above steps, the anomaly detection and defect identification solution can effectively identify anomalies and specific defect types in the operation of power grid equipment. Each step is closely linked to ensure the integrity and efficiency of the detection and identification process, providing reliable support for the health assessment and fault diagnosis of power grid equipment.

[0305] Step S5: construct a fault prediction model to evaluate the health status of power grid equipment and predict potential faults, including:

[0306] S501. Constructing comprehensive health assessment indicators based on temperature features, image features and vibration features , and the exponential smoothing method is used to calculate the health index for the comprehensive health assessment index , according to the health index Classify health levels;

[0307] Among them, the multimodal features extracted in step 2 are fused to construct a comprehensive health assessment index:

[0308] ,

[0309] in, , and are temperature, vibration and image feature vectors respectively;

[0310] The health index is calculated by using the exponential smoothing method to reflect the health status of the equipment;

[0311] ,

[0312] in, For time The health index, is the smoothing coefficient, For time The comprehensive feature vector of

[0313] The health level classification is based on the health index The equipment health status is divided into different levels, such as healthy, sub-healthy, fault warning and fault:

[0314] ,

[0315] in, , and It is the preset health level threshold.

[0316] S502. Use long short-term memory network to build a fault prediction model to predict the future health index of the equipment ;

[0317] Specifically, a long short-term memory network (LSTM) is used to build a fault prediction model to predict the future health index of the equipment:

[0318] ,

[0319] in, For the predicted time The health index, is the length of the historical observation window.

[0320] S503. Based on the predicted health index , estimate the probability of failure , and according to the probability Predict when equipment will fail ;

[0321] Specifically, according to the predicted health index , estimate the probability of failure;

[0322] ,

[0323] in, is the sigmoid function, and is the regression coefficient;

[0324] Specifically, according to the failure probability , predict the time when equipment will fail;

[0325] ,

[0326] in, is the failure rate.

[0327] S504, according to the health index and probability Combined, calculate the comprehensive health score ;

[0328] Specifically, the health index and fault prediction results are combined to calculate the comprehensive health score:

[0329] ,

[0330] in, For the comprehensive health score, and is the weight coefficient.

[0331] S505, based on comprehensive health score Develop equipment operation and maintenance strategies;

[0332] Specifically, according to the comprehensive health score, formulate the operation and maintenance strategy of the equipment, such as preventive maintenance, corrective maintenance, etc.:

[0333] ,

[0334] in, and It is the operation and maintenance policy threshold.

[0335] S506, verifying and optimizing the fault prediction model;

[0336] Specifically, the performance of the health assessment and fault prediction models is evaluated on the validation dataset, and indicators such as accuracy, recall, and F1 score are calculated:

[0337] ,

[0338] ,

[0339] ,

[0340] ;

[0341] According to the verification results, the model parameters are tuned to improve the model performance:

[0342] ;

[0343] Through the above steps, the health assessment and fault prediction solution can effectively assess the health status of power grid equipment and predict potential faults. Each step is closely linked to ensure the integrity and efficiency of the assessment and prediction process, providing reliable support for the operation and maintenance decision-making of power grid equipment.

[0344] Step S6: Construct a power grid equipment maintenance strategy and perform dynamic optimization and closed-loop control:

[0345] Collect multimodal data in real time and upload it to the central server, build a dynamic maintenance optimization model and train the dynamic maintenance optimization model to learn the optimal maintenance strategy, define a multi-objective optimization function to optimize the optimal maintenance strategy of power grid equipment, and provide real-time feedback and adjustment of the maintenance strategy based on the real-time operation data of the equipment;

[0346] Specifically, a sensor network is deployed to collect temperature, vibration and other relevant data in real time, and upload the data to a central server for real-time analysis.

[0347] Furthermore, the collected data is used to train deep reinforcement learning models to optimize maintenance strategies, including:

[0348] State space definition: Define the equipment operating status, including health index, failure probability and other related indicators:

[0349] ,

[0350] in, Indicates time The state vector of

[0351] Action space definition: Define executable maintenance actions, such as preventive maintenance, corrective maintenance, etc.:

[0352] ,

[0353] in, Indicates time Maintenance actions;

[0354] Reward function design: Design a reward function based on the equipment operating status and maintenance cost to guide the model to learn the optimal strategy:

[0355] ,

[0356] in, represents the reward function, Represents the reward value, Indicates maintenance action The cost, Indicates the change in health index;

[0357] Use the deep Q-network algorithm to train the dynamic maintenance optimization model and learn the optimal maintenance strategy:

[0358] ,

[0359] in, represents the Q-value function, Represents the discount factor.

[0360] Furthermore, a multi-objective optimization function is defined to balance the relationship between different objectives:

[0361] ,

[0362] in, is the total maintenance cost, is the average health index, is the average failure rate;

[0363] Use Genetic Algorithm (GA) or Particle Swarm Optimization (PSO) algorithm to search for Pareto optimal solution set:

[0364] ,

[0365] in, is the Pareto optimal solution set;

[0366] According to the specific application requirements, select the appropriate maintenance strategy from the Pareto optimal solution set:

[0367] ,

[0368] in, The selected maintenance strategy.

[0369] Furthermore, the actual effect of the maintenance strategy needs to be continuously verified and adjusted during operation to adapt to the actual operation of the equipment and environmental changes, including:

[0370] Monitor equipment operating status in real time and evaluate the actual effect of maintenance strategies:

[0371] ;

[0372] Continuously update and retrain deep reinforcement learning models based on actual operating data:

[0373] ;

[0374] Combine real-time data and feedback results to dynamically optimize maintenance strategies to ensure long-term stable operation of equipment:

[0375] ;

[0376] Through the above steps, the dynamic maintenance optimization solution can adjust and optimize the maintenance strategy in real time according to the actual operation and health status of the power grid equipment to ensure the long-term stable operation and efficient management of the equipment.

[0377] Step S7: perform system integration and application:

[0378] Integrate steps S1-S6 into the system framework, implement data interaction and function calls of each module through the API interface, and deploy them to the actual power grid environment.

[0379] Specifically, first, design the overall architecture of the system to ensure that each module can work together efficiently. The system is divided into data collection layer, data processing layer, model training layer, decision layer and application layer, as shown below:

[0380] Data collection layer: includes all sensors and data collection devices;

[0381] Data processing layer: performs data preprocessing, feature extraction and other operations;

[0382] Model training layer: conduct model training such as health assessment, fault prediction, and dynamic maintenance optimization;

[0383] Decision-making layer: Make maintenance decisions based on model output;

[0384] Application layer: actually executes the maintenance strategy and feeds back to the system.

[0385] Define the interfaces between modules to ensure smooth transmission of data and instructions, as shown below:

[0386] Data acquisition interface: sensor data upload interface;

[0387] Data processing interface: preprocessing and feature extraction result transmission interface;

[0388] Model interface: model training and prediction result transmission interface;

[0389] Decision interface: maintenance strategy transmission interface;

[0390] Application interface: maintenance execution feedback interface.

[0391] Furthermore, in order to ensure efficient data flow and processing, a data pipeline is built to achieve automated data flow:

[0392] A. Use sensor data to collect data in real time and transmit it to the central server through the MQTT protocol. Use data buffer to prevent data transmission delays;

[0393] B. Build a data processing pipeline through Apache Kafka for real-time data preprocessing and feature extraction. Use Spark Streaming for real-time data processing;

[0394] C. Use Hadoop HDFS and HBase for data storage and management to ensure efficient data access and management. Build a data warehouse to store historical data and real-time data.

[0395] Furthermore, the trained health assessment, fault prediction, and dynamic maintenance optimization models are deployed to the system, and Docker containers are used for model management and deployment. Kubernetes is used for container orchestration to ensure high availability and scalability of the model. Model performance is monitored in real time, and the model is retrained and optimized based on new data. Scheduled tasks are used to automatically trigger model training and updates to ensure that the model always maintains optimal performance;

[0396] Based on the output of the model, maintenance decisions are made and maintenance tasks are performed through the automated system. Maintenance strategies are dynamically generated based on health assessment and failure prediction results. Decisions are made using a rule-based decision engine that combines real-time and historical data. Maintenance strategies are transferred to the execution layer to automate maintenance tasks. Preventive and corrective maintenance tasks are performed using robots and automated equipment.

[0397] Furthermore, during the execution of maintenance tasks, real-time monitoring is performed and feedback data is transmitted back to the system for continuous optimization. The SCADA system is used to monitor the operating status of the equipment in real time. Sensor data, health assessment results, and fault prediction results are monitored. Based on the feedback data, maintenance strategies and model parameters are optimized to ensure continuous optimization and improvement of the system.

[0398] Preferably, a user-friendly interface is designed to facilitate operator use and understanding, and regular reports are generated. A dashboard is designed to display the equipment operating status, health assessment results, and maintenance strategies in real time. An interactive interface is provided to facilitate data query and decision-making by operators. Maintenance reports are generated regularly, including equipment operating status, health assessment results, fault prediction results, and maintenance execution status. Report tools are used to automatically generate and send reports for management personnel to review and make decisions.

[0399] In summary, the present invention solves the shortcomings of the existing technology in data limitations, privacy security, fault diagnosis accuracy and maintenance strategy optimization through the comprehensive application of advanced technologies such as multimodal data fusion, federated learning, self-supervised learning, knowledge graphs, generative adversarial networks and reinforcement learning, and provides an innovative and practical solution for health assessment, fault diagnosis and maintenance optimization of power grid equipment.

[0400] The above disclosures are only some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention are still within the scope of the invention.

Claims

1. A multi-modal power grid health assessment and analysis method, characterized in that: The method comprises: Step S1, deploying a variety of sensors to collect multimodal data of power grid equipment in real time, preprocessing the multimodal data, extracting features from the preprocessed multimodal data to form multimodal features, and fusing the multimodal features to form a comprehensive evaluation model; wherein the multimodal data includes temperature data, image data and vibration data; and the multimodal features include temperature features, image features and vibration features; Step S2: construct a federated learning model, perform local federated learning model training on multiple distributed nodes, aggregate model parameters using a federated averaging algorithm, and update the aggregated global model and distribute it to each node; Step S3: introduce a self-supervised learning mechanism into the operation data of the power grid equipment, automatically generate pseudo labels for the operation data and train the deep learning model, and at the same time collect the operation data, maintenance records and fault history of the power grid equipment to build a device knowledge graph; Step S4: Identify normal operation data and abnormal patterns, identify defects based on the abnormal patterns, and locate the fault point of the power grid equipment; Step S5: construct a fault prediction model to evaluate the health status of power grid equipment and predict potential faults; In step S2, a federated learning model is constructed, local federated learning model training is performed on multiple distributed nodes, model parameters are aggregated using a federated averaging algorithm, and the aggregated global model is updated and distributed to each node, specifically including: S201, initialize a global model on the central server and distribute its parameters to each edge node; S202, after receiving the global model parameters, each edge node uses local data to perform global model training, as shown below: in, represents the model parameters of the i-th node at the t-th iteration, η is the learning rate, is the loss function Parameters The gradient of Represents a local dataset; S203, after each edge node completes the local global model training, it sends the updated model parameters back to the central server; S204: After receiving the model parameters updated by all edge nodes, the central server performs weighted average aggregation to generate new global model parameters, which are expressed as follows: Among them, w (t+1) represents the global model parameters after the t+1th iteration, K is the number of edge nodes, N is the total number of data samples of all nodes, and N i is the number of data samples of the i-th node; S205, distributing the updated global model parameters to each edge node; S206, performing federated learning iterations until a predetermined number of iteration rounds is reached; S207. After the training is completed, the global model is verified and tested on the central server.

2. The multi-modal power grid health assessment and analysis method according to claim 1, characterized in that: The method further comprises: Step S6: Construct a power grid equipment maintenance strategy and perform dynamic optimization and closed-loop control: Collect multimodal data in real time and upload it to the central server, build a dynamic maintenance optimization model and train the dynamic maintenance optimization model to learn the optimal maintenance strategy, define a multi-objective optimization function to optimize the optimal maintenance strategy of power grid equipment, and provide real-time feedback and adjustment of the maintenance strategy based on the real-time operation data of the equipment; Step S7: perform system integration and application: Integrate steps S1-S6 into the system framework, implement data interaction and function calls of each module through the API interface, and deploy them to the actual power grid environment.

3. The multi-modal power grid health assessment and analysis method according to claim 1, characterized in that: In the step S1, the temperature features include statistical features and temperature frequency domain features; the image features include edge features and thermal image features; the vibration features include time domain features and vibration frequency domain features; The comprehensive assessment model includes: The temperature features, image features and vibration features are concatenated into a high-dimensional feature vector F, which is expressed as follows: F=[F T ,F V ,F I ], Among them, F T represents the temperature eigenvector, F V represents the vibration eigenvector, F I represents the image feature vector; Use principal component analysis PCA to reduce the dimension of high-dimensional feature vectors, as shown below: F pca =UΣV T , Among them, U, Σ, and V are eigenvectors, eigenvalues, and eigenvector matrices, respectively, and F pca represents the feature vector after dimensionality reduction; For the eigenvector F pca Min-Max normalization is performed to scale the eigenvalues ​​to the [0,1] interval, as shown below: Among them, F norm represents the normalized feature vector; For the eigenvector F norm Standardization is performed so that the mean of the eigenvector is 0 and the standard deviation is 1, as shown below: Among them, μ norm and σ norm They represent the mean and standard deviation of the normalized eigenvectors respectively.

4. The multi-modal power grid health assessment and analysis method according to claim 1, characterized in that: The verification and testing of the global model on the central server includes: The verification is expressed as follows: validate_model(w (T) ), Among them, validate_mode represents the validation dataset function, which means evaluating the global model w on the validation dataset (T) performance; The test is expressed as follows: test_model(in (T) ), Among them, test_model represents the test data set function, which means testing the model w on the test data set (T) performance and obtain the final result.

5. The multi-modal power grid health assessment and analysis method according to claim 1, characterized in that: The step S3 specifically includes: S301, construct a deep learning model, introduce a self-supervised learning mechanism, automatically generate pseudo labels for deep learning model training to form a multimodal deep learning model; S302, constructing a knowledge graph; S303: Integrate the deep learning model and the knowledge graph to diagnose and predict the faults of power grid equipment.

6. The multi-modal power grid health assessment and analysis method according to claim 5, characterized in that: The loss function of the deep learning model Defined as: Among them, sim(h(x i ),h(x i ′ )) represents the vector h(x i ) and h(x i ′ ), h(x i ) represents the sample x i The representation vector, h(x i ′ ) represents the sample x i ′ The representation vector, τ is the temperature parameter, sim(h(x i ),h(x j ′ )) represents the vector h(x i ) and h(x j ′ ), h(x j ′ ) represents the sample x j ′ The representation vector of The loss function of the multimodal deep learning model Defined as: Among them, h T (·) and h V (·) represent the feature extraction networks of temperature data and vibration data, respectively, T i and V i They are temperature and vibration data respectively; The triples of the knowledge graph are expressed as follows: in, The triple set representing the knowledge graph, (e i ,r j ,e k ) is a triple, representing entity e i With entity e k Through the relationship r j connected; Using graph neural network for knowledge graph reasoning is expressed as follows: in, represents the representation vector of node v at layer l+1, represents the representation vector of node v at layer l, represents the set of neighbor nodes of node v, c vu represents the normalization coefficient, W (l) represents the weight matrix of the lth layer, σ represents the activation function; The fusion of the deep learning model and the knowledge graph includes: concatenating the representation vector obtained by self-supervised learning with the node representation vector of the knowledge graph, as shown below: in, represents the fused representation vector, represents the representation vector obtained by self-supervised learning, Represents the representation vector obtained by knowledge graph reasoning; Design a fault diagnosis loss function and train the fault diagnosis model on the fused representation vector; wherein the fault diagnosis loss function It is expressed as follows: Among them, CrossEntropy represents the cross entropy loss function, f represents the fault diagnosis model, and y i Indicates a fault tag.

7. The multi-modal power grid health assessment and analysis method according to claim 6, characterized in that: The step S4 specifically includes: S401, using a generative adversarial network to generate normal operating data samples; wherein the generator G is used to generate simulated normal operating data samples, and the discriminator D is used to distinguish between real data and generated data; S402, using adversarial training to detect abnormal patterns in device operation data, setting a threshold θ, and if the abnormal score exceeds the threshold, marking the current sample as an abnormal point; S403, combine convolutional neural network to identify defects and locate equipment failure points: Extract and enhance the features of the abnormal points, build a defect classification model and use the multimodal features of step S1 to train the defect classification model, input the feature-extracted and enhanced abnormal point data into the defect classification model to identify the fault type, and analyze the identified fault type in combination with the knowledge graph to generate maintenance suggestions and repair plans; wherein the loss function of the defect classification model is It is expressed as follows: Among them, f CNN represents the defect classification model, y i represents the fault label and CrossEntropy represents the cross entropy loss function.

8. The multi-modal power grid health assessment and analysis method according to claim 3, characterized in that: The step S5 specifically includes: S501, constructing a comprehensive health assessment index F based on temperature features, image features and vibration features health , and the exponential smoothing method is used to calculate the health index H for the comprehensive health assessment index t , according to the health index H t Classify health levels; S502. Use long short-term memory network to build a fault prediction model to predict the future health index of the equipment S503. Based on the predicted health index Estimate the probability of failure And according to the probability Predict the time T before equipment failure fault ; S504, according to the health index and probability Combined, calculate the comprehensive health score S health ; S505, based on the comprehensive health score S health Develop equipment operation and maintenance strategies; S506: Verify and optimize the fault prediction model.

9. The multi-modal power grid health assessment and analysis method according to claim 2, characterized in that: The dynamic maintenance optimization model includes: The equipment operation state is defined as the state vector, and the executable maintenance action is defined as the action space. The reward function is designed as: r t =R(s t ,a t )=-C(a t )+ΔH t , Among them, R(s t ,a t ) represents the reward function, r t represents the reward value, C(a t ) indicates maintenance action a t Cost, ΔH t Indicates the change in health index; The dynamic maintenance optimization model is trained using the deep Q network algorithm to learn the optimal maintenance strategy, which is expressed as follows: Where Q represents the Q-value function and γ represents the discount factor.

10. The multi-modal power grid health assessment and analysis method according to claim 9, characterized in that: The multi-objective optimization function minF is expressed as follows: minF=[C total ,H avg ,P fault_avg ], Among them, C total is the total maintenance cost, H avg is the average health index, P fault_avg is the average failure rate.

Images